neuron-technologies/neuron
2
0
Fork 0
Code Issues 21 Pull Requests 4 Actions Packages Projects Releases Wiki Activity

Agent web-search/report pipeline: 8 fixes proven on the clean soul — patch ready for main (soul.c + chat.el upstream) #74

New Issue
Open
opened 2026-07-11 17:59:44 +00:00 by tim.lingo · 4 comments
tim.lingo commented 2026-07-11 17:59:44 +00:00
Member
Copy Link
Copy Source

Tim's 'research MSFT + build HTML report' failed silently (real server-side searches ran, zero shown, no report). Root-caused and FIXED on the clean :7798 soul, E2E-proven (full report on disk, all actions visible). Patch: neuron-container-build/soul-webfix-20260711.patch (185 lines vs origin/main dist/soul.c) — needs your .el-side mirror + regen for main/prod.

The 8 fixes, each E2E-verified:

  1. main regression: web_search request flag ignored — handle_chat_agentic never attaches the server-side search tool (older deployed binaries had it). Restored, honoring the app's flag.
  2. pause_turn never resumed — Anthropic's server-tool loop pauses mid-turn; docs require re-send-and-resume; failing = silently truncated answers. Resume loop added (cap 8→12).
  3. Server-side work invisible — server_tool_use blocks now counted generically by name into tools_used (future server tools appear automatically — code_execution already shows up via the new search variant's dynamic filtering).
  4. Tool version future-proofing — web_search_20250305 (deprecated) → config-driven via state key web_search_tool_version, default web_search_20260209; on API rejection falls back to 20250305, logs [soul] DRIFT: + sets state web_search_version_drift (verified with a bogus version end-to-end).
  5. Duplicate tool name — homegrown client web_search stripped when the server tool is attached (model was picking nondeterministically between the two).
  6. container_id contract — the 20260209 variant's code-execution filtering requires echoing container.id on follow-ups or the API 400s; now captured + carried.
  7. Honest tools_used + cap notice — client tools only logged when actually dispatched/bridged (a max_tokens-truncated tool block was being logged as 'used' without running); on cap-hit mid-action the reply now says so. max_tokens 4096→16384 for report-sized write_file inputs.
  8. LLM error logging — the loop was swallowing API error bodies entirely; first 220 chars now logged.

Also: runtime env EL_HTTP_TIMEOUT_MS=300000 needed for long agentic turns (60s default times out; knob already existed). launch-clean.sh sets it; prod launcher should too.

Architecture recommendation (your call): provider-facing literals (model IDs, tool versions, stop reasons, wire shapes) are sprinkled through chat.el — by VBD they belong in one accessor layer with config overrides. Today's drift class keeps recurring until then.

App-side companion: neuron-ui PR #123 (workspace grounding line). Monthly drift audit now in our BACKLOG (next due 2026-08-11).

— Tim + Neuron

Tim's 'research MSFT + build HTML report' failed silently (real server-side searches ran, zero shown, no report). Root-caused and FIXED on the clean :7798 soul, E2E-proven (full report on disk, all actions visible). Patch: `neuron-container-build/soul-webfix-20260711.patch` (185 lines vs origin/main dist/soul.c) — needs your .el-side mirror + regen for main/prod. **The 8 fixes, each E2E-verified:** 1. **main regression: `web_search` request flag ignored** — handle_chat_agentic never attaches the server-side search tool (older deployed binaries had it). Restored, honoring the app's flag. 2. **pause_turn never resumed** — Anthropic's server-tool loop pauses mid-turn; docs require re-send-and-resume; failing = silently truncated answers. Resume loop added (cap 8→12). 3. **Server-side work invisible** — `server_tool_use` blocks now counted generically by name into tools_used (future server tools appear automatically — code_execution already shows up via the new search variant's dynamic filtering). 4. **Tool version future-proofing** — `web_search_20250305` (deprecated) → config-driven via state key `web_search_tool_version`, default `web_search_20260209`; on API rejection falls back to 20250305, logs `[soul] DRIFT:` + sets state `web_search_version_drift` (verified with a bogus version end-to-end). 5. **Duplicate tool name** — homegrown client `web_search` stripped when the server tool is attached (model was picking nondeterministically between the two). 6. **container_id contract** — the 20260209 variant's code-execution filtering requires echoing `container.id` on follow-ups or the API 400s; now captured + carried. 7. **Honest tools_used + cap notice** — client tools only logged when actually dispatched/bridged (a max_tokens-truncated tool block was being logged as 'used' without running); on cap-hit mid-action the reply now says so. max_tokens 4096→16384 for report-sized write_file inputs. 8. **LLM error logging** — the loop was swallowing API error bodies entirely; first 220 chars now logged. **Also:** runtime env `EL_HTTP_TIMEOUT_MS=300000` needed for long agentic turns (60s default times out; knob already existed). launch-clean.sh sets it; prod launcher should too. **Architecture recommendation (your call):** provider-facing literals (model IDs, tool versions, stop reasons, wire shapes) are sprinkled through chat.el — by VBD they belong in one accessor layer with config overrides. Today's drift class keeps recurring until then. App-side companion: neuron-ui PR #123 (workspace grounding line). Monthly drift audit now in our BACKLOG (next due 2026-08-11). — Tim + Neuron
will.anderson was assigned by tim.lingo 2026-07-11 17:59:46 +00:00
tim.lingo commented 2026-07-11 18:08:12 +00:00
Author
Member
Copy Link
Copy Source

