Dartai Verdict Schema

Reviewer Verdict Schema (Canonical)

This document defines the single source of truth for the return shape every adversarial reviewer subagent emits. The main loop and orchestrator skills (e.g. dartai:adversarial-quality-loop, workflow:loop-orchestration) read only the fields defined here. Anything outside the schema is dropped at the boundary.

Why verdict-only: prose narrative across N reviewers per loop iteration bloats the main-thread context. The verdict block carries the same decision signal at ~80% token reduction. Detail still has a home — see evidence_path.

Schema

Verdict values

• pass — no blockers, gate may proceed.
• fail — at least one blocker; gate must not proceed; blockers MUST be non-empty.
• warn — no blockers but advisories warrant attention; gate proceeds; orchestrator may surface advisories to the operator.

Confidence values

• high — reviewer is sure; finding/no-finding is well-grounded.
• med — reviewer is reasonably sure; one or two ambiguities exist.
• low — reviewer is uncertain; orchestrator should consider re-dispatch with a different model or a tighter prompt, or escalate to a human.

Optional evidence_path

When a finding genuinely needs more than a one-liner (e.g. a long stack-trace, a multi-step exploit chain, or a table of duplicate sites), write the detail to a file under .dartai/reports/<task-id>/<reviewer-name>.md and put the relative path here. The main loop will not read it; downstream tools or operators may.

Wire format

Reviewer agents emit a single fenced YAML block as the final message body. No preamble, no postamble, no prose around it. Keep the entire block ≤30 lines.

```yaml
verdict: pass | fail | warn
confidence: high | med | low
blockers:
  - "<file:line> — <one-line description>"
advisories:
  - "<one-line nit or follow-up>"
evidence_path: ".dartai/reports/<task-id>/<reviewer-name>.md"  # optional
```

Examples

Pass with no notes:

verdict: pass
confidence: high
blockers: []
advisories: []

Pass with advisories:

verdict: pass
confidence: med
blockers: []
advisories:
  - "src/auth.ts:42 — consider extracting role-check helper (used 3×)"
  - "no test for 0-length input on parseToken; happy path only"

Fail with blockers:

verdict: fail
confidence: high
blockers:
  - "src/handler.ts:88 — SQL string-concat with user input (A03 injection)"
  - "test/handler.test.ts — missing assertion on error message"
advisories:
  - "src/handler.ts:120 — duplicate of util/parseQuery (LCI hit)"
evidence_path: ".dartai/reports/RXB0s0OPWWZF/code-quality-reviewer.md"

Warn (advisory-only):

verdict: warn
confidence: med
blockers: []
advisories:
  - "test distribution 70/20/10 — slightly under target on adversarial cases"

Length budget

The full block — including header, lists, and evidence_path — must be ≤30 lines. If you cannot fit findings in that budget:

1. Collapse multiple related findings into one line with a count (e.g. "3× weak assertions in test/foo.test.ts").
2. Move detail into evidence_path and reference it.
3. Never split a single reviewer's output across multiple blocks.

Consumer contract

The main loop and orchestrator skills:

• Parse verdict and blockers only.
• Treat advisories as informational; may surface or drop based on telemetry settings.
• Use confidence: low as a re-dispatch signal (per orchestrator policy).
• Read evidence_path only when explicitly requested by an operator.

Pass/fail gate semantics (preserved from prior prose-format reports):

• Any reviewer verdict: fail → gate fails; orchestrator collects blockers from all failing reviewers and routes them back to the implementer.
• All reviewers verdict: pass (or warn) → gate passes.
• Reviewers gated by enabled_when predicates are skipped when the diff does not match; skipped reviewers do not contribute to the AND.

Authoring rules for reviewer agents

When updating a reviewer agent's "Report Format" / "Return" section:

1. Replace any prose-narrative template (qa_report:, code_quality_report:, post_task_report:, etc.) with a reference to this schema.
2. Specify in the agent's instructions: emit the YAML block defined above as the final message body, no preamble.
3. Keep the agent's internal analysis structure (review areas, attack vectors, etc.) — that shapes how the reviewer thinks. Only the output shape is constrained by this schema.
4. Map the reviewer's old verdict tokens to the new ones:
   • PASS / pass → pass
   • FAIL / fail / REJECT / STOP → fail
   • NEEDS_WORK / WARNING → warn (when no blocker is gate-breaking) or fail (when it is)
5. If the reviewer needs to express depth beyond ≤30 lines, write to .dartai/reports/<task-id>/<reviewer-name>.md and set evidence_path.

Backward compatibility

Legacy reviewers emitting the old prose *_report: blocks remain readable but are deprecated. Orchestrators may log a one-time deprecation note and proceed by extracting verdict from the legacy block. New reviewer authors and edits MUST use the schema above.

Reviewers using this schema

Always-dispatched (dartai plugin):

• plugins/dartai/agents/qa-reviewer.md
• plugins/dartai/agents/code-quality-reviewer.md
• plugins/dartai/agents/post-task-reviewer.md

Always-dispatched (workflow plugin mirrors):

• plugins/workflow/agents/qa-reviewer.md
• plugins/workflow/agents/code-quality-reviewer.md
• plugins/workflow/agents/post-task-reviewer.md

Always-dispatched (compound-review plugin):

• plugins/compound-review/agents/correctness-reviewer.md
• plugins/compound-review/agents/maintainability-reviewer.md
• plugins/compound-review/agents/testing-reviewer.md

Conditional (compound-review plugin — dispatched when diff matches):

• plugins/compound-review/agents/typescript-strict-reviewer.md (when diff touches *.ts/*.tsx)
• plugins/compound-review/agents/cli-readiness-reviewer.md (when diff touches CLI paths)

Not yet migrated (rationalization-trap-reviewer) — still emits structured JSON per its bespoke schema. Orchestrators treat absence of verdict-file: on stdout line 1 as legacy format and read the JSON body directly.

