code: no structured bail-out when coder lacks sufficient context

[BorekZnovustvoritel]

What happens

When the coder agent cannot determine what to implement, it stops without committing. The post-script reports a generic failure. There is no mechanism for the coder to signal what context it needs or to bounce the issue back for clarification.

Context

Suggestion from a fullsend user (konflux-ci/mobster). During experimentation, multiple /fs-fix iterations were needed because the coder attempted changes without enough context. A structured "needs more context" output — removing ready-to-code and posting the coder's question on the issue — could shorten this loop and reduce wasted human review time.

---

This issue was authored by Claude (Cursor agent) on behalf of @BorekZnovustvoritel based on the Mobster onboarding feedback document.

[fullsend-ai-prioritize]

RICE Priority Score: 1.05

Score breakdown

Dimension	Score	Reasoning
Reach	1.5	This affects all users who trigger the coder agent on issues with insufficient context. The report comes from a real user (konflux-ci/mobster) and the problem is inherent to the platform workflow, so multiple orgs would benefit from a structured bail-out mechanism.
Impact	1	Medium improvement to workflow. Currently users must go through multiple /fs-fix iterations when context is lacking, wasting human review time. A structured bail-out would shorten this loop, but there is an existing workaround (manual re-triggering with more context).
Confidence	0.7	The issue is well-described with a concrete proposal (remove ready-to-code label, post coder's question on the issue) and grounded in real user feedback. The scope is reasonably clear, though implementation details around the coder's structured output format are not fully specified.
Effort	1	Requires adding a structured output path to the coder agent, updating the post-script to handle the new signal (remove label, post comment), and defining what 'needs more context' looks like. Touches the coder agent and post-script but is a well-scoped change.

Formula: (1.5 x 1 x 0.7) / 1 = 1.05

[rh-hemartin]

I see, so the triage agent did not do a good job in requesting more info I would guess. We can at least make the code agent make a better error that is surfaced, so users know that the better path is to add more info to triage and then start the code agent again.

[rh-hemartin]

/fs-triage

[fullsend-ai-triage]

🤖 Finished Triage · ✅ Success · Started 7:48 AM UTC · Completed 7:52 AM UTC
Commit: 9f4f2b6 · View workflow run →

[fullsend-ai-triage]

Triage Summary

Severity: Medium · Category: Feature

Problem

When the code agent cannot determine what to implement, it silently exits without committing. The post-script reports a generic "nothing to do" notice with no actionable feedback. There is no mechanism for the coder to communicate what context is missing.

Root Cause

The code-result.schema.json only defines a target_branch field — there is no structured way for the agent to signal "I need more context" vs. "I completed successfully." The post-script (post-code.sh) treats all no-commit exits identically.

Recommended Fix

1. Extend the schema: Add an optional action field (completed | needs-context) and an optional message field to code-result.schema.json
2. Update agent instructions: Have the code agent write a structured bail-out when it cannot proceed, including what specific context is needed
3. Update post-script: Detect the needs-context action, remove the ready-to-code label, post the coder's question as an issue comment, and exit cleanly

As noted in the comments, ideally the triage agent catches insufficient context earlier — but the code agent should still have a structured fallback.

Proposed Test Case

Scenario: Code agent bails out with needs-context
  Given a code-result.json with:
    {"action": "needs-context", "message": "Issue does not specify which API endpoint to modify"}
  When post-code.sh runs
  Then it should:
    - NOT attempt to push or create a PR
    - Post a comment on the issue with the coder's question
    - Remove the ready-to-code label
    - Exit with code 0
  Edge cases:
    - Missing message field → produce a useful default comment
    - Unknown action value → treat as error

---

Labels: Issue targets the code agent's structured output and post-script behavior; feature request for a new capability.