Hivemind turns agent traces into skills and shares with your team

Summary

• Surface a quantified user-facing insight from mine-local instead of the generic "N local skills" count at the unauthenticated SessionStart banner. Replaces "3 local skills from past mining runs live in ~/.claude/skills/" with "Hivemind found a pattern in your past sessions: You revisited 4 merged PRs in the last month… Minted skill verify-before-done to catch it next time."
• Driven by the install→signup drop-off problem (Slack thread 2026-05-20): users currently see no concrete value before being asked to authenticate. This is the first user-visible iteration; install-time value-show comes in a follow-up PR.
• Backward-compatible — legacy manifests and entries the gate didn't emit an insight for fall back to the existing count surface.

Changes

• LocalManifestEntry gains optional insight: string; new getLatestInsightEntry accessor picks the most-recent insight-bearing entry by created_at.
• mine-local gate prompt now asks for a concrete, quantified, second-person insight per minted skill. parseMultiVerdict trims, collapses empty/whitespace to undefined, ignores non-string values (defensive against malformed gate output).
• New pure renderer renderLocalMinedNote handles both branches (insight present → new copy; absent → legacy count). The Claude Code SessionStart hook delegates to it so the conditional copy lives in one unit-testable place.
• Tests: schema roundtrip, accessor on mixed manifests, both banner branches with negative patterns, parseMultiVerdict insight cases (5 new, all passing).

No new infra; no telemetry; no process.exit paths touched. Citation hallucinations are not a concern in this pipeline — mine-local controls the source session paths, the LLM is not asked to cite anything it didn't see.

Evidence

Ran mine-local --force --n 5 against real ~/.claude/projects/*.jsonl (40 sessions on disk, picked 5). Haiku produced 11/11 skills with concrete quantified insights, zero null. Samples:

verify-before-declaring-ready — "You pushed back on incomplete work 5+ times in this session because the assistant presented plans and code as done without validating them against real systems."

npm-global-install-corruption-diagnosis — "You hit the same hivemind: command not found + partial install twice in one day before tracing it to concurrent npm processes racing on the same deterministic backup-rename path."

posthog-cohort-staleness-funnel-trap — "You spent 30+ min analyzing a funnel showing 0 new signups because fresh Persons hadn't entered the cohort table yet."

End-to-end smoke (worktree session-start.js against a tmp HOME with the real manifest) renders the new banner verbatim under the un-authed branch.

Test plan

• [x] npm run typecheck clean
• [x] npm run build succeeds (all bundles)
• [x] npx vitest run tests/claude-code/{local-manifest,local-mined-banner,mine-local-helpers}.test.ts → 46 pass
• [x] Full npm test → 2895 pass (one unrelated flake in deeplake-fs.test.ts under bulk-run load; passes cleanly in isolation in 3.44s)
• [x] Smoke 1: tmp HOME + synthetic manifest with insight → banner renders new copy
• [x] Smoke 2: tmp HOME + synthetic manifest without insight → banner renders legacy count copy (1 vs N pluralization correct)
• [x] Smoke 3: real mine-local --force --n 5 against ~/.claude → 11/11 insights, banner renders against tmp HOME with the real manifest

Follow-ups (not in this PR)

• Install-time value-show: move the primary surface into runAuthGate so users who run hivemind install and never open Claude still see the value before the sign-in prompt. The banner change here is the secondary/second-touch surface.
• Pick by quality, not write order: when multiple insights are written in the same mine-local batch (~16ms span), getLatestInsightEntry's created_at ordering effectively picks at random within the batch. Worth a follow-up that scores by source-session recency or insight quality signals.
• Telemetry: wire hivemind_install_started + alias chain so we can measure the conversion lift this banner produces.

Summary by CodeRabbit

• New Features

  • Mined skills can include an optional one-line quantified insight (normalized, capped, omitted when empty).
  • Session-start notes and notification banners can surface the latest insight with truncation/dedup rules, falling back to legacy count messaging and preserving the sign-in CTA.
  • Notifications now distinguish user-visible-only content from model-safe additional context; delivery splits model vs user channels.
  • Local manifests persist insight metadata conditionally for backward compatibility.

• Tests

  • Added comprehensive tests for insight parsing, manifest read/write, banner rendering, notification routing, and session-start behavior.