tai/taicpp/docs/glue-issues.md

glue issues and gaps

Found while wiring src/glue. Each entry names the file and the smallest repro. Nothing here was patched around silently; items marked "glue side" are limits the glue itself has.

Open

1. No vault seal. src/seal is empty (another agent ports Seal). repl::Boot uses store::NullSeal when TAI_PASSPHRASE is unset: the banner says seal: off, python says on (default key). An explicit TAI_PASSPHRASE=x is refused: seal error: TAI_PASSPHRASE is set but this build links no vault seal (exit 2). Repro: TAI_PASSPHRASE=x build/detz --version works (before boot), TAI_PASSPHRASE=x build/detz hi exits 2. Closes with a store::SealPort implementation passed as BootOptions::seal_factory in src/glue/app_bundle.cpp (one line).
2. SysInfo.collect_sysinfo was ported by no module (tai.py 8978-9008). src/comms/comms.cpp only has a stub sys_info(). The glue carries SysInfoProbe (sequential, each check timed, 10 s ceiling); the python runs the checks on threads. python and venv checks report the native binary instead of an interpreter.
3. release tool cannot work natively. do_release rewrites the VERSION = line of the running script. The binary has no script. OpsAdapter::do_release throws and the tool reports the error; the native flow is make release PART=... MSG=... in taicpp/. ensure_self_backup at boot is not called for the same reason.
4. Ctrl-C does not kill the call in flight (glue side). AgentReplHost expires agent.deadline, the loop stops at the next step. A running HTTP call, shell command or sandbox exec finishes first. agent::Agent has no cancellation point inside ChatClient::complete.
5. MCP port is bound to the main agent (glue side). McpHostAdapter asks approvals, names the audit actor and runs the resolver subagent through the primary agent, and uses the primary store connection. A worker agent on its own thread that calls an MCP tool therefore shares the primary sqlite connection. tai.py passes the calling agent (app) instead. Needs the port to carry the caller, i.e. McpPort::call(..., Agent&).
6. docs/wiring/capmcp.md does not exist. The task listed it; cap/mcp were wired from the headers (cap_agent_port.h, mcp_serve_main.h, mcp_host.h) and cap.md/mcp.md.
7. src/cli: Command::run() is still the stub (--help/--version only) and Parser::parse maps argc <= 1 to --help, whereas tai.py starts the REPL. main.cpp therefore calls Parser::parseOpts and GlueApp::run (which prints help only for an explicit --help). The other agent that rewrites src/cli should drop Command or keep parseOpts and the Options fields stable.
8. Makefile: -I../../moloapp/src resolves from taicpp/ of the main checkout only. From a git worktree (.claude/worktrees/*/taicpp) it needs CXXFLAGS=... -I/home/retoor/projects/moloapp/src. make test builds tests/test_main.cpp, which still calls the removed placeholder API (llm::ChatClient("openai", ...), ModelHealth, Opencode, feed_sse, provider_info; docs/wiring/agent.md); not touched here.
9. probe_backend in the REPL banner runs sequentially with a 6 s timeout per backend, like python: a dead backend delays the prompt. Not changed.
10. --tui-local model turn not proven. The chat scene starts, draws, runs slash commands through the adapter and exits cleanly (driven over a pty); a full model turn through the TUI was started but not confirmed in the vault within the observation window. --tui (grr client) needs a grr server and was not started.
11. --mplex, --plexer-wrap, --install, --update-models with writes were wired to the module entry points; only --update-models (read-only discovery) and --demo-shot were run. --install writes into $HOME and was not run on a real home.

Fixed in the glue (one-line integration errors)

• None needed in other modules: every compile error was on the glue side.