Skip to content
Back to search

review-forge

review-forge is an Agent Skill for structured, auditable code review workflows

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/vikingmute--review-forge/badge.svg)](/listings/vikingmute--review-forge)

HTML

<a href="/listings/vikingmute--review-forge"><img src="/api/listings/vikingmute--review-forge/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

Install from GitHub with the Skills CLI:

npx skills add vikingmute/review-forge

Or use the full GitHub URL:

npx skills add https://github.com/vikingmute/review-forge

Install globally for a specific agent:

npx skills add vikingmute/review-forge -g -a codex

Install globally for multiple agents:

npx skills add vikingmute/review-forge -g -a codex -a cursor -a opencode

Update an installed copy:

npx skills update review-forge

Update only global installs:

npx skills update review-forge -g

For local development without publishing, clone it into the cross-client Agent Skills directory:

git clone https://github.com/vikingmute/review-forge.git ~/.agents/skills/review-forge

For a project-local install, copy or clone it into the project:

mkdir -p .agents/skills
git clone https://github.com/vikingmute/review-forge.git .agents/skills/review-forge

Some clients also support their own native skill locations. For Codex, this is also valid:

cp -R review-forge ~/.codex/skills/review-forge

Then invoke it from a compatible agent with:

Use $review-forge to review this branch.

If a client does not support explicit `$skill` syntax, ask it to use the `review-forge` skill or select it from the client's skills UI.

