---
name: saf-fix-pr
description: Apply authorized fixes to verified findings from one SDD task review; does not perform the review, redesign the feature, or publish changes.
compatibility: Requires Git and a compatible Agent Skills host.
---

# Fix SDD pull-request findings

## When to use

Use for repairs to verified findings from one task-scoped local change review. In `manual` and
`supervised`, the user explicitly requests the repair. In `autonomous`, the original bounded
delegation is standing authorization for verified local findings only. Read [the TLC baseline](../sdd-agentic-flow-shared/references/tlc-baseline.md), [engineering principles](../sdd-agentic-flow-shared/references/engineering-principles.md), and [safety rules](../sdd-agentic-flow-shared/references/workflow-safety.md).

## When not to use

Do not use for unverified comments, broad cleanup, feature redesign, sibling tasks, or automatic commits and pushes.

## Inputs

- One task reference and a review report, PR findings, or user-supplied evidence.
- Optional `.sdd-agentic-flow/config.yml` overrides, SDD artifacts, current diff, and validation commands.

## Workflow

1. Read `.sdd-agentic-flow/config.yml` when present; otherwise use canonical effective defaults, then resolve one task and its permitted scope.
2. Build a findings ledger, applying `../sdd-agentic-flow-shared/references/evidence-standard.md`. Fix only findings with reproducible evidence; classify preferences, missing evidence, and spec drift without changing them. In autonomous mode, reconcile intent-preserving spec drift through `saf-create-spec` rather than editing the contract here. Do not close findings by reclassifying missing evidence as preference or by silently changing delegated intent.
3. Apply `../sdd-agentic-flow-shared/references/engineering-principles.md`. Apply the smallest patch per actionable finding and add or update focused regression evidence. No opportunistic cleanup.
4. Run configured targeted checks and update the ledger under `../sdd-agentic-flow-shared/references/reviewability.md`: preserve finding IDs, record fix evidence and unresolved items, then hand off to `saf-review-pr` to confirm resolution. A repair claim alone does not close a finding.

## Safety

Preserve unrelated changes. Stop for SDD reconciliation, sibling scope, unsafe environments, or unresolved identity. Do not commit, push, amend, post comments, update PR metadata, mutate trackers, or make network/default mutations unless explicitly authorized.

## Output

Return the findings ledger, changes and checks, unresolved items, re-review scope, and next step. Include `Status`, `Next recommended skill`, and `Reason`. When actionable findings remain unresolved across a session or agent boundary, write or update `handoff.md` per `../sdd-agentic-flow-shared/references/handoff-standard.md`, referencing the findings ledger rather than duplicating it.

### Autonomy

Supports `manual`, `supervised`, and `autonomous` autonomy levels. In autonomous mode, advancing
back to `saf-review-pr` requires `fix-evidence` and every actionable finding addressed for reviewer
confirmation or explicitly deferred with a reason. An unresolved authority boundary or scope violation blocks the transition;
verified local findings do not require a new human confirmation. See
`../sdd-agentic-flow-shared/references/autonomy-guardrails.md`.

The sidecar's `findings_unresolved` blocks unaddressed actionable repairs; pending reviewer
confirmation alone does not block returning current fix evidence for re-review.