Verdict File Delivery (file-streaming channel) 裁決檔派送

Reviewer subagents return their decision via a verdict file written to disk, not via stdout body. The loop driver consumes the file content (read fresh, parse line-oriented), not the subagent transcript. This eliminates per-iteration child-context bloat and lets the driver replay any gate decision from the verdict file alone.

File layout 檔案佈局

The directory is .dartai/reports/<task-id>/ (consistent with existing evidence_path references). Filenames are fixed per reviewer role so the driver can scan a known set without listing.

workflow: orchestrator mirrors substitute .workflow/reports/<task-id>/ for its own state dir but the file shapes are identical.

File format (line-oriented, parseable by the loop driver)

Each verdict file is plain text, line-oriented:

verdict: pass|fail|warn
confidence: high|med|low
blocker: <file:line> <one-line description>
blocker: <file:line> <one-line description>
advisory: <one-line nit>
evidence: <inline body or relative path>

Rules:

• Line 1 MUST be verdict: <pass|fail|warn> — single token after the colon.
• Line 2 MUST be confidence: <high|med|low>.
• Lines 3+ are zero or more blocker: lines (required when verdict: fail), followed by zero or more advisory: lines.
• Trailing evidence: line is optional. If the value starts with ./ or is a path (no spaces, no leading prose), the driver treats it as a path reference; otherwise the rest of the file (after evidence:) is the inline evidence body.
• Lines starting with # are comments and ignored.
• Trailing whitespace and blank lines are ignored.

Example (.dartai/reports/abc123/quality.md):

verdict: fail
confidence: high
blocker: src/handler.ts:88 SQL string-concat with user input
blocker: src/handler.ts:120 duplicate of util/parseQuery (LCI hit)
advisory: src/auth.ts:42 consider extracting role-check helper
evidence: ./quality-evidence.md

Example (pass with no notes):

verdict: pass
confidence: high

Stdout contract 標準輸出契約

When a reviewer subagent uses the verdict-file channel, its stdout body MUST be ≤5 lines and serve only as a pointer plus a one-line summary:

verdict-file: .dartai/reports/<task-id>/<filename>.md
verdict: <pass|fail|warn> (<short reason if fail/warn, else empty>)

The first line is the file path; the second is the one-line verdict for human-readable progress logs. Reviewers MUST NOT inline the YAML block from the wire format above when delivering via file — the file is the canonical channel. The ≤5-line cap leaves room for an optional confidence: line and one advisory headline if the reviewer chooses, but no more.

Loop driver semantics 循環驅動語義

Drivers consuming the verdict-file channel:

1. Clear the reports dir at task start. Before dispatching reviewers for a new task, remove .dartai/reports/<task-id>/ (if present) and re-create it empty. This is the stale-verdict mitigation — a previous run's verdict file in the same path would otherwise be read as the current run's gate decision.

2. React to verdict files via Monitor, not stdout consumption. Spawn each reviewer subagent in the background; use Monitor over the verdict-file path (or the parent reports dir) to detect file changes. Parse the file fresh on each notification — do NOT consume the subagent's stdout body into driver context.

3. Prefer parse-on-completion over event polling. When the harness emits a subagent-completion notification, parse the verdict file at that point. File-system event polling is a fallback for harnesses without completion notifications. This is the missed-event mitigation — completion notifications are the durable signal; file-system events alone can drop under load. If both are available, parse on completion AND keep the Monitor stream open for late writes.

4. Gate on file content, not stdout body. The gate decision is verdict: from line 1 of the file. The subagent's transcript is discarded after parsing. Replay is possible from the file alone — re-running a gate decision means re-reading the verdict file, no need to re-dispatch the reviewer.

5. Aggregate to verdict-summary.kdl after all reviewers report. Run node ${CLAUDE_PLUGIN_ROOT}/scripts/aggregate-verdicts.js <task-id> (or the equivalent relative path). The aggregator reads all role-fixed verdict files, computes the overall gate verdict (any fail → fail; any warn and no fail → warn; else pass), and writes .dartai/reports/<task-id>/verdict-summary.kdl. The aggregator is the only consumer that reads multiple verdict files at once.

Backward compatibility 向後兼容

• Reviewers still implementing the in-stdout YAML wire format (above) remain readable. Orchestrators MUST accept either shape: file-channel reviewers expose verdict-file: on line 1 of stdout; legacy reviewers emit the fenced YAML block. The driver detects which by checking line 1 of stdout.
• Existing evidence_path references in agent files continue to work — evidence_path is the YAML-block field name; evidence: is the line-oriented field name in the verdict file. They carry the same payload.
• The fixed per-role filenames (qa.md, quality.md, security.md) replace the older <reviewer-name>.md convention for new gate decisions. Legacy paths are still readable for replay; new writes use the role-fixed names.