[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-75430":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":9,"language":10,"languages":9,"totalLinesOfCode":9,"stars":11,"forks":12,"watchers":13,"openIssues":14,"contributorsCount":15,"subscribersCount":15,"size":15,"stars1d":16,"stars7d":17,"stars30d":18,"stars90d":15,"forks30d":15,"starsTrendScore":19,"compositeScore":20,"rankGlobal":9,"rankLanguage":9,"license":21,"archived":22,"fork":22,"defaultBranch":23,"hasWiki":22,"hasPages":22,"topics":24,"createdAt":9,"pushedAt":9,"updatedAt":25,"readmeContent":26,"aiSummary":27,"trendingCount":15,"starSnapshotCount":15,"syncStatus":28,"lastSyncTime":29,"discoverSource":30},75430,"adamsreview","adamjgmiller\u002Fadamsreview","adamjgmiller","Multi-lens code review pipeline for Claude Code: deep review (Claude or Codex), auto-fix loop, interactive walkthrough, external-finding injection.",null,"Shell",234,6,3,5,0,4,14,79,12,2.54,"MIT License",false,"main",[],"2026-06-12 02:03:34","# adamsreview\n\nMulti-stage code review for Claude Code — parallel sub-agent detection, validation passes, persistent JSON state, and an automated fix loop that re-reviews and reverts regressions before committing.\n\nOn my own PRs, it's been catching dramatically more real bugs than Claude Code's built-in `\u002Freview`, `\u002Fultrareview`, CodeRabbit, Greptile, and Codex's built-in review — while producing fewer false positives. (Anecdotal, n=me.) Modeled after the built-in `\u002Freview` and extended into a six-command pipeline. Runs against your regular Claude Code subscription (Max plan recommended) — unlike `\u002Fultrareview`, which charges against your Extra Usage pool.\n\n```\n\u002Fplugin marketplace add adamjgmiller\u002Fadamsreview\n\u002Fplugin install adamsreview@adamsreview\n```\n\nThe six commands:\n\n- **`\u002Fadamsreview:review`** — multi-lens code review of a branch or PR. Up to seven parallel sub-agent lenses (correctness, security, UX, etc.) feed a dedup pass, a cheap-then-deep validation gate, and (optionally) a holistic Opus cross-cutting pass. High-confidence auto-fix proposals are pre-computed so `:fix` and `:walkthrough` can batch-accept them in one confirm. `--ensemble` adds a Codex CLI pass and PR bot-comment scrape on top of the internal Claude lenses.\n- **`\u002Fadamsreview:codex-review`** — Codex CLI peer to `:review`. Same artifact shape, drop-in for everything downstream (`:fix`, `:add`, `:walkthrough`, `:promote`). Effort tunable via `--effort low|medium|high|xhigh` (default `high`).\n- **`\u002Fadamsreview:add`** — inject externally-sourced findings (a Claude Code cloud `\u002Fultrareview` paste, an Opus once-over, a teammate's note) into the most recent review's artifact. Deduped against what's already there, validated by the same gates, re-published to the existing PR comment.\n- **`\u002Fadamsreview:walkthrough`** — interactive driver for findings `:fix` would skip. Uses the harness's `AskUserQuestion` UI to walk through uncertain or human-judgment items one by one — promote what you want auto-fixed, skip the rest. Pre-computed auto-fix proposals are batch-accepted up front; the remainder get per-finding briefing + options + recommendation. Posts a decisions log to the PR.\n- **`\u002Fadamsreview:fix`** — automated fix loop. Dispatches per-fix-group sub-agents in parallel, then re-reviews the work with Opus, **reverts any regressions, and commits the survivors** (one combined commit by default; `--granular-commits` for one per group).\n- **`\u002Fadamsreview:promote`** — human override that promotes a single finding to auto-fixable, bypassing the lane filter and score threshold.\n\nCommand files live at bare-stem paths under `commands\u002F`; shared phase fragments and prompt references live under `fragments\u002F`; helper scripts and the artifact schema live under `bin\u002F`. The plugin runtime auto-adds `bin\u002F` to `$PATH` on load — no symlinks, no install script.\n\n## Recommended flow\n\nOn a non-trivial PR, the commands work best in this order:\n\n1. **Review.