feat(tanya): Report per-criterion verdicts for engine acceptance criteria

Outcome: done
Changed: none
Verified by: make verify -> EXIT_CODE=0, "Ran 226 tests in 0.959s OK verification passed"; only pre-existing StarletteDeprecationWarning
Findings:
- C1 PASS - grep: 0 non-rsearch URL literals; test_research_client.py:695 asserts rsearch URL; test_research_integration.py:94-96 asserts all URLs start RSEARCH_BASE_URL
- C2 PASS - test_research_scheduling.py:131-163 max_concurrency==8, peak==8, 64 drained, pending 0; :459/:499 pool bound
- C3 PASS - test_research_pipeline.py:272-289 endpoints web/images->/search describe->/describe chat->/chat; test_research_engine.py:141-146 content_types web:3 images:3 chat:3 describe:1
- C4 PASS - test_research_dedup.py:107,149,161,186,213 URL/content/query dedup; test_research_scheduling.py:235,248,266 concurrency dedup
- C5 PASS - test_research_engine.py:79-101 closed round1 (new_urls=0,new_queries=0); :103-147 discovery then closure; test_research_dedup.py:236-291 closure matrix
- C6 PASS - AST 9 modules: 0 annotation problems; 21 logger.info + 10 logger.debug; retoor header 9/9
- C7 PASS - grep TODO|FIXME|placeholder|stub: 0 hits
- C8 PASS - make verify EXIT_CODE=0, 226 tests OK, verification passed; no new warning
- Adversarial - run("   \n\t") and run("") -> ValueError "research subject must not be empty", no API call, no hang
Open: none
Confidence: high - all 8 engine criteria executed first-hand with asserting tests and verify output

Typosaurus-Run: 4e2afb673c7f4578a12276d9181b982d
Typosaurus-Node: 3bb5cbfb04ca4194a0b509da6f861dbc
Typosaurus-Agent: @tanya
Refs: #31

.gitea/workflows/ci.yml

@@ -0,0 +1,24 @@
+# retoor <retoor@molodetz.nl>
+name: CI
+on:
+  push:
+    branches: [main, master]
+  pull_request:
+    branches: [main]
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - name: Set up Python 3.12
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install -e .
+      - name: Run tests
+        run: make verify
+
+

.gitignore

@@ -1,2 +1,6 @@
 __pycache__/
 *.pyc
+.env.json
+logs/
+
+

CLAUDE.md

@@ -1,3 +1,4 @@
+# retoor <retoor@molodetz.nl>
 # typosaurus-sandbox
 
 A minimal Python calculator used to verify the Typosaurus agent system.
@@ -26,3 +27,11 @@ A minimal Python calculator used to verify the Typosaurus agent system.
 ```
 make verify
 ```
+
+## CI
+
+- Workflow file: `.gitea/workflows/ci.yml`
+- Trigger: push to `main` or `master` branches
+- Steps: checkout, Python 3.12 setup, dependency install, `make verify`
+
+

Makefile

@@ -1,4 +1,8 @@
+# retoor <retoor@molodetz.nl>
+
 verify:
 	@PYTHONPATH=src python3 -m compileall -q src tests && PYTHONPATH=src python3 -m unittest discover -s tests -q && echo "verification passed"
 
+run:
+	@PYTHONPATH=src python3 -m typosaurus_sandbox
 

README.md

@@ -113,6 +113,90 @@ Response:
 {"result": 255}
 ```
 
+### POST /api/v1/calculator/average
+
+Compute the arithmetic mean of a list of values.
+
+Request:
+
+```json
+{"values": [1, 2, 3, 4, 5]}
+```
+
+Response:
+
+```json
+{"result": 3.0}
+```
+
+An empty list produces a 422 validation response.
+
+### POST /api/v1/calculator/median
+
+Compute the median of a list of values. Values are sorted internally; an even-length list returns the average of the two middle values as a float.
+
+Request:
+
+```json
+{"values": [1, 3, 5]}
+```
+
+Response:
+
+```json
+{"result": 3.0}
+```
+
+Request (even length):
+
+```json
+{"values": [1, 2, 3, 4]}
+```
+
+Response:
+
+```json
+{"result": 2.5}
+```
+
+An empty list produces a 422 validation response.
+
+### POST /api/v1/calculator/variance
+
+Compute the population variance of a list of values.
+
+Request:
+
+```json
+{"values": [1, 2, 3, 4, 5]}
+```
+
+Response:
+
+```json
+{"result": 2.0}
+```
+
+An empty list produces a 422 validation response.
+
+### POST /api/v1/calculator/percentage
+
+Compute what percentage `value` is of `total`.
+
+Request:
+
+```json
+{"value": 50, "total": 100}
+```
+
+Response:
+
+```json
+{"result": 50.0}
+```
+
+A zero `total` produces a 422 validation response.
+
 ## Verification
 
 ```sh
@@ -121,3 +205,4 @@ make verify
 
 Runs compile-all checks against all source and test files, then executes the full test suite.
 Zero warnings are tolerated.
+

app/__init__.py

@@ -0,0 +1,68 @@
+# retoor <retoor@molodetz.nl>
+
+import os
+
+from flask import Flask
+from flask import jsonify
+from flask import make_response
+from flask import request
+from flask.wrappers import Response
+
+from typosaurus_sandbox.domain.calculator import add
+from typosaurus_sandbox.domain.calculator import clamp
+from typosaurus_sandbox.domain.calculator import clamp_to_byte
+from typosaurus_sandbox.domain.calculator import subtract
+
+app = Flask(__name__)
+
+
+@app.route('/')
+def index() -> Response:
+    index_path = os.path.join(os.path.dirname(__file__), 'index.html')
+    with open(index_path) as f:
+        return make_response(f.read(), 200, {'Content-Type': 'text/html'})
+
+
+@app.route('/add', methods=['GET'])
+def add_route() -> Response:
+    try:
+        left = int(request.args['left'])
+        right = int(request.args['right'])
+    except (KeyError, TypeError, ValueError):
+        return make_response(jsonify({'error': 'Invalid or missing parameters'}), 400)
+    return make_response(jsonify({'result': add(left, right)}), 200)
+
+
+@app.route('/subtract', methods=['GET'])
+def subtract_route() -> Response:
+    try:
+        left = int(request.args['left'])
+        right = int(request.args['right'])
+    except (KeyError, TypeError, ValueError):
+        return make_response(jsonify({'error': 'Invalid or missing parameters'}), 400)
+    return make_response(jsonify({'result': subtract(left, right)}), 200)
+
+
+@app.route('/clamp', methods=['GET'])
+def clamp_route() -> Response:
+    try:
+        value = int(request.args['value'])
+        low = int(request.args['low'])
+        high = int(request.args['high'])
+    except (KeyError, TypeError, ValueError):
+        return make_response(jsonify({'error': 'Invalid or missing parameters'}), 400)
+    try:
+        result = clamp(value, low, high)
+    except ValueError:
+        return make_response(jsonify({'error': 'Invalid or missing parameters'}), 400)
+    return make_response(jsonify({'result': result}), 200)
+
+
+@app.route('/clamp_to_byte', methods=['GET'])
+def clamp_to_byte_route() -> Response:
+    try:
+        value = int(request.args['value'])
+    except (KeyError, TypeError, ValueError):
+        return make_response(jsonify({'error': 'Invalid or missing parameters'}), 400)
+    return make_response(jsonify({'result': clamp_to_byte(value)}), 200)
+

app/index.html

@@ -0,0 +1,69 @@
+<!-- retoor <retoor@molodetz.nl> -->
+<!DOCTYPE html>
+<html lang="en">
+<head>
+<meta charset="UTF-8">
+<meta name="viewport" content="width=device-width, initial-scale=1.0">
+<title>Calculator</title>
+</head>
+<body>
+<h1>Calculator</h1>
+
+<fieldset>
+<legend>Add / Subtract</legend>
+<input type="number" id="left" placeholder="Left operand">
+<input type="number" id="right" placeholder="Right operand">
+<button onclick="calculate('add')">Add</button>
+<button onclick="calculate('subtract')">Subtract</button>
+</fieldset>
+
+<fieldset>
+<legend>Clamp</legend>
+<input type="number" id="value" placeholder="Value">
+<input type="number" id="low" placeholder="Low">
+<input type="number" id="high" placeholder="High">
+<button onclick="calculate('clamp')">Clamp</button>
+</fieldset>
+
+<fieldset>
+<legend>Clamp to Byte</legend>
+<input type="number" id="byte_value" placeholder="Value">
+<button onclick="calculate('clamp_to_byte')">Clamp to Byte</button>
+</fieldset>
+
+<p id="output"></p>
+
+<script>
+function calculate(operation) {
+  const resultEl = document.getElementById('output');
+  let url;
+  if (operation === 'add' || operation === 'subtract') {
+    const left = document.getElementById('left').value;
+    const right = document.getElementById('right').value;
+    url = '/' + operation + '?left=' + encodeURIComponent(left) + '&right=' + encodeURIComponent(right);
+  } else if (operation === 'clamp') {
+    const value = document.getElementById('value').value;
+    const low = document.getElementById('low').value;
+    const high = document.getElementById('high').value;
+    url = '/clamp?value=' + encodeURIComponent(value) + '&low=' + encodeURIComponent(low) + '&high=' + encodeURIComponent(high);
+  } else if (operation === 'clamp_to_byte') {
+    const value = document.getElementById('byte_value').value;
+    url = '/clamp_to_byte?value=' + encodeURIComponent(value);
+  }
+  fetch(url)
+    .then(function(response) {
+      return response.json().then(function(data) {
+        if (!response.ok) {
+          resultEl.textContent = 'Error: ' + (data.error || 'Unknown error');
+        } else {
+          resultEl.textContent = 'Result: ' + data.result;
+        }
+      });
+    })
+    .catch(function() {
+      resultEl.textContent = 'Error: Network error';
+    });
+}
+</script>
+</body>
+</html>

deepresearch.md

@@ -0,0 +1,140 @@
+# retoor <retoor@molodetz.nl>
+# Deep Research Engine — Design, Optimality Argument and Verification Evidence
+
+This document describes the exhaustive deep research engine in `src/typosaurus_sandbox/research/`,
+the mathematical argument that recursive query expansion with URL/content deduplication and
+closure detection is the most aggressive feasible research strategy over the rsearch API, and the
+four recursive verification passes executed against it. Every claim is traceable to the run's
+verified nodes (fact sheet node d5d9e290; optimality node b042b1d23; tester nodes f7f10c64,
+fde105db, 2e6bd38b; engine node 1b0176bf) and to source path:line references.
+
+## 1. Scope and constraints
+
+- Only search API: `https://rsearch.app.molodetz.nl`; the client issues requests only to the
+  `/search`, `/chat` and `/describe` endpoints (client.py:211). `/search` is GET-only.
+- Content-type agnostic: web results, image results (`type=images`), describe and chat flow
+  through one asynchronous pipeline; no per-type special casing beyond parameter selection.
+- Native Python 3.12, standard library only (`asyncio`, `urllib`); no new dependency was added.
+- No artificial depth cap, page cap or time budget stops a run before closure; the engine stops
+  only when a full round adds zero new URLs and zero new queries (least fixed point).
+
+## 2. Architecture (module map)
+
+| Module | Public symbol | Path:line |
+|---|---|---|
+| config | `ResearchConfig` (base_url, TTLs, `max_concurrency=8`, default_count) | `src/typosaurus_sandbox/research/config.py:12` |
+| client | `RsearchClient`, `RsearchError` (search/chat/describe, `_request`) | `src/typosaurus_sandbox/research/client.py:72` |
+| cache | `TTLCache`, `CacheEntry` (thread-safe, monotonic expiry) | `src/typosaurus_sandbox/research/cache.py:20` |
+| envelopes | `SearchResponse`, `SearchResult`, `DeepReport`, `ChatResponse`, `DescribeResponse` | `src/typosaurus_sandbox/research/envelopes.py:103` |
+| frontier | `QueryFrontier`, `DedupStats`, URL normalization, content fingerprint | `src/typosaurus_sandbox/research/frontier.py:102` |
+| pipeline | `ResearchPipeline`, `WorkItem`, `PipelineReport` (bounded worker pool) | `src/typosaurus_sandbox/research/pipeline.py:125` |
+| engine | `ResearchEngine`, `ResearchReport`, `RoundSummary` (closure loop) | `src/typosaurus_sandbox/research/engine.py:85` |
+| entry | `main()` CLI | `src/typosaurus_sandbox/research/__main__.py:24` |
+
+## 3. Concurrency model
+
+- Bounded asyncio worker pool: `asyncio.Semaphore(pool_size)` with
+  `pool_size = max(1, max_concurrency)` and `max_concurrency = 8`
+  (config.py:18, pipeline.py:126-136).
+- `run()` drains the frontier through a bounded queue (pool * 4) with pool-size workers and
+  `None` sentinels; every request runs via `asyncio.to_thread` over `urllib` (no extra deps).
+- Pool size is logged at INFO; every request outcome (endpoint, query/url, status, cache hit)
+  at INFO, every extraction at DEBUG.
+
+## 4. Deduplication and closure strategy
+
+- Query dedup key: whitespace-collapsed `casefold` (frontier.py:28); length window 2-200 chars.
+- URL dedup: `normalize_url` lowercases scheme/host, applies IDNA, strips default port,
+  userinfo and fragment, collapses slashes (frontier.py:28).
+- Content dedup: SHA-256 fingerprint of whitespace-normalized text (frontier.py:61).
+- One `threading.Lock` guards all seen-sets and counters for concurrent worker access
+  (frontier.py:103).
+- Closure rule: a round that adds 0 new URLs and 0 new queries halts the run
+  (engine.py:178-183). The engine is closed-loop verified: a fixed-fixture fake client closed
+  in 3 rounds with all four content types, and a 4-level chain client closed in 5 rounds,
+  proving no depth cap (engine node 1b0176bf).
+
+## 5. Content-type agnosticism
+
+- One worker path serves all kinds: `web` -> `search(content=True)`, `images` ->
+  `search(type="images")`, `describe` -> GET `/describe?url=`, `chat` -> POST `/chat`
+  (pipeline.py:138-143, engine.py:106).
+- Extraction yields new URLs and new query seeds from titles, descriptions and `extra` fields
+  of every content type (frontier.py:66).
+
+## 6. Optimality argument
+
+Let `R(q)` be the set of result URLs returned by the aggregator for query `q`, `gen(u)` the
+query variants generated from URL/content `u`, and `S` the set of collected URLs.
+
+- Completeness: the process is coverage-complete for subject `t` iff it halts at the least
+  fixed point `S* = lfp(F)` with `F(S) = S ∪ ⋃_{u∈S, q∈gen(u)} R(q)`; the halt condition is
+  "a full round adds 0 new URLs and 0 new queries" (node b042b1d23).
+- Dominance: depth-`d` iteration reaches `F^d(S0) ⊆ S*`; the inclusion is strict whenever the
+  discovery chain exceeds `d`, so every fixed-depth strategy is incomplete. Closure iterates
+  `F` to its unique least fixed point (Knaster-Tarski), attaining the maximum reachable
+  coverage; any strategy that stops before the fixed point is strictly dominated.
+- Cost model: `Cost = Σ_{q∈Q_issued} c(q) + Σ_{u∈F_issued} c_c(u)`. Search (5 min) and content
+  (24 h) caches (config.py:16-17) make repeat queries near-free; the dominant cost is
+  `|Q_issued| + |F_issued|`, and query/URL dedup touches each element exactly once.
+- Stated assumptions and limits: single aggregator (rsearch only), no pagination API,
+  documented count bound 1-100 with the provider capping at 10, and content retrieval only
+  through the aggregator. Optimality is proven within these constraints.
+- Dated references (tier): rsearch docs https://rsearch.app.molodetz.nl/about (2026-08-07, 1);
+  Gemini https://blog.google/products-and-platforms/products/gemini/google-gemini-deep-research/
+  (2024-12-11, 1); OpenAI https://openai.com/index/introducing-deep-research/ (Feb-2025, 1) +
+  https://techcrunch.com/2025/02/02/openai-unveils-a-new-chatgpt-agent-for-deep-research/ (4);
+  Ntoulas 2005 ACM JCDL 10.1145/1065385.1065407 (3); Chakrabarti 1999 Computer Networks
+  10.1016/S1389-1286(99)00052-3 (3); Olston & Najork 2010 FnTIR 10.1561/1500000017 (3).
+
+## 7. Four recursive verification passes
+
+Each pass re-checks the previous pass's optimality claim ("recursive closure over the rsearch
+aggregator is the most aggressive feasible strategy") and records its own evidence. All four
+passes passed.
+
+- Pass 1 — Optimality argument: formal completeness criterion, cost model and Knaster-Tarski
+  dominance proof produced with seven dated, tiered sources (node b042b1d23, 2026-08-07).
+- Pass 2 — Engine matches the argument: all eight engine acceptance criteria executed with
+  pass verdicts and exact commands (node f7f10c64): rsearch-only source, bounded pool at
+  max_concurrency=8, one web/images/describe/chat pipeline, URL+content dedup (64 concurrent
+  same-query pushes -> 1 enqueued, 63 skipped), closure decision (NullClient probe closed in 1
+  round with 0 new URLs and 0 new queries), logging/annotations, no deferred markers, and
+  `make verify` -> "Ran 199 tests in 2.168s OK verification passed".
+- Pass 3 — Live probe coverage/cost (node fde105db, 2026-08-07): subject "python asyncio",
+  max_concurrency=8, count=10, 240 s guard: queries_issued=86, urls_seen=754, contents_seen=281,
+  164 network requests (search 105 / chat 46 / describe 13), X-AI-Cost-USD sum $0.002075, wall
+  elapsed 264.91 s. Adversarial subjects ("", spaces, tabs) raised ValueError
+  "research subject must not be empty" (engine.py:125) before any API call; urlopen delta 0.
+- Pass 4 — Closure and determinism (node 1b0176bf, confirmed by fact sheet d5d9e290):
+  fixed-fixture fake client closed in 3 rounds with all 4 content types; 4-level chain closed
+  in 5 rounds (no depth cap); live `python -m typosaurus_sandbox.research` logged INFO rounds,
+  closure and the typed report JSON; final gate `make verify` green (199 tests OK, git clean).
+
+## 8. Usage
+
+```sh
+python -m typosaurus_sandbox.research [subject]
+```
+
+Run a research session on `subject` (default "typosaurus sandbox") until closure; rounds and
+closure decisions are logged at INFO, and the typed `ResearchReport` JSON is logged at the end.
+Configuration (base_url, TTLs, max_concurrency, default_count) is loaded from `.env.json` under
+the `research` key with plug-and-play defaults (config.py:22).
+
+Verification gate:
+
+```sh
+make verify
+```
+
+## 9. Verification status
+
+- `make verify`: exit 0, "Ran 199 tests OK verification passed" (2026-08-07); only the
+  pre-existing Starlette deprecation warning from the FastAPI test client remains, none
+  introduced by the research package.
+- Re-run at document time: `make verify` exit 0, "Ran 217 tests in 2.162s OK", verification
+  passed; same pre-existing Starlette deprecation warning only.
+- The research package contains no deferred markers (grep verified, node f7f10c64).
+
+

requirements.txt

@@ -0,0 +1,3 @@
+fastapi
+uvicorn[standard]
+

src/calculator.py

@@ -1,34 +0,0 @@
-# retoor <retoor@molodetz.nl>
-
-from typing import Sequence
-
-
-def add(left: int, right: int) -> int:
-    return left + right
-
-
-def subtract(left: int, right: int) -> int:
-    return left - right
-
-
-def clamp(value: int, low: int, high: int) -> int:
-    if low > high:
-        raise ValueError
-    if value < low:
-        return low
-    if value > high:
-        return high
-    return value
-
-
-def clamp_to_byte(value: int) -> int:
-    return max(0, min(255, value))
-
-
-def variance(values: Sequence[float]) -> float:
-    if not values:
-        raise ValueError
-    mean = sum(values) / len(values)
-    return sum((x - mean) ** 2 for x in values) / len(values)
-
-

src/typosaurus_sandbox/__init__.py

@@ -1,5 +1,7 @@
 # retoor <retoor@molodetz.nl>
 
 from typosaurus_sandbox.app import App
+from typosaurus_sandbox.core import Config, setup_logging
+
+__all__ = ["App", "Config", "setup_logging"]
 
-__all__ = ["App"]

src/typosaurus_sandbox/__main__.py

@@ -1,7 +1,22 @@
 # retoor <retoor@molodetz.nl>
 
+import logging
+
 import uvicorn
 
 from typosaurus_sandbox.app import App
+from typosaurus_sandbox.core import Config, setup_logging
+
+logger = logging.getLogger(__name__)
+
+
+def main() -> None:
+    setup_logging()
+    config = Config.load()
+    logger.info("starting server on %s:%d", config.host, config.port)
+    uvicorn.run(App, host=config.host, port=config.port, log_level="info")
+
+
+if __name__ == "__main__":
+    main()
 
-uvicorn.run(App, host="127.0.0.1", port=8000, log_level="info")

src/typosaurus_sandbox/app.py

@@ -1,15 +1,26 @@
 # retoor <retoor@molodetz.nl>
 
+import logging
+
 from fastapi import FastAPI
 
 from typosaurus_sandbox.presentation.api.v1.calculator import calculator_router
 
+logger = logging.getLogger(__name__)
+
 App = FastAPI(title="typosaurus-sandbox")
 
 
+@App.on_event("startup")
+def on_startup() -> None:
+    logger.info("application startup complete")
+
+
 @App.get("/health")
 def health() -> dict[str, str]:
+    logger.debug("health check requested")
     return {"status": "ok"}
 
 
 App.include_router(calculator_router)
+

src/typosaurus_sandbox/core/__init__.py

@@ -0,0 +1,7 @@
+# retoor <retoor@molodetz.nl>
+
+from typosaurus_sandbox.core.config import Config
+from typosaurus_sandbox.core.logging import setup_logging
+
+__all__ = ["Config", "setup_logging"]
+

src/typosaurus_sandbox/core/config.py

@@ -0,0 +1,28 @@
+# retoor <retoor@molodetz.nl>
+
+import json
+import logging
+from dataclasses import dataclass
+from pathlib import Path
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class Config:
+    host: str = "127.0.0.1"
+    port: int = 8000
+
+    @classmethod
+    def load(cls) -> "Config":
+        config_path = Path(".env.json")
+        if not config_path.exists():
+            logger.info("no .env.json found, using defaults")
+            return cls()
+        with config_path.open() as f:
+            data = json.load(f)
+        host = data.get("host", cls.host)
+        port = data.get("port", cls.port)
+        logger.info("loaded config from .env.json: host=%s port=%s", host, port)
+        return cls(host=host, port=port)
+

src/typosaurus_sandbox/core/logging.py

@@ -0,0 +1,23 @@
+# retoor <retoor@molodetz.nl>
+
+import logging
+import logging.handlers
+from pathlib import Path
+
+
+def setup_logging() -> None:
+    log_dir = Path("logs")
+    log_dir.mkdir(exist_ok=True)
+
+    handler = logging.handlers.RotatingFileHandler(
+        log_dir / "typosaurus-sandbox.log",
+        maxBytes=10 * 1024 * 1024,
+        backupCount=5,
+    )
+    handler.setFormatter(
+        logging.Formatter("%(asctime)s [%(levelname)s] %(name)s: %(message)s")
+    )
+
+    logging.basicConfig(level=logging.DEBUG, handlers=[handler])
+    logging.getLogger(__name__).info("logging configured")
+

src/typosaurus_sandbox/domain/calculator/__init__.py

@@ -1,6 +1,6 @@
 # retoor <retoor@molodetz.nl>
 
-from typosaurus_sandbox.domain.calculator.operations import add, clamp, clamp_to_byte, subtract, variance
+from typosaurus_sandbox.domain.calculator.operations import add, average, clamp, clamp_to_byte, median, percentage, subtract, variance
 
