fix: hooks resolve .gitmem like the server; session_close names the payload path (GIT-99)

[nTEG-dev]

What

The hooks find the same .gitmem the server uses, and a close with nothing to close with says where it looked (GIT-99).

• hooks/scripts/resolve-root.sh resolves $GITMEM_DIR, then $GITMEM_HOME/.gitmem, then ~/.gitmem, and makes the result absolute. This mirrors src/services/gitmem-dir.ts. recall-check.sh, post-tool-use.sh and session-close-check.sh source it.
• Dead pids: registry entries on this host whose pid is dead are skipped. Entries without a pid, or from another host, are kept, since their pid can't be checked here. recall-check now uses the newest live session instead of sessions[0]. Liveness is checked with the kill -0 builtin (EPERM counts as alive), not ps. The first version of this PR used ps -p, and on an image without ps (e.g. node:*-slim) every session looked dead, so the Stop hook never blocked. That was found while building the jq-less CI job (GIT-83) and fixed in c18e82a.
• The Stop hook prints the resolved absolute payload path, JSON-escaped: WRITE structured payload to /abs/.gitmem/closing-payload.json (this exact absolute path — it is where session_close reads).
• session_close: a standard close with no payload file and no inline reflection returns closing-payload.json not found at <absolute path> instead of the "requires N answers" family. An unparseable payload is reported as found but unreadable, at its path.

Why

Since GIT-91 the server resolves its root independently of cwd, but the hooks still read .gitmem/... relative to their own cwd. So a hook running in a repo read a different store from the one the server writes. It saw no session (or someone else's), and it told the agent to write a payload the server never reads. The server then answered:

Standard close requires task_completion object. …
Standard close requires closing_reflection with 7 answers
Standard close requires human_corrections field (even if empty string)

That sent the agent back to re-answer questions it had already answered, when the real problem was the path.

⚠️ Found along the way: Stop enforcement was off without jq (fixed here, GIT-112)

Changelog (Release C, its own line): "Stop enforcement was inactive on machines without jq; fixed." A CI hook-test run on a runner without jq comes with GIT-83.

On a machine without jq, session-close-check.sh falls back to inline node -e "...{...,...}..." inside $( … ). Bash brace-expands that into two broken node invocations, so the session count is empty and the Stop hook never blocks. That's why 3 of 32 hook tests are red on main on this machine: session_start + >5 calls → blocks, session_start + >5 min → blocks, and No state dir + session → creates state dir. resolve-root.sh passes the path through the environment into single-quoted JS, so the new code isn't affected. All three pass on this branch.

How to test

Hooks: bash hooks/tests/test-hooks.sh

	main scripts	This branch
Existing 32	29 pass: the 3 above fail (no jq)	32 pass
New 12 (GIT-99)	2 pass: vacuous, since main never blocks	12 pass
Total	31/44	46/46 (44 + 2 for the pid check)

The new tests cover: precedence (GITMEM_DIR > GITMEM_HOME > HOME, never cwd/.gitmem); a relative GITMEM_DIR made absolute; the Stop hook reading the store from an unrelated cwd while ignoring a decoy cwd/.gitmem; the absolute path in the text; valid JSON, including a path with a quote in it; a dead-pid entry skipped; an other-host entry kept; newest-live ordering; and recall-check reading the live session's scars.

Venue (v1.8.0 setup.sql, vujndhcumgroztfkkdus; production not touched): flow now calls session_close (standard, no payload, no reflection) before the real close.

• main: exit 1. It returned the three "requires …" errors above.
• This branch: exit 0. It returned closing-payload.json not found at /var/folders/…/gitmem-venue-git99-flow2-flow-…/.gitmem/closing-payload.json. Write the closing payload to exactly that path (or pass closing_reflection inline), then call session_close again. The real close then persisted, relevance was readable, and thread counts matched.

Unit tests: 1279/1279 pass (+5, session-close-payload-path.test.ts). The two "missing/unreadable" tests fail on main.

Existing tests that encoded the old behaviour (changed, not deleted)

1. hooks/tests/test-hooks.sh: every registry test planted $TMPDIR/.gitmem and relied on cd $TMPDIR, i.e. cwd-relative resolution, which is exactly the bug. The suite now exports GITMEM_DIR="$TMPDIR/.gitmem": same store, named explicitly. Left unpointed, it would now read the developer's real ~/.gitmem. No assertion changed.
2. session-close-supabase.test.ts: updates generic title…, preserves custom session title and persists decision titles… called a standard close with no payload and no reflection. They only reached persistence because validateSessionClose is mocked to valid: true; unmocked, that input was already rejected. It's now rejected up front, so these tests pass a minimal closing_reflection and keep testing titles and decisions. No assertion changed.

Not in this PR (flagged)

• session-start.sh still reads .gitmem/config.json from cwd for its project hint. It isn't one of the three registry readers; it can source resolve-root.sh the same way.
• The Stop hook text still says "7 reflection questions"; the server's CLOSING_QUESTIONS has 9.
• The installer copies the scripts into <cwd>/.gitmem/hooks. They still work there: resolve-root.sh sits beside them and ignores cwd.

Checklist

• npm run test:unit passes (1279)
• hooks/tests/test-hooks.sh 46/46
• New tests added
• No schema change; works on a v1.8.0 store
• Docs: hooks/README.md doesn't describe root resolution

Breaking changes

Hooks no longer read <cwd>/.gitmem. A setup that relied on a project-local store must name it with GITMEM_DIR, which is already what the server requires since GIT-91.

🤖 Generated with Claude Code