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

SKILL.md · Manifest · 229 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. ---
  2. name: review-forge
  3. description: Use when orchestrating a code review workflow across multiple LLMs or agents, synthesizing review findings, producing checklist-driven fix scopes, applying selected fixes, running regression tests, independently verifying fixes, updating review status, or keeping code review process artifacts isolated under code_review/.
  4. ---
  5. # Review Forge
  6. Review Forge runs a conservative, feature-scoped code review workflow:
  7. 1. Inspect the target diff and project context.
  8. 2. Create one review file per model or agent inside the current feature folder.
  9. 3. Synthesize multiple model review files into a human-controlled checklist.
  10. 4. Fix only checked and confirmed items.
  11. 5. Run focused and regression tests after every fix batch.
  12. 6. Verify fixes independently when possible.
  13. 7. Update machine-readable status and localized report labels.
  14. 8. Keep all workflow artifacts isolated from product code.
  15. ## Commands
  16. - `review`: create one model-specific review file. Do not modify product code.
  17. - `synthesize`: merge model review files from one feature folder into one summary checklist. Do not modify product code.
  18. - `fix`: read checked summary items, confirm the selected scope, implement only those items, run tests, and update status.
  19. - `verify`: verify fixed items with an independent perspective, rerun or inspect tests, and update verification status.
  20. Legacy mode aliases are supported:
  21. - `review-only` means `review` unless the user asks to merge existing review files, then it means `synthesize`.
  22. - `review-and-fix` means `fix`.
  23. - `review-verify` means `verify`.
  24. If the user does not specify a command, infer the safest command from the request. When ambiguous, default to `review`.
  25. ## Inputs
  26. Resolve or ask for these inputs before doing work:
  27. - `target`: branch, commit range, pull request, or current working tree.
  28. - `base`: comparison branch or ref. If omitted, infer it using the default diff rules below.
  29. - `artifact_root`: default `code_review/`.
  30. - `feature`: feature name used as the artifact subfolder. If omitted, infer from branch name, PR title, changed package, or a short sanitized summary of the diff.
  31. - `feature_dir`: default `code_review/<feature>/`.
  32. - `command`: `review`, `synthesize`, `fix`, or `verify`.
  33. - `model`: model or agent name for a review file, such as `codex`, `opencode`, `cursor`, `claude`, or `gemini`.
  34. - `report_language`: default `auto`, meaning follow the user's prompt language. Explicit values such as `en`, `zh-CN`, or `ja` are allowed.
  35. - `auto_fix_allowed`: default `false` unless the user explicitly asks to fix.
  36. - `perspective`: optional review focus, such as correctness, security, maintainability, tests, or UX.
  37. - `test_command`: optional override for validation.
  38. ## Default Diff Rules
  39. When the user does not specify a target or base, infer the review scope in this order:
  40. 1. If the working tree or index has uncommitted changes, review the uncommitted diff.
  41. - Inspect both unstaged and staged changes.
  42. - Treat this as `target: working tree`.
  43. - No base branch is required unless additional context is needed.
  44. 2. If there are no uncommitted changes, review the current `HEAD` against `main`.
  45. - Prefer a merge-base comparison such as `main...HEAD` when supported.
  46. - If `main` does not exist, try `master...HEAD`.
  47. 3. If neither `main` nor `master` exists, inspect branch tracking metadata or PR metadata when available.
  48. 4. Ask the user for `base` only when no reasonable comparison ref can be inferred.
  49. Explicit user-provided PRs, branches, commit ranges, or base refs always override these defaults.
  50. ## Required Initial Inspection
  51. Before producing artifacts or modifying code, inspect:
  52. - repo status and uncommitted changes;
  53. - target diff against the chosen base;
  54. - existing `code_review/` feature folders and review artifacts;
  55. - `.gitignore` for `code_review/`;
  56. - package, build, and test configuration;
  57. - project-specific instructions for tests, lint, typecheck, or CI.
  58. Never revert unrelated user changes. If `code_review/` is not ignored, recommend adding it to `.gitignore`, but do not edit `.gitignore` unless the user asks.
  59. ## Artifact Layout
  60. Group every workflow by feature:
  61. ```text
  62. code_review/<feature>/
  63. <model>.md
  64. <model>-<perspective>.md
  65. summary.md
  66. fix-plan.md
  67. verify.md
  68. status.md
  69. ```
  70. Rules:
  71. - One model review pass should produce exactly one review file.
  72. - Name review files with the model or agent name first, for example `codex.md`, `opencode.md`, `cursor.md`, `claude.md`.
  73. - If the same model runs multiple perspectives, append a short perspective suffix, for example `codex-security.md`.
  74. - Do not include timestamps by default. Use stable names so the folder is easy to scan.