-__all__ = ["add", "subtract", "clamp", "clamp_to_byte", "variance"]
+__all__ = ["add", "average", "clamp", "clamp_to_byte", "median", "percentage", "subtract", "variance"]
 

src/typosaurus_sandbox/domain/calculator/operations.py

@@ -1,6 +1,6 @@
 # retoor <retoor@molodetz.nl>
 
-from typing import Sequence
+from typing import Sequence, Union
 
 
 def add(left: int, right: int) -> int:
@@ -25,9 +25,32 @@ def clamp_to_byte(value: int) -> int:
     return max(0, min(255, value))
 
 
+def average(values: list[int | float]) -> float:
+    if not values:
+        raise ValueError
+    return sum(values) / len(values)
+
+
+def median(values: list[float]) -> float:
+    if not values:
+        raise ValueError
+    sorted_values = sorted(values)
+    n = len(sorted_values)
+    mid = n // 2
+    if n % 2 == 1:
+        return sorted_values[mid]
+    return (sorted_values[mid - 1] + sorted_values[mid]) / 2.0
+
+
 def variance(values: Sequence[float]) -> float:
     if not values:
         raise ValueError
     mean = sum(values) / len(values)
     return sum((x - mean) ** 2 for x in values) / len(values)
 
+
+def percentage(value: Union[int, float], total: Union[int, float]) -> float:
+    if total == 0:
+        raise ValueError
+    return (value / total) * 100
+

src/typosaurus_sandbox/presentation/api/v1/calculator.py

@@ -1,9 +1,13 @@
 # retoor <retoor@molodetz.nl>
 
+import logging
+
 from fastapi import APIRouter, HTTPException
 from pydantic import BaseModel, Field
 
-from typosaurus_sandbox.domain.calculator import add, clamp, clamp_to_byte, subtract
+from typosaurus_sandbox.domain.calculator import add, average, clamp, clamp_to_byte, median, percentage, subtract, variance
+
+logger = logging.getLogger(__name__)
 
 calculator_router = APIRouter(prefix="/api/v1/calculator")
 
@@ -28,6 +32,15 @@ class ClampToByteRequest(BaseModel):
     value: int = Field(ge=-2147483648, le=2147483647)
 
 
+class ValuesRequest(BaseModel):
+    values: list[float]
+
+
+class PercentageRequest(BaseModel):
+    value: float
+    total: float
+
+
 class IntResult(BaseModel):
     result: int
 
@@ -38,11 +51,13 @@ class FloatResult(BaseModel):
 
 @calculator_router.post("/add", response_model=IntResult)
 def calculate_add(body: AddRequest) -> IntResult:
+    logger.debug("add %d + %d", body.left, body.right)
     return IntResult(result=add(body.left, body.right))
 
 
 @calculator_router.post("/subtract", response_model=IntResult)
 def calculate_subtract(body: SubtractRequest) -> IntResult:
+    logger.debug("subtract %d - %d", body.left, body.right)
     return IntResult(result=subtract(body.left, body.right))
 
 
@@ -57,4 +72,46 @@ def calculate_clamp(body: ClampRequest) -> FloatResult:
 
 @calculator_router.post("/clamp-to-byte", response_model=IntResult)
 def calculate_clamp_to_byte(body: ClampToByteRequest) -> IntResult:
+    logger.debug("clamp-to-byte %d", body.value)
     return IntResult(result=clamp_to_byte(body.value))
+
+
+@calculator_router.post("/average", response_model=FloatResult)
+def calculate_average(body: ValuesRequest) -> FloatResult:
+    logger.debug("average of %d values", len(body.values))
+    try:
+        result = average(body.values)
+    except ValueError:
+        raise HTTPException(status_code=422, detail="values list must not be empty")
+    return FloatResult(result=result)
+
+
+@calculator_router.post("/median", response_model=FloatResult)
+def calculate_median(body: ValuesRequest) -> FloatResult:
+    logger.debug("median of %d values", len(body.values))
+    try:
+        result = median(body.values)
+    except ValueError:
+        raise HTTPException(status_code=422, detail="values list must not be empty")
+    return FloatResult(result=result)
+
+
+@calculator_router.post("/variance", response_model=FloatResult)
+def calculate_variance(body: ValuesRequest) -> FloatResult:
+    logger.debug("variance of %d values", len(body.values))
+    try:
+        result = variance(body.values)
+    except ValueError:
+        raise HTTPException(status_code=422, detail="values list must not be empty")
+    return FloatResult(result=result)
+
+
+@calculator_router.post("/percentage", response_model=FloatResult)
+def calculate_percentage(body: PercentageRequest) -> FloatResult:
+    logger.debug("percentage %f of %f", body.value, body.total)
+    try:
+        result = percentage(body.value, body.total)
+    except ValueError:
+        raise HTTPException(status_code=422, detail="total must not be zero")
+    return FloatResult(result=result)
+

src/typosaurus_sandbox/research/__init__.py

@@ -0,0 +1,68 @@
+# retoor <retoor@molodetz.nl>
+
+from typosaurus_sandbox.research.cache import TTLCache
+from typosaurus_sandbox.research.client import RsearchClient, RsearchError
+from typosaurus_sandbox.research.config import ResearchConfig
+from typosaurus_sandbox.research.engine import ResearchEngine, ResearchReport, RoundSummary
+from typosaurus_sandbox.research.envelopes import (
+    ChatResponse,
+    ChatUsage,
+    DeepReport,
+    DescribeResponse,
+    SearchGrade,
+    SearchResponse,
+    SearchResult,
+)
+from typosaurus_sandbox.research.frontier import (
+    DedupStats,
+    QueryFrontier,
+    fingerprint_text,
+    normalize_url,
+    query_variants_from_result,
+)
+from typosaurus_sandbox.research.pipeline import (
+    ContentKind,
+    Extraction,
+    PipelineReport,
+    ResearchPipeline,
+    WorkItem,
+    WorkOutcome,
+    apply_extraction,
+    extract_response,
+)
+
+__all__ = [
+    "ChatResponse",
+    "ChatUsage",
+    "ContentKind",
+    "DedupStats",
+    "DeepReport",
+    "DescribeResponse",
+    "Extraction",
+    "PipelineReport",
+    "QueryFrontier",
+    "ResearchEngine",
+    "ResearchPipeline",
+    "ResearchReport",
+    "RoundSummary",
+    "RsearchClient",
+    "RsearchError",
+    "ResearchConfig",
+    "SearchGrade",
+    "SearchResponse",
+    "SearchResult",
+    "TTLCache",
+    "WorkItem",
+    "WorkOutcome",
+    "apply_extraction",
+    "extract_response",
+    "fingerprint_text",
+    "normalize_url",
+    "query_variants_from_result",
+]
+
+
+
+
+
+

src/typosaurus_sandbox/research/__main__.py

@@ -0,0 +1,42 @@
+# retoor <retoor@molodetz.nl>
+
+import argparse
+import asyncio
+import json
+import logging
+
+from typosaurus_sandbox.core.logging import setup_logging
+from typosaurus_sandbox.research.client import RsearchClient
+from typosaurus_sandbox.research.config import ResearchConfig
+from typosaurus_sandbox.research.engine import ResearchEngine
+
+logger = logging.getLogger(__name__)
+
+
+def _enable_console_logging() -> None:
+    root = logging.getLogger()
+    console = logging.StreamHandler()
+    console.setLevel(logging.INFO)
+    console.setFormatter(logging.Formatter("%(asctime)s [%(levelname)s] %(name)s: %(message)s"))
+    root.addHandler(console)
+
+
+def main(argv: list[str] | None = None) -> None:
+    setup_logging()
+    _enable_console_logging()
+    parser = argparse.ArgumentParser(
+        prog="typosaurus-sandbox-research",
+        description="Exhaustive deep research over the rsearch API until closure",
+    )
+    parser.add_argument("subject", nargs="?", default="typosaurus sandbox", help="subject to research until closure")
+    args = parser.parse_args(argv)
+    config = ResearchConfig.load()
+    client = RsearchClient(config)
+    logger.info("research session starting subject=%r base_url=%s", args.subject, config.base_url)
+    report = asyncio.run(ResearchEngine(client=client).run(args.subject))
+    logger.info("research report %s", json.dumps(report.to_dict(), indent=2))
+
+
+if __name__ == "__main__":
+    main()
+

src/typosaurus_sandbox/research/cache.py

@@ -0,0 +1,50 @@
+# retoor <retoor@molodetz.nl>
+
+import logging
+import threading
+import time
+from dataclasses import dataclass
+from typing import Generic, TypeVar
+
+logger = logging.getLogger(__name__)
+
+T = TypeVar("T")
+
+
+@dataclass
+class CacheEntry(Generic[T]):
+    value: T
+    expires_at: float
+
+
+class TTLCache(Generic[T]):
+    def __init__(self, name: str, ttl_seconds: float) -> None:
+        self._name = name
+        self._ttl_seconds = ttl_seconds
+        self._entries: dict[str, CacheEntry[T]] = {}
+        self._lock = threading.Lock()
+
+    def get(self, key: str) -> T | None:
+        with self._lock:
+            entry = self._entries.get(key)
+            if entry is None:
+                logger.debug("cache %s miss key=%s", self._name, key)
+                return None
+            if time.monotonic() >= entry.expires_at:
+                del self._entries[key]
+                logger.debug("cache %s expired key=%s", self._name, key)
+                return None
+            logger.debug("cache %s hit key=%s", self._name, key)
+            return entry.value
+
+    def set(self, key: str, value: T) -> None:
+        with self._lock:
+            self._entries[key] = CacheEntry(value=value, expires_at=time.monotonic() + self._ttl_seconds)
+            logger.debug("cache %s set key=%s ttl=%.0fs", self._name, key, self._ttl_seconds)
+
+    def clear(self) -> None:
+        with self._lock:
+            count = len(self._entries)
+            self._entries.clear()
+            logger.debug("cache %s cleared %d entries", self._name, count)
+

src/typosaurus_sandbox/research/client.py

@@ -0,0 +1,251 @@
+# retoor <retoor@molodetz.nl>
+
+import asyncio
+import hashlib
+import json
+import logging
+import secrets
+import urllib.error
+import urllib.parse
+import urllib.request
+from typing import Any
+
+from typosaurus_sandbox.research.cache import TTLCache
+from typosaurus_sandbox.research.config import ResearchConfig
+from typosaurus_sandbox.research.envelopes import ChatResponse, DescribeResponse, SearchResponse
+
+logger = logging.getLogger(__name__)
+
+MAX_ERROR_LENGTH = 200
+
+
+class RsearchError(RuntimeError):
+    def __init__(self, message: str, status_code: int | None = None) -> None:
+        super().__init__(message)
+        self.status_code = status_code
+
+
+def _multipart_body(field_name: str, filename: str, mime_type: str, payload: bytes) -> tuple[bytes, str]:
+    boundary = "----rsearch-" + secrets.token_hex(8)
+    head = (
+        f"--{boundary}\r\n".encode()
+        + f'Content-Disposition: form-data; name="{field_name}"; filename="{filename}"\r\n'.encode()
+        + f"Content-Type: {mime_type}\r\n\r\n".encode()
+    )
+    tail = b"\r\n--" + boundary.encode() + b"--\r\n"
+    return head + payload + tail, f"multipart/form-data; boundary={boundary}"
+
+
+def _content_hash(image_bytes: bytes) -> str:
+    return hashlib.sha256(image_bytes).hexdigest()
+
+
+def _search_params(
+    query: str,
+    *,
+    source: str | None,
+    count: int | None,
+    content: bool,
+    type: str | None,
+    deep: bool,
+    ai: bool,
+    cache: bool,
+) -> dict[str, str]:
+    params: dict[str, str] = {"query": query}
+    if source is not None:
+        params["source"] = source
+    if count is not None:
+        params["count"] = str(count)
+    if content:
+        params["content"] = "true"
+    if type is not None:
+        params["type"] = type
+    if deep:
+        params["deep"] = "true"
+    if ai:
+        params["ai"] = "true"
+    if not cache:
+        params["cache"] = "false"
+    return params
+
+
+class RsearchClient:
+    def __init__(self, config: ResearchConfig | None = None) -> None:
+        self._config = config if config is not None else ResearchConfig()
+        self._search_cache = TTLCache[SearchResponse]("search", self._config.search_cache_ttl_seconds)
+        self._content_cache = TTLCache[str]("content", self._config.content_cache_ttl_seconds)
+        self._describe_cache = TTLCache[DescribeResponse]("describe", self._config.content_cache_ttl_seconds)
+
+    @property
+    def config(self) -> ResearchConfig:
+        return self._config
+
+    def get_cached_content(self, url: str) -> str | None:
+        return self._content_cache.get(url)
+
+    async def search(
+        self,
+        query: str,
+        *,
+        source: str | None = None,
+        count: int | None = None,
+        content: bool = False,
+        type: str | None = None,
+        deep: bool = False,
+        ai: bool = False,
+        cache: bool = True,
+    ) -> SearchResponse:
+        params = _search_params(query, source=source, count=count, content=content, type=type, deep=deep, ai=ai, cache=cache)
+        key = urllib.parse.urlencode(sorted(params.items()))
+        if cache:
+            cached_response = self._search_cache.get(key)
+            if cached_response is not None:
+                return cached_response
+        timeout = self._config.deep_timeout_seconds if deep else self._config.request_timeout_seconds
+        status, data = await asyncio.to_thread(self._request, "GET", "/search", params, None, None, timeout)
+        response = SearchResponse.from_dict(data)
+        if cache:
+            self._search_cache.set(key, response)
+        if content:
+            for result in response.results:
+                if result.content:
+                    self._content_cache.set(result.url, result.content)
+        logger.info(
+            "search query=%r source=%s count=%s deep=%s ai=%s results=%d",
+            query,
+            response.source,
+            response.count,
+            deep,
+            ai,
+            len(response.results),
+        )
+        return response
+
+    def search_cached(
+        self,
+        query: str,
+        *,
+        source: str | None = None,
+        count: int | None = None,
+        content: bool = False,
+        type: str | None = None,
+        deep: bool = False,
+        ai: bool = False,
+        cache: bool = True,
+    ) -> SearchResponse | None:
+        if not cache:
+            return None
+        params = _search_params(query, source=source, count=count, content=content, type=type, deep=deep, ai=ai, cache=cache)
+        key = urllib.parse.urlencode(sorted(params.items()))
+        return self._search_cache.get(key)
+
+    def describe_cached(self, url: str) -> DescribeResponse | None:
+        return self._describe_cache.get(f"url:{url}")
+
+    async def chat(
+        self,
+        prompt: str,
+        *,
+        system: str | None = None,
+        json_mode: bool = False,
+        cache: bool = True,
+    ) -> ChatResponse:
+        payload: dict[str, Any] = {"prompt": prompt}
+        if system is not None:
+            payload["system"] = system
+        if json_mode:
+            payload["json"] = True
+        if not cache:
+            payload["cache"] = False
+        body = json.dumps(payload).encode()
+        headers = {"Content-Type": "application/json"}
+        status, data = await asyncio.to_thread(self._request, "POST", "/chat", None, body, headers, None)
+        response = ChatResponse.from_dict(data)
+        logger.info("chat prompt=%r cached=%s", prompt, response.cached)
+        return response
+
+    async def describe(self, url: str) -> DescribeResponse:
+        key = f"url:{url}"
+        cached = self._describe_cache.get(key)
+        if cached is not None:
+            return cached
+        status, data = await asyncio.to_thread(self._request, "GET", "/describe", {"url": url}, None, None, None)
+        response = DescribeResponse.from_dict(data)
+        self._describe_cache.set(key, response)
+        logger.info("describe url=%s", url)
+        return response
+
+    async def describe_upload(self, image_bytes: bytes, *, filename: str, mime_type: str) -> DescribeResponse:
+        body, content_type = _multipart_body("file", filename, mime_type, image_bytes)
+        headers = {"Content-Type": content_type}
+        return await self._describe_post(image_bytes, body, headers)
+
+    async def describe_raw(self, image_bytes: bytes, *, mime_type: str) -> DescribeResponse:
+        headers = {"Content-Type": mime_type}
+        return await self._describe_post(image_bytes, image_bytes, headers)
+
+    async def _describe_post(self, image_bytes: bytes, body: bytes, headers: dict[str, str]) -> DescribeResponse:
+        key = "hash:" + _content_hash(image_bytes)
+        cached = self._describe_cache.get(key)
+        if cached is not None:
+            return cached
+        status, data = await asyncio.to_thread(self._request, "POST", "/describe", None, body, headers, None)
+        response = DescribeResponse.from_dict(data)
+        self._describe_cache.set(key, response)
+        logger.info("describe post size=%d", len(image_bytes))
+        return response
+
+    @staticmethod
+    def _error_message(data: dict[str, Any]) -> str:
+        error = data.get("error")
+        if isinstance(error, str) and error:
+            return error
+        detail = data.get("detail")
+        if isinstance(detail, str) and detail:
+            return detail
+        title = data.get("title")
+        if isinstance(title, str) and title:
+            return title
+        return json.dumps(data)[:MAX_ERROR_LENGTH]
+
+    def _request(
+        self,
+        method: str,
+        path: str,
+        params: dict[str, str] | None = None,
+        payload: bytes | None = None,
+        headers: dict[str, str] | None = None,
+        timeout: float | None = None,
+    ) -> tuple[int, dict[str, Any]]:
+        timeout_seconds = timeout if timeout is not None else self._config.request_timeout_seconds
+        base_url = self._config.base_url
+        if base_url.endswith("/"):
+            base_url = base_url[:-1]
+        url = base_url + path
+        if params:
+            url = url + "?" + urllib.parse.urlencode(params)
+        request = urllib.request.Request(url, data=payload, method=method, headers=headers or {})
+        try:
+            with urllib.request.urlopen(request, timeout=timeout_seconds) as response:
+                status = response.status
+                body = response.read()
+        except urllib.error.HTTPError as exc:
+            status = exc.code
+            body = exc.read()
+        except urllib.error.URLError as exc:
+            raise RsearchError(f"connection failure for {method} {path}: {exc.reason}") from exc
+        if not body:
+            raise RsearchError(f"empty response for {method} {path}", status)
+        try:
+            data = json.loads(body)
+        except (json.JSONDecodeError, UnicodeDecodeError) as exc:
+            raise RsearchError(f"invalid JSON for {method} {path}: {exc}", status) from exc
+        if not isinstance(data, dict):
+            raise RsearchError(f"unexpected response shape for {method} {path}", status)
+        if status >= 400 or data.get("success") is False:
+            raise RsearchError(self._error_message(data), status)
+        return status, data
+
+
+
+

src/typosaurus_sandbox/research/config.py

@@ -0,0 +1,40 @@
+# retoor <retoor@molodetz.nl>
+
+import json
+import logging
+from dataclasses import dataclass
+from pathlib import Path
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class ResearchConfig:
+    base_url: str = "https://rsearch.app.molodetz.nl"
+    request_timeout_seconds: float = 30.0
+    deep_timeout_seconds: float = 180.0
+    search_cache_ttl_seconds: float = 300.0
+    content_cache_ttl_seconds: float = 86400.0
+    max_concurrency: int = 8
+    default_count: int = 10
+
+    @classmethod
+    def load(cls) -> "ResearchConfig":
+        config_path = Path(".env.json")
+        if not config_path.exists():
+            logger.info("no .env.json found, using default research config")
+            return cls()
+        with config_path.open() as f:
+            data = json.load(f)
+        research = data.get("research", {})
+        logger.info("loaded research config from .env.json")
+        return cls(
+            base_url=research.get("base_url", cls.base_url),
+            request_timeout_seconds=research.get("request_timeout_seconds", cls.request_timeout_seconds),
+            deep_timeout_seconds=research.get("deep_timeout_seconds", cls.deep_timeout_seconds),
+            search_cache_ttl_seconds=research.get("search_cache_ttl_seconds", cls.search_cache_ttl_seconds),
+            content_cache_ttl_seconds=research.get("content_cache_ttl_seconds", cls.content_cache_ttl_seconds),
+            max_concurrency=research.get("max_concurrency", cls.max_concurrency),
+            default_count=research.get("default_count", cls.default_count),
+        )
+

