Repository navigation
improve(skills): review retained conversation for durable learning - #139656
Merged
Merged
Conversation
Contributor
|
🦞👀 Pull request received. I will update this pull request when review starts. ClawSweeper review completeClawSweeper finished reviewing this revision. The review result is being finalized. |
Contributor
|
Codex review: needs maintainer review before merge. Reviewed September 5, 2026, 11:05 PM ET / September 6, 2026, 03:05 UTC. ClawSweeper reviewWhat this changesThe PR asks Skill Workshop to learn from the full retained conversation, removes repeated prompt mechanics and prose-pinning assertions, and corrects collection-review documentation. Merge readiness✅ Ready for maintainer review Keep open: the clarification remains useful against inspected main, and no actionable introduced defect was found. The maintainer-authored, protected PR is not a cleanup-close candidate. Priority: P3 Review scores
Verification
How this fits togetherSkill Workshop reviews saved conversation context after eligible foreground work and can stage one reusable-skill proposal. The existing configured pipeline controls whether that proposal is applied. flowchart TD
A[Eligible completed conversation] --> B[Saved transcript boundary]
B --> C[Detached Workshop review]
D[Existing skills and usage] --> C
C --> E{New durable learning?}
E -->|No| F[No change]
E -->|Yes| G[One pending proposal]
G --> H[Configured apply pipeline]
Before mergeNone. Agent review detailsSecurityNone. PR surfaceSource -12, Tests -38, Docs -6. Total -56 across 4 files. View PR surface stats
Review metricsNone. Technical reviewBest possible solution: Keep learning scope aligned with the retained conversation while preserving the existing bounded, proposal-only review pipeline. Do we have a high-confidence way to reproduce the issue? Not applicable as a deterministic bug reproduction: source confirms the instruction mismatch, while the reported paired trial succeeded with both prompts. Is this the best way to solve the issue? Yes: updating the existing review instruction and deferring patch mechanics to the tool contract is a narrow solution that preserves runtime enforcement. AGENTS.md: found and applied where relevant. Codex review notes: model internal, reasoning medium; reviewed against 77ee04bad2b0. LabelsLabel changes:
Label justifications:
EvidenceWhat I checked:
Likely related people:
Rating scale
Overall follows the weaker of proof and patch quality. Workflow
|
Collaborator
Author
|
Landed in 96feee2da216.
Merged the reviewed head through the guarded landing workflow with attribution verified. |
vincentkoc
added a commit
that referenced
this pull request
Sep 6, 2026
* commit 'a2886546df92c3a601db538b1b70125e3519a16f': test(reports): deduplicate native process census setup (#139596) fix(ui): avoid refetching unrelated catalog hosts (#139570) test(compaction): consolidate classification proof at its owner (#139564) fix(apple): harden native state file permissions (#139623) improve(skills): review retained conversation for durable learning (#139656) fix(docker): stage bundled plugin runtime dependencies under the packaged plugin root (#139117) test: close HTTP fixtures before destroying their sockets (#139595) refactor(agents): simplify attempt execution and recovery (#139597) test(gateway): share HTTP authority scenarios (#139611) test(gateway): distinguish final CLI results from draft text (#139605) fix(shell-env): preserve CLI terminal ownership during login imports (#139308)
github-actions Bot
pushed a commit
to Desicool/openclaw
that referenced
this pull request
Sep 6, 2026
…penclaw#139656) Clarify experience review's evidence scope to include the full retained conversation. Consolidate duplicate instructions without changing eligibility, transcript checkpoints, permissions, or the proposal/apply pipeline. Remove obsolete prompt fixtures and correct stale collection-review documentation. Production net -12 lines; total net -56 lines. Verified with 40 focused tests and seven local model-backed prompt scenarios, with independent source and semantic review. The paired baseline also passed; this is not a measured learning-rate improvement. Co-authored-by: Ayaan Zaidi <hi@obviy.us>
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
1 participant
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.
What Problem This Solves
Skill Workshop receives the retained conversation, but its experience prompt limited review to the latest turn, separating earlier corrections from later evidence that they worked.
Why This Change Was Made
The prompt connects requirements, corrections, attempts, and observed results across retained history. Duplicate patch instructions and an unused input are removed. The existing trigger, saved-transcript boundary, provider selection, permissions, and proposal/apply flow remain unchanged.
User Impact
Earlier lessons remain eligible after a routine final exchange. Review still abstains for covered lessons, routine work, unsupported guesses, and private facts, and can stage at most one proposal. No write access is added.
Evidence
All 40 focused scheduling, prompt, authoring, and real Workshop-tool tests passed. Tool tests retained read-before-write, pending-only mutation, stale-read rejection, and the single-mutation limit. Formatting, lint, and changed-page checks passed; full anchor checking found unrelated missing external pages. Seven synthetic multi-turn scenarios produced the expected proposals or abstentions with tool execution denied. The old prompt also captured the earlier correction in a paired check, so this clarifies instruction scope without claiming improved learning rate or a deterministic missed capture.