Make structured output repair errors actionable

[brynary]

Summary

• preserve structured schema validation details instead of flattening them to strings
• identify the exact JSON Pointer for required, type, enum, and additional-property errors
• include bounded schema context in repair prompts
• call out validation problems that remain unchanged across repair attempts
• use the same actionable errors for command-stage schema failures

Why

A repair prompt could previously say that rationale was required without saying that the missing field belonged at /findings/0/rationale. A model could add the field at the wrong level and repeat the same failure. The new prompt identifies the exact target and explicitly flags an unchanged error on the next attempt.

Same-name misplaced-field detection is intentionally out of scope.

Testing

• ulimit -n 4096 && cargo nextest run -p fabro-workflow — 1,323 passed, 31 skipped
• cargo +nightly-2026-04-14 clippy -p fabro-workflow --all-targets -- -D warnings
• cargo +nightly-2026-04-14 fmt --check --all
• git diff --check

[Copilot]

Pull request overview

Improves structured-output repair UX in fabro-workflow by preserving JSON Schema validation structure (instance/schema pointers + keyword/detail) so repair prompts can point to the exact JSON Pointer that needs correction and can detect repeated, unchanged validation failures across repair attempts.

Changes:

• Replaces flattened string validation errors with structured SchemaValidationIssue data (instance/schema/evaluation paths + keyword + typed details) and renders actionable messages including JSON Pointers and bounded schema fragments.
• Extends repair prompting to (optionally) reference the previous failure and call out at least one unchanged schema issue across repair attempts.
• Reuses the same actionable rendered validation errors for command-stage output_schema failures and adds tests verifying pointer reporting and repeated-error messaging.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
lib/components/fabro-workflow/src/handler/structured_output.rs	Introduces structured schema-validation issue modeling + richer rendering and repair messaging; adds targeted tests for pointer accuracy and repeated-issue prompts.
lib/components/fabro-workflow/src/handler/llm/api.rs	Threads previous validation error through LLM repair loops and adds an integration test asserting repeated-error callout behavior.
lib/components/fabro-workflow/src/handler/command.rs	Updates command-stage schema failure reporting to use the new rendered, schema-aware validation messages.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[brynary]

Ran a reuse / quality / efficiency review pass over the branch and applied the findings. Net -160/+106 on top of the original commit; behavior is the same or better.

Bug fixed. same_problem_as fell through to _ => true, so any two non-Required issues sharing an instance path, schema path and keyword compared equal. A model that removed one unexpected property and added a different one was told "at least one validation problem is unchanged from your previous repair" — the opposite of what happened. SchemaValidationIssue already derives PartialEq, so the 17-line comparison is now previous.contains(issue). Added a test that locks this in.

Messages got more actionable, not less. The hand-written Type and Enum rendering is gone — jsonschema already renders both, and its messages name the offending value where the hand-written ones only named its type. At JSON Pointer /line: "85" is not of type "integer" instead of expected JSON type "integer", but got string. Relatedly, error.masked() is back to error.to_string(): masking substitutes a placeholder for the bad value, which works against the whole point of this PR, and buys no privacy since the model's full response is already in the prompt.

The schema fragment is resolved when the issue is captured, not threaded through rendering as Option<&OutputSchemaKind>. That reverts the command.rs change entirely and drops the test-only messages() shim. The fragment is now attached only to Other, where it adds something; for required, type, enum and additionalProperties it repeated the prose verbatim, doubling the length of every repair prompt.

Also removed evaluation_path (dead except under $ref, where the pointer it printed doesn't resolve against the schema), the keyword field (already the last segment of the schema path), and expected_json_types / format_expected_types / pointer_code. Capped the model-controlled unexpected-property list at 10 — .take(5) bounded the number of issues but not the size of one, so a wide object under additionalProperties: false could put a megabyte of model output into the repair prompt. bounded_json now truncates in one bounded pass. And in the agent loop, the previous error is recorded only after the session accepts the repair; if the send fails and failover rebuilds from the original prompt, the model never saw the repair message.

All 1324 fabro-workflow tests pass; clippy and rustfmt clean.

Two follow-ups I did not take here, both pre-existing:

• On exhausted repairs, api.rs returns only output schema validation failed after N repair attempt(s) and drops the error detail this PR adds. command.rs includes it. Two tests pin the current string, so it's a separate change. handler/agent.rs:376 and handler/prompt.rs:194 are worse — they use if let Ok(..) else, which never binds the error at all.
• bounded_json is the sixth copy of truncate-to-N-chars-with-ellipsis in the workspace (fabro-util/src/text.rs, check_report.rs, run_title_generation.rs, run_title.rs, cli/.../events.rs, cli/.../list.rs — the last has a latent underflow panic). Worth one shared helper in fabro-util, but that's a cross-crate change.

[claude]

Claude Code Review

Claude Code Review is paused for this repository. To reconnect it, an admin of this repository's GitHub organization (or the account owner, for personal repositories) who can also manage your Claude organization's Code Review settings needs to re-link GitHub in Code Review settings. This is a one-time step.

_{Tip: disable this comment in your organization's Code Review settings.}

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.