src/typosaurus_sandbox/research/engine.py

@@ -0,0 +1,210 @@
+# retoor <retoor@molodetz.nl>
+
+import logging
+from dataclasses import dataclass, field
+from typing import AsyncIterator
+
+from typosaurus_sandbox.research.client import RsearchClient
+from typosaurus_sandbox.research.config import ResearchConfig
+from typosaurus_sandbox.research.frontier import QueryFrontier
+from typosaurus_sandbox.research.pipeline import PipelineReport, ResearchPipeline, WorkItem
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class RoundSummary:
+    number: int = 0
+    items_processed: int = 0
+    requests_succeeded: int = 0
+    requests_failed: int = 0
+    cache_hits: int = 0
+    content_types: dict[str, int] = field(default_factory=dict)
+    new_urls: int = 0
+    new_queries: int = 0
+    new_contents: int = 0
+    closed: bool = False
+
+    def to_dict(self) -> dict[str, int | dict[str, int] | bool]:
+        return {
+            "number": self.number,
+            "items_processed": self.items_processed,
+            "requests_succeeded": self.requests_succeeded,
+            "requests_failed": self.requests_failed,
+            "cache_hits": self.cache_hits,
+            "content_types": self.content_types,
+            "new_urls": self.new_urls,
+            "new_queries": self.new_queries,
+            "new_contents": self.new_contents,
+            "closed": self.closed,
+        }
+
+
+@dataclass
+class ResearchReport:
+    subject: str
+    rounds: list[RoundSummary] = field(default_factory=list)
+    total_rounds: int = 0
+    queries_generated: int = 0
+    queries_enqueued: int = 0
+    queries_issued: int = 0
+    queries_duplicates_skipped: int = 0
+    urls_collected: int = 0
+    urls_duplicates_skipped: int = 0
+    contents_seen: int = 0
+    content_duplicates_skipped: int = 0
+    content_types: dict[str, int] = field(default_factory=dict)
+    requests_succeeded: int = 0
+    requests_failed: int = 0
+    cache_hits: int = 0
+    cache_misses: int = 0
+    closed: bool = False
+
+    def to_dict(self) -> dict[str, object]:
+        return {
+            "subject": self.subject,
+            "rounds": [round_summary.to_dict() for round_summary in self.rounds],
+            "total_rounds": self.total_rounds,
+            "queries_generated": self.queries_generated,
+            "queries_enqueued": self.queries_enqueued,
+            "queries_issued": self.queries_issued,
+            "queries_duplicates_skipped": self.queries_duplicates_skipped,
+            "urls_collected": self.urls_collected,
+            "urls_duplicates_skipped": self.urls_duplicates_skipped,
+            "contents_seen": self.contents_seen,
+            "content_duplicates_skipped": self.content_duplicates_skipped,
+            "content_types": self.content_types,
+            "requests_succeeded": self.requests_succeeded,
+            "requests_failed": self.requests_failed,
+            "cache_hits": self.cache_hits,
+            "cache_misses": self.cache_misses,
+            "closed": self.closed,
+        }
+
+
+class ResearchEngine:
+    def __init__(
+        self,
+        client: RsearchClient | None = None,
+        frontier: QueryFrontier | None = None,
+        pipeline: ResearchPipeline | None = None,
+    ) -> None:
+        self._client = client if client is not None else RsearchClient()
+        self._config: ResearchConfig = self._client.config
+        self._frontier = frontier if frontier is not None else QueryFrontier()
+        self._pipeline = pipeline if pipeline is not None else ResearchPipeline(self._client, self._frontier)
+        self._described_marker = 0
+
+    @property
+    def frontier(self) -> QueryFrontier:
+        return self._frontier
+
+    @property
+    def pipeline(self) -> ResearchPipeline:
+        return self._pipeline
+
+    async def _round_items(self, pending_queries: int, urls_to_describe: list[str]) -> AsyncIterator[WorkItem]:
+        for _ in range(pending_queries):
+            query = self._frontier.pop_query()
+            if query is None:
+                break
+            yield WorkItem("web", query, deep=True, ai=True)
+            yield WorkItem("images", query)
+            yield WorkItem("chat", query)
+        for url in urls_to_describe:
+            yield WorkItem("describe", url)
+
+    @staticmethod
+    def _round_summary(number: int, pipeline_report: PipelineReport) -> RoundSummary:
+        summary = RoundSummary(number=number, items_processed=len(pipeline_report.outcomes))
+        for outcome in pipeline_report.outcomes:
+            if outcome.success:
+                summary.requests_succeeded += 1
+            else:
+                summary.requests_failed += 1
+            if outcome.cache_hit:
+                summary.cache_hits += 1
+            kind = outcome.item.kind
+            summary.content_types[kind] = summary.content_types.get(kind, 0) + 1
+        return summary
+
+    async def run(self, subject: str) -> ResearchReport:
+        cleaned_subject = " ".join(subject.split())
+        if not cleaned_subject:
+            raise ValueError("research subject must not be empty")
+        self._frontier.seed(cleaned_subject)
+        report = ResearchReport(subject=cleaned_subject)
+        round_number = 0
+        while True:
+            round_start = self._frontier.snapshot()
+            pending_queries = round_start.queries_enqueued - round_start.queries_issued
+            urls_to_describe = self._frontier.urls_since(self._described_marker)
+            self._described_marker = round_start.urls_seen
+            if pending_queries == 0 and not urls_to_describe:
+                logger.info("research closed, no pending queries or urls after round %d", round_number)
+                break
+            round_number += 1
+            logger.info(
+                "round %d start pending_queries=%d urls_to_describe=%d",
+                round_number,
+                pending_queries,
+                len(urls_to_describe),
+            )
+            pipeline_report = await self._pipeline.run(self._round_items(pending_queries, urls_to_describe))
+            summary = self._round_summary(round_number, pipeline_report)
+            round_end = self._frontier.snapshot()
+            summary.new_urls = round_end.urls_seen - round_start.urls_seen
+            summary.new_queries = round_end.queries_enqueued - round_start.queries_enqueued
+            summary.new_contents = round_end.content_seen - round_start.content_seen
+            summary.closed = (
+                summary.new_urls == 0
+                and summary.new_queries == 0
+                and summary.requests_failed == 0
+            )
+            report.rounds.append(summary)
+            logger.info(
+                "round %d finished new_urls=%d new_queries=%d new_contents=%d closed=%s",
+                round_number,
+                summary.new_urls,
+                summary.new_queries,
+                summary.new_contents,
+                summary.closed,
+            )
+            if summary.closed:
+                break
+        report.total_rounds = round_number
+        self._finalize(report)
+        report.closed = report.requests_failed == 0
+        logger.info(
+            "research complete subject=%r rounds=%d queries_issued=%d urls_collected=%d contents_seen=%d cache_hits=%d closed=%s",
+            report.subject,
+            report.total_rounds,
+            report.queries_issued,
+            report.urls_collected,
+            report.contents_seen,
+            report.cache_hits,
+            report.closed,
+        )
+        return report
+
+    def _finalize(self, report: ResearchReport) -> None:
+        stats = self._frontier.snapshot()
+        report.queries_generated = stats.queries_generated
+        report.queries_enqueued = stats.queries_enqueued
+        report.queries_issued = stats.queries_issued
+        report.queries_duplicates_skipped = stats.queries_duplicates_skipped
+        report.urls_collected = stats.urls_seen
+        report.urls_duplicates_skipped = stats.urls_duplicates_skipped
+        report.contents_seen = stats.content_seen
+        report.content_duplicates_skipped = stats.content_duplicates_skipped
+        total_items = 0
+        for summary in report.rounds:
+            total_items += summary.items_processed
+            report.requests_succeeded += summary.requests_succeeded
+            report.requests_failed += summary.requests_failed
+            report.cache_hits += summary.cache_hits
+            for kind, count in summary.content_types.items():
+                report.content_types[kind] = report.content_types.get(kind, 0) + count
+        report.cache_misses = total_items - report.cache_hits
+
+

src/typosaurus_sandbox/research/envelopes.py

@@ -0,0 +1,203 @@
+# retoor <retoor@molodetz.nl>
+
+from dataclasses import dataclass, field
+from typing import Any
+
+
+def _as_float(value: Any) -> float | None:
+    if value is None:
+        return None
+    try:
+        return float(value)
+    except (TypeError, ValueError):
+        return None
+
+
+@dataclass
+class SearchGrade:
+    overall: float = 0.0
+    relevance: float = 0.0
+    depth: float = 0.0
+    authority: float = 0.0
+    freshness: float = 0.0
+    word_count: int = 0
+    intent_hits: int = 0
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any] | None) -> "SearchGrade | None":
+        if data is None:
+            return None
+        return cls(
+            overall=float(data.get("overall", 0.0) or 0.0),
+            relevance=float(data.get("relevance", 0.0) or 0.0),
+            depth=float(data.get("depth", 0.0) or 0.0),
+            authority=float(data.get("authority", 0.0) or 0.0),
+            freshness=float(data.get("freshness", 0.0) or 0.0),
+            word_count=int(data.get("word_count", 0) or 0),
+            intent_hits=int(data.get("intent_hits", 0) or 0),
+        )
+
+
+@dataclass
+class SearchResult:
+    title: str = ""
+    url: str = ""
+    description: str = ""
+    source: str = ""
+    content: str | None = None
+    extra: dict[str, Any] = field(default_factory=dict)
+    index: int | None = None
+    grade: SearchGrade | None = None
+    query_origin: str | None = None
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> "SearchResult":
+        return cls(
+            title=data.get("title", ""),
+            url=data.get("url", ""),
+            description=data.get("description", ""),
+            source=data.get("source", ""),
+            content=data.get("content"),
+            extra=data.get("extra", {}),
+            index=data.get("index"),
+            grade=SearchGrade.from_dict(data.get("grade")),
+            query_origin=data.get("query_origin"),
+        )
+
+
+@dataclass
+class DeepReport:
+    query: str = ""
+    markdown: str = ""
+    sources: list[SearchResult] = field(default_factory=list)
+    graded_count: int = 0
+    total_count: int = 0
+    model: str = ""
+    elapsed: float = 0.0
+    cache_hit: bool = False
+    rounds: int = 0
+    queries_tried: list[str] = field(default_factory=list)
+    error: str | None = None
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any] | None) -> "DeepReport | None":
+        if data is None:
+            return None
+        sources = [SearchResult.from_dict(item) for item in data.get("sources", [])]
+        return cls(
+            query=data.get("query", ""),
+            markdown=data.get("markdown", ""),
+            sources=sources,
+            graded_count=int(data.get("graded_count", 0) or 0),
+            total_count=int(data.get("total_count", 0) or 0),
+            model=data.get("model", ""),
+            elapsed=_as_float(data.get("elapsed")) or 0.0,
+            cache_hit=bool(data.get("cache_hit", False)),
+            rounds=int(data.get("rounds", 0) or 0),
+            queries_tried=list(data.get("queries_tried", [])),
+            error=data.get("error"),
+        )
+
+
+@dataclass
+class SearchResponse:
+    query: str = ""
+    source: str = ""
+    count: int = 0
+    results: list[SearchResult] = field(default_factory=list)
+    success: bool = False
+    error: str | None = None
+    ai_response: str | None = None
+    ai_error: str | None = None
+    deep: DeepReport | None = None
+    timestamp: str | None = None
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> "SearchResponse":
+        results = [SearchResult.from_dict(item) for item in data.get("results", [])]
+        return cls(
+            query=data.get("query", ""),
+            source=data.get("source", ""),
+            count=int(data.get("count", 0) or 0),
+            results=results,
+            success=bool(data.get("success", False)),
+            error=data.get("error"),
+            ai_response=data.get("ai_response"),
+            ai_error=data.get("ai_error"),
+            deep=DeepReport.from_dict(data.get("deep")),
+            timestamp=data.get("timestamp"),
+        )
+
+
+@dataclass
+class ChatUsage:
+    prompt_tokens: int = 0
+    completion_tokens: int = 0
+    total_tokens: int = 0
+    cost_usd: float = 0.0
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any] | None) -> "ChatUsage | None":
+        if data is None:
+            return None
+        return cls(
+            prompt_tokens=int(data.get("prompt_tokens", 0) or 0),
+            completion_tokens=int(data.get("completion_tokens", 0) or 0),
+            total_tokens=int(data.get("total_tokens", 0) or 0),
+            cost_usd=float(data.get("cost_usd", 0.0) or 0.0),
+        )
+
+
+@dataclass
+class ChatResponse:
+    response: str = ""
+    prompt: str = ""
+    json_mode: bool = False
+    cached: bool = False
+    usage: ChatUsage | None = None
+    error: str | None = None
+    max_context_window: int | None = None
+    max_output_tokens: int | None = None
+    elapsed: float | None = None
+    timestamp: str | None = None
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> "ChatResponse":
+        return cls(
+            response=data.get("response", ""),
+            prompt=data.get("prompt", ""),
+            json_mode=bool(data.get("json_mode", False)),
+            cached=bool(data.get("cached", False)),
+            usage=ChatUsage.from_dict(data.get("usage")),
+            error=data.get("error"),
+            max_context_window=data.get("max_context_window"),
+            max_output_tokens=data.get("max_output_tokens"),
+            elapsed=_as_float(data.get("elapsed")),
+            timestamp=data.get("timestamp"),
+        )
+
+
+@dataclass
+class DescribeResponse:
+    description: str = ""
+    url: str | None = None
+    mime_type: str | None = None
+    size: int | None = None
+    elapsed: float | None = None
+    timestamp: str | None = None
+    success: bool = True
+    error: str | None = None
+
+    @classmethod
+    def from_dict(cls, data: dict[str, Any]) -> "DescribeResponse":
+        return cls(
+            description=data.get("description", ""),
+            url=data.get("url"),
+            mime_type=data.get("mime_type"),
+            size=data.get("size"),
+            elapsed=_as_float(data.get("elapsed")),
+            timestamp=data.get("timestamp"),
+            success=bool(data.get("success", True)),
+            error=data.get("error"),
+        )
+

src/typosaurus_sandbox/research/frontier.py

@@ -0,0 +1,234 @@
+# retoor <retoor@molodetz.nl>
+
+import asyncio
+import hashlib
+import logging
+import re
+import threading
+import urllib.parse
+from dataclasses import dataclass
+
+from typosaurus_sandbox.research.envelopes import SearchResult
+
+logger = logging.getLogger(__name__)
+
+MIN_QUERY_LENGTH = 2
+MAX_QUERY_LENGTH = 200
+DEFAULT_PORTS: dict[str, int] = {"http": 80, "https": 443}
+
+
+def _clean_text(value: str) -> str:
+    return " ".join(value.split())
+
+
+def _query_key(query: str) -> str:
+    return _clean_text(query).casefold()
+
+
+def normalize_url(url: str) -> str:
+    cleaned = _clean_text(url)
+    try:
+        parsed = urllib.parse.urlsplit(cleaned)
+    except ValueError:
+        return cleaned
+    scheme = parsed.scheme.lower()
+    if scheme not in DEFAULT_PORTS:
+        return cleaned
+    host = (parsed.hostname or "").lower()
+    if not host:
+        return cleaned
+    try:
+        host = host.encode("idna").decode("ascii")
+    except UnicodeError:
+        pass
+    port: int | None = None
+    try:
+        port = parsed.port
+    except ValueError:
+        port = None
+    if port is not None and DEFAULT_PORTS.get(scheme) == port:
+        port = None
+    display_host = f"[{host}]" if ":" in host else host
+    netloc = display_host if port is None else f"{display_host}:{port}"
+    path = re.sub(r"/{2,}", "/", parsed.path)
+    if len(path) > 1 and path.endswith("/"):
+        path = path[:-1]
+    if parsed.query:
+        return f"{scheme}://{netloc}{path}?{parsed.query}"
+    return f"{scheme}://{netloc}{path}"
+
+
+def fingerprint_text(text: str) -> str:
+    normalized = _clean_text(text)
+    return hashlib.sha256(normalized.encode("utf-8")).hexdigest()
+
+
+def query_variants_from_result(result: SearchResult) -> list[tuple[str, str]]:
+    variants: list[tuple[str, str]] = []
+    if result.title:
+        variants.append((result.title, "title"))
+    if result.description:
+        variants.append((result.description, "description"))
+    for value in result.extra.values():
+        if isinstance(value, str) and value:
+            variants.append((value, "extra"))
+    return variants
+
+
+@dataclass(frozen=True)
+class DedupStats:
+    queries_generated: int = 0
+    queries_enqueued: int = 0
+    queries_issued: int = 0
+    queries_duplicates_skipped: int = 0
+    urls_seen: int = 0
+    urls_duplicates_skipped: int = 0
+    content_seen: int = 0
+    content_duplicates_skipped: int = 0
+
+    def to_dict(self) -> dict[str, int]:
+        return {
+            "queries_generated": self.queries_generated,
+            "queries_enqueued": self.queries_enqueued,
+            "queries_issued": self.queries_issued,
+            "queries_duplicates_skipped": self.queries_duplicates_skipped,
+            "urls_seen": self.urls_seen,
+            "urls_duplicates_skipped": self.urls_duplicates_skipped,
+            "content_seen": self.content_seen,
+            "content_duplicates_skipped": self.content_duplicates_skipped,
+        }
+
+
+class QueryFrontier:
+    def __init__(self, subject: str | None = None) -> None:
+        self._lock = threading.Lock()
+        self._seen_queries: set[str] = set()
+        self._seen_urls: set[str] = set()
+        self._seen_url_order: list[str] = []
+        self._seen_content: set[str] = set()
+        self._origins: dict[str, str] = {}
+        self._pending: asyncio.Queue[str] = asyncio.Queue()
+        self._queries_generated = 0
+        self._queries_enqueued = 0
+        self._queries_issued = 0
+        self._queries_duplicates_skipped = 0
+        self._urls_seen = 0
+        self._urls_duplicates_skipped = 0
+        self._content_seen = 0
+        self._content_duplicates_skipped = 0
+        if subject:
+            self.seed(subject)
+
+    def seed(self, subject: str) -> None:
+        cleaned = _clean_text(subject)
+        if cleaned:
+            self.push_query(cleaned, "seed")
+            logger.info("frontier seeded subject=%r", cleaned)
+
+    def push_query(self, query: str, origin: str = "manual") -> bool:
+        cleaned = _clean_text(query)
+        if not MIN_QUERY_LENGTH <= len(cleaned) <= MAX_QUERY_LENGTH:
+            logger.debug("query variant invalid length=%d query=%r", len(cleaned), cleaned)
+            return False
+        key = _query_key(cleaned)
+        with self._lock:
+            self._queries_generated += 1
+            if key in self._seen_queries:
+                self._queries_duplicates_skipped += 1
+                logger.debug("query duplicate skipped origin=%s query=%r", origin, cleaned)
+                return False
+            self._seen_queries.add(key)
+            self._origins[key] = origin
+            self._queries_enqueued += 1
+        self._pending.put_nowait(cleaned)
+        logger.info("query enqueued origin=%s query=%r", origin, cleaned)
+        return True
+
+    def push_variants_from_result(self, result: SearchResult) -> int:
+        new_queries = 0
+        for text, origin in query_variants_from_result(result):
+            if self.push_query(text, origin):
+                new_queries += 1
+        return new_queries
+
+    def register_url(self, url: str) -> bool:
+        if not url:
+            return False
+        normalized = normalize_url(url)
+        with self._lock:
+            if normalized in self._seen_urls:
+                self._urls_duplicates_skipped += 1
+                logger.debug("url duplicate skipped url=%s", normalized)
+                return False
+            self._seen_urls.add(normalized)
+            self._seen_url_order.append(normalized)
+            self._urls_seen += 1
+        logger.info("url registered url=%s", normalized)
+        return True
+
+    def urls_since(self, seen_count: int) -> list[str]:
+        with self._lock:
+            return list(self._seen_url_order[seen_count:])
+
+    def register_content(self, text: str) -> bool:
+        if not text.strip():
+            return False
+        fingerprint = fingerprint_text(text)
+        with self._lock:
+            if fingerprint in self._seen_content:
+                self._content_duplicates_skipped += 1
+                logger.debug("content duplicate skipped fingerprint=%s", fingerprint)
+                return False
+            self._seen_content.add(fingerprint)
+            self._content_seen += 1
+        logger.info("content registered fingerprint=%s", fingerprint)
+        return True
+
+    def register_result(self, result: SearchResult) -> bool:
+        is_new = self.register_url(result.url)
+        if result.content:
+            self.register_content(result.content)
+        return is_new
+
+    async def get_query(self) -> str:
+        query = await self._pending.get()
+        with self._lock:
+            self._queries_issued += 1
+        logger.info("query issued query=%r", query)
+        return query
+
+    def pop_query(self) -> str | None:
+        try:
+            query = self._pending.get_nowait()
+        except asyncio.QueueEmpty:
+            return None
+        with self._lock:
+            self._queries_issued += 1
+        logger.info("query issued query=%r", query)
+        return query
+
+    def pending_count(self) -> int:
+        return self._pending.qsize()
+
+    def has_pending(self) -> bool:
+        return not self._pending.empty()
+
+    def origin_of(self, query: str) -> str | None:
+        with self._lock:
+            return self._origins.get(_query_key(query))
+
+    def snapshot(self) -> DedupStats:
+        with self._lock:
+            return DedupStats(
+                queries_generated=self._queries_generated,
+                queries_enqueued=self._queries_enqueued,
+                queries_issued=self._queries_issued,
+                queries_duplicates_skipped=self._queries_duplicates_skipped,
+                urls_seen=self._urls_seen,
+                urls_duplicates_skipped=self._urls_duplicates_skipped,
+                content_seen=self._content_seen,
+                content_duplicates_skipped=self._content_duplicates_skipped,
+            )
+
+
+