** `\u002Fadamsreview:review` — or `\u002Fadamsreview:review --ensemble` if you have the Codex CLI installed and want to pool a Codex pass plus a PR bot-comment scrape on top of the internal Claude lenses (higher token cost). **Or** `\u002Fadamsreview:codex-review [--effort \u003Clevel>]` for a Codex-driven peer review (drop-in for everything downstream; effort tunable; no `--ensemble`).\n2. **Add.** *(optional)* `\u002Fadamsreview:add \u003Cpaste...>` — if you ran a parallel review (cloud `\u002Fultrareview`, Opus once-over, manual scan, etc.) that surfaced bugs the original review missed, paste the result here. The findings are validated by Phase 4 and land in the same artifact, deduped against what's already there. Auto-eligible additions feed step 4; non-eligible ones surface in step 3.\n3. **Walkthrough.** *(optional)* `\u002Fadamsreview:walkthrough [threshold]` — step through findings the fix command would skip (deep-manual, deep-report, and the entire light lane including light `confirmed_mechanical`), restricted to those scoring at or above `$threshold` (default 60) so low-signal items don't pad the session. Step 4.5 batch-accepts all findings carrying a pre-computed auto-fix proposal in one confirm (the fast path); the rest get per-finding briefing + options + recommendation via the harness's `AskUserQuestion` UI. Promote the ones you want auto-fixed with tailored fix-hints, skip the rest. Posts a decisions log to the PR for audit. Pass a lower threshold (e.g. `\u002Fadamsreview:walkthrough 30`) and pick the **Full** tier at the preflight prompt to audit Phase-3-demoted `below_gate` findings too.\n4. **Fix.** `\u002Fadamsreview:fix` — applies every auto-eligible finding (including whatever was added in step 2 and promoted in step 3). Phase 7.5 surfaces any remaining auto-fix proposals (light-lane \u002F manual \u002F report findings) for one-confirm batch-accept before Phase 8 dispatch. Default: one combined commit for all surviving fixes; pass `--granular-commits` for one commit per fix group. Per-group Phase-9 outcome lands in the commit message either way.\n\nEach command is independent — you can go straight from review to fix if you only care about auto-eligible findings, or skip review entirely and run `:fix` against an existing artifact. Steps 2–4 can land days or weeks after step 1; the review artifact persists under `~\u002F.adams-reviews\u002F\u003Cslug>\u002F\u003Cbranch>\u002F`.\n\n`\u002Fadamsreview:promote \u003Cid>` remains useful for one-off manual promotions outside the walkthrough flow (e.g. promoting a `disproven` finding with `--force`, or conceptually looping over a set of IDs — `F003`, `F037`, `F039` — with `--defer-publish` on each so only the final invocation re-publishes to the PR).\n\n## Documents\n\n- **`CLAUDE.md`** — operational guide for Claude Code sessions working in this repo. Self-contained for routine work; read first on a fresh session.\n- **`docs\u002Fstate-and-gates.md`** — finding state model, score gates, deep\u002Flight lanes (the normative spec).\n- **`docs\u002Fpipeline.md`** — phase trees and token-tally semantics for every command.\n- **`docs\u002Fhelpers.md`** — helper-script inventory and the batched-helper pattern.\n- **`bin\u002Fschema-v1.json`** — JSON Schema for `artifact.json` (source of truth for artifact shape).\n- **`docs\u002Farchive\u002F`** — frozen design + build docs (2026-04-19 onward). `DESIGN.md` (rev 8) is the original normative spec; `BUILD.md` is the stage-by-stage journal. Not maintained; consult only for historical rationale.\n- **`plans\u002F`** — per-branch plan files. Active follow-ups live in GitHub issues; historical backlog at `plans\u002Fold-backlog.md` (frozen 2026-05-04).\n\n## Dependencies\n\n### Runtime\n\n| Tool | Version | Used by | Notes |\n|---|---|---|---|\n| `uv` | 0.7+ | `artifact-patch.py`, `artifact-render.py` | `brew install uv`. Scripts use a PEP 723 inline-script shebang (`#!