Fix relative schema path resolution in post-fix script under content-addressable cache

[fullsend-ai-retro]

What happened

On PR #5732, fix attempts 2, 3, and 4 (run 30488267454, run 30522112798, run 30525914229) all failed with the identical error: Cannot read schema /home/runner/work/.fullsend/.fullsend/.fullsend-cache/resources/sha256/<hash>/scripts/../schemas/fix-result.schema.json: [Errno 2] No such file or directory. The fix agent itself succeeded in all three runs (exit code 0). The failure occurred in post-fix.sh when it invoked process-fix-result.py, which resolves the schema path relative to its own location (scripts/../schemas/fix-result.schema.json). The content-addressable resource cache stores each fetched file under its own SHA256 hash path, so scripts/ and schemas/ are not sibling directories — the relative traversal resolves to a nonexistent path. This blocked all fix progress and required the human to manually rebase and re-trigger review.

What could go better

The process-fix-result.py script assumes a repository-like directory layout where scripts/ and schemas/ are siblings. This assumption held when resources were checked out as a git tree but breaks under the content-addressable cache that the fullsend harness uses to fetch agent resources. The bug is deterministic — it will affect every fix run that uses this script until the path resolution is fixed. Confidence is high: three consecutive identical failures with the same error message, and the log clearly shows the schema file was fetched successfully but stored at a different cache path.

Proposed change

Modify process-fix-result.py (or post-fix.sh which invokes it) in the fullsend-ai/agents repo to receive the schema file path via an environment variable or command-line argument instead of computing it relative to the script's own location. The harness already fetches schemas/fix-result.schema.json as a resource (visible in the run logs) — its resolved cache path should be passed to the script explicitly. Alternatively, validate-output-schema.sh could be updated to accept the schema path from the harness config rather than deriving it from $0. If FULLSEND_OUTPUT_SCHEMA is already set to the schema's cache path by the harness, the script should use that env var directly.

Validation criteria

The next 3 fix agent runs on fullsend-ai/fullsend should not fail with 'Cannot read schema ... No such file or directory'. Specifically, process-fix-result.py should successfully load the fix-result schema regardless of the resource cache layout.

---

Generated by retro agent from fullsend-ai/fullsend#5732

[fullsend-ai-prioritize]

RICE Priority Score: 10.8

Score breakdown

Dimension	Score	Reasoning
Reach	3	This is a deterministic bug that affects every fix agent run using the post-fix script under the content-addressable cache. The issue body explicitly states 'it will affect every fix run that uses this script until the path resolution is fixed,' making it platform-wide.
Impact	2	The fix agent itself succeeds (exit code 0) but post-fix processing fails, completely blocking fix progress. Users must manually rebase and re-trigger review. This blocks a core automated workflow — scoring as severely degrading/blocking.
Confidence	0.9	Very high confidence. Three consecutive identical failures with the same error message provide clear reproduction. The root cause is well-understood (relative path resolution breaks under content-addressable cache), and the proposed fix is concrete and well-scoped.
Effort	0.5	The fix is small and well-scoped: modify the script to accept the schema path via an environment variable (e.g., FULLSEND_OUTPUT_SCHEMA) or CLI argument instead of computing it relative to the script's location. The harness already fetches the schema and likely knows its cache path. This is a straightforward change to one or two files.

Formula: (3 x 2 x 0.9) / 0.5 = 10.8