[dbdefault]

Summary

Fixes #1278. A positive vote on a document that was never recalled left last_recalled_at empty. computeConfidence treated that as 180 days old, so recency was 0 and confidence fell below HOT_THRESHOLD, applying COLD_PENALTY and demoting the doc in recall search.

Changes

• Recency now uses the later of lastRecalledAt and lastUpvotedAt.
• Search index aggregation passes lastUpvotedAt into computeConfidence.
• Unit tests cover upvote-only recency and the search-index hot path.
• Docs (EN + zh-CN docs/guide/knowledge.md, skill-data/core/references/troubleshooting.md) now say a fresh upvote also refreshes recency.

Real CLI run

Built dist/ with npm run build and ran the CLI in a throwaway HOME (temp dir, local bare origin, usageReport: false). The team repo has two learnings that match the same query: retry-backoff.md and retry-jitter.md. Neither has ever been recalled. Only retry-backoff gets one upvote:

$ teamai pull
$ teamai recall feedback --positive retry-backoff
  ✔ Upvoted: retry-backoff
$ teamai pull
  ✔ Synced 2 learnings (index: 17ms)
$ jq -c '.entries[] | {filename, votes, confidence, hotness}' ~/.teamai/search-index.json
$ cat ~/.teamai/user-votes/alice.yaml
$ teamai recall --check "retry failed requests"

Build	retry-backoff.md in search-index.json	recall --check top hit
this PR (4ad6eb5)	"votes":1,"confidence":0.72,"hotness":1	score=6.3 title="Retry failed requests" (the voted doc)
main (8a94438)	"votes":1,"confidence":0.42,"hotness":0.3	score=5.9 title="Retry failed requests with jitter" (the unvoted doc)

In both runs the vote file showed recalled_count: 0, upvoted_count: 1, last_recalled_at: "" and a fresh last_upvoted_at. retry-jitter.md had confidence: null and hotness: 1 in both.

Impact on promote / prune / --confidence-writeback

computeConfidence also drives prune (src/maintenance/prune.ts), promote (src/maintenance/promote.ts, MIN_CONFIDENCE = 0.90) and --confidence-writeback. Taking the later of the two timestamps can only raise confidence, never lower it. That has two effects:

• prune/archive: a learning that was upvoted recently but never recalled no longer looks stale, so it is less likely to be archived.
• promote: a learning that already meets the vote, contributor and age thresholds, but whose last recall is old, can become promotable after one fresh upvote, because recency goes from 0 back to 1. Upvotes can still add at most the capped 30% vote factor.

I worked this out from the formula and did not add a promote/prune case in this PR. I'm happy to add one if you'd like it pinned down.

Test plan

• npx vitest run src/__tests__/confidence.test.ts src/__tests__/search-index-local-votes.test.ts (24 passed)
• npx tsc --noEmit, npm run lint
• Full npx vitest run: 11961 passed, 7 failed. The same 7 also fail on main 8a94438 on this machine (cursor-bubble-range, cursor-user-images, git-discovery, submodule-checkout, windows-test-list), so this PR doesn't cause them.
• Real CLI run above

Note: #1282 adds a test at the end of the same describe block in search-index-local-votes.test.ts. If it lands first, I'll rebase and keep both tests.

AI disclosure

This change was prepared with AI assistance (Claude / Grok coding agents). The values above come from real runs on my machine.

[github-actions]

• [P1 blocking] Missing required real-CLI verification. This PR changes runtime confidence calculation and search ranking, but its Test plan explicitly says build, e2e, and real-CLI verification were not run. This violates the first ## Code Review Rules item requiring an end-to-end / real-CLI record for runtime changes. After building, run one representative real-CLI scenario verifying that a recently upvoted, never-recalled document avoids the cold penalty, and document the result in the PR description.

No other confident findings in the current diff.

[SaulMoro]

This was generated by AI during triage.

@dbdefault Thanks, the fix works. We built 4ad6eb5d and origin/main (8a944380) and rebuilt the search index with the CLI in an isolated HOME. There were two learnings that match the same query, and one of them has a single upvote and no recall:

Build	Upvote date	Voted doc in the index	Ranking
main	now	confidence 0.42, hotness 0.3	below the unvoted doc
this PR	now	confidence 0.72, hotness 1	first
this PR	200 days ago	confidence 0.42, hotness 0.3	below the unvoted doc

tsc, lint and the related unit tests pass, and so does index-local-votes-1233.

Blocking

[P1 blocking] The description has no real-CLI record, and this change affects recall ranking at runtime (Code Review Rules, first item). One run is enough. For example, upvote a never-recalled learning, rebuild the index, and paste the doc's confidence and hotness from ~/.teamai/search-index.json along with the commands you ran.

Non-blocking

[P2 non-blocking] computeConfidence also drives prune (src/maintenance/prune.ts:38), promote (src/maintenance/promote.ts:26, MIN_CONFIDENCE = 0.90) and --confidence-writeback. Taking the later of the two timestamps can only raise confidence. As a result, a stale learning that already meets the vote thresholds can become promotable after one fresh upvote: 1.0 instead of 0.7. We worked this out from the formula and did not run it. Please mention it in the description. A promote or prune case for a fresh upvote would pin it down.

[P2 non-blocking] Two docs now contradict the change. docs/guide/knowledge.md:102 (and the same line in docs/guide/zh-CN/knowledge.md) and skill-data/core/references/troubleshooting.md:374-375 say upvotes cannot replace the recency factor. After this PR, an upvote does reset recency. These docs need a one-line update in both languages.