\u002Fusr\u002Fbin\u002Fenv -S uv run --quiet --script`) so `uv` fetches and caches `jsonschema` on first run — no venv, no global pip install |\n| `python3` | 3.10+ | invoked by `uv` | `uv` will install a matching Python if needed |\n| `bash` | 3.2+ | all `*.sh` helpers | Helpers are intentionally 3.2-portable (no `declare -A`, `mapfile`, `${var,,}`), so macOS's default `\u002Fbin\u002Fbash` works as-is. On Windows, Git for Windows ships bash 5+ via Git Bash and Claude Code auto-routes through it |\n| `jq` | 1.6+ | `artifact-read.sh`, log helpers | `brew install jq` |\n| `gh` | 2.x | `artifact-publish.sh`, `external-scrape.sh` | `brew install gh`, `gh auth login` |\n| `git` | 2.x | everywhere | standard |\n\n## Installation\n\n### macOS \u002F Linux\n\n1. Install deps: `brew install uv jq gh git` (macOS) or the distro equivalent. (macOS's default `\u002Fbin\u002Fbash` 3.2 is fine — helpers are 3.2-portable.)\n2. In a Claude Code session: `\u002Fplugin marketplace add adamjgmiller\u002Fadamsreview`\n3. In the same session: `\u002Fplugin install adamsreview@adamsreview`\n\n### Windows (native)\n\n1. Install [Git for Windows](https:\u002F\u002Fgit-scm.com\u002Fdownloads\u002Fwin) — provides Git Bash (bash 5+) and `git`, which Claude Code uses internally. Claude Code auto-routes `#!\u002Fusr\u002Fbin\u002Fenv bash` helpers through Git Bash; set `CLAUDE_CODE_GIT_BASH_PATH` if Git Bash lives in a non-default location (see *Troubleshooting*).\n2. Install [uv](https:\u002F\u002Fdocs.astral.sh\u002Fuv\u002F), [jq](https:\u002F\u002Fjqlang.github.io\u002Fjq\u002Fdownload\u002F), and the [GitHub CLI](https:\u002F\u002Fcli.github.com\u002F).\n3. In a Claude Code session: `\u002Fplugin marketplace add adamjgmiller\u002Fadamsreview` and `\u002Fplugin install adamsreview@adamsreview`.\n\n### Install from a local checkout\n\nIf you've cloned this repo and prefer running from source — or you want to pin to a specific commit — two paths work without the GitHub marketplace round-trip:\n\n- **Persistent install from a local path.** In a Claude Code session, run `\u002Fplugin marketplace add \u002Fpath\u002Fto\u002Fadamsreview` then `\u002Fplugin install adamsreview@adamsreview`. Same end state as the GitHub marketplace flow above — the plugin is registered under `~\u002F.claude\u002F` and survives restarts. Use `.` in place of the absolute path if your cwd is already the clone.\n- **One-shot via `--plugin-dir`.** `claude --plugin-dir \u002Fpath\u002Fto\u002Fadamsreview` launches Claude Code with the clone loaded as a plugin for that session only. Nothing is written to `~\u002F.claude\u002F`; re-launch without the flag and the plugin is gone. Handy for trying the plugin without any persistent state, or for running a specific checkout side-by-side with an installed version.\n\nBoth paths still require the runtime deps listed above (`uv`, `jq`, `gh`, `bash`, `git`).\n\n### Commands (post-install)\n\nAll invocations are plugin-namespaced:\n\n- `\u002Fadamsreview:review [--ensemble] [--full]`\n- `\u002Fadamsreview:codex-review [--effort \u003Clow|medium|high|xhigh>] [--full]`\n- `\u002Fadamsreview:add [\u003Cpaste...>] [--file \u003Cpath> --line \u003CN> --claim \"...\"]`\n- `\u002Fadamsreview:walkthrough [threshold]`\n- `\u002Fadamsreview:fix [threshold]`\n- `\u002Fadamsreview:promote \u003Cfinding_id> [--reason \"...\"] [--fix-hint \"...\"]`\n\n`--full` (on `:review` and `:codex-review`) opts out of the trivial-mode optimization, forcing every detection lens to run even on small or docs-only diffs. Useful when you want full coverage on a deliberately-small PR; otherwise the default trivial-mode classifier is the right call.\n\nNo separate Python dep install. First invocation of any `*.py` helper triggers `uv` to resolve declared deps (`jsonschema` etc.) and cache them — this can take a few seconds on a fresh machine (see *Troubleshooting*). Subsequent runs are fast.\n\n### Plugin-author iteration\n\nIf you're hacking on the plugin itself (not just using it), `scripts\u002Fdev-run.sh` launches Claude Code with the working tree loaded as a plugin via `claude --plugin-dir \"$(pwd)\"` — no marketplace install needed. For install-path simulation from a working tree, run `\u002Fplugin marketplace add .` inside a Claude Code session.