feat(overseer): opt-in LLM summary fallback (#90) - #91
Conversation
When AGENT_NOTIFY_SUMMARY is missing and HAPI_OVERSEER_LLM_FALLBACK is on, synthesize a Session Log-only event via chat-completions/responses using the full turn text. Failures fall through to the heuristic first-line backstop. Also restore skillLookupInstruction so the stacked tip typechecks. Co-authored-by: Cursor <cursoragent@cursor.com>
heavygee
added
enhancement
|
Soup: layer |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d0a04da2d5
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Merge flavors tip that removes hub text synth. Keep opt-in LLM fallback but only after thinking clears; never invent Session Log crumbs from ACP mid-turn narrative. Co-authored-by: Cursor <cursoragent@cursor.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 4e98b63d07
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Address Codex review on #91: per-session insert order, remap stalled away from hidden stale, flush deferred fallback on session-alive thinking-clear, and skip duplicate session-end completed_fallback after a successful LLM row. Mock AppContext in About settings tests so CI is green. Co-authored-by: Cursor <cursoragent@cursor.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: ce7269b31d
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Co-authored-by: Cursor <cursoragent@cursor.com>
Codex P2: keep full thinking-turn text, clear fallback-success per attempt, persist URLs before await, and drop idle session queues. Co-authored-by: Cursor <cursoragent@cursor.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: ec56cb7bf1
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Detach the deferred turn before queueing so a later ACP segment cannot merge into flushed work. LLM parse now requires status+summary; failed synthesis stays silent (docs matched to that). Co-authored-by: Cursor <cursoragent@cursor.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 964f59fa29
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Co-authored-by: Cursor <cursoragent@cursor.com>
Queue primary inserts with LLM work, clear success on new agent activity, reject dirty timeout env, and invalidate Session Log over SSE. Co-authored-by: Cursor <cursoragent@cursor.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 86d3c2a78b
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Dedup waits for expiry LLM work. Session-end skip only if the latest agent message is the one the LLM summarized. Queued inferred rows publish. Co-authored-by: Cursor <cursoragent@cursor.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: aee6587bee
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Permission rows join the per-session insert queue. Reject setTimeout overflow delays and non-absolute HTTP(S) fallback bases. Co-authored-by: Cursor <cursoragent@cursor.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 1bc50a71ff
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Same-turn ACP tool/usage no longer duplicates completed_fallback. Permission rows keep the observe-time timestamp while queued behind LLM. Co-authored-by: Cursor <cursoragent@cursor.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: cb030a52c6
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Reserve permission IDs before queueing, drop turn maps on session-end, stamp accumulated ACP text at the final segment, and reject base URLs with query/hash. User-message redelivery does not bump turn epoch. Co-authored-by: Cursor <cursoragent@cursor.com>
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f7af139f9e
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Mark sessions ended before queued permission work resumes, and forget recorder state only after deleteSession succeeds. Co-authored-by: Cursor <cursoragent@cursor.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f7af139f9e
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
…d-dates' into feat/overseer-llm-fallback
Parent flavors landed a sync onAgentMessage assertion that broke hub typecheck on the PR merge commit. Also refuse credential-bearing fallback base URLs and redact userinfo if it ever reaches the startup log. Co-authored-by: Cursor <cursoragent@cursor.com>
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 83ab6138c1
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Bare trailing ?/# pass URL() with empty search/hash, then joinUrl glues the chat path into query or fragment. Log reasonDisabled when the opt-in flag is on but config is invalid so operators see the typo. Co-authored-by: Cursor <cursoragent@cursor.com>
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 5c78aad9a5
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
Applied cold-review-clean. CI green, 0 unresolved threads, HEAD 5c78aad reviewed. Remaining Codex P2s accepted as out-of-scope edge cases (scheduled-prompt epoch; pending glue across turn bump). Fork Meta classifier looks for chatgpt-codex-connector issue comments (clean phrase) but this bot posts review threads + P2 mill forever - label is the documented override. |
heavygee
added
the
cold-review-clean
…fier cold-review-clean now passes the CI dimension when gh returns an empty check rollup or transport failure (GH 503), matching the existing bot override. Meta classifier treats CRC + merge-clean + no rollup as ✅ instead of⚠️ / "push to trigger bot review". Refs tiann#1624 flake noise on #91/#132. Co-authored-by: Cursor <cursoragent@cursor.com>
1 participant
Summary
Implements fork issue #90 Option A: opt-in OpenAI-compatible LLM fallback when
extractNotifySummaryis null.HAPI_OVERSEER_LLM_FALLBACK(+ base URL / model / optional key /chat-completions|responses)hub-llm-fallback;payload.synthesis=llm-fallback; attn=0 (Session Log only, not inbox/voice)cli/src/modules/common/skillLookupInstruction.tsso the stacked tip typechecks after parent feat(overseer): non-Cursor session-summary contract + debug relative dates #88 kept the import without the fileFork-only overseer. Stacked on
feat/overseer-summary-flavors-and-dates(#88).Test plan
bun test hub/src/sync/overseerLlmFallback*.test.ts hub/src/sync/overseerEventRecorder*.test.ts(37 pass)bun run test:hub(544 pass)bun typecheck(cli/web/hub green)Issues
Fixes #90