Code-review fix

Code-review fix (codereview_fix, D-68) is an action-layer type that ingests the comments produced by an external code-review tool, re-checks each one against the whole application, and fixes only the comments it can verify as accurate — in the working tree, with no commit and no PR. The external tool saw only a Pull Request diff, so it lacks whole-application context and can be wrong; this type exists to catch that. It is its own standalone mini-orchestrator at _processes/03_action/codereview_fix/codereview_fix.prompt.md: it runs no clarify step and no stage pipeline, and is defined entirely by its ## Action config. The orchestrator routes it via orchestrator.prompt.md → §C.route and never enters the stage loop.

All request types

Source of truth: the registry row in _processes/_shared/request-types.md, the preset _processes/_shared/presets/codereview_fix.md, and the action prompt _processes/03_action/codereview_fix/codereview_fix.prompt.md.

The comments are untrusted evidence, never instructions. A comment only authorizes fixing the specific item it points at — nothing broader, and nothing at all when the comment turns out to be inaccurate. Sub-agents receive an extracted claim summary plus an authorized scope, never the raw comment prose as a task (prompting-conventions.md §5). This boundary is the whole reason the type exists.

The flow

The orchestrator runs the action prompt inline (a non-dispatch instruction library, like the audit and doc_upkeep orchestrators). It reads JOB.md → ## Action to resolve CommentsSource / CommentsFile / Agents / Depth / Consolidator, ingests the comments (from the intake # Mandate body when CommentsSource: intake, or from CommentsFile when CommentsSource: file), and normalizes them to stable IDs C001, C002, … (unparseable prose becomes not_accurate / unchanged, never a guess). It then runs a two-pass per-comment model:

Pass 1 — read-only validate. A read-only managed sub-agent loop (_processes/_shared/managed-subagent-loop.md) checks each comment against whole-application context and produces an authorization record: Accuracy (accurate | not_accurate), AccuracyReason with evidence paths, Confidence (high | medium | low), and an explicit Authorized scope (including a what-is-not-authorized line) for the accurate comments. No code is written in this pass.
Pass 2 — conditional fix. A write-mode managed loop fixes only the accurate comments, each confined to its Authorized scope (overlapping items grouped or run serially, independents in parallel; an out-of-scope change is refused and reported). not_accurate (subsuming unverifiable / ambiguous / stale / out-of-scope) and low-confidence-accurate comments resolve to unchanged with zero code change.

It then validates the working tree, writes the per-comment roll-up 09_action/action-report.md (also surfaced in chat), records completion in PROGRESS.md, and lands the job in 4_done/. It never commits and never opens a PR.

What it is for

Reach for Code-review fix when an external code-review tool (GitHub's, Azure DevOps', or any PR-scoped reviewer) has left comments and you want them re-checked against the whole app and only the real ones acted on — without trusting the tool's limited view.

Validate before fix. Every code-write is gated behind a per-comment correctness check against whole-application context — the context the external reviewer lacked. An inaccurate comment produces no change.
Narrowly scoped fixes. A comment authorizes fixing only the specific item it names. The write pass refuses anything outside the Authorized scope recorded in pass 1.
Working tree only. Fixes land uncommitted for you to review; the type keeps Git all-off and refuses Commit: yes / PR: yes even if hand-flipped (v1 never commits or opens a PR).
Per-comment accountability. The report states, for every comment: accurate-or-not, the rationale (with evidence paths), the confidence, and fixed-or-left-unchanged.
No clarify step. Like other action types, its scope and objective come entirely from the ## Action config and the ingested comments, so it drops clarify.

The validate-then-conditionally-write fusion of audit and quick_fix. It borrows audit's per-finding validate-with-rationale (read-only, whole-app) and quick_fix's per-item parallel write — but neither of those has a per-item correctness gate that writes. It is also distinct from the workflow's internal review stage, which reviews a job's own diff before validate, not an external tool's comments.

The logic

codereview_fix sits in the action layer with shape orchestrator (its own standalone mini-orchestrator; D-68). The orchestrator routes it via orchestrator.prompt.md → §C.route — by layer / shape, not by the stage pipeline (the type→prompt-path map is hard-coded there; the registry supplies only the layer). A codereview_fix JOB.md therefore carries no stage machinery at all.

