feat(autoresearch): improve setup and run observability

[fercgomes]

Summary

• redesign the Autoresearch setup input with clearer metric direction, examples, help content, feedback contact, and accept edits defaults
• surface streamed research findings, hypotheses, iteration plans, approach labels, live activity, context usage, time breakdowns, and completed run summaries
• record completed metric reports while the agent is still working, including composer runs whose initial prompt is already active
• prioritize iteration metrics and results in the dashboard hierarchy
• add dedicated and configurable Storybook coverage for the full dashboard and individual states

Validation

• 155 core Autoresearch tests
• 25 UI Autoresearch tests
• core and UI typechecks
• static Storybook build
• Biome and diff checks

Created with Autoresearch.

---

Created with PostHog Code

[trunk-io]

😎 Merged manually by @fercgomes - details.

[github-actions]

React Doctor found no issues in the changed files. 🎉

_{Reviewed by React Doctor for commit 9339e76.}

[greptile-apps]

_{Reviews (1): Last reviewed commit: "feat(autoresearch): improve setup and ru..." | Re-trigger Greptile}

[greptile-apps]

Streaming Reports Become Iterations

When a turn is still streaming, this call parses any complete autoresearch block already present and recordMetricReport() immediately appends an iteration and can complete the run. If the agent later corrects the metric or emits the final report in the same turn, the dashboard records an interim value as a real iteration, burns the iteration budget, and can stop the run before the final answer arrives.

[fercgomes]

Fixed in 69f144d. Streaming now treats the tail report as provisional. A report is recorded mid-turn only after the transcript explicitly starts the next iteration, and the final unresolved report is committed only when the turn ends. Target and iteration-budget completion decisions also run only at turn completion. Added regression coverage for corrected tail reports and multi-iteration streaming.

[greptile-apps]

Historical Runs Include Later Activity

When a completed run is selected, AutoresearchPanel still passes the current task session events into this analyzer. Since the filter only checks event.ts >= startedAt, tool calls and agent chunks from a newer run or later manual chat are included in the old run's activity list, time breakdown, and parsed plan.

Suggested change

	const relevant = events.filter((event) => event.ts >= startedAt);
	const relevant = events.filter(
	(event) => event.ts >= startedAt && (endedAt === null || event.ts <= endedAt),
	);

[fercgomes]

Fixed in 69f144d. Historical activity is now bounded to events between the run start and end timestamps. Added a regression test proving later manual activity is excluded from the historical run timeline and time breakdown.

[fercgomes]

/trunk merge

[fercgomes]

/trunk cancel

[fercgomes]

/trunk merge

[fercgomes]

/trunk cancel