fix: last-response viewer follows /clear to the new Claude conversation

[TeigenZhang]

Summary

The last-response viewer (eye icon) keeps showing the pre-/clear transcript after a user runs /clear in the Claude CLI. This is because the Session's _claudeSessionId is only refreshed when Claude emits JSON on stdout, which never happens in the interactive PTY mode Codeman spawns.

This PR adds two complementary update paths:

• Session.adoptClaudeSessionId() — public setter mirroring the existing no-op-if-same guard. Called from POST /api/hook-event whenever Claude Code hooks carry data.session_id (works automatically once hooks are configured in the target case).

• /api/sessions/:id/last-response now resolves the active conversation id from ~/.claude/history.jsonl before reading the transcript. This is the only source-of-truth that does not require hooks, which matters because Codeman intentionally does not write hook config into arbitrary user repos (see the comment at session-routes.ts POST /api/sessions).

The history scan:

• Skips entries held by other Codeman sessions in the same cwd so concurrent tabs don't shadow each other.
• Requires the candidate's jsonl mtime to exceed the initial <session.id>.jsonl mtime, so a stale id inherited from a prior dead session is never adopted.

Why history.jsonl

Claude Code writes one line per user prompt to ~/.claude/history.jsonl with {sessionId, project, timestamp}. After /clear the new conversation uuid shows up there on the first prompt. ~/.claude/sessions/<pid>.json would be nicer but older Claude CLI versions don't update it after /clear, so it isn't reliable across versions.

Test plan

• Verified against a session that had been /cleared: claudeSessionId adopts the post-/clear uuid on the next eye-icon request, and the returned transcript reflects the new conversation.
• Verified non-cleared sessions are unaffected (claudeSessionId stays equal to session.id).
• Synthetic hook POST with a fake session_id flips claudeSessionId and a follow-up POST with the original id restores it.
• tsc --noEmit clean, npm run build succeeds.

Known limitation

If two concurrent Codeman sessions in the same cwd both run /clear before either has had its id re-resolved, the history scan can't disambiguate them (history.jsonl doesn't record which Codeman session typed). Single-session /clear and single-tab-per-cwd setups are unaffected.

[aakhter]

Nice fix for the /clear session tracking issue! One security concern I noticed:

Path traversal via candidateSid -- In resolveActiveClaudeSessionIdFromHistory, the sessionId parsed from ~/.claude/history.jsonl is used unsanitized in path.join:

const cs = await fs.stat(join(projectsDir, projDir, `${candidateSid}.jsonl`));

path.join does not neutralize .. segments, so a crafted sessionId like ../../etc/passwd would cause the server to stat arbitrary filesystem paths. The same value also flows into adoptClaudeSessionId() which stores it as trusted session state with no format validation.

Suggested fix -- Validate that session IDs are UUIDs before use. This can be enforced in adoptClaudeSessionId as a defensive invariant, plus at the point of extraction from history.jsonl:

// In Session class:
private static readonly SESSION_ID_RE = /^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/i;

adoptClaudeSessionId(newId: string): void {
    if (!newId || newId === this._claudeSessionId) return;
    if (!Session.SESSION_ID_RE.test(newId)) return;
    this._claudeSessionId = newId;
}

// In resolveActiveClaudeSessionIdFromHistory, after extracting candidateSid:
const SESSION_ID_RE = /^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/i;
if (!candidateSid || !SESSION_ID_RE.test(candidateSid)) return null;

This protects both the stat calls and any downstream consumers of _claudeSessionId.

[Ark0N]

Thank you, @TeigenZhang! Reviewed the PR and verified locally — typecheck, lint, and format all pass.

The two-pronged design is well thought out:

• Hook path (adoptClaudeSessionId from data.session_id) is the clean route once hooks exist.
• history.jsonl fallback correctly handles the "no hooks installed in user repo" case, and the mtime safety check + sibling-tab dedup are good guards against stale-id adoption.

The known limitation (concurrent /clear across same-cwd tabs) is honestly disclosed and an acceptable trade-off — single-session-per-cwd, the common case, works correctly.

Merging now.