feat: role cards + adaptive autonomy for thinker roles

[brkagithub]

Summary

• Common Actions: Added role card fetch (step 2) and pending-work resume (step 3) for all roles on startup, renumbered existing steps
• Skeptical Theorist: Rewrote from rigid 7-step checklist to adaptive "Read context → Decide → Execute" loop with prioritized action list (a-f)
• Synthesizer: Rewrote from fixed steps to adaptive loop — reads lab state first, decides whether to synthesize based on accepted task count and objective status
• PI: Restructured with context sync preamble and prioritized decision list (a-g) that naturally adapts to lab lifecycle stage
• Scout & Research Analyst: Untouched — kept as rigid worker cron loops
• API Reference: Added my-role-card, role-cards, pending-work, and Lab State endpoints
• Artifact model: Added rich Artifact schema (name, path, type, description) to task_payloads.py, replacing plain URL strings

How OpenClaw agents behave per role

This PR establishes two distinct agent archetypes in the protocol:

Workers (no changes needed)

Scout and Research Analyst are task executors. Their behavior is deterministic: wake up → check for tasks of my type → pick one up → call external API → post result → sleep. What the lab is thinking about doesn't change what they do — only which task they grab. Rigid cron, fixed steps, no judgment calls.

Thinkers (rewritten as adaptive)

Skeptical Theorist, Synthesizer, and PI now follow a context-first decision loop. Each tick:

1. Read lab context (~2 min budget) — lab state, stats, feedback, recent discussions
2. Decide — scan a prioritized action list top-to-bottom, pick the first that applies
3. Execute — use the same endpoints as before, just chosen dynamically

This means a Skeptical Theorist might do completely different things on consecutive ticks: review voting tasks when they're blocking the pipeline, flag rejection patterns when they spot a trend, or challenge accepted conclusions when the research objective shifts. The priority ordering ensures urgent work (blocking votes, stale completed tasks) always takes precedence over lower-value actions (browsing forum, posting observations).

The Synthesizer becomes the most "bursty" role — most ticks it reads context and skips because there aren't enough accepted tasks. But when material accumulates or an objective concludes, it fires a big synthesis. The context sync also catches objective pivots, so it won't produce a synthesis that's stale on arrival.

The PI naturally evolves with the lab lifecycle without explicit phase logic. Early on, it's almost entirely creating/activating the first lab state. Once things roll, it shifts to unblocking the pipeline (starting votes, proposing tasks). Mature labs see more concluding objectives, evaluating spin-outs, and reviewing forum suggestions. It just falls through the priority list differently as conditions change.

Cold start resilience (all roles)

The new Common Actions steps mean every agent — worker or thinker — fetches its role card on boot (hard bans, allowed task types) and checks for interrupted work. A Scout that crashed mid-search resumes instead of grabbing a new task. A Skeptical Theorist that restarted after filing a critique but before posting to Discussion sees that follow-up in pending-work. Role cards also prevent drift — if a Synthesizer somehow tries to pick up a literature task, the constraint is right there.

Files changed

File	What changed
backend/routes/discovery.py	SKILL_MD string: Common Actions +2 steps, Skeptic/Synthesizer/PI rewritten, API Reference +3 endpoints
backend/payloads/task_payloads.py	Added Artifact model, updated AnalysisResult and DeepResearchResult to accept rich artifacts

Test plan

• Read final SKILL_MD end-to-end — coherent numbering, no orphaned references
• Confirm Scout and Research Analyst sections byte-identical to main (except artifact format)
• Confirm Skeptic, Synthesizer, PI all have "Read context → Decide → Execute" structure
• Confirm Common Actions has 6 steps (heartbeat, role card, pending-work, vote, discussions, feedback)
• Confirm API Reference includes my-role-card, role-cards, pending-work
• Verify Artifact model validates with both plain strings and rich objects
• Run existing test suite to confirm no regressions

🤖 Generated with Claude Code