fix: track agent in boulder state to fix session continuation (fixes #927)

[kaizen403]

Summary

Fixes #927 - After session interruption during /start-work, when user types "continue", Prometheus (planner) was resuming instead of Sisyphus (executor), causing all subsequent delegate_task calls to get READ-ONLY directive injected.

Root Cause

The boulder.json file was missing an agent field to specify which agent should resume on continuation. The atlas hook hardcoded agent: "atlas" but this wasn't being persisted when the boulder state was created.

Changes

• Add optional agent field to BoulderState interface in types.ts
• Update createBoulderState() in storage.ts to accept optional agent parameter
• Update start-work hook to set agent: 'atlas' when creating boulder state
• Update atlas hook to use stored boulderState.agent (defaults to 'atlas') on continuation
• Add tests for new agent field functionality

Testing

• Added 2 new tests for createBoulderState with agent field
• All 59 related tests pass
• Full test suite: 2099 pass (10 pre-existing failures in MCP OAuth tests)
• Typecheck passes

---

Summary by cubic

Persist the active agent in boulder.json so “continue” resumes the correct executor (atlas/sisyphus) instead of the planner (Prometheus). Fixes #927 and prevents READ-ONLY mode from being injected into delegate_task calls.

• Bug Fixes
  • Added optional agent to BoulderState and createBoulderState(plan, session, agent?)
  • Set agent="atlas" when /start-work initializes boulder state
  • Idle handler now checks the last session agent against boulderState.agent (defaults to "atlas") and uses it for continuation; supports non-Atlas agents
  • Prometheus MD-only: prioritize boulder state agent over message files so sessions resume with the executor even after restarts
  • Added tests for agent field, backward compatibility, non-Atlas continuation, and boulder-over-message priority

^{Written for commit 38b40bc. Summary will update on new commits.}

[cubic-dev-ai]

1 issue found across 5 files

Confidence score: 3/5

• There is a concrete behavior mismatch in src/hooks/atlas/index.ts: continuation accepts boulderState.agent, but idle handler still early-returns unless last message agent is Atlas, which can block non-Atlas agents from continuing.
• Given the medium severity and user-facing impact (continuations failing for agents like Sisyphus), there’s some merge risk despite being localized.
• Pay close attention to src/hooks/atlas/index.ts - idle handler gating prevents non-Atlas continuations.

Prompt for AI agents (all issues)

Check if these issues are valid — if so, understand the root cause of each and fix them.


<file name="src/hooks/atlas/index.ts">

<violation number="1" location="src/hooks/atlas/index.ts:571">
P2: Continuation injection now accepts `boulderState.agent`, but the idle handler still returns early unless the last message agent is Atlas. This blocks continuation for non-Atlas agents (e.g., Sisyphus), making the new agent parameter ineffective and preventing session resumption.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[kaizen403]

@code-yeongyu hello again! please review this when you can :3 continuation now checks against boulderState.agent instead of hardcoding atlas

[code-yeongyu]

Hey @kaizen403, thanks for the PR! I have a question about the fix:

/start-work always creates boulder state with agent: "atlas". The existing isCallerOrchestrator() already checks lastAgent === "atlas", and the new logic checks lastAgent === boulderState.agent which is also "atlas". So the behavior should be identical for the current /start-work flow.

Could you clarify the exact reproduction scenario for Issue #927? Specifically:

1. After session interruption and typing "continue", what does findNearestMessageWithFields() return as the last agent?
2. Is the issue that the last agent becomes something other than "atlas" after interruption (e.g., due to compaction or message storage state)?

Want to make sure we understand the root cause before merging. Thanks!

[kaizen403]

@code-yeongyu you're right. and after spending quite a while. i found the actual root cause isn't in the atlas hook.. it's in prometheus-md-only.

to answer your questions:

1. after interruption, findNearestMessageWithFields() isn't the problem. it's findFirstMessageWithAgent() which returns the oldest message (from /plan), not the newest.

2. yes, after interruption the in-memory map is gone, so it falls back to message files and picks "prometheus" from the first message

i've updated the pr with the actual fix

[kaizen403]

so.. what's happening is when you run /plan then /start-work, the session has messages from both agents. the switch to atlas is saved in memory. when the session crashes, memory is wiped. on "continue", the system checks message files and finds prometheus was first - so it thinks prometheus is running. wrong.

what i fixed:
boulder.json already stores which agent started the work (atlas). i just made prometheus-md-only check boulder.json before falling back to message files.

check order: memory -> boulder.json -> message files

added 3 tests to cover this. all passing

[kaizen403]

appreciate you asking before merging btw. saved us from shipping an incomplete fix @code-yeongyu :3