docs(narrated-runs): engine notes for the regen — compiled-form fixes + debts

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-13 13:40:07 -05:00
parent aa67f86f90
commit a54770d606
1 changed files with 34 additions and 0 deletions
@@ -0,0 +1,34 @@
+# Narrated runs — engine notes for Will (2026-07-13)
+
+Source half: commit aa67f86 on feat/agent-phase1-soul (run-progress ledger,
+`/api/run-progress/<sid>` route, narration on the pause envelope, config display
+default). E2E-verified via the compiled test bed on Tim's clean profile.
+
+Compiled-form-only fixes (in `neuron-container-build/soul-narrated-runs-20260713.patch`,
+applies ON TOP of `soul-webfix-20260711.patch` — these need porting to chat.el when the
+webfix itself is ported):
+
+1. **pause_turn + tool_use interleave**: a pause_turn response can ALSO carry a client
+   tool_use; resuming verbatim leaves it unpaired → Anthropic 400 "tool_use ids were
+   found without tool_result". Fix: tool-bearing pause rounds are tool turns
+   (dispatch + pair); verbatim resume only when the round has no client tool.
+2. **Agentic toolset scope**: agentic_tools_all() fed EVERY connector/MCP tool (Notion,
+   code-execution…) into the loop. Code-execution flips the API into programmatic
+   tool calling, whose pairing protocol the single-tool manual loop does not speak —
+   source of the dangling-pair 400s AND the bash_code_execution workspace-dodge.
+   Fix: handle_chat_agentic declares builtins + ONE server web_search only.
+   Connector tools return when the loop gains real multi-tool/programmatic support.
+3. **disable_parallel_tool_use: true** on agentic requests — the loop captures only the
+   first tool_use per round; Opus-class models parallel-call. Enforce the invariant.
+4. **web_search server-tool default variant → web_search_20250305 (GA)**. The 20260209
+   variant couples to code-execution ⇒ programmatic mode (see #2, and the June note:
+   "inert unless code-execution attached").
+5. **Homegrown web_search removed** from the tool catalog (server-side is the one tool).
+
+Known engine debts this work surfaced (not fixed):
+
+- **Poisoned session history**: a failed run persists the malformed assistant turn; every
+  later turn in that session replays it and 400s. Needs history sanitation on load.
+- **Huge-history invalid-escape 400** (~346KB request) — likely the same poisoned blob.
+- **macOS note**: replacing a binary in place invalidates its ad-hoc signature (instant
+  silent SIGKILL, looks like exit 0). `rm + cp + codesign -f -s -` is the swap ritual.