src/typosaurus_sandbox/research/pipeline.py

@@ -0,0 +1,345 @@
+# retoor <retoor@molodetz.nl>
+
+import asyncio
+import logging
+import re
+from dataclasses import dataclass, field
+from typing import AsyncIterator, Literal
+
+from typosaurus_sandbox.research.client import RsearchClient, RsearchError
+from typosaurus_sandbox.research.config import ResearchConfig
+from typosaurus_sandbox.research.envelopes import ChatResponse, DescribeResponse, SearchResponse
+from typosaurus_sandbox.research.frontier import QueryFrontier, query_variants_from_result
+
+logger = logging.getLogger(__name__)
+
+ContentKind = Literal["web", "images", "describe", "chat"]
+
+URL_PATTERN = re.compile(r"https?://[^\s<>\"']+")
+
+RETRY_MAX_ATTEMPTS = 3
+RETRY_BACKOFF_BASE_SECONDS = 0.5
+RETRY_BACKOFF_MAX_SECONDS = 8.0
+TRANSIENT_STATUS_MIN = 500
+
+
+@dataclass(frozen=True)
+class WorkItem:
+    kind: ContentKind
+    value: str
+    deep: bool = False
+    ai: bool = False
+
+
+@dataclass(frozen=True)
+class Extraction:
+    urls: tuple[str, ...] = ()
+    query_seeds: tuple[tuple[str, str], ...] = ()
+    content_texts: tuple[str, ...] = ()
+
+
+@dataclass
+class WorkOutcome:
+    item: WorkItem
+    endpoint: str
+    success: bool
+    cache_hit: bool
+    status_code: int | None = None
+    error: str | None = None
+    urls_found: int = 0
+    queries_seeded: int = 0
+    contents_seen: int = 0
+
+
+@dataclass
+class PipelineReport:
+    outcomes: list[WorkOutcome] = field(default_factory=list)
+    requests_succeeded: int = 0
+    requests_failed: int = 0
+    urls_found: int = 0
+    queries_seeded: int = 0
+    contents_seen: int = 0
+
+
+def _urls_from_text(text: str) -> list[str]:
+    cleaned: list[str] = []
+    for match in URL_PATTERN.findall(text):
+        cleaned.append(match.rstrip(".,;:!?)]}\"'"))
+    return cleaned
+
+
+def extract_response(
+    item: WorkItem,
+    response: SearchResponse | ChatResponse | DescribeResponse,
+) -> Extraction:
+    urls: list[str] = []
+    query_seeds: list[tuple[str, str]] = []
+    content_texts: list[str] = []
+    if isinstance(response, SearchResponse):
+        for result in response.results:
+            if result.url:
+                urls.append(result.url)
+            query_seeds.extend(query_variants_from_result(result))
+            if result.content:
+                content_texts.append(result.content)
+        if response.ai_response:
+            content_texts.append(response.ai_response)
+            query_seeds.append((response.ai_response, "ai_response"))
+            urls.extend(_urls_from_text(response.ai_response))
+        if response.deep is not None:
+            for source in response.deep.sources:
+                if source.url:
+                    urls.append(source.url)
+                query_seeds.extend(query_variants_from_result(source))
+            if response.deep.markdown:
+                content_texts.append(response.deep.markdown)
+                urls.extend(_urls_from_text(response.deep.markdown))
+    elif isinstance(response, ChatResponse):
+        if response.response:
+            content_texts.append(response.response)
+            query_seeds.append((response.response, "chat"))
+            urls.extend(_urls_from_text(response.response))
+    elif isinstance(response, DescribeResponse):
+        if response.description:
+            content_texts.append(response.description)
+            query_seeds.append((response.description, "describe"))
+            urls.extend(_urls_from_text(response.description))
+    return Extraction(
+        urls=tuple(dict.fromkeys(urls)),
+        query_seeds=tuple(query_seeds),
+        content_texts=tuple(content_texts),
+    )
+
+
+def apply_extraction(frontier: QueryFrontier, extraction: Extraction) -> tuple[int, int, int]:
+    new_urls = 0
+    new_queries = 0
+    new_contents = 0
+    for url in extraction.urls:
+        if frontier.register_url(url):
+            new_urls += 1
+    for text, origin in extraction.query_seeds:
+        if frontier.push_query(text, origin):
+            new_queries += 1
+    for text in extraction.content_texts:
+        if frontier.register_content(text):
+            new_contents += 1
+    return new_urls, new_queries, new_contents
+
+
+class ResearchPipeline:
+    def __init__(self, client: RsearchClient, frontier: QueryFrontier, config: ResearchConfig | None = None) -> None:
+        self._client = client
+        self._frontier = frontier
+        self._config = config if config is not None else client.config
+        self._pool_size = max(1, self._config.max_concurrency)
+        self._semaphore = asyncio.Semaphore(self._pool_size)
+
+    @property
+    def pool_size(self) -> int:
+        return self._pool_size
+
+    @staticmethod
+    def _endpoint(item: WorkItem) -> str:
+        if item.kind in ("web", "images"):
+            return "/search"
+        if item.kind == "describe":
+            return "/describe"
+        return "/chat"
+
+    def _probe_cache(self, item: WorkItem) -> bool:
+        if item.kind == "web":
+            return (
+                self._client.search_cached(
+                    item.value,
+                    content=True,
+                    count=self._config.default_count,
+                    deep=item.deep,
+                    ai=item.ai,
+                )
+                is not None
+            )
+        if item.kind == "images":
+            return self._client.search_cached(item.value, type="images", count=self._config.default_count) is not None
+        if item.kind == "describe":
+            return self._client.describe_cached(item.value) is not None
+        return False
+
+    async def _fetch(self, item: WorkItem) -> SearchResponse | ChatResponse | DescribeResponse:
+        if item.kind == "web":
+            return await self._client.search(
+                item.value,
+                content=True,
+                count=self._config.default_count,
+                deep=item.deep,
+                ai=item.ai,
+            )
+        if item.kind == "images":
+            return await self._client.search(item.value, type="images", count=self._config.default_count)
+        if item.kind == "describe":
+            return await self._client.describe(item.value)
+        return await self._client.chat(item.value)
+
+    async def process(self, item: WorkItem) -> WorkOutcome:
+        async with self._semaphore:
+            return await self._handle(item)
+
+    async def _handle(self, item: WorkItem) -> WorkOutcome:
+        endpoint = self._endpoint(item)
+        cache_hit = self._probe_cache(item)
+        response: SearchResponse | ChatResponse | DescribeResponse | None = None
+        failure: RsearchError | None = None
+        for attempt in range(1, RETRY_MAX_ATTEMPTS + 1):
+            try:
+                response = await self._fetch(item)
+                failure = None
+                break
+            except RsearchError as exc:
+                failure = exc
+                if exc.status_code is None or exc.status_code < TRANSIENT_STATUS_MIN:
+                    break
+                if attempt == RETRY_MAX_ATTEMPTS:
+                    break
+                delay = min(RETRY_BACKOFF_BASE_SECONDS * (2 ** (attempt - 1)), RETRY_BACKOFF_MAX_SECONDS)
+                logger.warning(
+                    "transient request failure endpoint=%s kind=%s target=%r status=%s attempt=%d/%d retry_in=%.1fs",
+                    endpoint,
+                    item.kind,
+                    item.value,
+                    exc.status_code,
+                    attempt,
+                    RETRY_MAX_ATTEMPTS,
+                    delay,
+                )
+                await asyncio.sleep(delay)
+        if failure is not None:
+            logger.error(
+                "request failed endpoint=%s kind=%s target=%r status=%s cache_hit=%s error=%s",
+                endpoint,
+                item.kind,
+                item.value,
+                failure.status_code,
+                cache_hit,
+                failure,
+            )
+            return WorkOutcome(
+                item=item,
+                endpoint=endpoint,
+                success=False,
+                cache_hit=cache_hit,
+                status_code=failure.status_code,
+                error=str(failure),
+            )
+        if response is None:
+            return WorkOutcome(
+                item=item,
+                endpoint=endpoint,
+                success=False,
+                cache_hit=cache_hit,
+                status_code=None,
+                error="no response",
+            )
+        if isinstance(response, ChatResponse) and response.cached:
+            cache_hit = True
+        if isinstance(response, SearchResponse) and response.deep is not None and response.deep.cache_hit:
+            cache_hit = True
+        extraction = extract_response(item, response)
+        urls_found, queries_seeded, contents_seen = apply_extraction(self._frontier, extraction)
+        logger.debug(
+            "extraction endpoint=%s kind=%s target=%r urls=%s query_seeds=%d content_texts=%d",
+            endpoint,
+            item.kind,
+            item.value,
+            list(extraction.urls),
+            len(extraction.query_seeds),
+            len(extraction.content_texts),
+        )
+        outcome = WorkOutcome(
+            item=item,
+            endpoint=endpoint,
+            success=True,
+            cache_hit=cache_hit,
+            urls_found=urls_found,
+            queries_seeded=queries_seeded,
+            contents_seen=contents_seen,
+        )
+        logger.info(
+            "request done endpoint=%s kind=%s target=%r status=ok cache_hit=%s urls=%d queries=%d contents=%d",
+            endpoint,
+            item.kind,
+            item.value,
+            cache_hit,
+            urls_found,
+            queries_seeded,
+            contents_seen,
+        )
+        return outcome
+
+    async def run(self, item_source: AsyncIterator[WorkItem]) -> PipelineReport:
+        logger.info("worker pool size=%d max_concurrency=%d", self._pool_size, self._config.max_concurrency)
+        queue: asyncio.Queue[WorkItem | None] = asyncio.Queue(maxsize=self._pool_size * 4)
+        outcomes: list[WorkOutcome] = []
+
+        async def produce() -> None:
+            try:
+                async for item in item_source:
+                    await queue.put(item)
+            finally:
+                for _ in range(self._pool_size):
+                    await queue.put(None)
+
+        async def consume() -> None:
+            while True:
+                item = await queue.get()
+                if item is None:
+                    return
+                try:
+                    outcome = await self.process(item)
+                except Exception as exc:
+                    logger.error(
+                        "pool worker unexpected error kind=%s target=%r error=%s",
+                        item.kind,
+                        item.value,
+                        exc,
+                    )
+                    outcome = WorkOutcome(
+                        item=item,
+                        endpoint=self._endpoint(item),
+                        success=False,
+                        cache_hit=False,
+                        status_code=None,
+                        error=f"unexpected error: {exc}",
+                    )
+                outcomes.append(outcome)
+
+        producer_task = asyncio.create_task(produce())
+        worker_tasks = [asyncio.create_task(consume()) for _ in range(self._pool_size)]
+        try:
+            await producer_task
+        except Exception as exc:
+            logger.error("item source failed error=%s", exc)
+        await asyncio.gather(*worker_tasks)
+        report = self._build_report(outcomes)
+        logger.info(
+            "pipeline finished requests_succeeded=%d requests_failed=%d urls_found=%d queries_seeded=%d contents_seen=%d",
+            report.requests_succeeded,
+            report.requests_failed,
+            report.urls_found,
+            report.queries_seeded,
+            report.contents_seen,
+        )
+        return report
+
+    @staticmethod
+    def _build_report(outcomes: list[WorkOutcome]) -> PipelineReport:
+        report = PipelineReport(outcomes=outcomes)
+        for outcome in outcomes:
+            if outcome.success:
+                report.requests_succeeded += 1
+            else:
+                report.requests_failed += 1
+            report.urls_found += outcome.urls_found
+            report.queries_seeded += outcome.queries_seeded
+            report.contents_seen += outcome.contents_seen
+        return report
+

tests/test_api.py

@@ -1,6 +1,9 @@
 # retoor <retoor@molodetz.nl>
 
 import unittest
+import warnings
+
+warnings.filterwarnings("ignore", category=DeprecationWarning, module="starlette")
 
 from fastapi.testclient import TestClient
 
@@ -117,3 +120,135 @@ class TestCalculatorClampToByteEndpoint(unittest.TestCase):
     def test_clamp_to_byte_invalid_input_returns_422(self) -> None:
         response = client.post("/api/v1/calculator/clamp-to-byte", json={"value": "abc"})
         self.assertEqual(response.status_code, 422)
+
+
+class TestCalculatorAverageEndpoint(unittest.TestCase):
+
+    def test_average_positive_values(self) -> None:
+        response = client.post("/api/v1/calculator/average", json={"values": [1, 2, 3, 4, 5]})
+        self.assertEqual(response.status_code, 200)
+        self.assertEqual(response.json(), {"result": 3.0})
+
+    def test_average_single_value(self) -> None:
+        response = client.post("/api/v1/calculator/average", json={"values": [5]})
+        self.assertEqual(response.status_code, 200)
+        self.assertEqual(response.json(), {"result": 5.0})
+
+    def test_average_negative_values(self) -> None:
+        response = client.post("/api/v1/calculator/average", json={"values": [-10, -20, -30]})
+        self.assertEqual(response.status_code, 200)
+        self.assertEqual(response.json(), {"result": -20.0})
+
+    def test_average_empty_returns_422(self) -> None:
+        response = client.post("/api/v1/calculator/average", json={"values": []})
+        self.assertEqual(response.status_code, 422)
+
+    def test_average_invalid_input_returns_422(self) -> None:
+        response = client.post("/api/v1/calculator/average", json={"values": ["a", "b"]})
+        self.assertEqual(response.status_code, 422)
+
+    def test_average_missing_field_returns_422(self) -> None:
+        response = client.post("/api/v1/calculator/average", json={})
+        self.assertEqual(response.status_code, 422)
+
+
+class TestCalculatorMedianEndpoint(unittest.TestCase):
+
+    def test_median_odd_length(self) -> None:
+        response = client.post("/api/v1/calculator/median", json={"values": [1, 3, 5]})
+        self.assertEqual(response.status_code, 200)
+        self.assertEqual(response.json(), {"result": 3.0})
+
+    def test_median_even_length(self) -> None:
+        response = client.post("/api/v1/calculator/median", json={"values": [1, 2, 3, 4]})
+        self.assertEqual(response.status_code, 200)
+        self.assertEqual(response.json(), {"result": 2.5})
+
+    def test_median_single_element(self) -> None:
+        response = client.post("/api/v1/calculator/median", json={"values": [7]})
+        self.assertEqual(response.status_code, 200)
+        self.assertEqual(response.json(), {"result": 7.0})
+
+    def test_median_unsorted_input(self) -> None:
+        response = client.post("/api/v1/calculator/median", json={"values": [3, 1, 2]})
+        self.assertEqual(response.status_code, 200)
+        self.assertEqual(response.json(), {"result": 2.0})
+
+    def test_median_empty_returns_422(self) -> None:
+        response = client.post("/api/v1/calculator/median", json={"values": []})
+        self.assertEqual(response.status_code, 422)
+
+    def test_median_invalid_input_returns_422(self) -> None:
+        response = client.post("/api/v1/calculator/median", json={"values": ["a"]})
+        self.assertEqual(response.status_code, 422)
+
+    def test_median_missing_field_returns_422(self) -> None:
+        response = client.post("/api/v1/calculator/median", json={})
+        self.assertEqual(response.status_code, 422)
+
+
+class TestCalculatorVarianceEndpoint(unittest.TestCase):
+
+    def test_variance_known_set(self) -> None:
+        response = client.post("/api/v1/calculator/variance", json={"values": [1, 2, 3, 4, 5]})
+        self.assertEqual(response.status_code, 200)
+        self.assertEqual(response.json(), {"result": 2.0})
+
+    def test_variance_constant_values(self) -> None:
+        response = client.post("/api/v1/calculator/variance", json={"values": [1.0, 1.0, 1.0]})
+        self.assertEqual(response.status_code, 200)
+        self.assertEqual(response.json(), {"result": 0.0})
+
+    def test_variance_single_element(self) -> None:
+        response = client.post("/api/v1/calculator/variance", json={"values": [42.0]})
+        self.assertEqual(response.status_code, 200)
+        self.assertEqual(response.json(), {"result": 0.0})
+
+    def test_variance_empty_returns_422(self) -> None:
+        response = client.post("/api/v1/calculator/variance", json={"values": []})
+        self.assertEqual(response.status_code, 422)
+
+    def test_variance_invalid_input_returns_422(self) -> None:
+        response = client.post("/api/v1/calculator/variance", json={"values": ["a", "b", "c"]})
+        self.assertEqual(response.status_code, 422)
+
+    def test_variance_missing_field_returns_422(self) -> None:
+        response = client.post("/api/v1/calculator/variance", json={})
+        self.assertEqual(response.status_code, 422)
+
+
+class TestCalculatorPercentageEndpoint(unittest.TestCase):
+
+    def test_percentage_half(self) -> None:
+        response = client.post("/api/v1/calculator/percentage", json={"value": 50, "total": 100})
+        self.assertEqual(response.status_code, 200)
+        self.assertEqual(response.json(), {"result": 50.0})
+
+    def test_percentage_quarter(self) -> None:
+        response = client.post("/api/v1/calculator/percentage", json={"value": 25, "total": 100})
+        self.assertEqual(response.status_code, 200)
+        self.assertEqual(response.json(), {"result": 25.0})
+
+    def test_percentage_zero_value(self) -> None:
+        response = client.post("/api/v1/calculator/percentage", json={"value": 0, "total": 100})
+        self.assertEqual(response.status_code, 200)
+        self.assertEqual(response.json(), {"result": 0.0})
+
+    def test_percentage_total_zero_returns_422(self) -> None:
+        response = client.post("/api/v1/calculator/percentage", json={"value": 50, "total": 0})
+        self.assertEqual(response.status_code, 422)
+
+    def test_percentage_invalid_input_returns_422(self) -> None:
+        response = client.post("/api/v1/calculator/percentage", json={"value": "abc", "total": 100})
+        self.assertEqual(response.status_code, 422)
+
+    def test_percentage_missing_value_returns_422(self) -> None:
+        response = client.post("/api/v1/calculator/percentage", json={"total": 100})
+        self.assertEqual(response.status_code, 422)
+
+    def test_percentage_missing_total_returns_422(self) -> None:
+        response = client.post("/api/v1/calculator/percentage", json={"value": 50})
+        self.assertEqual(response.status_code, 422)
+
+
+

tests/test_calculator.py

@@ -3,7 +3,7 @@
 import math
 import unittest
 
-from typosaurus_sandbox.domain.calculator import add, clamp, clamp_to_byte, subtract, variance
+from typosaurus_sandbox.domain.calculator import add, average, clamp, clamp_to_byte, median, percentage, subtract, variance
 
 
 class TestAddFunction(unittest.TestCase):
@@ -18,6 +18,28 @@ class TestAddFunction(unittest.TestCase):
         self.assertEqual(add(-3, 5), 2)
 
 
+class TestAverageFunction(unittest.TestCase):
+
+    def test_empty_sequence_raises_value_error(self) -> None:
+        with self.assertRaises(ValueError):
+            average([])
+
+    def test_single_element(self) -> None:
+        self.assertEqual(average([5]), 5.0)
+
+    def test_positive_values(self) -> None:
+        self.assertEqual(average([1, 2, 3, 4, 5]), 3.0)
+
+    def test_negative_values(self) -> None:
+        self.assertEqual(average([-10, -20, -30]), -20.0)
+
+    def test_mixed_positive_and_negative(self) -> None:
+        self.assertEqual(average([-5, 0, 5]), 0.0)
+
+    def test_float_values(self) -> None:
+        self.assertEqual(average([1.5, 2.5, 3.0]), 7.0 / 3.0)
+
+
 class TestSubtractFunction(unittest.TestCase):
 
     def test_subtract_positive(self) -> None:
@@ -103,6 +125,32 @@ class TestClampFunction(unittest.TestCase):
         self.assertTrue(math.isnan(result))
 
 
+class TestMedianFunction(unittest.TestCase):
+
+    def test_odd_length_returns_middle_element(self) -> None:
+        self.assertEqual(median([1, 3, 5]), 3)
+
+    def test_even_length_returns_float_average_of_two_middle_values(self) -> None:
+        result = median([1, 2, 3, 4])
+        self.assertIsInstance(result, float)
+        self.assertEqual(result, 2.5)
+
+    def test_single_element_returns_that_element(self) -> None:
+        self.assertEqual(median([7]), 7)
+
+    def test_empty_list_raises_value_error(self) -> None:
+        with self.assertRaises(ValueError):
+            median([])
+
+    def test_unsorted_input_sorts_correctly(self) -> None:
+        self.assertEqual(median([3, 1, 2]), 2)
+
+    def test_unsorted_even_length_returns_float_average(self) -> None:
+        result = median([10, 30, 20, 40])
+        self.assertIsInstance(result, float)
+        self.assertEqual(result, 25.0)
+
+
 class TestVarianceFunction(unittest.TestCase):
 
     def test_empty_list_raises_value_error(self) -> None:
@@ -133,3 +181,46 @@ class TestVarianceFunction(unittest.TestCase):
     def test_tuple_input_returns_variance(self) -> None:
         self.assertEqual(variance((1, 2, 3, 4, 5)), 2.0)
 
+
+class TestPercentageFunction(unittest.TestCase):
+
+    def test_half_returns_50(self) -> None:
+        self.assertEqual(percentage(50, 100), 50.0)
+
+    def test_quarter_returns_25(self) -> None:
+        self.assertEqual(percentage(25, 100), 25.0)
+
+    def test_zero_value_returns_zero(self) -> None:
+        self.assertEqual(percentage(0, 100), 0.0)
+
+    def test_value_exceeds_total(self) -> None:
+        self.assertEqual(percentage(150, 100), 150.0)
+
+    def test_negative_value(self) -> None:
+        self.assertEqual(percentage(-50, 100), -50.0)
+
+    def test_negative_total(self) -> None:
+        self.assertEqual(percentage(50, -100), -50.0)
+
+    def test_both_negative(self) -> None:
+        self.assertEqual(percentage(-50, -100), 50.0)
+
+    def test_float_inputs(self) -> None:
+        result = percentage(33.0, 100.0)
+        self.assertAlmostEqual(result, 33.0)
+
+    def test_total_zero_raises_value_error(self) -> None:
+        with self.assertRaises(ValueError):
+            percentage(50, 0)
+
+    def test_total_zero_float_raises_value_error(self) -> None:
+        with self.assertRaises(ValueError):
+            percentage(50.0, 0.0)
+
+    def test_integer_inputs_return_float(self) -> None:
+        result = percentage(1, 4)
+        self.assertIsInstance(result, float)
+        self.assertEqual(result, 25.0)
+
+
+

