fix(agents): prevent orchestrator from rushing ahead of background agents

[mrosnerr]

Summary

• Tighten Sisyphus prompt's Background Result Collection instructions so the model waits for explore/librarian results instead of duplicating their work and delivering premature answers.

Changes

• src/agents/sisyphus.ts: Remove Continue working immediately instruction (which models interpreted as "do ALL work yourself, ignore agents"). Restructure Background Result Collection steps to: (1) launch agents, (2) do DIFFERENT independent work if any, (3) otherwise END RESPONSE and wait for system-reminder, (4-6) collect results and clean up.
• src/agents/dynamic-agent-prompt-builder.ts: Add not for files you already know to explore section header to reduce duplicate reads when the parent already knows file locations.

Testing

Tested locally with 3 models via opencode run --format json:

Scenario	Claude Opus 4.6	Gemini 2.5 Pro	GPT-4.1
Broad discovery (unknown files)	✅ Fires bg agents, waits	✅ 4 bg agents, waits	✅ 2 bg agents, waits
Known directory question	✅ Direct tools only	✅ Direct tools only	✅ Direct tools only

bun run typecheck  # ✅ passes
bun test           # ✅ 639 tests pass
bun run build      # ✅ succeeds

Related Issues

Closes #2124
Closes #1967

---

Summary by cubic

Make the orchestrator wait for explore/librarian background agents before proceeding, preventing duplicate work and premature answers (fixes #2124, #1967). Also scope the Explore Agent to discovery-only when file locations are already known.

• Bug Fixes
  • Sisyphus prompts (all variants): removed “Continue working immediately”; clarified to only do different independent work while waiting, otherwise end response and wait for the system reminder; then collect via background_output and clean up with background_cancel.
  • Explore prompt builder: added “not for files you already know” to reduce duplicate reads.

^{Written for commit f1f682c. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 2 files

Confidence score: 5/5

• Automated review surfaced no issues in the provided summaries.
• No files require special attention.

_{Auto-approved: Prompt adjustments tighten orchestration logic to prevent race conditions and duplicate work, with verified test results across multiple models.}

[github-actions]

All contributors have signed the CLA. Thank you! ✅
_{Posted by the CLA Assistant Lite bot.}

[mrosnerr]

I have read the CLA Document and I hereby sign the CLA

[specialfortificationgoods-cpu]

Hopefully this fix would help with the subagents in general, ie rushing sisyphus-junior subagents

[acamq]

LGTM