Summary

• Add global prompt_context configuration with defaults enabled: true, recent_events: 8, and max_events: 16.
• Project long request histories for model prompts by keeping the recent tail and backfilling the newest older evidence for roles missing from that tail.
• Apply the shared projection to normal role tasks, contract repairs, research decisions, and version-controller prompts.
• Add a content-free compaction notice with total/included/omitted counts and the canonical request-record path.
• Document compaction, backfill, an exact input/output example, rollback, and telemetry comparison guidance outside the README.

Behavior and boundaries

• Persisted .teams_runtime/requests/<request_id>.json records are unchanged and retain every event.
• Selected events are copied without payload truncation, summarization, or mutation.
• Deep Research payload shaping and session rollover are unchanged.
• Config changes require role-service restart.
• Rollback is prompt_context.enabled: false, which restores full event-history prompt inclusion.

Selection policy

1. Include the final recent_events entries by list position.
2. Seed roles already represented in that tail.
3. Scan older events newest-first.
4. Add the newest qualifying evidence for each missing role until max_events is reached.
5. Render selected events in their original chronological order.

Role evidence is a role_report in type or legacy event_type, or an event with non-empty payload.role and payload.status. Identity uses payload.role, then actor.

Validation

• python -m unittest discover -s teams_runtime/tests (769 tests passed)
• python -m unittest teams_runtime.tests.test_prompt_context (9 tests passed)
• python -m compileall -q teams_runtime
• git diff --check
• Synthetic 100-event prompt remains at or below 16 events and 30% of the disabled-policy prompt size.

Tracks #2.