tests/test_research_client.py

@@ -0,0 +1,700 @@
+# retoor <retoor@molodetz.nl>
+
+import io
+import json
+import unittest
+import urllib.error
+from typing import Any
+from unittest import mock
+
+from typosaurus_sandbox.research.client import RsearchClient, RsearchError
+from typosaurus_sandbox.research.envelopes import (
+    ChatResponse,
+    ChatUsage,
+    DeepReport,
+    DescribeResponse,
+    SearchGrade,
+    SearchResponse,
+    SearchResult,
+)
+
+WEB_RESPONSE: dict[str, Any] = {
+    "query": "asyncio python",
+    "source": "duckduckgo",
+    "count": 3,
+    "success": True,
+    "error": None,
+    "timestamp": "2026-08-07T12:00:00Z",
+    "results": [
+        {
+            "title": "asyncio documentation",
+            "url": "https://docs.python.org/3/library/asyncio.html",
+            "description": "Asynchronous I/O event loop.",
+            "source": "docs.python.org",
+            "extra": {"rank": 1},
+            "index": 0,
+        },
+        {
+            "title": "asyncio in Python",
+            "url": "https://example.com/asyncio",
+            "description": "Tutorial on asyncio.",
+            "source": "example.com",
+            "extra": {"rank": 2},
+            "index": 1,
+        },
+    ],
+}
+
+AI_MEMORY_RESPONSE: dict[str, Any] = {
+    "query": "python history",
+    "source": "ai",
+    "count": 0,
+    "success": True,
+    "error": None,
+    "results": [],
+    "ai_response": "From memory: Python was released in 1991 by Guido van Rossum.",
+    "ai_error": None,
+}
+
+AI_PROVIDER_RESPONSE: dict[str, Any] = {
+    "query": "quantum computing",
+    "source": "google",
+    "count": 1,
+    "success": True,
+    "error": None,
+    "results": [
+        {
+            "title": "Quantum computing overview",
+            "url": "https://example.com/quantum",
+            "description": "Overview of quantum computing.",
+            "source": "example.com",
+            "extra": {},
+            "index": 0,
+        }
+    ],
+    "ai_response": "Quantum computing uses qubits. [citation:1]",
+    "ai_error": None,
+}
+
+GRADED_RESPONSE: dict[str, Any] = {
+    "query": "deep research",
+    "source": "google",
+    "count": 1,
+    "success": True,
+    "error": None,
+    "results": [
+        {
+            "title": "Deep research systems",
+            "url": "https://example.com/deep-research",
+            "description": "Survey of deep research systems.",
+            "source": "example.com",
+            "extra": {},
+            "index": 0,
+            "grade": {
+                "overall": 9.2,
+                "relevance": 8.8,
+                "depth": 9.0,
+                "authority": 9.5,
+                "freshness": 7.0,
+                "word_count": 1200,
+                "intent_hits": 4,
+            },
+        }
+    ],
+}
+
+DEEP_RESPONSE: dict[str, Any] = {
+    "query": "deep research systems",
+    "source": "google",
+    "count": 8,
+    "success": True,
+    "error": None,
+    "results": [
+        {
+            "title": "Deep research systems",
+            "url": "https://example.com/deep-research",
+            "description": "Survey of deep research systems.",
+            "source": "example.com",
+            "extra": {},
+            "index": 0,
+        }
+    ],
+    "deep": {
+        "query": "deep research systems",
+        "markdown": "# Deep research\n\nA survey.",
+        "sources": [
+            {
+                "title": "Deep research systems",
+                "url": "https://example.com/deep-research",
+                "description": "Survey of deep research systems.",
+                "source": "example.com",
+                "extra": {},
+                "grade": {
+                    "overall": 9.2,
+                    "relevance": 8.8,
+                    "depth": 9.0,
+                    "authority": 9.5,
+                    "freshness": 7.0,
+                    "word_count": 1200,
+                    "intent_hits": 4,
+                },
+            }
+        ],
+        "graded_count": 8,
+        "total_count": 10,
+        "model": "gemma-3-12b-it",
+        "elapsed": 166.96,
+        "cache_hit": False,
+        "rounds": 3,
+        "queries_tried": ["deep research systems", "deep research architecture"],
+        "error": None,
+    },
+}
+
+IMAGES_RESPONSE: dict[str, Any] = {
+    "query": "aurora borealis",
+    "source": "wikimedia",
+    "count": 2,
+    "success": True,
+    "error": None,
+    "results": [
+        {
+            "title": "Aurora borealis over Norway",
+            "url": "https://commons.wikimedia.org/wiki/File:Aurora.jpg",
+            "description": "Photograph of the aurora borealis.",
+            "source": "wikimedia",
+            "extra": {
+                "thumbnail": "https://upload.wikimedia.org/thumb.jpg",
+                "dimensions": {"width": 1920, "height": 1080},
+                "mime": "image/jpeg",
+                "license": "CC BY-SA 4.0",
+            },
+            "index": 0,
+        }
+    ],
+}
+
+CHAT_RESPONSE: dict[str, Any] = {
+    "response": "The answer.",
+    "prompt": "question",
+    "json_mode": True,
+    "cached": False,
+    "error": None,
+    "usage": {
+        "prompt_tokens": 120,
+        "completion_tokens": 80,
+        "total_tokens": 200,
+        "cost_usd": 0.0012,
+    },
+}
+
+DESCRIBE_RESPONSE: dict[str, Any] = {
+    "url": "https://example.com/page",
+    "description": "Page description",
+    "elapsed": 1.23,
+    "timestamp": "2026-08-07T12:00:00Z",
+}
+
+SEARCH_EMPTY_OK: dict[str, Any] = {
+    "query": "q",
+    "source": "s",
+    "count": 1,
+    "success": True,
+    "error": None,
+    "results": [],
+}
+
+
+def _recorded_request(fixture: dict[str, Any]) -> tuple[list[tuple[Any, ...]], Any]:
+    recorded: list[tuple[Any, ...]] = []
+
+    def fake(
+        method: str,
+        path: str,
+        params: dict[str, str] | None,
+        payload: bytes | None,
+        headers: dict[str, str] | None,
+        timeout: float | None,
+    ) -> tuple[int, dict[str, Any]]:
+        recorded.append((method, path, params, payload, headers, timeout))
+        return 200, fixture
+
+    return recorded, fake
+
+
+def _raising_request(message: str, status_code: int) -> Any:
+    def fake(
+        method: str,
+        path: str,
+        params: dict[str, str] | None,
+        payload: bytes | None,
+        headers: dict[str, str] | None,
+        timeout: float | None,
+    ) -> tuple[int, dict[str, Any]]:
+        raise RsearchError(message, status_code)
+
+    return fake
+
+
+class _FakeResponse:
+    def __init__(self, status: int, body: bytes) -> None:
+        self.status = status
+        self._body = body
+
+    def __enter__(self) -> "_FakeResponse":
+        return self
+
+    def __exit__(self, *args: object) -> None:
+        return None
+
+    def read(self) -> bytes:
+        return self._body
+
+
+class TestSearchResponseParsing(unittest.TestCase):
+
+    def test_web_results_parse_into_search_response(self) -> None:
+        response = SearchResponse.from_dict(WEB_RESPONSE)
+        self.assertEqual(response.query, "asyncio python")
+        self.assertEqual(response.source, "duckduckgo")
+        self.assertEqual(response.count, 3)
+        self.assertTrue(response.success)
+        self.assertIsNone(response.error)
+        self.assertEqual(response.timestamp, "2026-08-07T12:00:00Z")
+        self.assertEqual(len(response.results), 2)
+        first = response.results[0]
+        self.assertIsInstance(first, SearchResult)
+        self.assertEqual(first.title, "asyncio documentation")
+        self.assertEqual(first.url, "https://docs.python.org/3/library/asyncio.html")
+        self.assertEqual(first.description, "Asynchronous I/O event loop.")
+        self.assertEqual(first.source, "docs.python.org")
+        self.assertEqual(first.extra, {"rank": 1})
+        self.assertEqual(first.index, 0)
+        self.assertIsNone(first.content)
+        self.assertIsNone(first.grade)
+        self.assertIsNone(first.query_origin)
+        self.assertIsNone(response.ai_response)
+        self.assertIsNone(response.deep)
+
+    def test_ai_memory_variant_parses(self) -> None:
+        response = SearchResponse.from_dict(AI_MEMORY_RESPONSE)
+        self.assertEqual(response.source, "ai")
+        self.assertEqual(response.results, [])
+        self.assertIn("From memory", response.ai_response)
+        self.assertIsNone(response.ai_error)
+
+    def test_ai_provider_variant_parses(self) -> None:
+        response = SearchResponse.from_dict(AI_PROVIDER_RESPONSE)
+        self.assertEqual(response.source, "google")
+        self.assertEqual(len(response.results), 1)
+        self.assertIn("[citation:1]", response.ai_response)
+        self.assertIsNone(response.ai_error)
+
+    def test_deep_block_parses_into_deep_report(self) -> None:
+        response = SearchResponse.from_dict(DEEP_RESPONSE)
+        self.assertIsNotNone(response.deep)
+        deep = response.deep
+        self.assertIsInstance(deep, DeepReport)
+        self.assertEqual(deep.query, "deep research systems")
+        self.assertEqual(deep.markdown, "# Deep research\n\nA survey.")
+        self.assertEqual(deep.graded_count, 8)
+        self.assertEqual(deep.total_count, 10)
+        self.assertEqual(deep.model, "gemma-3-12b-it")
+        self.assertEqual(deep.elapsed, 166.96)
+        self.assertFalse(deep.cache_hit)
+        self.assertEqual(deep.rounds, 3)
+        self.assertEqual(deep.queries_tried, ["deep research systems", "deep research architecture"])
+        self.assertIsNone(deep.error)
+        self.assertEqual(len(deep.sources), 1)
+        source = deep.sources[0]
+        self.assertIsInstance(source, SearchResult)
+        self.assertEqual(source.url, "https://example.com/deep-research")
+        self.assertIsInstance(source.grade, SearchGrade)
+        self.assertEqual(source.grade.overall, 9.2)
+
+    def test_images_results_parse_extra_metadata(self) -> None:
+        response = SearchResponse.from_dict(IMAGES_RESPONSE)
+        self.assertEqual(response.source, "wikimedia")
+        result = response.results[0]
+        self.assertEqual(result.extra["mime"], "image/jpeg")
+        self.assertEqual(result.extra["dimensions"], {"width": 1920, "height": 1080})
+        self.assertEqual(result.extra["license"], "CC BY-SA 4.0")
+        self.assertIn("thumbnail", result.extra)
+
+    def test_result_grade_parses_into_search_grade(self) -> None:
+        response = SearchResponse.from_dict(GRADED_RESPONSE)
+        grade = response.results[0].grade
+        self.assertIsInstance(grade, SearchGrade)
+        self.assertEqual(grade.overall, 9.2)
+        self.assertEqual(grade.relevance, 8.8)
+        self.assertEqual(grade.depth, 9.0)
+        self.assertEqual(grade.authority, 9.5)
+        self.assertEqual(grade.freshness, 7.0)
+        self.assertEqual(grade.word_count, 1200)
+        self.assertEqual(grade.intent_hits, 4)
+
+    def test_sparse_body_parses_with_defaults(self) -> None:
+        response = SearchResponse.from_dict({"query": "x", "success": True})
+        self.assertEqual(response.source, "")
+        self.assertEqual(response.count, 0)
+        self.assertEqual(response.results, [])
+        self.assertIsNone(response.error)
+        self.assertIsNone(response.deep)
+        self.assertIsNone(response.ai_response)
+
+
+class TestSearchRequestConstruction(unittest.IsolatedAsyncioTestCase):
+
+    async def test_search_forwards_all_parameters(self) -> None:
+        client = RsearchClient()
+        recorded, fake = _recorded_request(SEARCH_EMPTY_OK)
+        client._request = fake
+        await client.search(
+            "query text",
+            source="google",
+            count=7,
+            content=True,
+            type="images",
+            deep=True,
+            ai=True,
+            cache=False,
+        )
+        method, path, params, payload, headers, timeout = recorded[0]
+        self.assertEqual(method, "GET")
+        self.assertEqual(path, "/search")
+        self.assertEqual(
+            params,
+            {
+                "query": "query text",
+                "source": "google",
+                "count": "7",
+                "content": "true",
+                "type": "images",
+                "deep": "true",
+                "ai": "true",
+                "cache": "false",
+            },
+        )
+        self.assertIsNone(payload)
+        self.assertIsNone(headers)
+        self.assertEqual(timeout, 180.0)
+
+    async def test_search_without_deep_uses_request_timeout(self) -> None:
+        client = RsearchClient()
+        recorded, fake = _recorded_request(SEARCH_EMPTY_OK)
+        client._request = fake
+        await client.search("q")
+        self.assertEqual(recorded[0][5], 30.0)
+
+    async def test_count_none_omits_count_parameter(self) -> None:
+        client = RsearchClient()
+        recorded, fake = _recorded_request(SEARCH_EMPTY_OK)
+        client._request = fake
+        await client.search("q")
+        self.assertNotIn("count", recorded[0][2])
+
+    async def test_count_zero_forwarded_and_server_clamp_parsed(self) -> None:
+        client = RsearchClient()
+        recorded, fake = _recorded_request(
+            {"query": "q", "source": "s", "count": 1, "success": True, "error": None, "results": []}
+        )
+        client._request = fake
+        response = await client.search("q", count=0)
+        self.assertEqual(recorded[0][2]["count"], "0")
+        self.assertEqual(response.count, 1)
+
+    async def test_count_above_limit_forwarded_and_server_clamp_parsed(self) -> None:
+        client = RsearchClient()
+        recorded, fake = _recorded_request(
+            {"query": "q", "source": "s", "count": 10, "success": True, "error": None, "results": []}
+        )
+        client._request = fake
+        response = await client.search("q", count=25)
+        self.assertEqual(recorded[0][2]["count"], "25")
+        self.assertEqual(response.count, 10)
+
+    async def test_invalid_count_forwarded_and_server_clamp_parsed(self) -> None:
+        client = RsearchClient()
+        recorded, fake = _recorded_request(
+            {"query": "q", "source": "s", "count": 10, "success": True, "error": None, "results": []}
+        )
+        client._request = fake
+        response = await client.search("q", count="not-a-number")
+        self.assertEqual(recorded[0][2]["count"], "not-a-number")
+        self.assertEqual(response.count, 10)
+
+    async def test_search_cache_hit_skips_second_request(self) -> None:
+        client = RsearchClient()
+        recorded, fake = _recorded_request(SEARCH_EMPTY_OK)
+        client._request = fake
+        await client.search("cached query")
+        await client.search("cached query")
+        self.assertEqual(len(recorded), 1)
+
+    async def test_search_cache_disabled_repeats_request(self) -> None:
+        client = RsearchClient()
+        recorded, fake = _recorded_request(SEARCH_EMPTY_OK)
+        client._request = fake
+        await client.search("uncached query", cache=False)
+        await client.search("uncached query", cache=False)
+        self.assertEqual(len(recorded), 2)
+
+    async def test_search_with_content_populates_content_cache(self) -> None:
+        client = RsearchClient()
+        fixture = {
+            "query": "q",
+            "source": "s",
+            "count": 1,
+            "success": True,
+            "error": None,
+            "results": [
+                {
+                    "title": "t",
+                    "url": "https://example.com/a",
+                    "description": "d",
+                    "source": "s",
+                    "extra": {},
+                    "content": "full page text",
+                }
+            ],
+        }
+        recorded, fake = _recorded_request(fixture)
+        client._request = fake
+        await client.search("q", content=True)
+        self.assertEqual(len(recorded), 1)
+        self.assertEqual(client.get_cached_content("https://example.com/a"), "full page text")
+
+    async def test_search_error_in_body_surfaces_rsearch_error(self) -> None:
+        client = RsearchClient()
+        client._request = _raising_request("Empty query", 400)
+        with self.assertRaises(RsearchError) as ctx:
+            await client.search("")
+        self.assertEqual(ctx.exception.status_code, 400)
+        self.assertEqual(str(ctx.exception), "Empty query")
+
+
+class TestChatResponseParsing(unittest.IsolatedAsyncioTestCase):
+
+    async def test_chat_response_parses_envelope(self) -> None:
+        client = RsearchClient()
+        recorded, fake = _recorded_request(CHAT_RESPONSE)
+        client._request = fake
+        response = await client.chat("question", json_mode=True)
+        method, path, params, payload, headers, timeout = recorded[0]
+        self.assertEqual(method, "POST")
+        self.assertEqual(path, "/chat")
+        self.assertEqual(json.loads(payload), {"prompt": "question", "json": True})
+        self.assertEqual(headers, {"Content-Type": "application/json"})
+        self.assertIsNone(params)
+        self.assertIsNone(timeout)
+        self.assertIsInstance(response, ChatResponse)
+        self.assertEqual(response.response, "The answer.")
+        self.assertEqual(response.prompt, "question")
+        self.assertTrue(response.json_mode)
+        self.assertFalse(response.cached)
+        self.assertIsNone(response.error)
+        self.assertIsInstance(response.usage, ChatUsage)
+        self.assertEqual(response.usage.prompt_tokens, 120)
+        self.assertEqual(response.usage.completion_tokens, 80)
+        self.assertEqual(response.usage.total_tokens, 200)
+        self.assertEqual(response.usage.cost_usd, 0.0012)
+
+    async def test_chat_request_accepts_system_and_disables_cache(self) -> None:
+        client = RsearchClient()
+        recorded, fake = _recorded_request(CHAT_RESPONSE)
+        client._request = fake
+        await client.chat("q", system="sys", cache=False)
+        body = json.loads(recorded[0][3])
+        self.assertEqual(body, {"prompt": "q", "system": "sys", "cache": False})
+
+    async def test_chat_error_raises_mapped_rsearch_error(self) -> None:
+        client = RsearchClient()
+        client._request = _raising_request("No prompt provided", 400)
+        with self.assertRaises(RsearchError) as ctx:
+            await client.chat("")
+        self.assertEqual(ctx.exception.status_code, 400)
+        self.assertEqual(str(ctx.exception), "No prompt provided")
+
+
+class TestDescribeResponseParsing(unittest.IsolatedAsyncioTestCase):
+
+    async def test_describe_get_parses_envelope(self) -> None:
+        client = RsearchClient()
+        recorded, fake = _recorded_request(DESCRIBE_RESPONSE)
+        client._request = fake
+        response = await client.describe("https://example.com/page")
+        method, path, params, payload, headers, timeout = recorded[0]
+        self.assertEqual(method, "GET")
+        self.assertEqual(path, "/describe")
+        self.assertEqual(params, {"url": "https://example.com/page"})
+        self.assertIsNone(payload)
+        self.assertIsNone(headers)
+        self.assertIsNone(timeout)
+        self.assertIsInstance(response, DescribeResponse)
+        self.assertEqual(response.description, "Page description")
+        self.assertEqual(response.url, "https://example.com/page")
+        self.assertEqual(response.elapsed, 1.23)
+        self.assertEqual(response.timestamp, "2026-08-07T12:00:00Z")
+        self.assertIsNone(response.mime_type)
+        self.assertIsNone(response.size)
+        self.assertTrue(response.success)
+
+    async def test_describe_get_uses_cache(self) -> None:
+        client = RsearchClient()
+        recorded, fake = _recorded_request(DESCRIBE_RESPONSE)
+        client._request = fake
+        await client.describe("https://example.com/page")
+        await client.describe("https://example.com/page")
+        self.assertEqual(len(recorded), 1)
+
+    async def test_describe_error_raises_mapped_rsearch_error(self) -> None:
+        client = RsearchClient()
+        client._request = _raising_request("No url provided", 400)
+        with self.assertRaises(RsearchError) as ctx:
+            await client.describe("")
+        self.assertEqual(ctx.exception.status_code, 400)
+        self.assertEqual(str(ctx.exception), "No url provided")
+
+    async def test_describe_raw_posts_bytes_with_content_type(self) -> None:
+        client = RsearchClient()
+        recorded, fake = _recorded_request(DESCRIBE_RESPONSE)
+        client._request = fake
+        image = b"\x89PNG\r\n\x1a\npayload"
+        await client.describe_raw(image, mime_type="image/png")
+        method, path, params, payload, headers, timeout = recorded[0]
+        self.assertEqual(method, "POST")
+        self.assertEqual(path, "/describe")
+        self.assertEqual(payload, image)
+        self.assertEqual(headers, {"Content-Type": "image/png"})
+
+    async def test_describe_upload_builds_multipart_body(self) -> None:
+        client = RsearchClient()
+        recorded, fake = _recorded_request(DESCRIBE_RESPONSE)
+        client._request = fake
+        image = b"\x89PNGpayload"
+        await client.describe_upload(image, filename="photo.png", mime_type="image/png")
+        method, path, params, payload, headers, timeout = recorded[0]
+        self.assertEqual(method, "POST")
+        self.assertEqual(path, "/describe")
+        self.assertIn(b'name="file"; filename="photo.png"', payload)
+        self.assertIn(b"Content-Type: image/png", payload)
+        self.assertIn(image, payload)
+        self.assertIn("multipart/form-data; boundary=", headers["Content-Type"])
+
+    async def test_describe_raw_reuses_cache_by_hash(self) -> None:
+        client = RsearchClient()
+        recorded, fake = _recorded_request(DESCRIBE_RESPONSE)
+        client._request = fake
+        image = b"\x89PNGpayload"
+        await client.describe_raw(image, mime_type="image/png")
+        await client.describe_raw(image, mime_type="image/png")
+        self.assertEqual(len(recorded), 1)
+
+
+class TestErrorInBodyHandling(unittest.TestCase):
+
+    def test_empty_query_error_in_body_maps_to_rsearch_error(self) -> None:
+        client = RsearchClient()
+        body = b'{"success": false, "error": "Empty query"}'
+        error = urllib.error.HTTPError(
+            "https://rsearch.app.molodetz.nl/search", 400, "Bad Request", {}, io.BytesIO(body)
+        )
+        with mock.patch("urllib.request.urlopen", side_effect=error):
+            with self.assertRaises(RsearchError) as ctx:
+                client._request("GET", "/search", {"query": ""}, None, None, 30.0)
+        self.assertEqual(ctx.exception.status_code, 400)
+        self.assertEqual(str(ctx.exception), "Empty query")
+
+    def test_providers_exhausted_503_maps_to_rsearch_error(self) -> None:
+        client = RsearchClient()
+        body = b'{"success": false, "error": "All providers are exhausted, please try again later"}'
+        error = urllib.error.HTTPError(
+            "https://rsearch.app.molodetz.nl/search", 503, "Service Unavailable", {}, io.BytesIO(body)
+        )
+        with mock.patch("urllib.request.urlopen", side_effect=error):
+            with self.assertRaises(RsearchError) as ctx:
+                client._request("GET", "/search", {"query": "q"}, None, None, 30.0)
+        self.assertEqual(ctx.exception.status_code, 503)
+        self.assertEqual(str(ctx.exception), "All providers are exhausted, please try again later")
+
+    def test_success_false_body_with_http_200_raises(self) -> None:
+        client = RsearchClient()
+        fake = _FakeResponse(200, b'{"success": false, "error": "Empty query"}')
+        with mock.patch("urllib.request.urlopen", return_value=fake):
+            with self.assertRaises(RsearchError) as ctx:
+                client._request("GET", "/search", {"query": ""}, None, None, 30.0)
+        self.assertEqual(ctx.exception.status_code, 200)
+        self.assertEqual(str(ctx.exception), "Empty query")
+
+    def test_detail_field_falls_back_for_error_message(self) -> None:
+        client = RsearchClient()
+        body = b'{"detail": "No url provided"}'
+        error = urllib.error.HTTPError(
+            "https://rsearch.app.molodetz.nl/describe", 400, "Bad Request", {}, io.BytesIO(body)
+        )
+        with mock.patch("urllib.request.urlopen", side_effect=error):
+            with self.assertRaises(RsearchError) as ctx:
+                client._request("GET", "/describe", {"url": "x"}, None, None, 30.0)
+        self.assertEqual(ctx.exception.status_code, 400)
+        self.assertEqual(str(ctx.exception), "No url provided")
+
+    def test_title_field_falls_back_for_error_message(self) -> None:
+        client = RsearchClient()
+        body = b'{"title": "Provider error"}'
+        error = urllib.error.HTTPError(
+            "https://rsearch.app.molodetz.nl/search", 502, "Bad Gateway", {}, io.BytesIO(body)
+        )
+        with mock.patch("urllib.request.urlopen", side_effect=error):
+            with self.assertRaises(RsearchError) as ctx:
+                client._request("GET", "/search", {"query": "q"}, None, None, 30.0)
+        self.assertEqual(ctx.exception.status_code, 502)
+        self.assertEqual(str(ctx.exception), "Provider error")
+
+    def test_empty_body_raises_rsearch_error(self) -> None:
+        client = RsearchClient()
+        fake = _FakeResponse(200, b"")
+        with mock.patch("urllib.request.urlopen", return_value=fake):
+            with self.assertRaises(RsearchError) as ctx:
+                client._request("GET", "/search", {"query": "q"}, None, None, 30.0)
+        self.assertEqual(ctx.exception.status_code, 200)
+        self.assertIn("empty response", str(ctx.exception))
+
+    def test_invalid_json_body_raises_rsearch_error(self) -> None:
+        client = RsearchClient()
+        fake = _FakeResponse(200, b"<html>not json</html>")
+        with mock.patch("urllib.request.urlopen", return_value=fake):
+            with self.assertRaises(RsearchError) as ctx:
+                client._request("GET", "/search", {"query": "q"}, None, None, 30.0)
+        self.assertEqual(ctx.exception.status_code, 200)
+        self.assertIn("invalid JSON", str(ctx.exception))
+
+    def test_non_dict_body_raises_rsearch_error(self) -> None:
+        client = RsearchClient()
+        fake = _FakeResponse(200, b'["not", "a", "dict"]')
+        with mock.patch("urllib.request.urlopen", return_value=fake):
+            with self.assertRaises(RsearchError) as ctx:
+                client._request("GET", "/search", {"query": "q"}, None, None, 30.0)
+        self.assertEqual(ctx.exception.status_code, 200)
+
+    def test_connection_failure_raises_rsearch_error(self) -> None:
+        client = RsearchClient()
+        with mock.patch("urllib.request.urlopen", side_effect=urllib.error.URLError("connection refused")):
+            with self.assertRaises(RsearchError) as ctx:
+                client._request("GET", "/search", {"query": "q"}, None, None, 30.0)
+        self.assertIn("connection failure", str(ctx.exception))
+
+    def test_successful_request_returns_status_and_body(self) -> None:
+        client = RsearchClient()
+        fake = _FakeResponse(200, b'{"success": true, "query": "q", "count": 1, "results": []}')
+        with mock.patch("urllib.request.urlopen", return_value=fake) as urlopen:
+            status, data = client._request("GET", "/search", {"query": "q"}, None, None, 30.0)
+        self.assertEqual(status, 200)
+        self.assertEqual(data, {"success": True, "query": "q", "count": 1, "results": []})
+        request = urlopen.call_args[0][0]
+        self.assertEqual(request.get_method(), "GET")
+        self.assertEqual(request.get_full_url(), "https://rsearch.app.molodetz.nl/search?query=q")
+
+
+if __name__ == "__main__":
+    unittest.main()
+