Addendum: the sandbox soul now also implements GET/POST /api/safety-contact per issue #97's spec and the app's exact contract (DaemonClient.kt shapes; 988 auto-fill; RFC3339 set_at; persist path env-overridable via SOUL_SAFETY_CONTACT_PATH so clean-profile testing never writes ~/.neuron). Round-trip verified: unset GET -> {configured:false}; POST 988 -> full record + ok; GET -> persists. This unblocked the Hard Bell gate (BUG-5) on the clean profile — main's soul.c still needs your .el-side implementation for prod. Updated patch: neuron-container-build/soul-webfix-20260711.patch superseded by the full diff in ~/soul-fix-build/soul.c (regenerating the .patch file now includes safety-contact + all 8 web fixes).

Addendum: the sandbox soul now also implements GET/POST /api/safety-contact per issue #97's spec and the app's exact contract (DaemonClient.kt shapes; 988 auto-fill; RFC3339 set_at; persist path env-overridable via SOUL_SAFETY_CONTACT_PATH so clean-profile testing never writes ~/.neuron). Round-trip verified: unset GET -> {configured:false}; POST 988 -> full record + ok; GET -> persists. This unblocked the Hard Bell gate (BUG-5) on the clean profile — main's soul.c still needs your .el-side implementation for prod. Updated patch: neuron-container-build/soul-webfix-20260711.patch superseded by the full diff in ~/soul-fix-build/soul.c (regenerating the .patch file now includes safety-contact + all 8 web fixes).
tim.lingo commented 2026-07-11 18:21:46 +00:00
Author
Member
Copy Link
Copy Source

Second addendum — RUNTIME BUG FOUND during the safety-contact work, worth its own attention: HTTP responses that embed an fs_read string get clamped to the FILE's byte count (Content-Length pinned to file size, body truncated mid-JSON). Reproduced deterministically; survived three dodges (fresh str_slice copy, full rebuild-from-parsed-fields — truncation persisted as long as fs_read ran in the request). Workaround shipped: read the file via exec_capture('cat ...') whose buffer path flows through responses intact. This same aliasing may explain the historical /api/graph/edges emptiness and other truncation ghosts. Suspect: response writer measuring an aliased/stale allocation from fs_read. el_runtime-level fix belongs with you. Sandbox soul now serves valid safety-contact JSON both verbs (strict-validated); Hard Bell gate cleared; tamper-notice dialog fired correctly for the out-of-band change — L1 behaving beautifully.

