Content
62%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
The review orchestration itself is excellently specified — a clear phase sequence with strong validation checkpoints, feedback loops, and concrete templates. The weaknesses are structural: substantial duplicated reliability/compatibility guidance inflating the token budget, and no progressive disclosure, with large harness-specific reference material inlined in the main file.
Suggestions
Deduplicate the subagent reliability guidance: state the wait/back-off, WORKING/BLOCKED signaling, alive-not-silent, and fallback rules once — the compatibility section and the reliability section currently repeat nearly identical paragraphs, and "If a code block below conflicts with this section, this section wins" appears three times.
Move the Codex/OpenCode tool-compatibility table, the multi_agent_v1/v2 routing rules, and the full gate-reviewer spawn prompt template into a referenced bundle file (e.g. references/harness-compat.md and references/gate-reviewer-prompt.md), keeping SKILL.md to the phase workflow and verdict logic.
Trim restatements of the same rule across sections — e.g. the worktree lock/remove teardown and BLOCKED/INCONCLUSIVE handling each appear in both the orchestrator overview and Phase 3 — consolidating them into the phase where the action occurs.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The core phases are dense, but the document is noticeably verbose: the Codex compatibility and subagent-reliability sections repeat the same rules multiple times ("If a code block below conflicts with this section, this section wins" appears three times; the back-off/timeout and "fallback only when the child is completed without the deliverable, ack-only after followup, explicitly BLOCKED:, or no longer running" guidance appears nearly verbatim in both the compatibility and reliability sections), padding out roughly a third of the body with duplication. | 2 / 5 |
Actionability | Mostly executable guidance — copy-paste bash auto-collection commands, a complete gate-reviewer prompt template with placeholders, QA matrix and report formats, and concrete worktree commands — but the pivotal reviewer-launch example uses OpenCode `task(...)` syntax that must be translated per the compatibility table, and the web QA path references an external `browser` skill's internals without inline commands. | 4 / 5 |
Workflow Clarity | The five phases are clearly sequenced with explicit validation checkpoints and feedback loops: "Any FAIL ends the review here", artifact-required QA rows, INCONCLUSIVE lane handling with one bounded respawn, teardown ordering ("run `git worktree unlock` followed by `git worktree remove`"), and re-review as a fresh delta-scoped pass — plus an 8-item reviewer checklist. | 5 / 5 |
Progressive Disclosure | The body has reasonable section structure (headers, tables, phase organization) but is a 325-line monolith with no bundle files at all; content that clearly belongs in separate references — the ~58-line harness tool-compatibility table, the subagent reliability rules, and the full spawn-prompt template — is inlined in SKILL.md. | 3 / 5 |
Total | 14 / 20 Passed |