  75. - Use timestamps only when preserving multiple historical runs is explicitly requested.
  76. Command output boundaries:
  77. - `review`: create or overwrite exactly one model review file, such as `codex.md`. Do not create `summary.md`, `fix-plan.md`, `verify.md`, or `status.md`.
  78. - `synthesize`: read model review files and create or overwrite `summary.md`. Do not create `status.md`.
  79. - `fix`: read checked items from `summary.md`, create or update `fix-plan.md`, and create or update `status.md` after code changes and tests.
  80. - `verify`: create or update `verify.md` and update `status.md` with verification evidence.
  81. Legacy timestamped names such as `STATUS_20260529.md`, `REVIEW_SUMMARY_20260529.md`, or `CODE_REVIEW_codex_20260529.md` should not be created unless the user explicitly asks to preserve dated historical runs.
  82. Use the templates in `assets/templates/` unless the repository already has a stronger local convention.
  83. ## Review Collection
  84. In `review` command, create a single file for the current model or agent. If the user does not provide `model`, infer it from the host when obvious; otherwise use a clear name such as `codex`.
  85. For multi-model workflows, each model should run `review` separately and write into the same `feature_dir`. Use distinct perspectives when helpful:
  86. - correctness and edge cases;
  87. - security and data safety;
  88. - tests and regression risk;
  89. - maintainability and architecture;
  90. - product or UX behavior when relevant.
  91. Each finding must include severity, evidence, affected files, risk, and a concrete suggested fix. Avoid vague style preferences unless they affect maintainability or user-facing behavior.
  92. ## Synthesize Command
  93. In `synthesize` command, read review files from one `feature_dir` and produce `summary.md`:
  94. - select all model review files in the folder by default;
  95. - exclude `summary.md`, `fix-plan.md`, `verify.md`, and `status.md`;
  96. - merge duplicate findings;
  97. - preserve meaningful disagreement;
  98. - record reviewer agreement count;
  99. - include suggested tests;
  100. - use checkboxes for human fix approval.
  101. Checkbox semantics are stable:
  102. - unchecked: not approved for fixing;
  103. - checked: approved for fixing.
  104. Do not fix unchecked issues opportunistically.
  105. ## Fix Rules
  106. In `fix` command:
  107. 1. Read `summary.md` from the selected `feature_dir`.
  108. 2. Select only checked items whose status allows fixing.
  109. 3. Confirm the fix scope unless the user already explicitly approved it.
  110. 4. Create or update `fix-plan.md`.
  111. 5. Modify only the files needed for selected items.
  112. 6. Run tests.
  113. 7. Update `status.md` and, when appropriate, the summary status fields.
  114. If a checked item reveals a larger issue, stop and ask before expanding scope.
  115. ## Test Rules
  116. Tests are mandatory after fixes unless impossible.
  117. Run validation in this order:
  118. 1. The narrowest relevant test for the touched behavior, when discoverable.
  119. 2. Typecheck, lint, or build checks relevant to the touched files.
  120. 3. The broader project regression command when discoverable and reasonably scoped.
  121. If tests cannot be run, record:
  122. - status enum `test_blocked`;
  123. - the command that should have run;
  124. - the blocking reason;
  125. - residual risk.
  126. If tests fail, do not claim the item is fixed or verified. Mark affected items `verification_failed` or `partially_fixed`, record the failing command, and summarize the failure.
  127. ## Verification Rules
  128. In `verify` command:
  129. - verify from a perspective independent of the fixer when possible;
  130. - check whether each selected issue was actually addressed;
  131. - rerun or inspect test evidence;
  132. - look for regressions caused by the fix;
  133. - update status only when evidence supports it.
  134. Verification requires either passing test evidence or a clear `test_blocked` explanation.
  135. ## Language Policy
  136. Write skill instructions, template field names, and machine-readable status enums in English.
  137. Generated prose reports should use `report_language`:
  138. - `auto`: follow the user's prompt language;
  139. - explicit language values: write prose in that language;
  140. - keep status enums in English even inside localized reports.
  141. Status display labels may be localized.
  142. ## Status Enums
  143. Use these exact machine-readable status values:
  144. - `open`
  145. - `approved_for_fix`
  146. - `fixed`
  147. - `partially_fixed`
  148. - `wont_fix`
  149. - `risk_accepted`
  150. - `verified`
  151. - `verification_failed`
  152. - `test_blocked`
  153. Each status entry should include both:
  154. - `status`: one enum above;
  155. - `status_label`: localized human-readable label, for example `Fixed ✅` or `已修复 ✅`.
  156. ## Completion Criteria
  157. A Review Forge workflow is complete only when:
  158. - review artifacts are under `feature_dir`;
  159. - checked items are the only fixed items;
  160. - test evidence or `test_blocked` rationale is recorded after fixes;
  161. - verification status does not overclaim;
  162. - process artifacts remain isolated from product code.
