agents: split GPT-5 prompt and retry behavior#65597
Conversation
openclaw-barnacle
Bot
added
agents
Greptile SummaryThis PR splits two concerns out of a larger GPT-5 runtime work item: (1) the GPT-5 prompt overlay, now organized into three exported named constants ( Confidence Score: 5/5Safe to merge — logic is correct, regex patterns are well-tested, no breaking changes. All findings are P2 or lower. The regex patterns (SINGLE_ACTION_NARRATIVE_RE, PLANNING_ONLY_PROMISE_RE) behave correctly across the tested edge cases including the 'let me know' exclusion and 'i can do that' cue. The control flow in resolvePlanningOnlyRetryInstruction correctly gates the single-action-narrative bypass so that the outer text.includes('```') and other null returns still fire as expected. No files require special attention. Reviews (1): Last reviewed commit: "agents: split GPT-5 prompt and retry beh..." | Re-trigger Greptile |
![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: 2c0d6da010
ℹ️ About Codex in GitHub
Codex has been enabled to automatically 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 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 007c83dad4
ℹ️ About Codex in GitHub
Codex has been enabled to automatically 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 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
007c83d to
6b40925
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 012963962d
ℹ️ About Codex in GitHub
Codex has been enabled to automatically 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 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 57cf940866
ℹ️ About Codex in GitHub
Codex has been enabled to automatically 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 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 8d8d473cc6
ℹ️ About Codex in GitHub
Codex has been enabled to automatically 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 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 5569a935cd
ℹ️ About Codex in GitHub
Codex has been enabled to automatically 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 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| const SINGLE_ACTION_EXPLICIT_CONTINUATION_RE = | ||
| /\b(?:going to|first[, ]+i(?:'ll| will)|next[, ]+i(?:'ll| will)|then[, ]+i(?:'ll| will)|i can do that next|let me (?!know\b)\w+(?:\s+\w+){0,3}\s+(?:next|then|first)\b)/i; |
There was a problem hiding this comment.
SINGLE_ACTION_EXPLICIT_CONTINUATION_RE currently matches bare going to, so factual answer text can be misread as plan-only narration. In resolvePlanningOnlyRetryInstruction, a turn with one safe tool call (read/grep/etc.) then an answer like “The parser is going to crash on empty input” now satisfies the single-action bypass and returns PLANNING_ONLY_RETRY_INSTRUCTION, even though the assistant already delivered content. That creates unnecessary retries and can escalate to strict-agentic blocked exits on valid completions; this cue should be limited to first-person continuation patterns (for example, I'm going to/I'll ... next) rather than generic going to.
Useful? React with 👍 / 👎.
* agents: split GPT-5 prompt and retry behavior * agents: fix GPT-5 review follow-ups * agents: address GPT-5 review follow-ups * agents: avoid replaying side-effectful GPT retries * agents: mark subagent control as mutating * agents: fail closed on single-action retries * commands: stabilize channel legacy doctor migration test * agents: narrow single-action retry promise trigger
* agents: split GPT-5 prompt and retry behavior * agents: fix GPT-5 review follow-ups * agents: address GPT-5 review follow-ups * agents: avoid replaying side-effectful GPT retries * agents: mark subagent control as mutating * agents: fail closed on single-action retries * commands: stabilize channel legacy doctor migration test * agents: narrow single-action retry promise trigger
* agents: split GPT-5 prompt and retry behavior * agents: fix GPT-5 review follow-ups * agents: address GPT-5 review follow-ups * agents: avoid replaying side-effectful GPT retries * agents: mark subagent control as mutating * agents: fail closed on single-action retries * commands: stabilize channel legacy doctor migration test * agents: narrow single-action retry promise trigger
* agents: split GPT-5 prompt and retry behavior * agents: fix GPT-5 review follow-ups * agents: address GPT-5 review follow-ups * agents: avoid replaying side-effectful GPT retries * agents: mark subagent control as mutating * agents: fail closed on single-action retries * commands: stabilize channel legacy doctor migration test * agents: narrow single-action retry promise trigger
* agents: split GPT-5 prompt and retry behavior * agents: fix GPT-5 review follow-ups * agents: address GPT-5 review follow-ups * agents: avoid replaying side-effectful GPT retries * agents: mark subagent control as mutating * agents: fail closed on single-action retries * commands: stabilize channel legacy doctor migration test * agents: narrow single-action retry promise trigger
* agents: split GPT-5 prompt and retry behavior * agents: fix GPT-5 review follow-ups * agents: address GPT-5 review follow-ups * agents: avoid replaying side-effectful GPT retries * agents: mark subagent control as mutating * agents: fail closed on single-action retries * commands: stabilize channel legacy doctor migration test * agents: narrow single-action retry promise trigger
1 participant
Summary
I can do thatcue andlet me knowhandoff exclusionExcluded From This PR
src/agents/pi-embedded-runner/run.tsTesting