feat(agent): make next-action recommendations repo-specific

[grnlemon]

Closes #16.

Summary

• buildRepoDecision now receives the contributor's languageSet + labelHistory and computes per-repo languageMatch (from syncState.primaryLanguage ∩ profile.github.topLanguages) and labelFit (from repo.registryConfig.labelMultipliers ∩ profile.registeredRepoActivity.dominantLabels).
• whyThisHelpsFor / nextActionsFor rewritten as per-recommendation helpers taking a RepoCopyContext. Direct-PR copy names matching language + label overlap + queue size + prior merged count. Issue-discovery copy names the lane + open-issue count. Cleanup-first copy names the contributor's open PR count. Maintainer-lane copy uses repo-owner framing only.
• New publicNextActionsFor emits a sanitized variant (no raw counts) for GitHub-public surfaces. RepoDecision + DecisionAction gain publicNextActions: string[]; orchestrator's publicSafeSummary now pulls from decision.publicNextActions[0].
• recommendationText now prefers action.nextActions[0] (the new per-repo copy) over the hardcoded action-kind fallback, so the API/MCP "recommendation" headline is also concrete.
• priorityFor exempts the open_pr_pressure blocker from the cleanup_first penalty calc — fixes an ordering inversion where a cleanup_first repo with one critical blocker was outscored by a comparable pursue repo. AC feat(scoring): add situational score projections #3 (cleanup-first outranks new work) now holds end-to-end.

Acceptance criteria → coverage

• Agent plan names specific repos and work types — per-repo whyThisHelps[0] / nextActions[0] reference repo + lane + concrete signals; asserted in fixtures for direct-PR, issue-discovery, maintainer-lane.
• Maintainer-owned repo history does not inflate normal contributor-lane evidence — maintainer-lane fixture asserts the contributor's open-PR count is absent from nextActions and publicNextActions even when outcome.openPullRequests = 8.
• Cleanup-first outranks new work when open PR pressure blocks scoreability — paired-fixture regression asserts pressure.priorityScore > pursueRepo.priorityScore with the priorityFor fix.
• Direct-PR / issue-discovery / split / maintainer lanes produce different reasoning — lane-specific assertions across 4 fixtures; copy distinct on language hint, lane wording, repo-owner framing, queue/PR counts.
• Recommendations are deterministic — same-input twice returns identical repoDecisions + topActions ordering (snapshot regression test).

Public/private boundary

• Private (nextActions, whyThisHelps) may include raw counts (open PR count, merged PR count, open issue count, queue size).
• Public (publicNextActions) deliberately omits counts and contributor history numerics; regression test asserts no digits appear in publicNextActions[0] for issue-discovery and maintainer-lane fixtures.
• actionFromDecisionAction.publicSafeSummary now sources from publicNextActions instead of nextActions, closing the prior leak path (sanitizePublicSummary only redacts wallet/reward words — not digits).

Validation

• npm run typecheck
• npx vitest run test/unit test/integration — 270 tests pass
• npx vitest run --config vitest.workers.config.ts — workers suite green
• npm run changelog — refreshed CHANGELOG.md (single entry added)

Safety

• Backend-only change
• No secrets, wallet details, user PATs, raw trust scores, or private rankings exposed
• Public text avoids compensation-seeking or optimization-tactic language; new publicNextActions audited to avoid leaking counts
• OpenAPI/MCP behavior updated where needed (ContributorDecisionPackSchema per-decision shape is z.record(z.unknown()) — additive, no schema break)
• Public docs/changelogs updated where needed

Relationship to #15 / PR #26

PR #26 (issue #15) touched freshness / serving helper and is on a separate branch. This PR touches buildRepoDecision / per-recommendation copy / priorityFor / orchestrator publicSafeSummary. Surfaces overlap (decision-pack.ts, agent-orchestrator.ts, the same test files) but at different functions; expect a rebase-clean merge if PR #26 lands first.

Open questions for review

• Copy tone: I leaned on "pick a narrow change in {language} (target labels: {a, b}); queue has N open PRs" for direct-PR. Happy to revise wording in any direction — open to a per-lane copy review.
• priorityFor adjustment: I exempted the open_pr_pressure blocker only when recommendation === "cleanup_first". Alternative: raise the cleanup_first base from 75 → 105 (clamped at 100) — same effect for most repos. Exemption is the more defensible choice but happy to flip.
• labelFit source: I used repo.registryConfig.labelMultipliers ∩ profile.dominantLabels. Could also be dominantLabels directly. Open to either.

[JSONbored]

PR #30 - feat(agent): make next-action recommendations repo-specific
Signal: REQUEST CHANGES

Required changes:

• Fix branch coverage drop: CI reports 94.94% branches, below 95%.
• Add tests for languageMatch false/true, labelFit empty/non-empty, publicNextActions for every recommendation type, and cleanup_first priority without open_pr_pressure penalty.
• Keep private numeric lane/share reasoning out of publicNextActions.
• Re-run validate after branch tests are added.
• Expect possible conflicts with #26 because both modify decision-pack and agent-orchestrator.

[grnlemon]

Thanks for the review. Addressed each item in 0c61546:

• Branch coverage: 94.94 → 95.03% (above the 95% gate). New tests target the previously uncovered branches in decision-pack.ts: scoreBlockersFor with outcome undefined, whyThisHelpsFor/nextActionsFor cleanup-first + watch (split) with undefined outcome, withSnapshotMetadata both-generatedAt-null fallback.
• languageMatch true/false + labelFit empty/non-empty: new dedicated test exercising all four combinations; asserts copy includes/omits in TypeScript and target labels: … accordingly.
• publicNextActions per recommendation type: new tier-loop test covering cleanup_first, maintainer_lane, pursue (direct), pursue (split), watch, avoid_for_now. Each tier asserts noStructuralCountLeak + no decimals + no share|emission|priority words.
• cleanup_first priority without open_pr_pressure penalty: new test triggers cleanup_first via outcome.openPullRequests = 3 (no blocker emitted) and asserts it still outranks a paired pursue baseline, locking the priorityFor exemption invariant from both sides.
• No private numeric lane/share reasoning in publicNextActions: the tier-loop regex /\b\d+(\.\d+)?\b/ plus /share|emission|priority/i is now applied to every tier's publicNextActions (and reused via noStructuralCountLeak on six per-lane fixtures).
• Re-ran validate: npm run typecheck clean, 279/279 unit + integration tests pass, vitest --coverage reports branches at 95.03%.
• Conflicts with feat(agent): serve stale decision packs with freshness marker and background rebuild #26: acknowledged. Happy to rebase whichever lands second; the surfaces overlap (decision-pack.ts, agent-orchestrator.ts) but at different functions, so I'm expecting a rebase-clean (not auto-clean) merge.

Pushed to the same branch.

[grnlemon]

@JSONbored could you please review the update again?

[grnlemon]

Thank you!