fix(openai-agents): persist turn memory + expand ~/ in send_* actions

[dylanneve1]

Summary

Two production bugs hit while testing the openai-agents backend against OpenRouter.

1. No turn-to-turn memory

run() was always called without a session, so the Agents SDK couldn't carry the multi-turn ledger forward. Visible symptom: model wrote and sent an SVG file in turn 1; turn 2 ("convert it to PNG") had the model say "Telegram user client is disconnected, can't access the file" — pure hallucination because it had no record of what it just did.

Fix: per-chat MemorySession (from @openai/agents) stored in the backend's state map, passed into every run(). The SDK records model outputs, tool calls, tool results, and reasoning automatically — no Talon-side bookkeeping needed.

/reset now calls a new QueryBackend.resetChat(chatId) hook that clears the MemorySession. Stateless backends (claude-sdk, codex, kilo, opencode) ignore the hook. Session map is LRU-capped at 1000 entries.

2. ~/ not expanded in send_* actions

Models emit paths like ~/.talon/workspace/robot.svg (because that's how Talon documents the workspace). Node's fs.statSync doesn't expand ~/ — visible symptom: send_file failed: ENOENT: no such file or directory, stat '~/.talon/workspace/robot.svg'.

Fix: extract the existing expandPath from openai-agents/builtins.ts into a shared src/util/fs-path.ts. Use it in Telegram + Discord send_file / send_photo / send_video / send_animation / send_voice / send_audio / create_sticker_set / add_sticker_to_set, plus the builtin tool set (deduped).

Test plan

• npm run typecheck clean
• npm run format:check clean
• npx vitest run src/__tests__/fs-path.test.ts src/__tests__/openai-agents — 81 passed (75 openai-agents + 6 new fs-path)
• Manual smoke: bot restart, ask it to write a file, ask it to do something with that file in the next turn — model now remembers without hallucinating disconnects.
• Manual smoke: ask the bot to send_file an ~/.talon/workspace/... path — ENOENT gone.

🤖 Generated with Claude Code

[claudiusthebot]

APPROVE — two real production bugs found during live OpenRouter testing, both fixed correctly. Full diff read; CI 35/36 (one Windows job still running — expected given the functional timeout bump in this PR).

---

Bug 1: MemorySession per-chat

The root cause (run() always starting sessionless → model blind to previous turns) was correctly identified. The fix is the right one: per-chat MemorySession in a Map<string, MemorySession> in state.ts, the same instance passed to every run() for that chat.

LRU eviction via delete(chatId); set(chatId, existing) on access is the standard Map-based LRU trick — correct and zero-dependency. Cap at 1000 sessions is a sensible bound for a self-hosted bot.

clearChatSession on /reset → wired correctly in Telegram (commands.ts), Telegram admin (admin.ts), Discord (commands.ts), Discord admin (admin.ts). Optional-chaining everywhere (gateway?.backend?.resetChat?.(chatId)) is safe.

QueryBackend.resetChat as an optional hook (stateless backends ignore, openai-agents implements) is the right interface shape — doesn't break existing backends.

---

Bug 2: ~/ not expanded in send_* actions

Extracting expandFsPath to src/util/fs-path.ts is the right call. The Telegram + Discord action handlers both now route file_path through it before hitting statSync/readFileSync. Sticker tools (create_sticker_set, add_sticker_to_set) also patched.

Test coverage on expandFsPath is thorough — the ~foo case (NOT tilde-relative, should resolve as relative filename) is the subtle edge case that's easy to miss. Good that it's explicitly tested.

---

Zod → JSON Schema migration in builtins.ts

This is the most non-obvious fix and the most important one for OpenRouter compatibility.

The problem: @openai/agents's tool() factory forces strict: true for Zod schemas, which forces every declared property into required. That's spec-correct for OpenAI's own models but breaks non-OpenAI models routed through chat_completions (Trinity, Owl, etc.) — they drop optional fields like timeout_ms/offset/limit and the SDK rejects the call as "Invalid JSON input."

The fix (plain JSON Schema + strict: false + explicit required arrays with only the truly-required fields) is correct. Zod was convenient but wrong for a backend intended to run against arbitrary providers. The per-tool interface types for execute(input) keep the TypeScript surface clean.

One note: the same pattern should probably be applied to any future builtins added to this file.

---

Flow-violation contract

Handler now mirrors claude-sdk: trailing prose without end_turn/send → detectFlowViolation → one reminder retry → silent drop on second miss. incrementTurns correctly deferred until after the retry-or-drop decision. _retried guard on session naming prevents the FLOW_VIOLATION_REMINDER text getting captured as the session name. All correct.

The mcpBundle.servers.length > 0 guard (skip enforcement with frontend: "terminal", which has no delivery tools) is the right sentinel — production frontends always wire at least one MCP server.

---

CI matrix expansion

openai-agents added to the live-backend matrix. The install-backend-cli.mjs short-circuit (no external CLI needed) is clean. The dummy OpenAI server (dummy-openai-server.ts) is a solid in-process mock — scripted SSE responses, request recording, proper tool-call + tool-result round-trip support. 393-line integration test covers all 5 scenarios (endpoint discovery, plain text, tool dispatch, multi-turn memory, resetChat).

---

Functional timeout bump: 240s → 480s

Second bump in ~2 weeks (180s → 240s → 480s). Not a PR concern — Windows npm install latency is a CI infrastructure issue — but the trend is worth watching. If it keeps climbing, the right fix is probably caching the npm tarball between CI runs rather than expanding the timeout indefinitely.

---

35/36 CI green at review time (Windows Node 22 Tests in-progress, which is the job this PR's timeout bump targets). When it finishes, this should be 36/36. Clean to merge whenever you're ready.