[P3 nit] The timestamp comparison is a string compare. That is safe today because every writer uses toISOString().

Note: #1282 (approved, not merged) adds a test at the end of the same describe block in src/__tests__/search-index-local-votes.test.ts. Whichever of the two PRs lands second has to keep both tests. src/utils/search-index.ts merges cleanly.

[dbdefault]

Thanks @SaulMoro for the careful triage and for reproducing it on both builds.

• [P1] Real CLI record: Added a "Real CLI run" section to the description. I built the PR and main, upvoted a never-recalled learning in a throwaway HOME, and rebuilt the index with teamai pull. The voted doc shows confidence 0.72, hotness 1 and ranks first on this PR, and confidence 0.42, hotness 0.3 on main. The exact commands are included.
• [P2] promote/prune: Explained in the description. The change can only raise confidence, so a stale learning that already meets the thresholds can become promotable after one fresh upvote, and recently upvoted learnings are less likely to be pruned. I can add a promote/prune test if you'd like.
• [P2] Docs: Updated docs/guide/knowledge.md, docs/guide/zh-CN/knowledge.md and skill-data/core/references/troubleshooting.md in 7dc5701.
• [P3]: Noted. The comparison relies on every writer using toISOString().
• Re fix(recall): preserve confirmed votes in narrow clones #1282: whichever PR lands second will keep both tests.

[github-actions]

No confident findings in the specified diff.

Resolved: The earlier [P1 blocking] missing real-CLI verification finding is fixed. The PR description now records a build and a representative real-CLI comparison confirming that a recently upvoted, never-recalled document avoids the cold penalty.

The verification names an older commit (4ad6eb5), not the reviewed head (7dc5701); under the review rules, that is a note, not a blocking finding.

[SaulMoro]

This was generated by AI during triage.

@dbdefault Thanks. This re-reviews 7dc5701b against our comment at 4ad6eb5d.

• [P1 blocking] Real-CLI record: resolved. The description now has the recall feedback --positive run on both builds, and its numbers match ours (0.42/0.3 on main, 0.72/1 on this PR). The record names 4ad6eb5, but 7dc5701 only changes docs, so it still applies.
• [P2 non-blocking] promote / prune / --confidence-writeback: resolved. The description now explains the effect. A dedicated test is optional.
• [P2 non-blocking] Docs: resolved. docs/guide/knowledge.md, docs/guide/zh-CN/knowledge.md and skill-data/core/references/troubleshooting.md now say that recency uses the later of the last recall and the last upvote, and the English and Chinese versions match.

Approving. The order with #1282 is unchanged: whichever PR lands second keeps both tests in search-index-local-votes.test.ts.

[SaulMoro]

This was generated by AI during triage.

@dbdefault CI is red on 7dc5701b. Every Lint & Test job and the Windows job fail on windows-test-list.test.ts and codex-team-hooks.test.ts. Those failures came from main, not from this PR, and #1287 fixed them in c15fc0db.

Please rebase or merge current main so CI runs again. The code review from the earlier approval still holds, and we'll approve again once CI is green. If #1282 lands first, keep both tests in search-index-local-votes.test.ts, as you planned.

[dbdefault]

@SaulMoro Thanks — merged latest main (including c15fc0d from #1287) with a normal merge commit, no other changes: 6689e95. Merge was clean; build and search-index-local-votes.test.ts (14/14) pass locally.

[github-actions]

No confident findings in the specified diff.

Resolved: The earlier [P1 blocking] missing real-CLI verification finding is fixed. The PR description records a build and a representative real-CLI comparison demonstrating that a recently upvoted, never-recalled document avoids the cold penalty.

The verification names an older commit (4ad6eb5), not the reviewed head (6689e95); under the review rules, this is a note, not a blocking finding.

[SaulMoro]

This was generated by AI during triage.

@dbdefault #1282 has landed (c4a61fc3), so this branch now conflicts with main in src/__tests__/search-index-local-votes.test.ts. Both PRs add a test at the end of the same describe block. Please rebase onto current main and keep both tests: #1282's keeps unrelated same-member v1 scores while applying local feedback and your keeps an upvote-only document hot when last_upvoted_at is recent (#1278). src/utils/search-index.ts merges cleanly.

We tried that resolution locally on top of #1282: tsc and lint pass, and so do the vote, confidence and search-index tests. Once the rebased head has green CI, we'll approve and merge.

[SaulMoro]

@dbdefault #1282 landed first, so we resolved the conflict on your branch ourselves (2be9947a, a merge of current main). Both tests in search-index-local-votes.test.ts are kept and nothing else changed. We'll approve and merge as soon as CI is green, so there's nothing left for you to do here. Thanks a lot for the fix and for turning the review around so quickly.

[github-actions]

No confident findings in the specified diff.

Resolved: The earlier [P1 blocking] missing real-CLI verification finding is fixed. The PR description documents a build and a representative real-CLI comparison showing that a recently upvoted, never-recalled document avoids the cold penalty.

Note: Verification names 4ad6eb5, not reviewed head 2be9947. Under the trusted review rules, this is not blocking.

Review was read-only; no PR code was run, built, or installed.

[SaulMoro]

This was generated by AI during triage.

Approving 2be9947a. The only change since our earlier approval is the merge of current main. It keeps both tests in search-index-local-votes.test.ts, and tsc, lint and the related tests pass locally. We'll merge once CI is green.