tests/test_research_dedup.py

@@ -0,0 +1,310 @@
+# retoor <retoor@molodetz.nl>
+
+import unittest
+
+from typosaurus_sandbox.research.envelopes import SearchResult
+from typosaurus_sandbox.research.frontier import (
+    MAX_QUERY_LENGTH,
+    MIN_QUERY_LENGTH,
+    DedupStats,
+    QueryFrontier,
+    fingerprint_text,
+    normalize_url,
+    query_variants_from_result,
+)
+
+
+def _round_halts(frontier: QueryFrontier, before: DedupStats) -> bool:
+    after = frontier.snapshot()
+    new_urls = after.urls_seen - before.urls_seen
+    new_queries = after.queries_enqueued - before.queries_enqueued
+    return new_urls == 0 and new_queries == 0
+
+
+class TestNormalizeUrl(unittest.TestCase):
+
+    def test_lowercases_scheme_and_host_and_strips_default_port(self) -> None:
+        self.assertEqual(
+            normalize_url("HTTPS://Example.COM:443/Path//To//Page/"),
+            "https://example.com/Path/To/Page",
+        )
+
+    def test_strips_userinfo_and_fragment_keeps_query(self) -> None:
+        self.assertEqual(
+            normalize_url("https://user:pass@example.com:8443/a?x=1#frag"),
+            "https://example.com:8443/a?x=1",
+        )
+
+    def test_fragment_dropped_with_default_port(self) -> None:
+        self.assertEqual(normalize_url("https://example.com/a?x=1#sec"), "https://example.com/a?x=1")
+
+    def test_non_default_port_preserved(self) -> None:
+        self.assertEqual(normalize_url("https://example.com:80/x"), "https://example.com:80/x")
+
+    def test_idna_encodes_non_ascii_host(self) -> None:
+        self.assertEqual(normalize_url("https://MÜNCHEN.example/"), "https://xn--mnchen-3ya.example/")
+
+    def test_http_and_https_remain_distinct(self) -> None:
+        self.assertNotEqual(normalize_url("http://example.com/x"), normalize_url("https://example.com/x"))
+
+    def test_non_http_scheme_returned_cleaned(self) -> None:
+        self.assertEqual(normalize_url("not a url"), "not a url")
+
+    def test_blank_url_normalizes_to_empty(self) -> None:
+        self.assertEqual(normalize_url("   "), "")
+
+
+class TestFingerprintText(unittest.TestCase):
+
+    def test_whitespace_variants_produce_identical_fingerprint(self) -> None:
+        self.assertEqual(fingerprint_text("identical   body\n\n"), fingerprint_text("identical body"))
+
+    def test_distinct_text_produces_distinct_fingerprint(self) -> None:
+        self.assertNotEqual(fingerprint_text("first text"), fingerprint_text("second text"))
+
+    def test_fingerprint_is_sha256_hex(self) -> None:
+        digest = fingerprint_text("sample")
+        self.assertEqual(len(digest), 64)
+        int(digest, 16)
+
+
+class TestQueryVariantsFromResult(unittest.TestCase):
+
+    def test_title_description_and_string_extra_become_variants(self) -> None:
+        result = SearchResult(
+            title="Deep research",
+            description="Survey of deep research systems",
+            url="https://a.example",
+            extra={"tag": "research methods", "rank": 3},
+        )
+        self.assertEqual(
+            query_variants_from_result(result),
+            [
+                ("Deep research", "title"),
+                ("Survey of deep research systems", "description"),
+                ("research methods", "extra"),
+            ],
+        )
+
+    def test_non_string_extra_values_ignored(self) -> None:
+        result = SearchResult(title="t", url="https://a.example", extra={"rank": 3, "ok": True})
+        self.assertEqual(query_variants_from_result(result), [("t", "title")])
+
+    def test_empty_fields_produce_no_variants(self) -> None:
+        result = SearchResult(url="https://a.example")
+        self.assertEqual(query_variants_from_result(result), [])
+
+
+class TestUrlDeduplication(unittest.IsolatedAsyncioTestCase):
+
+    async def test_first_registration_records_url_once(self) -> None:
+        frontier = QueryFrontier()
+        self.assertTrue(frontier.register_url("https://example.com/page"))
+        stats = frontier.snapshot()
+        self.assertEqual(stats.urls_seen, 1)
+        self.assertEqual(stats.urls_duplicates_skipped, 0)
+
+    async def test_same_url_registered_twice_rejects_second(self) -> None:
+        frontier = QueryFrontier()
+        self.assertTrue(frontier.register_url("https://example.com/page"))
+        self.assertFalse(frontier.register_url("https://example.com/page"))
+        stats = frontier.snapshot()
+        self.assertEqual(stats.urls_seen, 1)
+        self.assertEqual(stats.urls_duplicates_skipped, 1)
+
+    async def test_normalized_variants_of_same_url_rejected(self) -> None:
+        frontier = QueryFrontier()
+        self.assertTrue(frontier.register_url("HTTPS://Example.COM:443/a//b/"))
+        self.assertFalse(frontier.register_url("https://example.com/a/b"))
+        stats = frontier.snapshot()
+        self.assertEqual(stats.urls_seen, 1)
+        self.assertEqual(stats.urls_duplicates_skipped, 1)
+
+    async def test_duplicate_urls_across_responses_recorded_once(self) -> None:
+        frontier = QueryFrontier()
+        first = SearchResult(url="https://example.com/page", title="first title", description="first description")
+        second = SearchResult(url="https://example.com/page", title="second title", description="second description")
+        self.assertTrue(frontier.register_result(first))
+        self.assertFalse(frontier.register_result(second))
+        stats = frontier.snapshot()
+        self.assertEqual(stats.urls_seen, 1)
+        self.assertEqual(stats.urls_duplicates_skipped, 1)
+
+    async def test_empty_url_rejected(self) -> None:
+        frontier = QueryFrontier()
+        self.assertFalse(frontier.register_url(""))
+        self.assertEqual(frontier.snapshot().urls_seen, 0)
+
+    async def test_whitespace_url_normalized_and_deduplicated(self) -> None:
+        frontier = QueryFrontier()
+        self.assertTrue(frontier.register_url("   "))
+        self.assertFalse(frontier.register_url("   "))
+        stats = frontier.snapshot()
+        self.assertEqual(stats.urls_seen, 1)
+        self.assertEqual(stats.urls_duplicates_skipped, 1)
+
+
+class TestContentDeduplication(unittest.IsolatedAsyncioTestCase):
+
+    async def test_identical_content_different_urls_rejects_second_occurrence(self) -> None:
+        frontier = QueryFrontier()
+        first = SearchResult(url="https://a.example/1", title="t1", description="d1", content="identical body")
+        second = SearchResult(url="https://b.example/2", title="t2", description="d2", content="identical body")
+        self.assertTrue(frontier.register_result(first))
+        self.assertTrue(frontier.register_result(second))
+        stats = frontier.snapshot()
+        self.assertEqual(stats.urls_seen, 2)
+        self.assertEqual(stats.content_seen, 1)
+        self.assertEqual(stats.content_duplicates_skipped, 1)
+        self.assertFalse(frontier.register_content("identical body"))
+
+    async def test_near_identical_whitespace_content_rejected(self) -> None:
+        frontier = QueryFrontier()
+        self.assertTrue(frontier.register_content("  Deep  research  system \n"))
+        self.assertFalse(frontier.register_content("Deep research system"))
+        stats = frontier.snapshot()
+        self.assertEqual(stats.content_seen, 1)
+        self.assertEqual(stats.content_duplicates_skipped, 1)
+
+    async def test_blank_content_rejected(self) -> None:
+        frontier = QueryFrontier()
+        self.assertFalse(frontier.register_content(""))
+        self.assertFalse(frontier.register_content("   \n "))
+        self.assertEqual(frontier.snapshot().content_seen, 0)
+
+    async def test_result_without_content_registers_url_only(self) -> None:
+        frontier = QueryFrontier()
+        result = SearchResult(url="https://a.example", title="t", description="d")
+        self.assertTrue(frontier.register_result(result))
+        stats = frontier.snapshot()
+        self.assertEqual(stats.urls_seen, 1)
+        self.assertEqual(stats.content_seen, 0)
+
+
+class TestQueryDeduplication(unittest.IsolatedAsyncioTestCase):
+
+    async def test_duplicate_query_rejected(self) -> None:
+        frontier = QueryFrontier()
+        self.assertTrue(frontier.push_query("deep research", "manual"))
+        self.assertFalse(frontier.push_query("deep research", "manual"))
+        stats = frontier.snapshot()
+        self.assertEqual(stats.queries_enqueued, 1)
+        self.assertEqual(stats.queries_duplicates_skipped, 1)
+
+    async def test_query_dedup_ignores_case_and_whitespace(self) -> None:
+        frontier = QueryFrontier()
+        self.assertTrue(frontier.push_query("  Deep   RESEARCH "))
+        self.assertFalse(frontier.push_query("deep research"))
+        self.assertEqual(frontier.snapshot().queries_enqueued, 1)
+
+    async def test_variants_from_result_deduplicated_across_fields(self) -> None:
+        frontier = QueryFrontier()
+        result = SearchResult(
+            title="Python asyncio",
+            description="python asyncio",
+            url="https://a.example",
+            extra={"tag": "  PYTHON   ASYNCIO "},
+        )
+        self.assertEqual(frontier.push_variants_from_result(result), 1)
+        stats = frontier.snapshot()
+        self.assertEqual(stats.queries_enqueued, 1)
+        self.assertEqual(stats.queries_duplicates_skipped, 2)
+
+    async def test_duplicate_query_never_issued_twice(self) -> None:
+        frontier = QueryFrontier("asyncio python")
+        self.assertEqual(frontier.pop_query(), "asyncio python")
+        self.assertFalse(frontier.push_query("ASYNCIO  python"))
+        self.assertIsNone(frontier.pop_query())
+        self.assertEqual(frontier.snapshot().queries_issued, 1)
+
+    async def test_query_length_window_enforced(self) -> None:
+        frontier = QueryFrontier()
+        self.assertFalse(frontier.push_query("a" * (MIN_QUERY_LENGTH - 1)))
+        self.assertTrue(frontier.push_query("a" * MIN_QUERY_LENGTH))
+        self.assertTrue(frontier.push_query("b" * MAX_QUERY_LENGTH))
+        self.assertFalse(frontier.push_query("c" * (MAX_QUERY_LENGTH + 1)))
+        self.assertEqual(frontier.snapshot().queries_enqueued, 2)
+
+    async def test_reseed_same_subject_enqueues_once(self) -> None:
+        frontier = QueryFrontier("subject alpha")
+        frontier.seed("SUBJECT   ALPHA")
+        stats = frontier.snapshot()
+        self.assertEqual(stats.queries_enqueued, 1)
+        self.assertEqual(stats.queries_duplicates_skipped, 1)
+
+
+class TestClosureDecision(unittest.IsolatedAsyncioTestCase):
+
+    async def test_round_with_no_new_urls_and_no_new_queries_halts(self) -> None:
+        frontier = QueryFrontier("subject alpha")
+        frontier.pop_query()
+        discovered = SearchResult(
+            url="https://a.example/page", title="alpha discovery", description="alpha details", content="body text"
+        )
+        self.assertTrue(frontier.register_result(discovered))
+        before = frontier.snapshot()
+        self.assertFalse(frontier.register_result(discovered))
+        self.assertFalse(frontier.push_query("SUBJECT ALPHA"))
+        after = frontier.snapshot()
+        self.assertEqual(after.urls_seen - before.urls_seen, 0)
+        self.assertEqual(after.queries_enqueued - before.queries_enqueued, 0)
+        self.assertTrue(_round_halts(frontier, before))
+
+    async def test_round_adding_new_url_continues(self) -> None:
+        frontier = QueryFrontier("subject beta")
+        frontier.pop_query()
+        before = frontier.snapshot()
+        self.assertTrue(frontier.register_url("https://new.example/x"))
+        after = frontier.snapshot()
+        self.assertEqual(after.urls_seen - before.urls_seen, 1)
+        self.assertFalse(_round_halts(frontier, before))
+
+    async def test_round_adding_new_query_continues(self) -> None:
+        frontier = QueryFrontier("subject gamma")
+        frontier.pop_query()
+        before = frontier.snapshot()
+        result = SearchResult(url="https://a.example", title="gamma subtopic", description="")
+        self.assertEqual(frontier.push_variants_from_result(result), 1)
+        after = frontier.snapshot()
+        self.assertEqual(after.queries_enqueued - before.queries_enqueued, 1)
+        self.assertFalse(_round_halts(frontier, before))
+
+    async def test_round_adding_url_and_query_continues(self) -> None:
+        frontier = QueryFrontier("subject delta")
+        frontier.pop_query()
+        before = frontier.snapshot()
+        result = SearchResult(url="https://b.example/page", title="delta subtopic", description="delta details")
+        self.assertTrue(frontier.register_result(result))
+        self.assertEqual(frontier.push_variants_from_result(result), 2)
+        after = frontier.snapshot()
+        self.assertEqual(after.urls_seen - before.urls_seen, 1)
+        self.assertEqual(after.queries_enqueued - before.queries_enqueued, 2)
+        self.assertFalse(_round_halts(frontier, before))
+
+    async def test_empty_result_round_halts_and_exhausts_pending(self) -> None:
+        frontier = QueryFrontier("lonely topic")
+        before = frontier.snapshot()
+        self.assertEqual(frontier.pop_query(), "lonely topic")
+        self.assertFalse(frontier.has_pending())
+        self.assertTrue(_round_halts(frontier, before))
+
+    async def test_snapshot_reports_round_deltas_for_closure_accounting(self) -> None:
+        frontier = QueryFrontier("subject epsilon")
+        before = frontier.snapshot()
+        self.assertEqual(frontier.pop_query(), "subject epsilon")
+        result = SearchResult(
+            url="https://c.example/page", title="epsilon topic", description="epsilon details", content="epsilon body"
+        )
+        self.assertTrue(frontier.register_result(result))
+        after = frontier.snapshot()
+        self.assertEqual(after.urls_seen - before.urls_seen, 1)
+        self.assertEqual(after.content_seen - before.content_seen, 1)
+        self.assertEqual(after.queries_issued - before.queries_issued, 1)
+        self.assertEqual(len(after.to_dict()), 8)
+
+
+if __name__ == "__main__":
+    unittest.main()
+
+
+

tests/test_research_engine.py

@@ -0,0 +1,305 @@
+# retoor <retoor@molodetz.nl>
+
+import unittest
+from collections.abc import Callable
+from unittest.mock import patch
+
+from typosaurus_sandbox.research.client import RsearchError
+from typosaurus_sandbox.research.config import ResearchConfig
+from typosaurus_sandbox.research.engine import ResearchEngine
+from typosaurus_sandbox.research.envelopes import ChatResponse, DescribeResponse, SearchResponse, SearchResult
+
+
+async def _no_sleep(delay: float) -> None:
+    return None
+
+
+class FakeResearchClient:
+    def __init__(
+        self,
+        *,
+        web_results: list[SearchResult] | None = None,
+        web_result_factory: Callable[[str], list[SearchResult]] | None = None,
+        chat_text: str = "",
+        describe_text: str = "",
+    ) -> None:
+        self.config = ResearchConfig(max_concurrency=4, default_count=5)
+        self._web_results = web_results if web_results is not None else []
+        self._web_result_factory = web_result_factory
+        self._chat_text = chat_text
+        self._describe_text = describe_text
+        self.calls: list[tuple[str, str, str | None, bool, bool]] = []
+
+    def search_cached(
+        self,
+        query: str,
+        *,
+        source: str | None = None,
+        count: int | None = None,
+        content: bool = False,
+        type: str | None = None,
+        deep: bool = False,
+        ai: bool = False,
+        cache: bool = True,
+    ) -> None:
+        return None
+
+    def describe_cached(self, url: str) -> None:
+        return None
+
+    async def search(
+        self,
+        query: str,
+        *,
+        source: str | None = None,
+        count: int | None = None,
+        content: bool = False,
+        type: str | None = None,
+        deep: bool = False,
+        ai: bool = False,
+        cache: bool = True,
+    ) -> SearchResponse:
+        self.calls.append(("search", query, type, deep, ai))
+        if type == "images":
+            return SearchResponse(query=query, source="wikimedia", count=0, success=True, results=[])
+        results = self._web_result_factory(query) if self._web_result_factory is not None else list(self._web_results)
+        return SearchResponse(query=query, source="duckduckgo", count=len(results), success=True, results=results)
+
+    async def chat(self, prompt: str, *, system: str | None = None, json_mode: bool = False, cache: bool = True) -> ChatResponse:
+        self.calls.append(("chat", prompt, None, False, False))
+        return ChatResponse(response=self._chat_text, prompt=prompt)
+
+    async def describe(self, url: str) -> DescribeResponse:
+        self.calls.append(("describe", url, None, False, False))
+        return DescribeResponse(description=self._describe_text, url=url)
+
+
+class TestEngineClosureDetection(unittest.IsolatedAsyncioTestCase):
+
+    async def test_run_closes_after_single_round_when_nothing_new(self) -> None:
+        client = FakeResearchClient(web_results=[], chat_text="", describe_text="")
+        engine = ResearchEngine(client=client)
+        report = await engine.run("  deep   research ")
+        self.assertEqual(report.subject, "deep research")
+        self.assertTrue(report.closed)
+        self.assertEqual(report.total_rounds, 1)
+        self.assertEqual(len(report.rounds), 1)
+        first = report.rounds[0]
+        self.assertEqual(first.number, 1)
+        self.assertEqual(first.items_processed, 3)
+        self.assertEqual(first.requests_succeeded, 3)
+        self.assertEqual(first.requests_failed, 0)
+        self.assertEqual(first.new_urls, 0)
+        self.assertEqual(first.new_queries, 0)
+        self.assertTrue(first.closed)
+        self.assertEqual(report.queries_issued, 1)
+        self.assertEqual(report.queries_enqueued, 1)
+        self.assertEqual(report.urls_collected, 0)
+        self.assertEqual(report.contents_seen, 0)
+        self.assertEqual(report.content_types, {"web": 1, "images": 1, "chat": 1})
+        self.assertEqual(len(client.calls), 3)
+        self.assertEqual({call[0] for call in client.calls}, {"search", "chat"})
+
+    async def test_run_discovery_rounds_then_closes(self) -> None:
+        client = FakeResearchClient(
+            web_results=[
+                SearchResult(
+                    title="topic alpha",
+                    url="https://example.com/alpha",
+                    description="alpha details",
+                    content="alpha body",
+                )
+            ],
+            chat_text="",
+            describe_text="",
+        )
+        engine = ResearchEngine(client=client)
+        report = await engine.run("deep research")
+        self.assertTrue(report.closed)
+        self.assertEqual(report.total_rounds, 2)
+        first = report.rounds[0]
+        self.assertEqual(first.new_urls, 1)
+        self.assertEqual(first.new_queries, 2)
+        self.assertEqual(first.new_contents, 1)
+        self.assertFalse(first.closed)
+        second = report.rounds[1]
+        self.assertEqual(second.new_urls, 0)
+        self.assertEqual(second.new_queries, 0)
+        self.assertTrue(second.closed)
+        self.assertEqual(report.queries_generated, 7)
+        self.assertEqual(report.queries_enqueued, 3)
+        self.assertEqual(report.queries_issued, 3)
+        self.assertEqual(report.queries_duplicates_skipped, 4)
+        self.assertEqual(report.urls_collected, 1)
+        self.assertEqual(report.urls_duplicates_skipped, 2)
+        self.assertEqual(report.contents_seen, 1)
+        self.assertEqual(report.content_duplicates_skipped, 2)
+        self.assertEqual(report.requests_succeeded, 10)
+        self.assertEqual(report.requests_failed, 0)
+        self.assertEqual(report.cache_hits, 0)
+        self.assertEqual(report.cache_misses, 10)
+        self.assertEqual(report.content_types, {"web": 3, "images": 3, "chat": 3, "describe": 1})
+        self.assertEqual(len(client.calls), 10)
+        self.assertEqual(sum(1 for kind, _, type_value, _, _ in client.calls if kind == "search" and type_value is None), 3)
+        self.assertEqual(sum(1 for kind, _, type_value, _, _ in client.calls if kind == "search" and type_value == "images"), 3)
+        self.assertEqual(sum(1 for kind, _, _, _, _ in client.calls if kind == "chat"), 3)
+        self.assertEqual(sum(1 for kind, _, _, _, _ in client.calls if kind == "describe"), 1)
+        self.assertIn(("describe", "https://example.com/alpha", None, False, False), client.calls)
+
+    async def test_new_content_alone_does_not_prevent_closure(self) -> None:
+        def factory(query: str) -> list[SearchResult]:
+            return [
+                SearchResult(
+                    title="dup title",
+                    url="https://example.com/dup",
+                    description="dup details",
+                    content=f"body for {query}",
+                )
+            ]
+
+        client = FakeResearchClient(web_result_factory=factory, chat_text="", describe_text="")
+        engine = ResearchEngine(client=client)
+        report = await engine.run("subject")
+        self.assertTrue(report.closed)
+        self.assertEqual(report.total_rounds, 2)
+        first = report.rounds[0]
+        self.assertEqual(first.new_urls, 1)
+        self.assertEqual(first.new_queries, 2)
+        self.assertFalse(first.closed)
+        second = report.rounds[1]
+        self.assertEqual(second.new_urls, 0)
+        self.assertEqual(second.new_queries, 0)
+        self.assertEqual(second.new_contents, 2)
+        self.assertTrue(second.closed)
+        self.assertEqual(report.contents_seen, 3)
+
+    async def test_round_summary_dict_is_serialisable(self) -> None:
+        client = FakeResearchClient(web_results=[], chat_text="", describe_text="")
+        engine = ResearchEngine(client=client)
+        report = await engine.run("serialisable subject")
+        summary_dict = report.rounds[0].to_dict()
+        self.assertEqual(summary_dict["number"], 1)
+        self.assertTrue(summary_dict["closed"])
+        report_dict = report.to_dict()
+        self.assertEqual(report_dict["subject"], "serialisable subject")
+        self.assertEqual(report_dict["total_rounds"], 1)
+        self.assertTrue(report_dict["closed"])
+
+
+class FailingWebClient(FakeResearchClient):
+    def __init__(
+        self,
+        *,
+        web_results: list[SearchResult] | None = None,
+        web_result_factory: Callable[[str], list[SearchResult]] | None = None,
+        chat_text: str = "",
+        describe_text: str = "",
+        failures_before_success: int = 0,
+    ) -> None:
+        super().__init__(
+            web_results=web_results,
+            web_result_factory=web_result_factory,
+            chat_text=chat_text,
+            describe_text=describe_text,
+        )
+        self._web_failures_left = failures_before_success
+
+    async def search(
+        self,
+        query: str,
+        *,
+        source: str | None = None,
+        count: int | None = None,
+        content: bool = False,
+        type: str | None = None,
+        deep: bool = False,
+        ai: bool = False,
+        cache: bool = True,
+    ) -> SearchResponse:
+        self.calls.append(("search", query, type, deep, ai))
+        if type == "images":
+            return SearchResponse(query=query, source="wikimedia", count=0, success=True, results=[])
+        if self._web_failures_left > 0:
+            self._web_failures_left -= 1
+            raise RsearchError("search failed", 503)
+        results = self._web_result_factory(query) if self._web_result_factory is not None else list(self._web_results)
+        return SearchResponse(query=query, source="duckduckgo", count=len(results), success=True, results=results)
+
+
+class TestEngineDeepAiWiring(unittest.IsolatedAsyncioTestCase):
+
+    async def test_web_search_work_items_issue_deep_and_ai_for_seed_and_subtopics(self) -> None:
+        def factory(query: str) -> list[SearchResult]:
+            return [
+                SearchResult(
+                    title="subtopic alpha",
+                    url="https://example.com/subtopic",
+                    description="subtopic details",
+                    content="subtopic body",
+                )
+            ]
+
+        client = FakeResearchClient(web_result_factory=factory, chat_text="", describe_text="")
+        engine = ResearchEngine(client=client)
+        report = await engine.run("seed topic")
+        self.assertTrue(report.closed)
+        web_calls = [call for call in client.calls if call[0] == "search" and call[2] is None]
+        self.assertEqual(len(web_calls), 3)
+        self.assertEqual({call[1] for call in web_calls}, {"seed topic", "subtopic alpha", "subtopic details"})
+        self.assertTrue(all(call[3] and call[4] for call in web_calls))
+
+
+class TestEngineClosureOnFailures(unittest.IsolatedAsyncioTestCase):
+
+    async def test_round_and_report_not_closed_when_request_failed(self) -> None:
+        client = FailingWebClient(failures_before_success=100)
+        engine = ResearchEngine(client=client)
+        with patch("typosaurus_sandbox.research.pipeline.asyncio.sleep", side_effect=_no_sleep):
+            report = await engine.run("subject")
+        self.assertFalse(report.closed)
+        self.assertEqual(report.requests_failed, 1)
+        self.assertEqual(report.requests_succeeded, 2)
+        self.assertEqual(report.total_rounds, 1)
+        first = report.rounds[0]
+        self.assertEqual(first.requests_failed, 1)
+        self.assertEqual(first.requests_succeeded, 2)
+        self.assertEqual(first.new_urls, 0)
+        self.assertEqual(first.new_queries, 0)
+        self.assertFalse(first.closed)
+
+    async def test_later_closed_round_keeps_report_unclosed_after_earlier_failure(self) -> None:
+        client = FailingWebClient(
+            chat_text="Reference at https://chat.example.org/note",
+            describe_text="",
+            failures_before_success=3,
+        )
+        engine = ResearchEngine(client=client)
+        with patch("typosaurus_sandbox.research.pipeline.asyncio.sleep", side_effect=_no_sleep):
+            report = await engine.run("subject")
+        self.assertEqual(report.total_rounds, 2)
+        first = report.rounds[0]
+        self.assertEqual(first.requests_failed, 1)
+        self.assertFalse(first.closed)
+        second = report.rounds[1]
+        self.assertEqual(second.requests_failed, 0)
+        self.assertTrue(second.closed)
+        self.assertEqual(report.requests_failed, 1)
+        self.assertFalse(report.closed)
+
+
+class TestEngineInputValidation(unittest.IsolatedAsyncioTestCase):
+
+    async def test_empty_subject_raises_without_requests(self) -> None:
+        client = FakeResearchClient(web_results=[], chat_text="", describe_text="")
+        engine = ResearchEngine(client=client)
+        with self.assertRaises(ValueError) as ctx:
+            await engine.run("   \n\t ")
+        self.assertEqual(str(ctx.exception), "research subject must not be empty")
+        self.assertEqual(client.calls, [])
+
+
+if __name__ == "__main__":
+    unittest.main()
+
+
+

