feat: expanded view toggle and AI-generated session titles

[amahpour]

Summary

• Expanded view toggle: "Expanded" button in the topbar switches all tiles to show last 5 transcript entries inline (no chevron). Grid widens in expanded mode. State persists via localStorage.
• AI session summaries: Every N user messages (default 5, configurable), runs claude -p in the background to generate an evolving 3-8 word session title. Respects title lock, disables gracefully if claude CLI is unavailable.
• Manual rename auto-locks: Editing a session title now sets display_name_locked: true so AI summaries don't overwrite user renames.
• Settings UI: New "AI Summaries" tab with configurable interval.

Closes #16

Test plan

• Open dashboard, click "Expanded" button — tiles show last 5 transcript lines inline
• Toggle off — compact view returns with chevron-based preview
• Refresh page — expanded state persists
• Start a Claude Code session, send 5+ user messages — verify title updates in server logs and on dashboard
• Manually rename a session — verify lock icon activates and AI does not overwrite
• Open Settings > AI Summaries — change interval, save, verify it persists
• make check passes (ruff, mypy, pytest — 256 tests)

🤖 Generated with Claude Code

[github-actions]

Unit Test Coverage

File	Stmts	Miss	Cover	Missing
__init__.py	0	0	100%
db.py	163	2	98%	133–134
hooks.py	192	19	90%	222–230, 236–242, 258, 262, 265
main.py	37	0	100%
pr_lookup.py	113	2	98%	53–54
sessions.py	0	0	100%
watcher.py	544	28	94%	26–29, 345, 516–517, 520–521, 636, 705, 786, 791–797, 844–852
routes
__init__.py	0	0	100%
api.py	144	0	100%
ws.py	41	3	92%	47, 64–65
TOTAL	1234	54	95%

Tests	Skipped	Failures	Errors	Time
308	0 💤	0 ❌	0 🔥	4.640s ⏱️

[amahpour]

Codex Adversarial Review

Target: branch diff against main
Verdict: ⚠️ needs-attention

No-ship: the new AI summary worker can leak untracked claude subprocesses under failure, and it stores model-emitted PR URLs without validation, turning transcript content into trusted dashboard links.

---

Findings

[high] Timed-out or cancelled AI summaries leave claude children running

Location: server/watcher.py:521-580

_generate_ai_summary starts a claude -p subprocess and wraps proc.communicate() in asyncio.wait_for(). On timeout it only logs and exits; on task cancellation the coroutine also exits through finally without terminating the child. In asyncio that does not reap the subprocess for you. The result is orphaned claude processes that keep consuming local resources and possibly billable background work, while _summary_tasks has already been cleared so later triggers can start even more of them.

Recommendation: Track the subprocess handle and, in timeout/cancellation/finally paths, explicitly kill() or terminate() it and await proc.communicate() before dropping _summary_tasks. Add a regression test that asserts the child is reaped on timeout and watcher shutdown.

---

[high] AI-extracted pr_url is trusted enough to become a stored dashboard link

Location: server/watcher.py:500-556

The response parser accepts any string for pr_url, and _generate_ai_summary persists it with no scheme, host, or path validation. That new value is later rendered as the session card's PR link, so any URL mentioned in the conversation can be mislabeled as a trusted PR/MR and clicked from the dashboard; a javascript: or arbitrary external URL would be especially bad. This is a new trust-boundary violation from transcript/model output into persisted UI navigation.

Recommendation: Validate AI-derived metadata before storing it: require https, restrict hosts to the repo's GitHub/GitLab remote or an allowlist, and require /pull/<id> or /merge_requests/<id> path shapes. Reject anything else, and consider a frontend safe-link check as defense in depth.

---

Next Steps

• Reap summary subprocesses on timeout and cancellation before merging this feature.
• Add strict server-side validation for AI-populated pr_url values before they are persisted or rendered.

---

🤖 Generated by Codex adversarial review via Claude Code