Hivemind turns agent traces into skills and shares with your team

Summary

Per-session savings notification at session start using the published LoCoMo benchmark multiplier:

🐝 Hivemind has saved you ~12.9k tokens
   14 sessions · 63 memory searches

Formula (plain math, X − Y = Z)

Y = memorySearchBytes / 4
    bytes returned from Bash tool calls grep'ing ~/.deeplake/memory/
    (memory summaries + sessions raw — both live under that root),
    summed across all records in ~/.deeplake/usage-stats.jsonl,
    divided by 4 to approximate tokens (BPE rule-of-thumb)

X = 1.7 · Y
    counterfactual: tokens to do the same work without hivemind,
    per the published LoCoMo benchmark at deeplake.ai/hivemind —
    1,008 vs 1,700 tokens / Q with Claude Haiku via `claude -p`,
    hybrid lexical + semantic retrieval

Z = X − Y = 0.7 · Y
    tokens saved (the headline number)

No LLM calls, no SQL, no new backend endpoints. Pure local file IO + arithmetic.

What's new

| File | Purpose |
|---|---|
| src/notifications/transcript-parser.ts | Walk session JSONL at SessionEnd, extract memorySearchBytes (bytes Claude received from memory-grep tool calls — matched via tool_use_id to tool_result content) and memorySearchCount. |
| src/notifications/usage-tracker.ts | Append/read ~/.deeplake/usage-stats.jsonl. Minimal record shape {endedAt, sessionId, memorySearchBytes, memorySearchCount}. Backward-compat reader for legacy records. |
| src/notifications/sources/local-usage.ts | At SessionStart, sum across all records, apply Z = 0.7 · Y / 4, render. dedupKey = {session: sessionId} so each new session refires with updated numbers. |

What's modified

• src/notifications/index.ts — thread sessionId from hook stdin, wire fetchLocalUsageNotifications into the source list, filter through tryClaim to prevent duplicate emission from the dual SessionStart hook registration (settings.json + marketplace hooks.json both fire in parallel).
• src/notifications/state.ts — tryClaim() uses openSync(path, "wx") (O_CREAT|O_EXCL semantics) on per-(id, dedupKey-hash) claim files. First process wins, second gets EEXIST and skips. Fail-open on non-EEXIST FS errors.
• src/hooks/session-end.ts — parse transcript_path from stdin, call recordSessionUsage() before the wiki-worker lock check.
• src/hooks/session-notifications.ts — extract session_id from stdin (previously discarded).
• src/cli/install-claude.ts — syncHivemindHooksToSettings(): rewrites hivemind-owned hook entries in ~/.claude/settings.json from the marketplace's canonical hooks.json on every install/update. Fixes a real install bug — users whose first install predated session-notifications.js weren't seeing it fire after hivemind update because settings.json was never re-synced.

What's deferred

• Channel 3 (skill auto-pull bytes in Y) — filed as #127. The current Y is a conservative lower bound; adding skill bytes will jump the displayed number.
• Channel 4 (SessionStart instructions block) — explicitly excluded; it's hivemind explaining itself, not user past content.
• $-figure rendering — model pricing varies; skip for v1.

Test plan

• [x] Unit tests: 2263 / 2263 passing
• [x] Per-file coverage thresholds met
  • transcript-parser.ts: 90/75/90/90
  • usage-tracker.ts: 90/75/90/90
  • sources/local-usage.ts: 90/80/90/90
  • state.ts: 90/75/90/90 (branches bumped from 80 for fail-open paths)
• [x] Live E2E on test machine:
  • Single hook fires renders 🐝 Hivemind has saved you ~12.9k tokens / 14 sessions · 63 memory searches
  • Concurrent race (2 parallel processes, same session_id): exactly ONE emission. The other process's stdout is empty (claim file held).
  • New session_id: refires with cumulative numbers.
  • Re-invoke same session: silent (state dedup).
• [x] Backward-compat verified: legacy records (predating memorySearchCount) are now counted in "N sessions" instead of silently dropped.

Verification math (live numbers from test machine)

total memorySearchBytes across 14 records = 73,474
Y = 73,474 / 4 = 18,368 tokens delivered
Z = 0.7 × 18,368 = 12,858 tokens saved → rounds to ~12.9k ✓
memorySearchCount total = 63 → "63 memory searches" ✓
records = 14 → "14 sessions" ✓

Citation

LoCoMo benchmark at deeplake.ai/hivemind: 100 QA, Claude Haiku via claude -p, hybrid lexical + semantic retrieval, 1,008 vs 1,700 tokens / Q → ratio 1.687 → rounded to 1.7×. Also: 25% cheaper ($6.65 vs $8.94 / 100 QA), 31% fewer turns (6.2 vs 8.9).

Summary by CodeRabbit

• New Features
  • Per-session memory usage tracking records and automatically reports lookup activity metrics for sessions
  • Token savings recap notifications display efficiency improvements achieved through hivemind memory integration
  • Improved plugin installation with enhanced configuration management and hook synchronization capabilities
  • Streamlined and reliable installation workflows across multiple plugins and system integrations