\n\n### Review state location\n\n`\u002Fadamsreview:review` writes per-run state (artifact, trace, phase logs, token logs) under `~\u002F.adams-reviews\u002F\u003Crepo-slug>\u002F\u003Cbranch>\u002F\u003Creview_id>\u002F`. Override with `export ADAMS_REVIEW_REVIEWS_ROOT=\u002Fsome\u002Fother\u002Fpath` if you want state elsewhere.\n\n**Why not `~\u002F.claude\u002Freviews\u002F`?** Claude Code hardcodes a sensitive-file permission prompt for writes to `~\u002F.claude\u002F...` that survives even `bypassPermissions` mode, and `~\u002F.claude\u002Freviews` is not on the short list of exempt subdirs (`.claude\u002Fcommands`, `.claude\u002Fagents`, `.claude\u002Fskills`). Keeping review state outside `~\u002F.claude\u002F` avoids dozens of permission prompts per run.\n\n**Migrating from pre-Stage-2.5 state.** If you have reviews under `~\u002F.claude\u002Freviews\u002F`, either:\n\n```bash\n# Option A: move state to the new canonical root (recommended).\nmv ~\u002F.claude\u002Freviews ~\u002F.adams-reviews\n\n# Option B: keep state at the old location via the env var (accepts the prompts).\nexport ADAMS_REVIEW_REVIEWS_ROOT=~\u002F.claude\u002Freviews\n```\n\n### Token counts: what they measure\n\nThe rendered report can surface two numbers:\n\n- **Sub-agent tokens** — rolled up from the per-review `tokens.jsonl` log. Counts every dispatched sub-agent (lenses, validators, fix agents, post-fix reviewer, etc.) for this specific review. Precise. Always shown.\n- **Orchestrator tokens** — rolled up from the Claude Code session transcripts under `~\u002F.claude\u002Fprojects\u002F\u003Ccwd-slug>\u002F`, filtered to assistant turns with `timestamp >= review_started_at`. Captures the main-session spend that `subagent_tokens` deliberately excludes. **Opt-in** — see below.\n\nWhen both are populated they're complementary (no overlap), and together estimate total cost.\n\n#### Orchestrator tokens are opt-in\n\nmacOS Sequoia and Tahoe show an \"*kitty (or your terminal) would like to access data from other apps*\" prompt the first time a shell helper reads files marked with the `com.apple.provenance` extended attribute. Every Claude Code transcript carries one, and `bin\u002Forchestrator-tokens.sh` reads them — so the helper would trigger the prompt on the first lifecycle command of every review and (because the TCC cache for this gate is partial) repeatedly thereafter. To avoid pestering users, the helper defaults to skip.\n\nTo enable, do **either**:\n\n- **Recommended** — grant your terminal app **Full Disk Access** (System Settings → Privacy & Security → Full Disk Access → `+` your terminal). One toggle, permanent, silences this prompt class for everything launched from your terminal. Then `export ADAMS_REVIEW_TALLY_ORCHESTRATOR=1` in your shell rc (`~\u002F.zshrc`, `~\u002F.bashrc`, etc.) so the helper actually runs.\n- **Just opt in without FDA** — `export ADAMS_REVIEW_TALLY_ORCHESTRATOR=1` and accept the macOS prompts when they fire. Each grant survives until the next OS update or terminal-app update; choose this if you want narrower permissions at the cost of clicking *Allow* periodically.\n\nWhen opted out (the default), the helper exits 0 with one `orchestrator-tally: skipped` line and leaves the artifact's `orchestrator_tokens` field absent. The PR comment shows only **Sub-agent tokens** — still the precise per-review counter and the primary cost signal. Sub-agent tokens log under `~\u002F.adams-reviews\u002F`, which carries no provenance xattr, so that path triggers no prompts.\n\n**Stale-data behavior.** If you opt in for the initial review and then opt out before running `\u002Fadamsreview:fix`, the helper preserves the previously-written `orchestrator_tokens` value rather than wiping it. The rendered line shows the last-measured value, not a freshly-skipped zero — meaning it can under-report subsequent fix-time activity. Re-opt-in on the next lifecycle command refreshes. The reverse direction (opt out for review, opt in for fix) has no staleness: the fix-time tally's `--since review_started_at` window covers the full review→fix arc, so the first opt-in write captures everything.