tai.py (8643 lines) is now a thin facade over tai_pkg/ (33 feature modules, one class or feature per file). tai-orig.py preserves the pre-split single-file version byte for byte (sha256 verified against snapshot); import tai and ./tai.py behavior unchanged.
Equivalence proof, all run on this tree immediately pre-commit: differential battery tai-orig.py vs facade 3718 checks 0 failures (surface, pure fuzz, tool dispatch, store/audit, release ops, fake-HOME install roundtrip, CLI runs); test_tai.py 261 OK; test_seal.py 16 OK; --version/--help byte-identical; dir(tai) zero removals.
Fix included: opencode wire-protocol schema echo leaked into replies; opencode_parse now scans with JSONDecoder.raw_decode capped at 200 scans (13 OpencodeParseTests). Test count 261 = 237 HEAD + 8 pre-existing uncommitted ToolArgumentsTests (now committed with this change, disclosed) + 13 OpencodeParseTests + 3 EmbeddedStringsTests. Patch-target-only test migration (24 hunks), zero assertion changes. tai_pkg clean-regenerated after mid-verification tamper was found and reverted.
The Backends section was two lines about the primary HTTP gateway and
didn't mention the five CLI-wrapped backends, the box last-resort
fallback, the infinite-retry design, or the codex/claude/gemini/muse/grok
status distinction from the previous commit. Document all of it with the
actual measured facts (pollinations' queue cap, zen's current 403, the
Node-tarball-vs-apt size difference, opencode's zero-credential free
tier) rather than a restated feature list.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
codex_health()/gemini_health()/claude_health()/grok_health()/muse_health()
ran only `<cli> --version` and printed "ok, subscription" on any zero exit
-- which proves the binary runs, nothing about whether it's authenticated.
Confirmed as a real false claim: `codex login status` reported "Not logged
in" (exit 1) on a machine whose banner nonetheless said "codex: ok,
subscription".
codex and claude each expose a real, cheap, read-only status check
(`codex login status`, `claude auth status` -- the latter returns JSON
with a loggedIn field) and now use it. grok, muse, and gemini have no
such command (confirmed against gemini --help and muse --help), so they
now honestly report "login unverified" instead of asserting a subscription
this check never measured.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- Plexer: always-active terminal multiplexing. NativePlexerController/
TmuxPlexerController share one interface, chosen by which env var is
already set (TAI_PLEXER_SOCKET vs TMUX_PANE); NativePlexerBackend is
the pty.fork()-based wrapper (raw mode, SIGWINCH resize forwarding,
a Unix-socket control protocol, continuous session logging, a
Ctrl+A prefix key distinct from tmux's own Ctrl+B). Auto-wraps new
interactive shells via a .bashrc hook mirroring the existing
command-not-found install pattern, gated on $TMUX/$TAI_PLEXER_ACTIVE/
a $HOME/.tai-no-plexer bypass. New plexer_send/read/type/status
tools, env-gated on a detected controller.
- tests/container.py + docker-compose.yml + tests/docker/Dockerfile:
a disposable podman/docker image for the tests that must never touch
a real machine (installing the plexer hook into a real ~/.bashrc, a
live tmux session). Required suite (test_seal.py, test_tai.py,
tests/plexer.py, tests/bashrc.py) gates the result; tests/rag.py
runs as best-effort since its real-LLM dependency can't be satisfied
inside a disposable container with no nested engine.
- tests/bashrc.py: real end-to-end hook install + bash -i auto-wrap,
anti-recursion, tmux-active skip, non-interactive skip, a live
TmuxPlexerController test, and an install/uninstall round-trip.
- tests/plexer.py: NativePlexerBackend under a real pty, plus a
dedicated robustness pass (file permissions, >64KB payload framing,
a malformed control request, a genuine self-deadlock fix in the
write-to-child path).
- tai_test.py: the shared async process-driving framework (pipe/pty
transport, group-wide kill-and-reap, a StdinBridge for --interactive)
both test files build on.
