Wynelson94/longhand

Critical fix: if your DB has NULL project_ids or missing sessions, recall has been lying to you.

Two related failure modes could cause recall_project_status("project-name") to return "No session history found" even when transcripts existed on disk:

1. Project inference used first-event cwd only. When Claude Code launches from your home directory and you cd into a project mid-session, the first event's cwd is $HOME — not the project. Longhand attributed those sessions to $HOME (project_id → NULL), so recall_project_status missed them entirely.
2. No reconciliation between disk and DB. SessionEnd hook failures (silent exits, stale hook commands, Claude Code exit paths that skip hooks) left transcripts in ~/.claude/projects/ that never got ingested. There was no retry, no detection, no backstop.

Recently dogfooded: a 2,526-line bsoi-mesh-kit work session was on disk, four transcripts referenced the project, and recall_project_status returned empty narrative. The fix replays disk → DB with improved attribution.

Upgrade

pip install --upgrade longhand==0.6.0
longhand reconcile --fix   # replay historical sessions with new inference

reconcile is idempotent. Without --fix it prints a summary of drift; with --fix it re-ingests the NULL-project and missing-session buckets. Typical throughput: ~1–2 sessions/sec.

What changed

New: longhand reconcile [--fix]

Walks ~/.claude/projects/*.jsonl, diffs against the sessions table, buckets results into:

• fully indexed — session row exists with a project_id
• ingested but project_id IS NULL — project inference missed (pre-0.6.0 behaviour)
• missing from sessions — hook didn't fire or failed silently

--fix re-ingests the last two buckets using the improved inference.

Project inference: mode-of-cwd instead of first-event-cwd

build_session now tallies every event's cwd, filters out $HOME and paths that don't resolve to a project marker (.git, pyproject.toml, package.json, etc.), and picks the mode. Multi-project sessions attribute to where the work actually happened. Single-project and home-only sessions are unchanged.

Staleness surface on recall_project_status

New fields on the MCP response:

• session_count_indexed — rows in sessions for this project_id
• session_count_on_disk — plus any on-disk JSONLs referencing it that aren't ingested yet
• last_ingested_at, last_transcript_mtime
• stale: bool + stale_reason: str — true only when there are transcripts truly missing from the DB (actionable via reconcile --fix). Sessions attributed to sibling projects (expected for multi-project sessions) do NOT trigger drift.

When stale=True, the narrative gets a ⚠ prefix pointing at longhand reconcile --fix.

Bonus: discover_sessions filter fix

discover_sessions() was rglobing all JSONLs under ~/.claude/projects, including /subagents/*.jsonl (subagent transcripts referenced from parent sessions) and pytest temp dirs. These were being treated as top-level sessions. On my machine this was inflating the "missing" count from 28 → 650. Now filtered out.

Recall skill rewritten

~/.claude/skills/recall/SKILL.md previously mandated search_in_context first — but that tool requires a session_id the caller usually doesn't have. The skill now branches: use search_in_context when the query includes a session ID, otherwise start with match_project → recall_project_status and surface any drift loudly instead of silently falling back to file reads.

Tests

182 passing (was 174). +8 new:

• Three parser tests for mode-cwd inference (home-filter, marker-walk fallback, multi-project tiebreak)
• Three CLI tests for reconcile (missing detection, --fix re-ingestion, NULL-project detection)
• Two MCP tests for staleness surface (drift detected, no false-positive when in sync)

Commits

• c83879f v0.6.0: recall sees sessions it previously missed
• 32f7602 fix: sync .claude-plugin/plugin.json to 0.6.0

Full Changelog: https://github.com/Wynelson94/longhand/compare/v0.5.13...v0.6.0