Conversation
Refined review of PR #8 (codex/auto-start-proxy-port-fallback) with 2 must-fix, 2 should-fix, and 1 note. No code changes — review only. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Squash of PR #8 — centralizes startup logic into `ocx ensure`, adds port fallback when busy, `codexAutoStart` dashboard toggle, and `ocx uninstall` for clean removal. Co-Authored-By: 이완우 <Ingwannu@users.noreply.github.com> Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- server.ts: remove redundant dynamic import of saveConfig, use static import - cli.ts: log sync failures in handleEnsure instead of silent catch - Dashboard.tsx: set error state on autostart toggle save failure Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: Ingwannu <ingwannu@users.noreply.github.com>
Dashboard now has two new panels to configure the web-search and vision sidecar models (+ reasoning effort for search) without editing config.json. API: GET/PUT /api/sidecar-settings. Supports gpt-5.3-codex-spark. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Merge origin/main into dev — kept dev's sidecar model settings, improved error logging, and main's shim internal-command guard. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This was referenced Jul 29, 2026
lidge-jun
added a commit
that referenced
this pull request
Aug 4, 2026
…claimed it did Six findings, all from executed probes or an independent review that ran the code rather than reading it. The one that matters: I wrote that `injectCodexConfig` was a production caller WP11 could serve on the apply direction. It is not. Production reaches it directly from sync.ts and cli/init.ts, never through convergence.ts — the only module this phase was scoped to touch — so the entry point I planned would have been an export nothing calls. It also awaits history mid-function and there is no runtime producer of AdmissionSnapshot at all. My own falsification test for the narrowing fired, so WP11 becomes mechanism-only and the caller moves to WP12, recorded as such instead of implied away. The rest: - The coordinator refuses to open on any routed home (probed both directions: clean opens, routed refuses legacy-ambiguous), so moving the catalog commit under N would regress every applied install. - The residue guard reads the ambient CODEX_HOME while the lock keys on a caller-supplied one — a routed home locked while a clean one was checked. The obvious fix is itself a TOCTOU, and it also refuses a symlinked default home against itself. - ACL success is cached by pathname, not file identity: unlink, recreate at the same name, and the replacement is credited with the old file's hardening. Absence-as-guarantee #15, and it is live in shipped code. - Deadline exhaustion is classified as permanent refusal, discarding ETIMEDOUT. - One citation was wrong and two were short; the type fence redeclares a brand that convergence-types and transition-state already own.
3 tasks done
wxj123654
added a commit
to wxj123654/opencodex
that referenced
this pull request
Sep 16, 2026
…reasoning replay A bundle of live-calibrated fixes for the Cascade wire: - session-identity.ts: stable per-conversation ids (lidge-jun#16 cascadeId, lidge-jun#15 model config id + monotonic turn counter, lidge-jun#22 exchange id rotated per user exchange and absent on the first turn) so a long agent loop no longer reads as N brand-new sessions to the velocity limiter - loop-fuse.ts: conversation-scoped step budget modeled on the language server's own mechanism — steer on empty turns, fail closed at the cap - images.ts: normalize replayed images under the upstream ~14MB nginx body cap so one oversized screenshot cannot 413 every later turn - proto.ts: ASSISTANT source value, ModelConfig (lidge-jun#15), prompt_cache_key (lidge-jun#27), and a 413 error that names the real cause - client.ts: 300s upstream idle budget matching the turn budget, and an abort-raced read so a silent stream still trips the idle path - chat/inbound.ts: replay assistant reasoning_content as a reasoning item before the message so downstream providers see the thinking the model actually produced - tests: session identity, loop fuse, images, proto, and inbound reasoning-replay coverage; layout maps registered
agentHits
pushed a commit
to agentHits/opencodex
that referenced
this pull request
Sep 17, 2026
feat: sidecar model settings, stream fixes, autostart fallback
agentHits
pushed a commit
to agentHits/opencodex
that referenced
this pull request
Sep 17, 2026
…claimed it did Six findings, all from executed probes or an independent review that ran the code rather than reading it. The one that matters: I wrote that `injectCodexConfig` was a production caller WP11 could serve on the apply direction. It is not. Production reaches it directly from sync.ts and cli/init.ts, never through convergence.ts — the only module this phase was scoped to touch — so the entry point I planned would have been an export nothing calls. It also awaits history mid-function and there is no runtime producer of AdmissionSnapshot at all. My own falsification test for the narrowing fired, so WP11 becomes mechanism-only and the caller moves to WP12, recorded as such instead of implied away. The rest: - The coordinator refuses to open on any routed home (probed both directions: clean opens, routed refuses legacy-ambiguous), so moving the catalog commit under N would regress every applied install. - The residue guard reads the ambient CODEX_HOME while the lock keys on a caller-supplied one — a routed home locked while a clean one was checked. The obvious fix is itself a TOCTOU, and it also refuses a symlinked default home against itself. - ACL success is cached by pathname, not file identity: unlink, recreate at the same name, and the replacement is credited with the old file's hardening. Absence-as-guarantee lidge-jun#15, and it is live in shipped code. - Deadline exhaustion is classified as permanent refusal, discarding ETIMEDOUT. - One citation was wrong and two were short; the type fence redeclares a brand that convergence-types and transition-state already own.