- ChatClient last-resort backend: when there's no local opencode CLI
but the sandbox box is already running, install opencode into it
(official Node tarball, not apt -- apt's nodejs/npm pull in
webpack/babel/jest as transitive deps) and route through it instead
of giving up. Only reachable after the first retry round already
failed everything else; never creates the box itself. TmuxPlexerController.status()
fixed to use `list-panes` instead of `display-message -p`, which
silently fell back to the default pane instead of erroring against
a dead target.
- Fix 4 pre-existing test_tai.py bugs surfaced by running it in a
clean environment for the first time: two stale mock signatures
(subprocess.run's stdout= kwarg, Chat.complete's deadline= kwarg),
one $TMUX-dependent catalog assertion, one tty-dependent color
assertion.
- research/: a sourced research archive (Source Dossier/Citation
Card/Conclusion tiers) backing the above design decisions, with its
own audit.py consistency checker.
- .gitignore: fix a blanket *.md rule that was about to silently drop
the entire research/ archive from version control; add __pycache__/
and the one-off debug transcript this session started from.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- install_skill: adopt an existing skill from a local dir, SKILL.md,
.zip, or git URL; symlinks when the source already lives in a
recognized external skills folder (.claude/skills, .agents/skills)
instead of vendoring a duplicate, with install provenance recorded
per skill. Skill discovery now also reads those external folders
read-only, so skills placed by other tools are visible with no
install step at all.
- export_terminal_log: full tmux scrollback export, visible only
inside a live tmux session (TOOL_ENV_GATES, a new env-conditional
layer on top of the existing keyword-based lazy tool loading).
- Chunked lexical RAG (SQLite FTS5 + bm25, no vectors/embeddings):
new chunks table wired into record_save, shell/subagent output
spill, and terminal log export, searchable via the existing search
tool (kind chunk) with full chunk text returned, not just a
snippet. delete_record cleans up a record's chunks with it.
- Fix a regression where create_skill's default scope silently
became context-dependent instead of always "project".
- Rewrite README.md to document all of the above plus the existing
context/tool-selection and search/memory internals in depth.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
web_fetch is now a real HTTP client: methods, custom headers, string
bodies, configurable timeout, status plus content-type plus size on
every response, content-aware rendering (HTML to text, JSON raw),
and server error bodies surfaced on HTTP failures. Shell, schema
descriptions, and the system prompt steer all HTTP(S) to web_fetch;
curl/wget shell use earns a model-side hint instead of a refusal.
Shell output streams live in a rolling 4-line window for the active
agent only, with an exit line showing return code, elapsed time,
line count, and byte size. Workers, pipes, and capture stay silent.
Fixed step budget replaced by a progress budget: productive steps
(unseen actions, unseen results, user interaction) reset the stall
counter; identical repeats get a loop warning at 3 and a stop at 5;
wider stalls stop at the patience budget; 500-step total cap
backstops everything. 210 + 16 tests green.
Per-profile bots (system plus history, shared vault): create_bot with
nicknames, /bots and /bot with resume, @mention one-turn routing filed
in both histories.
Unified retrieval: one trigger-synced FTS5 index over records, events,
and audit with safe tokenized MATCH, porter stemming, BM25 rank, and
marked snippets. New search tool with kind filters and graph expansion;
recall, record_search, and audit rank through the same index.
Sealed search without weakening the seal: each boot decrypts events
into a :memory: FTS5 index (newest 10000, incrementally synced, never
on disk), so sealed stores rank like plaintext ones. LIKE fallback kept
for partial tokens. Documented in README under Sealed search with the
researched alternatives (SQLCipher, blind indexes, SSE) and rejections.
Also in this batch: sysinfo, create_skill, markdown renderer with
colors, scheduler with subagent rule, --yolo/--auto, denial-loop fix,
read-before-write guard, WAL plus timeouts, records/graph/spillover,
audit time travel, self-backup and releases, lazy skill blueprints,
smart tagging, profile isolation, lazy tool payload, install
subsystem, and multi-profile rotation fix. 176 + 16 tests green.
retoor
