fix: code agent missing structured output instructions and self-validation

[waynesun09]

What happens

The code agent fails validation with output/code-result.json not found because it never writes the required structured output file. This was first observed on run 28621163275 — the first run after agents were registered from this repo (commit e0b4d16b on .fullsend).

Both validation iterations (1 of 2, 2 of 2) the agent completed its code task (exit 0) but never produced code-result.json. The agent made code changes and committed them but didn't write the structured output.

Root cause

The code agent's instructions are missing the structured output guidance that review and fix agents already have:

1. agents/code.md — no structured output section

The code agent definition has zero mentions of code-result.json, FULLSEND_OUTPUT_DIR, FULLSEND_OUTPUT_FILE, fullsend-check-output, or the output schema. It simply says: "Follow the code-implementation skill for the step-by-step procedure."

Compare to agents/fix.md which has a dedicated "Structured output" section with:

• "You MUST produce a JSON file..."
• "Without this file, the post-script cannot communicate your work back to the reviewer."
• Schema field documentation
• fullsend-check-output validation instructions

2. skills/code-implementation/SKILL.md — output buried and no self-validation

The instruction to write code-result.json is buried in step 3 ("Discover repo conventions") among ~60 lines of unrelated instructions about build commands, linter config, and branch conventions. Problems:

• Not a dedicated step — lost in context
• No "MANDATORY" or "MUST" language
• No fullsend-check-output validation (unlike review and fix skills)
• No reminder at step 10 (commit) to verify the output file exists
• No mention that the harness validation_loop will reject the run without it

Agent comparison

Agent	Agent def has output section	Skill has dedicated output step	fullsend-check-output in skill	Status
Review	Yes (detailed)	Yes (step 7)	Yes	Working
Fix	Yes ("MUST")	Yes (step 9, MANDATORY)	Yes	Working
Code	No	No (buried in step 3)	No	Broken

Proposed fix

A. Add "Structured output" section to agents/code.md

Between "Failure handling" and "Detailed implementation procedure", add a section modeled on agents/fix.md:

## Structured output

You MUST produce a JSON file at `$FULLSEND_OUTPUT_DIR/$FULLSEND_OUTPUT_FILE`
(typically `code-result.json`) that documents the target branch for PR creation.
The post-script reads this file to create the PR against the correct branch.
Without this file, the harness validation loop will reject the run.

Write this file early — during step 3 of the skill, after determining the
target branch — so it is available even if the agent hits a timeout later.

Include the schema fields, an example, and fullsend-check-output instructions.

B. Add output verification step to skills/code-implementation/SKILL.md

1. Add MANDATORY language around the existing output writing block in step 3
2. Add fullsend-check-output call after the write
3. Add a new step 11 ("Verify structured output") as a final check before exit, with a recovery path if the file is missing

Context

• This repo's harness/code.yaml has validation_loop: with max_iterations: 2 — this was added during the agents extraction (PR feat: extract code, fix, review, retro, prioritize agents from fullsend #12) and didn't exist in the old scaffold's harness/code.yaml. The validation correctly catches the missing file, but the agent doesn't know it needs to produce it.
• Related: Code agent missing fullsend-check-output self-validation before exit fullsend#2777 (same issue filed against the scaffold before extraction)
• Related: Review agent should guarantee output file is written before exit fullsend#2886 (same class of bug for review agent)

[fullsend-ai-triage]

🤖 Finished Triage · ✅ Success · Started 9:52 PM UTC · Completed 9:55 PM UTC
Commit: 0a95cac · View workflow run →

[fullsend-ai-triage]

Triage Summary

Severity: High — the code agent is non-functional; every run fails validation.

Root cause: The code agent definition (agents/code.md) has no structured output section, and the code implementation skill (skills/code-implementation/SKILL.md) has no output-writing instructions. The agent doesn't know it needs to produce code-result.json. This stands in contrast to the fix and review agents, which both have explicit output sections and fullsend-check-output validation.

Proposed fix (two parts):

1. agents/code.md — Add a "Structured output" section (modeled on agents/fix.md) documenting the requirement to write the output JSON and validate it with fullsend-check-output.
2. skills/code-implementation/SKILL.md — Add MANDATORY output-writing instructions, fullsend-check-output validation, and a final verification step before exit.

Use agents/fix.md and skills/fix-review/SKILL.md as reference implementations for the pattern.

Test case:

1. Verify agents/code.md contains a "Structured output" section with
   FULLSEND_OUTPUT_DIR, FULLSEND_OUTPUT_FILE, and fullsend-check-output
2. Verify skills/code-implementation/SKILL.md has a dedicated output
   step with MUST/MANDATORY language and fullsend-check-output
3. Verify a final verification step checks for the output file before exit
4. End-to-end: run code agent, confirm code-result.json is produced
   and passes schema validation

Related issues: fullsend-ai/fullsend#2777 (same issue in old scaffold) and fullsend-ai/fullsend#2886 (same class of bug for review agent) are tracked upstream.

[fullsend-ai-coder]

🤖 Finished Code · ✅ Success · Started 9:56 PM UTC · Completed 10:00 PM UTC
Commit: 0a95cac · View workflow run →