Content
56%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 body is a dense, highly concrete orchestration manual with strong validation gates and failure-path discipline, but it pays for that with structural repetition, scattered time-sensitive version notes, and detail that belongs in its own reference files. Organization of the core pipeline across many sections and one broken/absent reference keep structure and clarity just above midpoint.
Suggestions
Deduplicate the codex-companion `--effort max` handling and the dedicated worker model-route explanation into a single Codex/backend section, and link it from the other mentions instead of restating it three times.
Move secondary detail (Codex Native Orchestration, Mode 1 Sub-Lead, CC version notes and dated operator rulings) into a reference file, keeping SKILL.md to the pipeline, options table, and per-backend quick flows.
Fix or remove the ../../docs/long-running-harness.md link (target absent from the bundle) and either bundle or summarize the missing scripts/schemas (breezing-brief.sh, brief-card.v1.json) that the body instructs Claude to invoke; also relocate or delete the orphaned harness_ast_search bullets in the Mode 1 section.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | At ~480 lines the body restates the same facts repeatedly — the codex-companion `--effort max` / companion 1.0.6 override appears three times (backend table, "Natural Language Backend Triggers" area, and "Codex Mode"), the dedicated `worker` model route is explained twice, and [references/lean-path-detail.md] is linked six times. Scattered version/date specifics (CC 2.1.111/2.1.108/2.1.217/2.1.219, "2026-07-24 operator 裁定", "2026-08-22") are not isolated in a deprecated/old-patterns section, which the guidelines penalize. This is "noticeably verbose; several unnecessary explanations or padded sections" (anchor 2); it is below anchor 3 because the repetition is structural, not incidental. | 2 / 5 |
Actionability | Commands are concrete and copy-paste ready ("bash \"${HARNESS_PLUGIN_ROOT}/scripts/resolve-impl-backend.sh\"", "cursor-companion.sh task --write --workspace <wt>", the mktemp/stdin Codex example, `bin/harness work-mode on/off`, exact JSON for breezing-active.json, banner formats). Minor gaps keep it below anchor 5: placeholders like `<wt>` and `{base_ref}` are never resolved, `scripts/breezing-brief.sh` is sometimes invoked without the plugin-root prefix, and a few flows (e.g., Mode 1, model routing) defer to files outside the bundle. | 4 / 5 |
Workflow Clarity | The default pipeline is a clear numbered plan → work → review-gate → finalize sequence with explicit validation (self_review machine verification before Reviewer spawn, integrated review gate before Phase C finalization, iterate-until-APPROVE with a 3-attempt cap, escalation on non-convergence, `work-mode off` and state-file cleanup on all paths including failure). It misses anchor 5 because the full picture must be assembled from scattered sections (Default Pipeline, Execution, Review Policy, Phase 0, cursor flow), the Flow Summary is thin, and an orphaned English bullet list about harness_ast_search sits unrelated inside the Mode 1 section, breaking the reading flow. | 4 / 5 |
Progressive Disclosure | The two existing reference files (lean-path-detail.md, monitor-and-learning.md) are well-signaled with markdown links and stated purposes, but large bodies of detail that clearly belong in references remain inline (Codex Native Orchestration, Mode 1/Sub-Lead, backend selection tables, CC version notes), and one linked path (../../docs/long-running-harness.md) does not exist in the bundle while referenced scripts/templates (breezing-brief.sh, brief-card.v1.json schema) have no scripts/ or assets/ directories here. This fits anchor 3 ("Some structure but could be better organized; content that should be separate is inline") more than anchor 4. | 3 / 5 |
Total | 13 / 20 Passed |