\n\n#### Orchestrator tokens can over-count (when opted in)\n\nThe transcript scan is a pure time-window filter, so any Claude Code turn in the same working directory between `review_started_at` and the last tally gets counted — even if it's unrelated work. In practice that means:\n\n- **Clean:** review → fix back-to-back, or review → new review on updated codebase (each review's `review_started_at` excludes the prior one's turns).\n- **Over-counts:** review → unrelated work in the same cwd → fix (the unrelated turns land in the fix run's re-tally).\n- **Mitigation:** run the lifecycle commands close together, or do unrelated work in a different worktree (different cwd → different transcript directory → not scanned).\n\nSub-agent tokens don't have this problem — their log is per-review. If you need a precise total, trust sub-agent tokens and treat orchestrator tokens as a rough ceiling. See `bin\u002Forchestrator-tokens.sh` header for the full list of caveats.\n\n### Why `uv` instead of plain pip\n\nPEP 668 (Python 3.12+ with Homebrew) marks system and user site-packages as externally managed and refuses direct `pip install`. The original plan assumed plain pip; `uv`'s inline-script dep spec is the cleanest workaround: each Python helper is self-contained, runs without activation ceremony, and its dep list lives next to the code that imports it. Tradeoff: requires `uv` on the machine running the scripts.\n\n## Layout\n\n```\nadamsreview\u002F                           ← this repo (plugin root)\n├── CLAUDE.md                          ← operational guide (read first)\n├── README.md                          ← this file\n├── .claude-plugin\u002F\n│   ├── plugin.json                    ← plugin manifest (name: adamsreview)\n│   └── marketplace.json               ← single-plugin marketplace\n├── .gitattributes                     ← LF enforcement\n├── docs\u002F\n│   └── archive\u002F                       ← frozen historical references (not maintained)\n│       ├── README.md                  ← frozen-as-of banner\n│       ├── DESIGN.md                  ← original normative design (rev 8)\n│       └── BUILD.md                   ← build journal (Stages 1–3 + hardening + walkthrough)\n├── plans\u002F                             ← per-branch plan files (umbrella + optional PRD\u002FPLAN\u002FJOURNAL)\n├── test\u002F                              ← smoke harness + fixtures\n├── commands\u002F                          ← bare-stem command files (plugin namespacing)\n│   ├── review.md                      ← \u002Fadamsreview:review\n│   ├── codex-review.md                ← \u002Fadamsreview:codex-review\n│   ├── add.md                         ← \u002Fadamsreview:add\n│   ├── walkthrough.md                 ← \u002Fadamsreview:walkthrough\n│   ├── fix.md                         ← \u002Fadamsreview:fix\n│   └── promote.md                     ← \u002Fadamsreview:promote\n├── fragments\u002F                         ← shared phase fragments + prompt references\n│   ├── _prelude-shared.md             ← loaded by every command\n│   ├── promote-core.md                ← shared precondition + patch (promote + walkthrough)\n│   ├── 00-preflight.md … 10-post-fix-and-commit.md\n│   │     (incl. 06b-auto-fix-hint.md for Phase 5.5)\n│   ├── 02-ensemble-adapter.md         ← --ensemble pooling\n│   ├── 01-codex-detection.md          ← Codex Phase 1 variant\n│   ├── 05-codex-validation.md         ← Codex Phase 4 variant\n│   ├── 06-codex-cross-cutting.md      ← Codex Phase 5 variant\n│   ├── lens-prompts\u002F                  ← per-lens detection prompts\n│   └── lens-*-reference.md\n├── bin\u002F                               ← helper scripts (plugin runtime auto-adds to $PATH)\n│   ├── include                        ← `!include \u003Cfragment>.md` wrapper\n│   ├── schema-v1.json                 ← JSON Schema for artifact.json\n│   ├── _common.py                     ← shared Python helpers\n│   ├── artifact-patch.py              ← machine-state writer\n│   ├── artifact-render.py             ← JSON → Markdown\n│   ├── artifact-validate.sh           ← schema check (bash