fix: audit findings — index the polled panel query, stabilise its trigger, unify the terminal rule, report every live fan-out

[saucam]

A second full audit of the collaborative-sessions feature, after the panel UI landed in #271. Four findings, all in code #271 introduced. I'd intended to fold them into that PR, but it was squash-merged before this push landed, so they arrive as their own.

Net delta vs main: 8 files, +247/−65.

---

HIGH — the panel poll scanned the tenant's entire dispatch history

dispatchRecentGroups filters on created_by, and no index covered it. Measured over 20k tasks rather than guessed:

before  SEARCH dispatch_tasks USING INDEX idx_dispatch_tenant (account_id=? AND project_id=?)
        17.84 ms per call
after   SEARCH dispatch_tasks USING INDEX idx_dispatch_panels (account_id=? AND project_id=? AND created_by=?)
         0.091 ms per call   — 196×

That query runs on a 3-second timer for as long as a collaboration is focused, and its cost scaled with the tenant's total task history rather than with the number of panels — so it got worse the longer a daemon lived.

Fixed with a partial index on (account_id, project_id, created_by, created_at DESC) WHERE group_id IS NOT NULL. Partial so it stays small: only grouped rows are ever read this way.

Pinned by a test that asserts the query plan, deliberately not a duration — a timing assertion in CI is a flake generator.

HIGH — the poll re-fired on every session.info_update, not every 3 s

The effect derived its goal id by reading focusedSession()?.collaboration, which subscribed it to that store property. mergeSession reassigns every field on each info_update with no equality guard (unlike ingestSessionList, which has jsonEqual), so collaboration got a fresh object reference on every status flip. The effect then tore down and rebuilt its interval, and each rebuild fired an immediate poll.

The two findings multiplied: during a tool-heavy turn this produced a burst of the 17.8 ms query above on the daemon's event loop.

Now a createMemo returning a plain string — a memo only notifies when the value actually changes, so a fresh collaboration object with the same id is inert.

MEDIUM — the terminal-status rule was defined twice

TERMINAL in dispatch.ts drove the barrier; a second literal TERMINAL_TASK_STATUS in session-manager.ts drove the panel UI's settled. Correctness-critical duplication: add a status and the sidebar would quietly disagree with the barrier about whether a fan-out had finished. One exported definition now, imported by both.

MEDIUM — only one live fan-out was reported

Concurrent panels are legal — an orchestrator can start a second while the first resolves, which is precisely why the dispatcher watches a set of tasks per session. livePanel() returned the first unjoined one, so a child participating only in the other live panel showed no badge, and the progress bar described one of two.

livePanels() returns all of them, liveProgress() sums across them (one panel's "1/3" misstates the outstanding work when two are running), and livePanelMember searches every live fan-out. The chip reads ⇉ panels ×2 when there's more than one.

---

Verification

suite	before	after
daemon	2200	2202
web	382	387

biome clean; both typechecks clean; web eslint 30 warnings / 0 errors (unchanged); bun run build OK. Re-verified after rebasing onto main at e49cf5f, not only where it was authored.

All four mutation-checked — index removed, terminal rule diverged, member lookup narrowed to the newest panel, progress reporting one panel instead of the sum. Each fails exactly the test written for it.

Two things the tests themselves taught me

• The new index makes group_id undroppable, which broke the migration tests' artificial downgrade() helper. The real migration only ever adds columns, so the fix belonged in the helper — the drop ordering is an artifact of simulating a downgrade.
• My index-coverage block used it( in a file that imports test(. It threw ReferenceError: it is not defined and the entire describe was skipped while the suite still reported green. Caught only because the pass count didn't move: 21 → 21 → 23 once fixed.

🤖 Generated with Claude Code

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.