Second addendum — RUNTIME BUG FOUND during the safety-contact work, worth its own attention: **HTTP responses that embed an fs_read string get clamped to the FILE's byte count** (Content-Length pinned to file size, body truncated mid-JSON). Reproduced deterministically; survived three dodges (fresh str_slice copy, full rebuild-from-parsed-fields — truncation persisted as long as fs_read ran in the request). Workaround shipped: read the file via exec_capture('cat ...') whose buffer path flows through responses intact. This same aliasing may explain the historical /api/graph/edges emptiness and other truncation ghosts. Suspect: response writer measuring an aliased/stale allocation from fs_read. el_runtime-level fix belongs with you. Sandbox soul now serves valid safety-contact JSON both verbs (strict-validated); Hard Bell gate cleared; tamper-notice dialog fired correctly for the out-of-band change — L1 behaving beautifully.
tim.lingo commented 2026-07-11 18:21:59 +00:00
Author
Member
Copy Link
Copy Source

Second addendum — RUNTIME BUG FOUND during the safety-contact work: HTTP responses embedding an fs_read string get clamped to the FILE's byte count (Content-Length pinned to file size, body truncated mid-JSON). Deterministic; survived three dodges (str_slice copy, rebuild-from-parsed-fields) — truncation persisted whenever fs_read ran in-request. Workaround: read via exec_capture('cat ...'), whose buffers flow through responses intact. May explain historical /api/graph/edges emptiness + other truncation ghosts. el_runtime-level fix belongs with you. Sandbox now serves valid safety-contact JSON both verbs (strict-validated); Hard Bell gate cleared; tamper-notice fired correctly for the out-of-band change — L1 behaving exactly as designed.

Second addendum — RUNTIME BUG FOUND during the safety-contact work: HTTP responses embedding an fs_read string get clamped to the FILE's byte count (Content-Length pinned to file size, body truncated mid-JSON). Deterministic; survived three dodges (str_slice copy, rebuild-from-parsed-fields) — truncation persisted whenever fs_read ran in-request. Workaround: read via exec_capture('cat ...'), whose buffers flow through responses intact. May explain historical /api/graph/edges emptiness + other truncation ghosts. el_runtime-level fix belongs with you. Sandbox now serves valid safety-contact JSON both verbs (strict-validated); Hard Bell gate cleared; tamper-notice fired correctly for the out-of-band change — L1 behaving exactly as designed.
tim.lingo commented 2026-07-11 18:22:36 +00:00
Author
Member
Copy Link
Copy Source

Second addendum — RUNTIME BUG FOUND during the safety-contact work, worth separate attention:

HTTP responses that embed an fs_read string get clamped to the FILE's byte count — Content-Length pinned to the file's size, body truncated mid-JSON. Deterministic repro; survived three dodges (fresh str_slice copy, full rebuild-from-parsed-fields) — the truncation persisted as long as fs_read ran anywhere in the request. Workaround shipped: read the file via exec_capture("cat ..."), whose buffer path provably flows through responses intact. This same aliasing may explain the historical /api/graph/edges emptiness and other truncation ghosts. Suspect: the response writer measuring a stale/aliased allocation from fs_read. The el_runtime-level fix belongs with you.

Status: sandbox soul now serves valid safety-contact JSON on both verbs (strict-validated); the Hard Bell gate cleared; and the tamper-notice dialog fired correctly for the out-of-band contact change — the L1 layer behaving exactly as designed. Full updated patch: neuron-container-build/soul-webfix-20260711.patch.

Second addendum — RUNTIME BUG FOUND during the safety-contact work, worth separate attention: **HTTP responses that embed an `fs_read` string get clamped to the FILE's byte count** — Content-Length pinned to the file's size, body truncated mid-JSON. Deterministic repro; survived three dodges (fresh `str_slice` copy, full rebuild-from-parsed-fields) — the truncation persisted as long as `fs_read` ran anywhere in the request. Workaround shipped: read the file via `exec_capture("cat ...")`, whose buffer path provably flows through responses intact. This same aliasing may explain the historical `/api/graph/edges` emptiness and other truncation ghosts. Suspect: the response writer measuring a stale/aliased allocation from `fs_read`. The el_runtime-level fix belongs with you. Status: sandbox soul now serves valid safety-contact JSON on both verbs (strict-validated); the Hard Bell gate cleared; and the tamper-notice dialog fired correctly for the out-of-band contact change — the L1 layer behaving exactly as designed. Full updated patch: `neuron-container-build/soul-webfix-20260711.patch`.
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
No items
No labels
Milestone
No items
No Milestone
Projects
Clear projects
No projects
Assignees
Clear assignees
tim.lingo (Tim Lingo) will.anderson (Will Anderson)
No Assignees will.anderson
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: neuron-technologies/neuron#74
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?