Review Artifacts Persistence Protocol

Any code review agent that produces a structured verdict follows this protocol to enable CI integration and cross-agent artifact compatibility.

Folder Structure

.copilot-tracking/
  reviews/
    code-reviews/
      <sanitized-branch>/
        review.md              # full markdown review output
        metadata.json          # machine-readable summary (see schema below)
        diff-state.json         # shared subagent input (branch, base, files, depth)
        dispatch-manifest.json  # canonical loop state (phase gates, next actions, board items)
        dispatch-board.md       # human-readable enumerated dispatch board
        walkthrough.md          # factual Register 1 orientation narrative
        emission-record.json    # selected emission mode, target, status, outcome
        explanations/
          <board-item>-<symbol>.md   # per-item Register 1 explanation artifacts
        walkback/
          <board-item>-research.md   # per-item Register 2 investigation artifacts

Sanitize the branch name by replacing every / with - (e.g. feat/my-feature → feat-my-feature).

The review.md, metadata.json, diff-state.json, and dispatch-manifest.json artifacts are always produced. The orientation-first artifacts (dispatch-board.md, walkthrough.md, explanations/, walkback/, and emission-record.json) are produced only when the review runs in interactive orientation-first mode; omit them in non-interactive workflow runs.

metadata.json Schema

{
  "schema_version": "1",
  "branch": "<original branch name, e.g. feat/my-feature>",
  "head_commit": "<full SHA of HEAD at time of review>",
  "reviewed_at": "<ISO 8601 UTC timestamp, e.g. 2026-02-27T10:00:00Z>",
  "verdict": "<normalized verdict - see table below>",
  "files_changed": ["<workspace-relative paths of source files in the diff>"],
  "findings_count": {
    "critical": 0,
    "high": 0,
    "medium": 0,
    "low": 0
  },
  "reviewer": "<agent or prompt name, e.g. code-review>",
  "artifacts": {
    "dispatch_manifest": "dispatch-manifest.json",
    "dispatch_board": "dispatch-board.md",
    "walkthrough": "walkthrough.md",
    "emission_record": "emission-record.json",
    "explanations": ["explanations/<board-item>-<symbol>.md"],
    "walkbacks": ["walkback/<board-item>-research.md"]
  }
}

The artifacts object records the orientation-first artifacts produced during the review. Omit any key whose artifact was not produced (for example, omit artifacts entirely for a non-interactive workflow run, or omit explanations when no per-item explanation was requested). The explanations and walkbacks values are arrays of review-folder-relative paths, one entry per board item that was explained or investigated.

Verdict Normalization

Agent Output Verdict	verdict value
✅ Approve	approve
💬 Approve with comments	approve_with_comments
❌ Request changes	request_changes

Orientation-First Artifacts

Interactive orientation-first reviews persist the following artifacts alongside review.md and metadata.json. Each is referenced from review.md so the markdown report links to the supporting evidence.

• dispatch-manifest.json — the canonical loop state: phaseGates, currentPhase, nextActions, and boardItems. This is the machine-readable source of truth for the human-steered walk-back loop.
• dispatch-board.md — the human-readable enumerated board rendered from the manifest boardItems: id, area, status, register, summary, openable links, and selectable symbols.
• walkthrough.md — the factual Register 1 orientation narrative (diff summary, runway summary, and appendices) presented before any findings register. It contains no severity grades or verdicts.
• explanations/<board-item>-<symbol>.md — per-item Register 1 explanation artifacts written by the explainer when a human asks a shallow factual question about a symbol. Each includes the answer, source file reference, relevant code excerpt, and follow-on symbols.
• walkback/<board-item>-research.md — per-item Register 2 investigation artifacts written by the walk-back researcher when a human asks a deep investigative question. Each is anchored to its board item.
• emission-record.json — the selected emission mode (native or canonical), target (PR, MR, ADO, or review artifact), status (completed or skipped), and a short outcome summary.

Change-Risk Decision State

diff-state.json is the canonical owner of the evidence and depth decision used by every review lane:

• reviewStateProvenance records human-confirmed or automation-derived origins for scope, changeRiskEvidence, recommendedDepth, depthTier, and depthRationale. It describes origin, not confidence.
• changeRiskEvidence contains exactly one entry for each of change-scope, path-criticality, history, test-presence, coverage, and rollback.
• Each entry contains signal, availability, and evidence. The availability field records the evidence state, not confidence in a claim. Its value is observed, unavailable, or qualitative; unavailable inputs remain explicit and are never fabricated.
• recommendedDepth records the advisory basic, standard, or comprehensive recommendation. Use standard when evidence is incomplete or inconclusive unless observed evidence supports comprehensive review.
• In interactive mode, existing depthTier remains the human-selected verification depth, and depthRationale explains that selection and any difference from recommendedDepth. Mark scope, evidence, depth, and rationale as human-confirmed; keep the advisory recommendation automation-derived.
• In workflow mode, mark all five provenance values automation-derived. Host-supplied or default values are not represented as human-confirmed.

In interactive mode, write these values after human scope and depth confirmation. In workflow mode, write them after automated context preparation. Complete either write before perspective dispatch. Subagents consume the values and provenance as shared context but do not recalculate or overwrite them.

Writing Rules

• Always overwrite any existing review.md and metadata.json for the branch: only the latest review per branch is retained.
• Obtain the HEAD commit SHA with git rev-parse HEAD immediately before writing artifacts.
• Obtain the current UTC timestamp immediately before writing artifacts:
  • In POSIX-compatible shells, use date -u +%Y-%m-%dT%H:%M:%SZ.
  • In PowerShell, use Get-Date -AsUtc -Format "yyyy-MM-ddTHH:mm:ssZ".
• files_changed must list only source files present in the diff (additions, modifications, or deletions). Filter by relevance - e.g. .py, .sh, .ts, .tf - excluding lock files, binaries, and build output.
• Do not write artifacts if the diff was empty and the review was aborted.
• The reviewer field must use the kebab-case form of the agent's or prompt's name from its frontmatter (e.g. Code Review → code-review).
• Write orientation-first artifacts only when the review runs in interactive orientation-first mode; record each produced artifact under the artifacts key in metadata.json and link it from review.md.
• Keep walkthrough.md and explanations/ artifacts factual (Register 1): no severity grades or verdicts. Keep walkback/ artifacts in the structured investigation register (Register 2).
• Create the explanations/ and walkback/ subfolders only when at least one explanation or investigation artifact is written.
• Every review.md ends with a Disclaimer and Human Review section: the verbatim ## Code-Review CAUTION disclaimer from disclaimer-language.instructions.md followed by an unchecked - [ ] Reviewed and validated by a qualified human reviewer checkbox. This section is always present, is always the final section, and the agent never checks the checkbox; only a human may convert [ ] to [x].
• When the review scope targets a pull request or merge request, review.md includes a mandatory human-editable PR Comment Draft section with an unchecked posting checkbox. This section is the only place the general PR or MR comment is authored; the agent never reproduces the full drafted comment body in the conversational summary. The agent never checks the posting box; only the human may convert [ ] to [x], and that check is the gate that authorizes posting the general PR or MR comment.