tests/test_research_integration.py

@@ -0,0 +1,112 @@
+# retoor <retoor@molodetz.nl>
+
+import asyncio
+import json
+import unittest
+from typing import Any, AsyncIterator
+from unittest import mock
+
+from typosaurus_sandbox.research.client import RsearchClient
+from typosaurus_sandbox.research.config import ResearchConfig
+from typosaurus_sandbox.research.frontier import QueryFrontier
+from typosaurus_sandbox.research.pipeline import PipelineReport, ResearchPipeline, WorkItem
+
+PROBE_SUBJECT = "python asyncio"
+RSEARCH_BASE_URL = "https://rsearch.app.molodetz.nl"
+RUN_TIMEOUT_SECONDS = 60.0
+
+SEARCH_FIXTURE: dict[str, Any] = {
+    "query": PROBE_SUBJECT,
+    "source": "duckduckgo",
+    "count": 2,
+    "success": True,
+    "error": None,
+    "results": [
+        {
+            "title": "asyncio documentation",
+            "url": "https://docs.python.org/3/library/asyncio.html",
+            "description": "Asynchronous I/O event loop.",
+            "source": "docs.python.org",
+            "extra": {},
+            "index": 0,
+            "content": "The asyncio module provides infrastructure for writing single-threaded concurrent code.",
+        },
+        {
+            "title": "asyncio in Python",
+            "url": "https://example.com/asyncio",
+            "description": "Tutorial on asyncio.",
+            "source": "example.com",
+            "extra": {},
+            "index": 1,
+            "content": "A tutorial covering the asyncio event loop and coroutines.",
+        },
+    ],
+}
+
+
+class _FakeResponse:
+    def __init__(self, status: int, body: bytes) -> None:
+        self.status = status
+        self._body = body
+
+    def __enter__(self) -> "_FakeResponse":
+        return self
+
+    def __exit__(self, *args: object) -> None:
+        return None
+
+    def read(self) -> bytes:
+        return self._body
+
+
+class TestBoundedOfflineProbe(unittest.TestCase):
+
+    def test_bounded_probe_runs_against_mocked_transport_only(self) -> None:
+        config = ResearchConfig(
+            base_url=RSEARCH_BASE_URL,
+            max_concurrency=2,
+            default_count=2,
+            request_timeout_seconds=30.0,
+        )
+        self.assertEqual(config.base_url, RSEARCH_BASE_URL)
+        client = RsearchClient(config)
+        frontier = QueryFrontier(PROBE_SUBJECT)
+        requested: list[str] = []
+
+        def fake_urlopen(request: Any, timeout: float | None = None) -> _FakeResponse:
+            requested.append(request.get_full_url())
+            return _FakeResponse(200, json.dumps(SEARCH_FIXTURE).encode())
+
+        with mock.patch("urllib.request.urlopen", side_effect=fake_urlopen):
+            first = asyncio.run(self._bounded_run(client, frontier))
+            first_request_count = len(requested)
+            second = asyncio.run(self._bounded_run(client, frontier))
+            second_request_count = len(requested)
+
+        self.assertGreaterEqual(first.requests_succeeded, 1)
+        self.assertGreaterEqual(first.urls_found, 1)
+        self.assertGreaterEqual(first.contents_seen, 1)
+        self.assertFalse(any(outcome.cache_hit for outcome in first.outcomes))
+        stats = frontier.snapshot()
+        self.assertGreaterEqual(stats.urls_seen, 1)
+        self.assertGreaterEqual(stats.content_seen, 1)
+        self.assertGreaterEqual(first_request_count, 1)
+        for url in requested:
+            self.assertTrue(url.startswith(RSEARCH_BASE_URL), url)
+        self.assertTrue(any("/search" in url for url in requested))
+        self.assertEqual(second.requests_succeeded, 1)
+        self.assertTrue(any(outcome.cache_hit for outcome in second.outcomes))
+        self.assertEqual(second_request_count, first_request_count)
+
+    async def _bounded_run(self, client: RsearchClient, frontier: QueryFrontier) -> PipelineReport:
+        async def items() -> AsyncIterator[WorkItem]:
+            yield WorkItem("web", PROBE_SUBJECT)
+
+        pipeline = ResearchPipeline(client, frontier)
+        return await asyncio.wait_for(pipeline.run(items()), timeout=RUN_TIMEOUT_SECONDS)
+
+
+if __name__ == "__main__":
+    unittest.main()
+
+

tests/test_research_pipeline.py

@@ -0,0 +1,356 @@
+# retoor <retoor@molodetz.nl>
+
+import unittest
+from typing import Any, AsyncIterator
+from unittest.mock import patch
+
+from typosaurus_sandbox.research.client import RsearchError
+from typosaurus_sandbox.research.config import ResearchConfig
+from typosaurus_sandbox.research.envelopes import ChatResponse, DeepReport, DescribeResponse, SearchResponse, SearchResult
+from typosaurus_sandbox.research.frontier import QueryFrontier
+from typosaurus_sandbox.research.pipeline import (
+    RETRY_MAX_ATTEMPTS,
+    Extraction,
+    ResearchPipeline,
+    WorkItem,
+    apply_extraction,
+    extract_response,
+)
+
+
+async def _no_sleep(delay: float) -> None:
+    return None
+
+
+class StubResearchClient:
+    def __init__(self) -> None:
+        self.config = ResearchConfig(max_concurrency=4, default_count=5)
+        self.cache_hits: dict[tuple[str, str], Any] = {}
+        self.error_on: set[tuple[str, str]] = set()
+        self.explode_on: set[tuple[str, str]] = set()
+        self.failures_remaining: dict[tuple[str, str], int] = {}
+        self.chat_response = ChatResponse(response="chat answer")
+        self.describe_response = DescribeResponse(description="described page")
+        self.calls: list[tuple[str, str]] = []
+
+    def _maybe_fail(self, key: tuple[str, str], error: RsearchError) -> None:
+        if key in self.explode_on:
+            raise ValueError("unexpected boom")
+        if key in self.error_on:
+            raise error
+        remaining = self.failures_remaining.get(key, 0)
+        if remaining > 0:
+            self.failures_remaining[key] = remaining - 1
+            raise error
+
+    def search_cached(
+        self,
+        query: str,
+        *,
+        source: str | None = None,
+        count: int | None = None,
+        content: bool = False,
+        type: str | None = None,
+        deep: bool = False,
+        ai: bool = False,
+        cache: bool = True,
+    ) -> Any:
+        if type == "images":
+            return self.cache_hits.get(("images", query))
+        return self.cache_hits.get(("web", query))
+
+    def describe_cached(self, url: str) -> Any:
+        return self.cache_hits.get(("describe", url))
+
+    async def search(
+        self,
+        query: str,
+        *,
+        source: str | None = None,
+        count: int | None = None,
+        content: bool = False,
+        type: str | None = None,
+        deep: bool = False,
+        ai: bool = False,
+        cache: bool = True,
+    ) -> SearchResponse:
+        self.calls.append(("search", query))
+        self._maybe_fail(("search", query), RsearchError("search failed", 503))
+        return SearchResponse(
+            query=query,
+            source="duckduckgo",
+            count=1,
+            success=True,
+            results=[SearchResult(title=query, url=f"https://example.com/{query}", description="details", content="body")],
+        )
+
+    async def chat(self, prompt: str, *, system: str | None = None, json_mode: bool = False, cache: bool = True) -> ChatResponse:
+        self.calls.append(("chat", prompt))
+        if ("chat", prompt) in self.error_on:
+            raise RsearchError("chat failed", 400)
+        return self.chat_response
+
+    async def describe(self, url: str) -> DescribeResponse:
+        self.calls.append(("describe", url))
+        if ("describe", url) in self.error_on:
+            raise RsearchError("describe failed", 500)
+        return self.describe_response
+
+
+class TestExtractResponse(unittest.TestCase):
+
+    def test_web_results_extract_urls_seeds_and_content(self) -> None:
+        item = WorkItem("web", "query")
+        response = SearchResponse(
+            query="query",
+            source="duckduckgo",
+            count=2,
+            success=True,
+            results=[
+                SearchResult(title="First", url="https://a.example/1", description="First details", content="first body"),
+                SearchResult(title="Second", url="https://b.example/2", description="Second details", content=None),
+            ],
+        )
+        extraction = extract_response(item, response)
+        self.assertEqual(extraction.urls, ("https://a.example/1", "https://b.example/2"))
+        self.assertEqual(
+            extraction.query_seeds,
+            (
+                ("First", "title"),
+                ("First details", "description"),
+                ("Second", "title"),
+                ("Second details", "description"),
+            ),
+        )
+        self.assertEqual(extraction.content_texts, ("first body",))
+
+    def test_ai_response_adds_content_seed_and_urls(self) -> None:
+        item = WorkItem("web", "query")
+        response = SearchResponse(
+            query="query",
+            source="ai",
+            count=0,
+            success=True,
+            results=[],
+            ai_response="Overview at https://docs.example.org/x and https://blog.example.org/y.",
+        )
+        extraction = extract_response(item, response)
+        self.assertEqual(extraction.urls, ("https://docs.example.org/x", "https://blog.example.org/y"))
+        self.assertEqual(
+            extraction.query_seeds,
+            (("Overview at https://docs.example.org/x and https://blog.example.org/y.", "ai_response"),),
+        )
+        self.assertEqual(
+            extraction.content_texts,
+            ("Overview at https://docs.example.org/x and https://blog.example.org/y.",),
+        )
+
+    def test_deep_report_sources_and_markdown_extracted(self) -> None:
+        item = WorkItem("web", "query")
+        response = SearchResponse(
+            query="query",
+            source="google",
+            count=1,
+            success=True,
+            results=[],
+            deep=DeepReport(
+                query="query",
+                markdown="# Deep\n\nSee https://deep.example.org/report for details.",
+                sources=[SearchResult(title="Deep source", url="https://deep.example.org/source", description="Deep details")],
+            ),
+        )
+        extraction = extract_response(item, response)
+        self.assertEqual(extraction.urls, ("https://deep.example.org/source", "https://deep.example.org/report"))
+        self.assertEqual(
+            extraction.query_seeds,
+            (("Deep source", "title"), ("Deep details", "description")),
+        )
+        self.assertEqual(extraction.content_texts, ("# Deep\n\nSee https://deep.example.org/report for details.",))
+
+    def test_chat_response_extracts_content_seed_and_urls(self) -> None:
+        item = WorkItem("chat", "prompt")
+        response = ChatResponse(response="Answer at https://chat.example.org/a.")
+        extraction = extract_response(item, response)
+        self.assertEqual(extraction.urls, ("https://chat.example.org/a",))
+        self.assertEqual(extraction.query_seeds, (("Answer at https://chat.example.org/a.", "chat"),))
+        self.assertEqual(extraction.content_texts, ("Answer at https://chat.example.org/a.",))
+
+    def test_describe_response_extracts_description_seed_and_urls(self) -> None:
+        item = WorkItem("describe", "https://page.example.org/x")
+        response = DescribeResponse(description="Image shows a cat. More at https://gallery.example.org/cat.")
+        extraction = extract_response(item, response)
+        self.assertEqual(extraction.urls, ("https://gallery.example.org/cat",))
+        self.assertEqual(
+            extraction.query_seeds,
+            (("Image shows a cat. More at https://gallery.example.org/cat.", "describe"),),
+        )
+        self.assertEqual(extraction.content_texts, ("Image shows a cat. More at https://gallery.example.org/cat.",))
+
+
+class TestApplyExtraction(unittest.IsolatedAsyncioTestCase):
+
+    async def test_registers_each_kind_and_returns_counts(self) -> None:
+        frontier = QueryFrontier()
+        extraction = Extraction(
+            urls=("https://x.example/1", "https://y.example/2"),
+            query_seeds=(("variant one", "title"), ("variant two", "description")),
+            content_texts=("body one", "body two"),
+        )
+        self.assertEqual(apply_extraction(frontier, extraction), (2, 2, 2))
+        self.assertEqual(apply_extraction(frontier, extraction), (0, 0, 0))
+        stats = frontier.snapshot()
+        self.assertEqual(stats.urls_seen, 2)
+        self.assertEqual(stats.queries_enqueued, 2)
+        self.assertEqual(stats.content_seen, 2)
+
+
+class TestPipelineCacheProbe(unittest.IsolatedAsyncioTestCase):
+
+    async def test_web_probe_reflects_search_cache(self) -> None:
+        client = StubResearchClient()
+        frontier = QueryFrontier()
+        pipeline = ResearchPipeline(client, frontier)
+        self.assertFalse(pipeline._probe_cache(WorkItem("web", "alpha")))
+        client.cache_hits[("web", "alpha")] = SearchResponse(query="alpha", success=True)
+        self.assertTrue(pipeline._probe_cache(WorkItem("web", "alpha")))
+
+    async def test_images_probe_uses_images_search_cache(self) -> None:
+        client = StubResearchClient()
+        frontier = QueryFrontier()
+        pipeline = ResearchPipeline(client, frontier)
+        self.assertFalse(pipeline._probe_cache(WorkItem("images", "alpha")))
+        client.cache_hits[("images", "alpha")] = SearchResponse(query="alpha", success=True)
+        self.assertTrue(pipeline._probe_cache(WorkItem("images", "alpha")))
+
+    async def test_describe_probe_reflects_describe_cache(self) -> None:
+        client = StubResearchClient()
+        frontier = QueryFrontier()
+        pipeline = ResearchPipeline(client, frontier)
+        self.assertFalse(pipeline._probe_cache(WorkItem("describe", "https://z.example/page")))
+        client.cache_hits[("describe", "https://z.example/page")] = DescribeResponse(description="cached")
+        self.assertTrue(pipeline._probe_cache(WorkItem("describe", "https://z.example/page")))
+
+    async def test_chat_never_probes_cache(self) -> None:
+        client = StubResearchClient()
+        frontier = QueryFrontier()
+        pipeline = ResearchPipeline(client, frontier)
+        self.assertFalse(pipeline._probe_cache(WorkItem("chat", "question")))
+
+
+class TestPipelineOutcomes(unittest.IsolatedAsyncioTestCase):
+
+    async def test_chat_cached_response_marks_outcome_cache_hit(self) -> None:
+        client = StubResearchClient()
+        client.chat_response = ChatResponse(response="cached answer", cached=True)
+        frontier = QueryFrontier()
+        pipeline = ResearchPipeline(client, frontier)
+        outcome = await pipeline.process(WorkItem("chat", "question"))
+        self.assertTrue(outcome.success)
+        self.assertTrue(outcome.cache_hit)
+        self.assertEqual(outcome.endpoint, "/chat")
+
+    async def test_search_error_produces_failure_outcome_and_pool_survives(self) -> None:
+        client = StubResearchClient()
+        client.error_on.add(("search", "bad"))
+        frontier = QueryFrontier()
+
+        async def items() -> AsyncIterator[WorkItem]:
+            yield WorkItem("web", "bad")
+            yield WorkItem("web", "good")
+
+        pipeline = ResearchPipeline(client, frontier)
+        with patch("typosaurus_sandbox.research.pipeline.asyncio.sleep", side_effect=_no_sleep):
+            report = await pipeline.run(items())
+        self.assertEqual(report.requests_succeeded, 1)
+        self.assertEqual(report.requests_failed, 1)
+        self.assertEqual(len(report.outcomes), 2)
+        failed = next(outcome for outcome in report.outcomes if not outcome.success)
+        self.assertEqual(failed.endpoint, "/search")
+        self.assertEqual(failed.status_code, 503)
+        self.assertEqual(failed.error, "search failed")
+
+    async def test_run_drains_all_four_content_types(self) -> None:
+        client = StubResearchClient()
+        frontier = QueryFrontier()
+
+        async def items() -> AsyncIterator[WorkItem]:
+            yield WorkItem("web", "alpha")
+            yield WorkItem("images", "alpha")
+            yield WorkItem("describe", "https://z.example/page")
+            yield WorkItem("chat", "question")
+
+        pipeline = ResearchPipeline(client, frontier)
+        report = await pipeline.run(items())
+        self.assertEqual(report.requests_succeeded, 4)
+        self.assertEqual(report.requests_failed, 0)
+        self.assertEqual(len(report.outcomes), 4)
+        self.assertTrue(all(outcome.success for outcome in report.outcomes))
+        endpoints = {outcome.item.kind: outcome.endpoint for outcome in report.outcomes}
+        self.assertEqual(endpoints, {"web": "/search", "images": "/search", "describe": "/describe", "chat": "/chat"})
+
+
+class TestPipelineRetryAndFailureAccounting(unittest.IsolatedAsyncioTestCase):
+
+    async def test_transient_failure_retried_with_backoff_then_succeeds(self) -> None:
+        client = StubResearchClient()
+        client.failures_remaining[("search", "flaky")] = 2
+        frontier = QueryFrontier()
+        pipeline = ResearchPipeline(client, frontier)
+        delays: list[float] = []
+
+        async def fake_sleep(delay: float) -> None:
+            delays.append(delay)
+
+        with patch("typosaurus_sandbox.research.pipeline.asyncio.sleep", side_effect=fake_sleep):
+            outcome = await pipeline.process(WorkItem("web", "flaky"))
+        self.assertTrue(outcome.success)
+        self.assertEqual(client.calls.count(("search", "flaky")), 3)
+        self.assertEqual(delays, [0.5, 1.0])
+
+    async def test_transient_failure_exhausts_retries_and_reports_failure(self) -> None:
+        client = StubResearchClient()
+        client.failures_remaining[("search", "persistent")] = 100
+        frontier = QueryFrontier()
+        pipeline = ResearchPipeline(client, frontier)
+        with patch("typosaurus_sandbox.research.pipeline.asyncio.sleep", side_effect=_no_sleep):
+            outcome = await pipeline.process(WorkItem("web", "persistent"))
+        self.assertFalse(outcome.success)
+        self.assertEqual(outcome.status_code, 503)
+        self.assertEqual(outcome.error, "search failed")
+        self.assertEqual(client.calls.count(("search", "persistent")), RETRY_MAX_ATTEMPTS)
+
+    async def test_client_error_is_not_retried(self) -> None:
+        client = StubResearchClient()
+        client.error_on.add(("chat", "bad request"))
+        frontier = QueryFrontier()
+        pipeline = ResearchPipeline(client, frontier)
+        with patch("typosaurus_sandbox.research.pipeline.asyncio.sleep", side_effect=_no_sleep):
+            outcome = await pipeline.process(WorkItem("chat", "bad request"))
+        self.assertFalse(outcome.success)
+        self.assertEqual(outcome.status_code, 400)
+        self.assertEqual(client.calls.count(("chat", "bad request")), 1)
+
+    async def test_unexpected_exception_recorded_as_failure_outcome(self) -> None:
+        client = StubResearchClient()
+        client.explode_on.add(("search", "boom"))
+        frontier = QueryFrontier()
+
+        async def items() -> AsyncIterator[WorkItem]:
+            yield WorkItem("web", "boom")
+            yield WorkItem("web", "fine")
+
+        pipeline = ResearchPipeline(client, frontier)
+        report = await pipeline.run(items())
+        self.assertEqual(report.requests_succeeded, 1)
+        self.assertEqual(report.requests_failed, 1)
+        self.assertEqual(len(report.outcomes), 2)
+        failed = next(outcome for outcome in report.outcomes if not outcome.success)
+        self.assertEqual(failed.endpoint, "/search")
+        self.assertIsNone(failed.status_code)
+        self.assertIn("boom", failed.error or "")
+        self.assertEqual(failed.item.value, "boom")
+
+
+if __name__ == "__main__":
+    unittest.main()
+

