Context
The skill-learning-from-outcomes feature has landed (ADR docs/adr/skill-learning-from-outcomes.md, docs/skill-learning.md). The backend loop works end-to-end (verified live), but the user-facing surface is thin: today a user only sees a daily digest land in a "Skill review" channel. They cannot act on proposals in-app, cannot see what changed in a skill, and there is no over-time view. This issue tracks the full Skill review experience (web first; mobile follow-up).
What a user sees today (the gap)
- Digest only. The daily digest renders as a chat block in the auto-provisioned "Skill review" channel (
feature: "skill-review") and flags unread like any channel. That's the only proactive surface.
- No in-app actions. Approve/Reject a proposal is CLI-only (
gini improvement). useImprovements() exists in web/src/lib/queries.ts but no component renders it.
- Skill changes are invisible / misleading. Outcome-driven edits change the
SKILL.md body, but SkillVersion/previousVersions doesn't carry body, so the Skills page shows only a version bump, not a diff — and the existing Rollback button is wrong for body edits (see the linked correctness bug).
- No "improvements over time."
skillOutcomes / learningFindings are API-only; nothing renders them. There's no trend, no per-skill reliability, no changelog. (successCount/failureCount are validation counts, not the learning signal — don't surface them as quality.)
Scope
1. Actionable proposals
- List pending
ImprovementProposals (kind:"skill", payload.mode:"edit"): show the bounded diff (payload.edits / candidateBody vs baseBody) + rationale + evidence (sourceTaskId).
- Approve / Reject in-app (
POST /api/improvements/:id/approve|reject).
- Revert an applied edit in-app — needs a new
POST /api/improvements/:id/revert route wrapping the existing revertImprovement (re-installs baseBody). Without it there is no correct in-app undo for a body edit.
2. Findings
- Open
LearningFindings (environment / credential / model_ignored / bundled_skill) with a dismiss action.
3. Improvements over time
- Per-skill reliability / failure trend from
SkillOutcome rows over a recent window, plus a recent-outcomes feed and a record of accepted edits (a real changelog).
- This needs a defined skill score — see the scoring note below; the honest metric is reliability + adjudication-confidence, not a fake accuracy number.
4. Skill change history (pairs with the linked bug)
- Show the body diff for an applied edit and a correct revert. Requires capturing the prior body (see linked bug) —
previousVersions currently can't represent body edits.
Backend available today
GET /api/learning/outcomes, GET /api/learning/findingsGET /api/improvements + POST /api/improvements/:id/approve|reject- Digest renders in the
feature:"skill-review" channel.
Still needed (backend)
POST /api/improvements/:id/revert (wrap revertImprovement).- Body-level skill history (linked bug) so the panel can show a diff and revert correctly.
- (Optional) a per-skill score/aggregation endpoint, or compute reliability client-side from
/api/learning/outcomes.
Notes
- Dual client: keep web + mobile renderers in sync if chat-block content is surfaced (ADR
chat-block-protocol.md). Web-first is fine.
- Consistent with the existing Jobs/Agents panels.
Related correctness bug: skill body edits are not represented in version history and the Rollback button doesn't undo them.
Context
The skill-learning-from-outcomes feature has landed (ADR
docs/adr/skill-learning-from-outcomes.md,docs/skill-learning.md). The backend loop works end-to-end (verified live), but the user-facing surface is thin: today a user only sees a daily digest land in a "Skill review" channel. They cannot act on proposals in-app, cannot see what changed in a skill, and there is no over-time view. This issue tracks the full Skill review experience (web first; mobile follow-up).What a user sees today (the gap)
feature: "skill-review") and flags unread like any channel. That's the only proactive surface.gini improvement).useImprovements()exists inweb/src/lib/queries.tsbut no component renders it.SKILL.mdbody, butSkillVersion/previousVersionsdoesn't carrybody, so the Skills page shows only a version bump, not a diff — and the existing Rollback button is wrong for body edits (see the linked correctness bug).skillOutcomes/learningFindingsare API-only; nothing renders them. There's no trend, no per-skill reliability, no changelog. (successCount/failureCountare validation counts, not the learning signal — don't surface them as quality.)Scope
1. Actionable proposals
ImprovementProposals (kind:"skill",payload.mode:"edit"): show the bounded diff (payload.edits/candidateBodyvsbaseBody) + rationale + evidence (sourceTaskId).POST /api/improvements/:id/approve|reject).POST /api/improvements/:id/revertroute wrapping the existingrevertImprovement(re-installsbaseBody). Without it there is no correct in-app undo for a body edit.2. Findings
LearningFindings (environment/credential/model_ignored/bundled_skill) with a dismiss action.3. Improvements over time
SkillOutcomerows over a recent window, plus a recent-outcomes feed and a record of accepted edits (a real changelog).4. Skill change history (pairs with the linked bug)
previousVersionscurrently can't represent body edits.Backend available today
GET /api/learning/outcomes,GET /api/learning/findingsGET /api/improvements+POST /api/improvements/:id/approve|rejectfeature:"skill-review"channel.Still needed (backend)
POST /api/improvements/:id/revert(wraprevertImprovement)./api/learning/outcomes.Notes
chat-block-protocol.md). Web-first is fine.Related correctness bug: skill body edits are not represented in version history and the Rollback button doesn't undo them.