wrapper)\n│   ├── artifact-read.sh               ← jq wrapper\n│   ├── artifact-publish.sh            ← PR comment post\u002Fpatch\n│   ├── artifact-seed.sh               ← initial artifact scaffold\n│   ├── claude-md-paths.sh             ← walk-up CLAUDE.md finder\n│   ├── staleness.sh                   ← git diff intersection\n│   ├── freshness-gate.sh              ← trace freshness check\n│   ├── trivial-check.sh               ← trivial-mode classifier\n│   ├── codex-poll.sh                  ← Codex CLI invocation + watchdog\n│   ├── parse-validator-result.py      ← validator-output parser\n│   ├── parse-with-repair.py           ← JSON-with-repair parser\n│   ├── source-family-map.py           ← source-family lookup\n│   ├── log-phase.sh                   ← trace.md + phases.jsonl appender\n│   ├── log-tokens.sh                  ← tokens.jsonl appender\n│   └── (other helpers: group-fixes.py, repo-slug.sh, comment-freshness.sh,\n│      origin-crosscheck.sh, prior-fix-diff.sh, external-scrape.sh,\n│      assign-finding-ids.sh, line-range-check.sh, tally-subagent-tokens.sh,\n│      orchestrator-tokens.sh)\n├── hooks\u002F\n│   ├── hooks.json                     ← SessionStart registration\n│   └── dep-check.sh                   ← soft dep-missing warning at session start\n└── scripts\u002F\n    └── dev-run.sh                     ← `claude --plugin-dir` wrapper (plugin-author iteration)\n```\n\nNo symlinks, no install script. The plugin runtime discovers `commands\u002F`, `fragments\u002F`, `bin\u002F`, and `hooks\u002F` by convention once the plugin is installed via `\u002Fplugin install adamsreview@adamsreview`.\n\n## Troubleshooting\n\n### First invocation is slow\n\nThe Python helpers (`artifact-patch.py`, `artifact-render.py`) use a PEP 723 inline-script shebang (`#!\u002Fusr\u002Fbin\u002Fenv -S uv run --quiet --script`). On a fresh machine, the first run pauses for a few seconds while `uv` resolves a matching Python interpreter and fetches the `jsonschema` dep into its cache. Subsequent runs hit the cache and are effectively instant. This is a one-time cost per machine, not per review.\n\n### `--ensemble` mode requirements\n\n`\u002Fadamsreview:review --ensemble` additionally requires the `codex` Claude Code plugin (the local Codex CLI is invoked through the plugin's `codex-companion.mjs`, not as a standalone CLI on `$PATH`). Without the plugin installed, the readiness gate prompts you to either continue without Codex (in PR mode that means PR-comment scraping only; in local mode it means internal lenses only — Phase 1.5 has no work to do) or stop and run `\u002Fcodex:setup` first. The default (non-ensemble) mode has no such requirement.\n\n### Windows: Git Bash not found\n\nClaude Code auto-discovers Git Bash on Windows and routes `#!\u002Fusr\u002Fbin\u002Fenv bash` helpers through it. If the auto-discovery fails (non-default Git Bash install path, portable install, etc.), set `CLAUDE_CODE_GIT_BASH_PATH` to the absolute path of `bash.exe` before launching Claude Code — for example:\n\n```\nset CLAUDE_CODE_GIT_BASH_PATH=C:\\Program Files\\Git\\bin\\bash.exe\n```\n\nor the `$env:CLAUDE_CODE_GIT_BASH_PATH = ...` equivalent in PowerShell.\n\n## Status\n\nCurrent release: **v0.4.0** (auto-fix-hint feature: Phase 5.5 + Phase 7.5 + Step 4.5). All six commands are shipped and in daily use. `\u002Fadamsreview:codex-review` landed in v0.3.0; recent releases have focused on hardening (anti-serialization callouts at fan-out sites, parallel-dispatch correctness, JSON-pipeline backslash safety) and the auto-fix-hint flow that lets `:fix` and `:walkthrough` batch-accept Sonnet-proposed fixes in one confirm.\n\nActive follow-ups live in GitHub issues. Frozen historical context: `plans\u002Fold-backlog.md` and `docs\u002Farchive\u002F`.\n","adamsreview 是一个用于 Claude Code 的多阶段代码审查工具，它通过并行子代理检测、验证通道、持久化 JSON 状态和自动化修复循环来提高代码质量。其核心功能包括多视角代码审查（支持多达七个并行子代理）、自动修复建议预计算以及交互式审查流程。该工具特别适合于需要深度代码审查的场景，如复杂项目的拉取请求审查。相比 Claude Code 内置的审查功能，adamsreview 能够更有效地捕捉真实错误同时减少误报。此外，它还支持将外部发现注入到最新的审查结果中，并提供了一种机制让用户可以逐个处理不确定或需要人工判断的问题。",2,"2026-06-11 03:52:44","CREATED_QUERY"]