6 of 7 tasks
lidge-jun
pushed a commit
that referenced
this pull request
Sep 28, 2026
…tory overflow - Send the leading system text as GetChatMessage #2 instead of folding it into the first user prompt. Live on swe-1-6 with a ~6.7k-token system prompt the turn-2 cache ratio is unchanged (6688/6715 vs 7072/7097) and the prompt is smaller; swe-1-6, swe-2-medium, gemini and claude all obey #2. - Mark failed tool results with ChatMessagePrompt #9 tool_result_is_error instead of an in-band "ERROR:" prefix. - Rewrite JSON-Schema type arrays to anyOf in tool parameters for Gemini uids, which Cognition refuses with invalid_argument on every turn. - Surface a pre-output invalid_argument on a history near or past the model's input window as context_length_exceeded so Codex compacts; small requests with the same code stay a plain 400. - Correct the #15 (CortexTrajectoryReference), #17 prompt_id and #22 execution_id comments. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 29688bf)
lidge-jun
pushed a commit
that referenced
this pull request
Sep 28, 2026
…tory overflow - Send the leading system text as GetChatMessage #2 instead of folding it into the first user prompt. Live on swe-1-6 with a ~6.7k-token system prompt the turn-2 cache ratio is unchanged (6688/6715 vs 7072/7097) and the prompt is smaller; swe-1-6, swe-2-medium, gemini and claude all obey #2. - Mark failed tool results with ChatMessagePrompt #9 tool_result_is_error instead of an in-band "ERROR:" prefix. - Rewrite JSON-Schema type arrays to anyOf in tool parameters for Gemini uids, which Cognition refuses with invalid_argument on every turn. - Surface a pre-output invalid_argument on a history near or past the model's input window as context_length_exceeded so Codex compacts; small requests with the same code stay a plain 400. - Correct the #15 (CortexTrajectoryReference), #17 prompt_id and #22 execution_id comments. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 29688bf)
lidge-jun
added a commit
that referenced
this pull request
Sep 28, 2026
…flag tool errors, system prompt in #2 (carry #6092) (#6178) * fix(devin): system prompt in #2, tool-error flag, Gemini schemas, history overflow - Send the leading system text as GetChatMessage #2 instead of folding it into the first user prompt. Live on swe-1-6 with a ~6.7k-token system prompt the turn-2 cache ratio is unchanged (6688/6715 vs 7072/7097) and the prompt is smaller; swe-1-6, swe-2-medium, gemini and claude all obey #2. - Mark failed tool results with ChatMessagePrompt #9 tool_result_is_error instead of an in-band "ERROR:" prefix. - Rewrite JSON-Schema type arrays to anyOf in tool parameters for Gemini uids, which Cognition refuses with invalid_argument on every turn. - Surface a pre-output invalid_argument on a history near or past the model's input window as context_length_exceeded so Codex compacts; small requests with the same code stay a plain 400. - Correct the #15 (CortexTrajectoryReference), #17 prompt_id and #22 execution_id comments. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 29688bf) * fix(devin): measured overflow boundary, per-type Gemini branches, tool-error marker - Overflow: binary-searched live on swe-1-6 (200k window) the refusal sits at ~200-203k real tokens for prose and JSON alike, and a 32000 output cap does not move it. Characters per token ran 1.33-5.53 across samples, so the chars/4 estimate is replaced by a word-piece count (1.00-1.54x real) at 95% of the window: every sample at the window is caught, none at 60% is. - Gemini schemas: `{type:[T,"null"], ...}` becomes anyOf branches that carry the type-specific keywords (items, properties, ...); an existing anyOf is folded in rather than nested under allOf. Draft-7 `dependencies` is walked as a schema map with name lists left as data. - Tool errors keep the in-band ERROR: marker beside #9: with a neutral result flagged as an error only gemini reported a failure; swe-1-6, gpt-6-sol-low and gpt-5-6-luna-low read it as success. - A request with only system text keeps it as a user prompt instead of sending no prompts. - Docs: describe the context_length_exceeded reclassification beside the allowDevinInvalidArgument recovery option. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit d84edae) * fix(devin): keep outer constraints when folding a Gemini anyOf Folding each existing anyOf branch into the outer keywords let a branch override a contradicting outer keyword, so `{maxLength: 5, anyOf: [{maxLength: 50}]}` loosened to 50. On any such disagreement, keep both constraints under allOf instead; live, gemini-3-8-flash-medium accepts allOf. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit db28e2d) * docs(devin): match the collapseSystemIntoUser example to the #2 system prompt The example still showed the leading system messages folded into the first user turn, which now travel in request #2. Show a mid-conversation system run instead, with the real blank-line join, and note that a system-only request passes through whole. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit b2d9ded) * docs(devin): show the blank-line join in the collapse example Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit 3232330) * fix(devin): keep both constraints when a Gemini type union and anyOf are disjoint When no existing anyOf branch shared a type with the type array, the fallback kept the anyOf and dropped the type union, so the branches admitted types the node never allowed. Keep both under allOf, the same form the conflict path uses; live, gemini-3-8-flash-medium accepts it. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> (cherry picked from commit b87b332) * fix(devin): classify overflow from the selected catalog window Co-authored-by: Sayo <hi@sayo.wtf> * fix(devin): preserve null constraints in Gemini schema rewrite Co-authored-by: Sayo <hi@sayo.wtf> * docs(devin): record malformed-schema overflow ambiguity Co-authored-by: Sayo <hi@sayo.wtf> * fix(devin): preserve outer schema constraints on nullable types Co-authored-by: Sayo <hi@sayo.wtf> * fix(devin): estimate overflow from transmitted tool descriptions Co-authored-by: Sayo <hi@sayo.wtf> * docs(structure): keep the effective Devin family default in the adapter row Co-authored-by: Sayo <hi@sayo.wtf> --------- Co-authored-by: Sayo <hi@sayo.wtf> Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
7 tasks done
3 tasks done
lidge-jun
added a commit
that referenced
this pull request
Oct 4, 2026
Named Devin conversations retain an upstream trajectory across sequential turns and request-scoped adapters. Scope retained IDs by resolved credential, tenant host, and conversation; use fresh IDs for overlaps and when every retained slot is active. A 256-entry store evicts only inactive entries, and idempotent release in `finally` covers success, failure, and cancellation. The optional wire field preserves per-request allocation for unnamed calls. Refines #6488 at `a0b199fc6b96367fb174f6488a7a45eb58eccd54`. Carries the optional #15.1 field, named continuity across adapters, unnamed fallback, overlap isolation, and failure release regressions. Replaces the source's map-plus-live-set with one bounded store, hashes a structured identity tuple, and prefers own-thread identity over a shared parent. Adds credential/host/alias/parent/cancellation/eviction/overflow/idempotence/retry coverage and documents the contract. The original author's live cache measurements are original source evidence; this replacement does not claim new measured savings. Independent of #6554; neither PR's runtime commits are required by the other. Coordinator owns source disposition. Co-authored-by: Hanqing Zhao <hanqing@gatech.edu> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Verified current head f09d987 with CI 37174376575 and scoped independent technical/security review. Co-authored-by: Hanqing Zhao <hanqing@gatech.edu> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Changes
gui/src/pages/Dashboard.tsx— sidecar model config UI (search model, vision model, reasoning effort)src/server.ts—/api/configGET/PUT endpoints for sidecar settingssrc/config.ts,src/types.ts— sidecar model config typessrc/adapters/openai-chat.ts— chat tool choice fixsrc/service.ts— stream timeout handling improvementssrc/cli.ts— autostart ensure fallback + uninstall commandsrc/ports.ts— port availability check + fallback logicTest plan
npm testpasses🤖 Generated with Claude Code