README.md · README · 168 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. # Review Forge
  2. [![skills.sh](https://skills.sh/b/vikingmute/review-forge)](https://skills.sh/vikingmute/review-forge)
  3. `review-forge` is an Agent Skill for structured, auditable code review
  4. workflows. It orchestrates multi-pass LLM review, finding synthesis,
  5. human-approved fix selection, implementation, regression testing, independent
  6. verification, and status tracking.
  7. ## What It Does
  8. - Reviews local changes, branches, PRs, or commit ranges.
  9. - Synthesizes multiple review perspectives into one fix checklist.
  10. - Uses checkboxes as the human approval boundary for fixes.
  11. - Fixes only approved items.
  12. - Requires tests after fixes unless testing is blocked and documented.
  13. - Verifies fixes independently when possible.
  14. - Keeps process artifacts isolated under `code_review/`.
  15. ## Commands
  16. - `review`: create one model-specific review file.
  17. - `synthesize`: merge review files in one feature folder into `summary.md`.
  18. - `fix`: fix checked items, run tests, and update status.
  19. - `verify`: verify fixed items, inspect or rerun tests, and update status.
  20. ## Default Review Scope
  21. When no target or base is specified:
  22. 1. If there are uncommitted or staged changes, Review Forge reviews the working
  23. tree diff.
  24. 2. If the working tree is clean, it reviews `main...HEAD`.
  25. 3. If `main` does not exist, it tries `master...HEAD`.
  26. 4. If no reasonable base can be inferred, it asks for one.
  27. Explicit PRs, branches, commit ranges, or base refs always override these
  28. defaults.
  29. ## Artifacts
  30. Review Forge groups workflow files by feature under `code_review/`:
  31. ```text
  32. code_review/<feature>/
  33. codex.md
  34. opencode.md
  35. cursor.md
  36. summary.md
  37. fix-plan.md
  38. verify.md
  39. status.md
  40. ```
  41. Each model review should produce one file named after the model or agent. If
  42. the same model runs multiple perspectives, use a suffix such as
  43. `codex-security.md`.
  44. Command outputs are intentionally simple:
  45. - `review` creates one model review file only, for example `codex.md`.
  46. - `synthesize` creates `summary.md` only.
  47. - `fix` creates or updates `fix-plan.md` and `status.md`.
  48. - `verify` creates or updates `verify.md` and `status.md`.
  49. Timestamped files such as `STATUS_20260529.md` are legacy-style outputs and
  50. should not be created unless you explicitly ask to preserve historical runs.
  51. Add `code_review/` to the target repository's `.gitignore` unless you
  52. intentionally want to commit review process files.
  53. ## Language Policy
  54. The skill instructions, template field names, and machine-readable status enums
  55. are in English. Generated reports can use `report_language: auto` to follow the
  56. user's prompt language, or explicit values such as `en`, `zh-CN`, or `ja`.
  57. Status values remain stable English enums, while display labels may be localized.
  58. ## Installation
  59. Install from GitHub with the Skills CLI:
  60. ```sh
  61. npx skills add vikingmute/review-forge
  62. ```
  63. Or use the full GitHub URL:
  64. ```sh
  65. npx skills add https://github.com/vikingmute/review-forge
  66. ```
  67. Install globally for a specific agent:
  68. ```sh
  69. npx skills add vikingmute/review-forge -g -a codex
  70. ```
  71. Install globally for multiple agents:
  72. ```sh
  73. npx skills add vikingmute/review-forge -g -a codex -a cursor -a opencode
  74. ```
  75. Update an installed copy:
  76. ```sh
  77. npx skills update review-forge
  78. ```
  79. Update only global installs:
  80. ```sh
  81. npx skills update review-forge -g
  82. ```
  83. For local development without publishing, clone it into the cross-client Agent
  84. Skills directory:
  85. ```sh
  86. git clone https://github.com/vikingmute/review-forge.git ~/.agents/skills/review-forge
  87. ```
  88. For a project-local install, copy or clone it into the project:
  89. ```sh
  90. mkdir -p .agents/skills
  91. git clone https://github.com/vikingmute/review-forge.git .agents/skills/review-forge
  92. ```
  93. Some clients also support their own native skill locations. For Codex, this is
  94. also valid:
  95. ```sh
  96. cp -R review-forge ~/.codex/skills/review-forge
  97. ```
  98. Then invoke it from a compatible agent with:
  99. ```text
  100. Use $review-forge to review this branch.
  101. ```
  102. If a client does not support explicit `$skill` syntax, ask it to use the
  103. `review-forge` skill or select it from the client's skills UI.
  104. ## Example Prompts
  105. ```text
  106. Use $review-forge review feature: checkout-refactor model: codex.
  107. ```
  108. ```text
  109. Use $review-forge review feature: checkout-refactor model: opencode perspective: security.
  110. ```
  111. ```text
  112. Use $review-forge synthesize feature: checkout-refactor.
  113. ```
  114. ```text
  115. Use $review-forge fix feature: checkout-refactor.
  116. ```
  117. ```text
  118. Use $review-forge verify feature: checkout-refactor.
  119. ```

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.