tests/test_research_scheduling.py

@@ -0,0 +1,519 @@
+# retoor <retoor@molodetz.nl>
+
+import asyncio
+import unittest
+from concurrent.futures import ThreadPoolExecutor
+from typing import Any, AsyncIterator
+from unittest import mock
+
+from typosaurus_sandbox.research.cache import TTLCache
+from typosaurus_sandbox.research.client import RsearchClient
+from typosaurus_sandbox.research.config import ResearchConfig
+from typosaurus_sandbox.research.envelopes import ChatResponse, DescribeResponse, SearchResponse, SearchResult
+from typosaurus_sandbox.research.frontier import QueryFrontier
+from typosaurus_sandbox.research.pipeline import ResearchPipeline, WorkItem
+
+SEARCH_FIXTURE: dict[str, Any] = {
+    "query": "subject",
+    "source": "duckduckgo",
+    "count": 1,
+    "success": True,
+    "error": None,
+    "results": [
+        {
+            "title": "Result",
+            "url": "https://example.com/result",
+            "description": "Description",
+            "source": "example.com",
+            "extra": {},
+            "index": 0,
+        }
+    ],
+}
+
+CHAT_FIXTURE: dict[str, Any] = {
+    "response": "Answer",
+    "prompt": "prompt",
+    "json_mode": False,
+    "cached": False,
+    "error": None,
+    "usage": {"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15, "cost_usd": 0.0001},
+}
+
+DESCRIBE_FIXTURE: dict[str, Any] = {
+    "url": "https://example.com/page",
+    "description": "Page description",
+    "elapsed": 0.5,
+    "timestamp": "2026-08-07T12:00:00Z",
+}
+
+IMAGES_FIXTURE: dict[str, Any] = {
+    "query": "subject",
+    "source": "wikimedia",
+    "count": 1,
+    "success": True,
+    "error": None,
+    "results": [
+        {
+            "title": "Aurora borealis over Norway",
+            "url": "https://example.com/image",
+            "description": "Photograph of the aurora borealis.",
+            "source": "wikimedia",
+            "extra": {},
+            "index": 0,
+        }
+    ],
+}
+
+
+class _FakeClient:
+    def __init__(self) -> None:
+        self.config = ResearchConfig()
+        self.search_calls: list[tuple[str, dict[str, Any]]] = []
+        self.chat_calls: list[str] = []
+        self.describe_calls: list[str] = []
+        self.delay_seconds = 0.0
+        self.active = 0
+        self.peak = 0
+
+    def search_cached(
+        self,
+        query: str,
+        *,
+        source: str | None = None,
+        count: int | None = None,
+        content: bool = False,
+        type: str | None = None,
+        deep: bool = False,
+        ai: bool = False,
+        cache: bool = True,
+    ) -> None:
+        return None
+
+    def describe_cached(self, url: str) -> None:
+        return None
+
+    async def search(
+        self,
+        query: str,
+        *,
+        source: str | None = None,
+        count: int | None = None,
+        content: bool = False,
+        type: str | None = None,
+        deep: bool = False,
+        ai: bool = False,
+        cache: bool = True,
+    ) -> SearchResponse:
+        self.search_calls.append((query, {"type": type, "content": content, "count": count}))
+        self.active += 1
+        self.peak = max(self.peak, self.active)
+        try:
+            if self.delay_seconds:
+                await asyncio.sleep(self.delay_seconds)
+            if type == "images":
+                return SearchResponse.from_dict(IMAGES_FIXTURE)
+            return SearchResponse.from_dict(SEARCH_FIXTURE)
+        finally:
+            self.active -= 1
+
+    async def chat(self, prompt: str, *, system: str | None = None, json_mode: bool = False, cache: bool = True) -> ChatResponse:
+        self.chat_calls.append(prompt)
+        return ChatResponse.from_dict(CHAT_FIXTURE)
+
+    async def describe(self, url: str) -> DescribeResponse:
+        self.describe_calls.append(url)
+        return DescribeResponse.from_dict(DESCRIBE_FIXTURE)
+
+
+class TestFrontierScheduling(unittest.IsolatedAsyncioTestCase):
+
+    async def test_configured_concurrency_bounds_pool_and_drains_frontier(self) -> None:
+        config = ResearchConfig()
+        self.assertEqual(config.max_concurrency, 8)
+        frontier = QueryFrontier()
+        for i in range(64):
+            self.assertTrue(frontier.push_query(f"query {i}", "seed"))
+        issued: list[str] = []
+        active = 0
+        peak = 0
+
+        async def worker() -> None:
+            nonlocal active, peak
+            active += 1
+            peak = max(peak, active)
+            try:
+                while True:
+                    query = frontier.pop_query()
+                    if query is None:
+                        return
+                    issued.append(query)
+                    await asyncio.sleep(0)
+            finally:
+                active -= 1
+
+        await asyncio.gather(*(worker() for _ in range(config.max_concurrency)))
+        stats = frontier.snapshot()
+        self.assertEqual(peak, config.max_concurrency)
+        self.assertEqual(stats.queries_enqueued, 64)
+        self.assertEqual(stats.queries_issued, 64)
+        self.assertEqual(len(issued), 64)
+        self.assertEqual(len(set(issued)), 64)
+        self.assertEqual(frontier.pending_count(), 0)
+
+    async def test_concurrent_pools_never_issue_same_query_twice(self) -> None:
+        frontier = QueryFrontier()
+        for i in range(50):
+            frontier.push_query(f"variant {i}", "seed")
+        issued: list[str] = []
+
+        async def pool(size: int) -> None:
+            async def pull() -> None:
+                while True:
+                    query = frontier.pop_query()
+                    if query is None:
+                        return
+                    issued.append(query)
+                    await asyncio.sleep(0)
+
+            await asyncio.gather(*(pull() for _ in range(size)))
+
+        await asyncio.gather(pool(4), pool(4))
+        stats = frontier.snapshot()
+        self.assertEqual(stats.queries_enqueued, 50)
+        self.assertEqual(stats.queries_issued, 50)
+        self.assertEqual(len(issued), 50)
+        self.assertEqual(len(set(issued)), 50)
+
+    async def test_queries_enqueued_while_pool_running_are_drained(self) -> None:
+        frontier = QueryFrontier()
+        for i in range(8):
+            frontier.push_query(f"early {i}", "seed")
+        issued: list[str] = []
+        stop = asyncio.Event()
+
+        async def worker() -> None:
+            while not stop.is_set() or frontier.has_pending():
+                query = frontier.pop_query()
+                if query is None:
+                    await asyncio.sleep(0)
+                    continue
+                issued.append(query)
+                await asyncio.sleep(0)
+
+        workers = [asyncio.create_task(worker()) for _ in range(4)]
+        await asyncio.sleep(0)
+        for i in range(5):
+            frontier.push_query(f"late {i}", "result")
+        stop.set()
+        await asyncio.gather(*workers)
+        stats = frontier.snapshot()
+        self.assertEqual(stats.queries_issued, 13)
+        self.assertEqual(len(set(issued)), 13)
+        self.assertEqual(frontier.pending_count(), 0)
+
+    def test_snapshot_accounting_is_consistent(self) -> None:
+        frontier = QueryFrontier("subject")
+        self.assertFalse(frontier.push_query("subject"))
+        self.assertTrue(frontier.push_query("second query"))
+        self.assertTrue(frontier.register_url("https://example.com/a"))
+        self.assertFalse(frontier.register_url("https://example.com/a"))
+        self.assertTrue(frontier.register_content("body text"))
+        self.assertFalse(frontier.register_content("body text"))
+        stats = frontier.snapshot()
+        self.assertEqual(stats.queries_generated, 3)
+        self.assertEqual(stats.queries_enqueued, 2)
+        self.assertEqual(stats.queries_duplicates_skipped, 1)
+        self.assertEqual(stats.urls_seen, 1)
+        self.assertEqual(stats.urls_duplicates_skipped, 1)
+        self.assertEqual(stats.content_seen, 1)
+        self.assertEqual(stats.content_duplicates_skipped, 1)
+
+
+class TestFrontierConcurrencyDedup(unittest.IsolatedAsyncioTestCase):
+
+    async def test_duplicate_query_pushes_under_concurrency_enqueue_once(self) -> None:
+        frontier = QueryFrontier()
+
+        async def push() -> bool:
+            return frontier.push_query("same query", "origin")
+
+        results = await asyncio.gather(*(push() for _ in range(64)))
+        stats = frontier.snapshot()
+        self.assertEqual(results.count(True), 1)
+        self.assertEqual(stats.queries_generated, 64)
+        self.assertEqual(stats.queries_enqueued, 1)
+        self.assertEqual(stats.queries_duplicates_skipped, 63)
+
+    async def test_concurrent_url_registration_dedups(self) -> None:
+        frontier = QueryFrontier()
+        urls = [
+            "https://example.com/page",
+            "https://EXAMPLE.com/page",
+            "https://example.com/page/",
+        ] * 21 + ["https://example.com/page"]
+
+        def storm() -> list[bool]:
+            with ThreadPoolExecutor(max_workers=16) as pool:
+                return list(pool.map(frontier.register_url, urls))
+
+        results = await asyncio.to_thread(storm)
+        stats = frontier.snapshot()
+        self.assertEqual(results.count(True), 1)
+        self.assertEqual(stats.urls_seen, 1)
+        self.assertEqual(stats.urls_duplicates_skipped, 63)
+
+    async def test_concurrent_content_registration_dedups(self) -> None:
+        frontier = QueryFrontier()
+
+        def storm() -> list[bool]:
+            with ThreadPoolExecutor(max_workers=16) as pool:
+                return list(pool.map(frontier.register_content, ["identical page body"] * 64))
+
+        results = await asyncio.to_thread(storm)
+        stats = frontier.snapshot()
+        self.assertEqual(results.count(True), 1)
+        self.assertEqual(stats.content_seen, 1)
+        self.assertEqual(stats.content_duplicates_skipped, 63)
+
+    async def test_overlapping_results_registered_once_under_concurrency(self) -> None:
+        frontier = QueryFrontier()
+        for i in range(32):
+            frontier.push_query(f"query {i}", "seed")
+        issued: list[str] = []
+
+        async def worker() -> None:
+            while True:
+                query = frontier.pop_query()
+                if query is None:
+                    return
+                issued.append(query)
+                frontier.register_result(
+                    SearchResult(title=query, url="https://example.com/shared", description="", source="s", extra={})
+                )
+                await asyncio.sleep(0)
+
+        await asyncio.gather(*(worker() for _ in range(8)))
+        stats = frontier.snapshot()
+        self.assertEqual(len(issued), 32)
+        self.assertEqual(stats.urls_seen, 1)
+        self.assertEqual(stats.urls_duplicates_skipped, 31)
+
+    async def test_same_content_different_urls_registered_once(self) -> None:
+        frontier = QueryFrontier()
+
+        async def register(index: int) -> bool:
+            return frontier.register_result(
+                SearchResult(
+                    title=f"title {index}",
+                    url=f"https://example.com/page/{index}",
+                    description="",
+                    source="s",
+                    content="identical page body",
+                    extra={},
+                )
+            )
+
+        results = await asyncio.gather(*(register(i) for i in range(16)))
+        stats = frontier.snapshot()
+        self.assertEqual(results.count(True), 16)
+        self.assertEqual(stats.urls_seen, 16)
+        self.assertEqual(stats.content_seen, 1)
+        self.assertEqual(stats.content_duplicates_skipped, 15)
+
+
+class TestTTLCacheBehaviour(unittest.TestCase):
+
+    def test_repeat_key_returns_cached_value(self) -> None:
+        cache = TTLCache[str]("repeat", ttl_seconds=60.0)
+        cache.set("key", "value")
+        self.assertEqual(cache.get("key"), "value")
+        self.assertIs(cache.get("key"), cache.get("key"))
+
+    def test_unknown_key_returns_none(self) -> None:
+        cache = TTLCache[str]("missing", ttl_seconds=60.0)
+        self.assertIsNone(cache.get("absent"))
+
+    def test_zero_ttl_boundary_immediately_expired(self) -> None:
+        cache = TTLCache[str]("zero", ttl_seconds=0.0)
+        cache.set("key", "value")
+        self.assertIsNone(cache.get("key"))
+
+    def test_negative_ttl_never_returns_value(self) -> None:
+        cache = TTLCache[str]("negative", ttl_seconds=-1.0)
+        cache.set("key", "value")
+        self.assertIsNone(cache.get("key"))
+
+    def test_fresh_entry_survives_within_ttl(self) -> None:
+        cache = TTLCache[str]("fresh", ttl_seconds=60.0)
+        cache.set("key", "value")
+        self.assertEqual(cache.get("key"), "value")
+
+    def test_set_overwrites_existing_entry(self) -> None:
+        cache = TTLCache[str]("overwrite", ttl_seconds=60.0)
+        cache.set("key", "first")
+        cache.set("key", "second")
+        self.assertEqual(cache.get("key"), "second")
+
+    def test_clear_removes_all_entries(self) -> None:
+        cache = TTLCache[str]("clear", ttl_seconds=60.0)
+        for i in range(10):
+            cache.set(f"key-{i}", f"value-{i}")
+        cache.clear()
+        for i in range(10):
+            self.assertIsNone(cache.get(f"key-{i}"))
+
+    def test_entry_expires_after_ttl_elapses(self) -> None:
+        cache = TTLCache[str]("expiry", ttl_seconds=10.0)
+        with mock.patch("typosaurus_sandbox.research.cache.time.monotonic", side_effect=[100.0, 100.0, 111.0]):
+            cache.set("key", "value")
+            self.assertEqual(cache.get("key"), "value")
+            self.assertIsNone(cache.get("key"))
+
+    def test_entry_expires_exactly_at_ttl_boundary(self) -> None:
+        cache = TTLCache[str]("boundary", ttl_seconds=10.0)
+        with mock.patch("typosaurus_sandbox.research.cache.time.monotonic", side_effect=[100.0, 100.0, 110.0]):
+            cache.set("key", "value")
+            self.assertEqual(cache.get("key"), "value")
+            self.assertIsNone(cache.get("key"))
+
+
+class TestTTLCacheConcurrency(unittest.TestCase):
+
+    def test_concurrent_distinct_keys_all_retrievable(self) -> None:
+        cache = TTLCache[str]("concurrent", ttl_seconds=60.0)
+        keys = [f"key-{i}" for i in range(256)]
+
+        def worker(key: str) -> None:
+            cache.set(key, key + "-value")
+            self.assertEqual(cache.get(key), key + "-value")
+
+        with ThreadPoolExecutor(max_workers=16) as pool:
+            list(pool.map(worker, keys))
+        for key in keys:
+            self.assertEqual(cache.get(key), key + "-value")
+
+    def test_concurrent_same_key_sets_single_consistent_value(self) -> None:
+        cache = TTLCache[str]("storm", ttl_seconds=60.0)
+        values = [f"value-{i}" for i in range(128)]
+
+        def worker(value: str) -> None:
+            cache.set("shared", value)
+            self.assertIn(cache.get("shared"), values)
+
+        with ThreadPoolExecutor(max_workers=16) as pool:
+            list(pool.map(worker, values))
+        self.assertIn(cache.get("shared"), values)
+        self.assertEqual(len(cache._entries), 1)
+
+
+class TestPipelineSingleMechanism(unittest.IsolatedAsyncioTestCase):
+
+    async def test_web_images_chat_describe_dispatch_through_single_request_mechanism(self) -> None:
+        client = RsearchClient()
+        recorded: list[tuple[Any, ...]] = []
+
+        def fake(
+            method: str,
+            path: str,
+            params: dict[str, str] | None,
+            payload: bytes | None,
+            headers: dict[str, str] | None,
+            timeout: float | None,
+        ) -> tuple[int, dict[str, Any]]:
+            recorded.append((method, path, params, payload, headers, timeout))
+            if path == "/search":
+                return 200, SEARCH_FIXTURE
+            if path == "/chat":
+                return 200, CHAT_FIXTURE
+            return 200, DESCRIBE_FIXTURE
+
+        client._request = fake
+        await client.search("subject")
+        await client.search("subject", type="images")
+        await client.chat("prompt")
+        await client.describe("https://example.com/page")
+        await client.describe_raw(b"\x89PNGpayload", mime_type="image/png")
+
+        self.assertEqual(len(recorded), 5)
+        search_calls = [call for call in recorded if call[1] == "/search"]
+        self.assertEqual(len(search_calls), 2)
+        self.assertEqual(search_calls[0][0], "GET")
+        self.assertEqual(search_calls[1][0], "GET")
+        web_params = dict(search_calls[0][2] or {})
+        images_params = dict(search_calls[1][2] or {})
+        self.assertEqual(web_params, {"query": "subject"})
+        self.assertEqual(images_params, {"query": "subject", "type": "images"})
+        self.assertEqual({k: v for k, v in images_params.items() if k != "type"}, web_params)
+        chat_calls = [call for call in recorded if call[1] == "/chat"]
+        describe_calls = [call for call in recorded if call[1] == "/describe"]
+        self.assertEqual(len(chat_calls), 1)
+        self.assertEqual(chat_calls[0][0], "POST")
+        self.assertEqual(len(describe_calls), 2)
+        self.assertEqual(describe_calls[0][0], "GET")
+        self.assertEqual(describe_calls[1][0], "POST")
+
+
+class TestResearchPipeline(unittest.IsolatedAsyncioTestCase):
+
+    async def test_pipeline_pool_size_bounded_by_configured_concurrency(self) -> None:
+        client = _FakeClient()
+        frontier = QueryFrontier()
+        default_pipeline = ResearchPipeline(client, frontier)
+        self.assertEqual(default_pipeline.pool_size, 8)
+        self.assertEqual(default_pipeline.pool_size, ResearchConfig().max_concurrency)
+        narrow_pipeline = ResearchPipeline(client, frontier, ResearchConfig(max_concurrency=3))
+        self.assertEqual(narrow_pipeline.pool_size, 3)
+        floor_pipeline = ResearchPipeline(client, frontier, ResearchConfig(max_concurrency=0))
+        self.assertEqual(floor_pipeline.pool_size, 1)
+
+    async def test_pipeline_drains_all_work_items_and_dedups(self) -> None:
+        client = _FakeClient()
+        frontier = QueryFrontier()
+
+        async def items() -> AsyncIterator[WorkItem]:
+            yield WorkItem("web", "subject")
+            yield WorkItem("web", "subject")
+            yield WorkItem("images", "subject")
+            yield WorkItem("describe", "https://example.com/page")
+            yield WorkItem("chat", "prompt")
+
+        pipeline = ResearchPipeline(client, frontier, ResearchConfig(max_concurrency=4))
+        report = await pipeline.run(items())
+        self.assertEqual(report.requests_succeeded, 5)
+        self.assertEqual(report.requests_failed, 0)
+        self.assertEqual(len(report.outcomes), 5)
+        self.assertTrue(all(outcome.success for outcome in report.outcomes))
+        endpoints = {outcome.item.kind: outcome.endpoint for outcome in report.outcomes}
+        self.assertEqual(endpoints, {"web": "/search", "images": "/search", "describe": "/describe", "chat": "/chat"})
+        self.assertEqual(len(client.search_calls), 3)
+        self.assertEqual(len(client.describe_calls), 1)
+        self.assertEqual(len(client.chat_calls), 1)
+        stats = frontier.snapshot()
+        self.assertEqual(stats.urls_seen, 2)
+        self.assertEqual(stats.urls_duplicates_skipped, 1)
+        self.assertEqual(stats.queries_enqueued, 6)
+        self.assertEqual(stats.queries_duplicates_skipped, 2)
+        self.assertEqual(stats.content_seen, 2)
+
+    async def test_pipeline_concurrency_bounded_by_pool_size(self) -> None:
+        client = _FakeClient()
+        client.delay_seconds = 0.02
+        frontier = QueryFrontier()
+
+        async def items() -> AsyncIterator[WorkItem]:
+            for i in range(12):
+                yield WorkItem("web", f"subject {i}")
+
+        pipeline = ResearchPipeline(client, frontier, ResearchConfig(max_concurrency=4))
+        report = await pipeline.run(items())
+        self.assertEqual(report.requests_succeeded, 12)
+        self.assertGreaterEqual(client.peak, 2)
+        self.assertLessEqual(client.peak, 4)
+
+
+if __name__ == "__main__":
+    unittest.main()
+
+
+