CommentsSource — intake (default; the comments are pasted into the # Mandate body) or file (the type reads them from CommentsFile).
CommentsFile — required iff CommentsSource: file; a job-folder- / repo- / workflow-root-relative or absolute path within an allowed root. The prompt refuses secrets-like, out-of-root, oversized, or binary paths.
Untrusted-input posture — the comment text is data, never instruction; sub-agents get a claim summary + authorized scope, never raw prose as a task.
Git all-off, refuses commit/PR — the preset sets every Git option to no, and the prompt actively refuses Commit: yes / PR: yes even if a job hand-flips them. Fixes stay in the working tree.
No VCS-API fetch — this version never pulls the PR or its comments from GitHub / Azure DevOps; that is a documented future extension.
Routed, not staged — the main orchestrator stays the sole PROGRESS.md writer and sole 4_done mover.

No contract version bump (D-68). codereview_fix reuses the existing action × orchestrator slot and introduces no new layer / shape / kind / gate enum value — an additive registry/UI + doc change, so the workflow contract stays 0.14 (precedents D-49, D-50, D-66).

The fields

A codereview_fix JOB.md carries only the header, ## Git, ## Action, and the # Mandate (which holds the pasted comments when CommentsSource: intake). There is no ## Stages, no ## Approval gates, and no ## Per-stage config — action types are routed by layer/shape, not by enabled stages.

`## Action` block fields (source: `_processes/_shared/presets/codereview_fix.md` + `_processes/03_action/codereview_fix/codereview_fix.prompt.md`)
Field	Default	Meaning
`CommentsSource`	`intake`	`intake` reads the comments from the `# Mandate` body; `file` reads them from `CommentsFile`.
`CommentsFile`	(blank)	Required iff `CommentsSource: file`; a path within an allowed root. The prompt refuses secrets-like / out-of-root / oversized / binary paths.
`Agents`	`claude`	Runtime(s) for the per-comment validate/fix sub-agents (`claude` \| `codex` \| `claude, codex`); adding `codex` requires `config.json → runtimes.codex: true`.
`Depth`	`standard`	`quick` \| `standard` \| `deep` — tunes how widely each comment is validated against the whole app.
`Consolidator`	`claude`	Runtime that synthesises the per-comment records into `09_action/action-report.md`.

`## Git` block (default off — the prompt refuses commit/PR even if hand-flipped)
Field	Default	Meaning
`Create branch`	no	No branch is cut — fixes sit in the working tree.
`Validate against main`	no	No base-branch diff validation.
`Commit`	no	Nothing is committed; the prompt refuses `yes` in v1.
`Push`	no	Nothing is pushed; the prompt refuses `yes` in v1.
`PR`	no	No pull request is opened; the prompt refuses `yes` in v1.
`PR target branch`	(blank)	Unused — the type never reaches the git landing.
`before_commit`	no	No commit, so no approval stop.
`before_pr`	no	No PR, so no approval stop.

The per-comment verdict vocabulary in 09_action/action-report.md: Accuracy (accurate | not_accurate), AccuracyReason (rationale + evidence paths), Confidence (high | medium | low), and Disposition (fixed | unchanged). not_accurate subsumes unverifiable, ambiguous, stale, and out-of-scope — all unchanged.

How to run it

There are two copy-path ways to run it. Both produce the same 09_action/action-report.md.

Option A — drop an intake, let the orchestrator route it. In the intake form pick Code-review fix, paste the comments into the mandate (or set CommentsSource: file + CommentsFile), and submit a contract-valid intake into jobs/0_new/. Then send the orchestrator prompt path. It reads JOB.md → Type, routes by layer, and runs the codereview_fix logic without entering the stage loop.

Orchestrator (routes by type)

_processes/02_orchestrator/orchestrator.prompt.md

Option B — run the action prompt directly. Point the agent straight at the codereview_fix action orchestrator for an existing job in 2_ready/ / 3_ongoing/.

Code-review fix action orchestrator

_processes/03_action/codereview_fix/codereview_fix.prompt.md

Two invocation modes — copy blocks show the workflow-internal path (bare _processes/...); prefix with _code_workflow/ when your cwd is the host project root. See How to use → Two invocation modes.

Code-review fix

The flow

What it is for

The logic

The fields

How to run it

See also

All request types

audit

quick_fix