Miucr

miucr: owned AI code-review CLI (v0.65.0)

miucr (the miu-cr project) is a fast, pure-Go (CGO_ENABLED=0) AI code reviewer. It keeps the correctness-critical parts deterministic (file selection, context assembly, line-anchoring, severity gating, dedupe) and uses the LLM only for judgment (finding bugs, proposing fixes). It runs five review ways:

• Local review: miucr review over a staged diff, a commit, or a ref range.
• GitHub PR review: miucr review --pr (dry-run by default; --post reacts 👀, upserts ONE summary issue comment, and posts inline comments as a PR review).
• serve daemon: HMAC webhook (default) and/or opt-in poll trigger; optional REST API + GitHub App auth.
• MCP server: miucr mcp exposes review_run / review_get over stdio to any agent host.
• Evaluation: miucr eval compares miu-cr and other reviewer commands against JSON expected findings.

Review behavior worth knowing (design choices that prevent noise):

• One upserted summary, posted first. --post writes ONE summary issue comment, edited in place on re-runs (never stacked), and acknowledges the PR with 👀. If no summary exists yet, it also creates a short Review running placeholder before the LLM starts; on later commits it keeps the prior completed summary visible until the new result is ready. The final summary replaces that same comment and is posted BEFORE the inline review so it anchors on top (overview → details). Inline findings are a separate PR review. review_id is NOT shown in the comment (it only resolves on the local store; it stays in the JSON envelope). On a fatal review failure after miucr's internal retries, --post upserts that SAME summary comment with a visible GitHub alert instead of failing silently: operational/provider/infrastructure failures (agent.unavailable, provider.rate_limited, quota.exceeded, review.timeout, github.unavailable, etc.) render as [!WARNING], while unknown internal.error failures render as [!CAUTION]; a later successful run replaces it with findings.
• The summary is a per-finding lifecycle ledger, not just the latest run. Below a concise ≤5-bullet "What changed" summary, it renders two always-visible tracking tables - ⚠️ Open (N) and ✅ Resolved (N) - each finding tracked by its line-independent fingerprint across commits: a Priority column (P0-P4), status (open / resolved / reopened), the origin commit it was first raised on and the resolved commit it disappeared on (both linked), priority before→after for escalations, and first-seen / resolved timestamps. A clean review uses natural all-clear prose like Review passed! No findings on the first review pass. or Review passed! 3 findings resolved. Good cleanup. The deterministic note gets stronger as more findings are cleared and adapts to focused vs broad diffs. The footer is Last reviewed commit + review attempts + the miu-cr release. Lifecycle state is storeless: it lives in a hidden <!-- miu-cr-ledger:<base64> --> marker inside the comment (like the runs counter), so it survives ephemeral CI with no DB. A finding resolves only when it is absent AND its file is still in the diff (absence off-diff ≠ fix).
• Inline comments persist; resolving the threads is left to you / your coding agent. Each finding's inline comment is posted ONCE and deduped across re-runs via a hidden <!-- miucr:fp=… --> marker, so a re-review never re-posts or deletes it. miucr does NOT click "Resolve conversation": when a finding is fixed, GitHub auto-marks the thread Outdated and the summary ledger moves it to ✅ Resolved, but the inline thread itself stays open for the developer/coding agent to resolve. A host config may opt into thread_resolution_sync.mode: poll to mirror manual GitHub "Resolve conversation" state into the summary (the Resolved row shows a clean 💬 conversation marker linked to the discussion thread, distinct from a commit-resolved row's open → resolved SHA arrow); this never starts an LLM review and never feeds approval. So an agent acting on a review should read the inline threads, apply the fixes, reply on each handled inline thread with what changed and why, then resolve the thread.
• Repeat-run stability is deterministic inputs + low-variance generation. For the same repo/ref/config, file selection, context assembly, rules, anchoring, gating, fingerprints, and comment dedupe are deterministic. SDK-backed Anthropic/OpenAI calls use temperature: 0; exact model output can still vary, so PR posting is idempotent rather than duplicate-prone.
• One-click suggestions and approvals are explicit write actions. --suggest emits a native GitHub ```suggestion block ONLY for findings at or above the suggestion floor (medium) when the patch *deterministically* replaces the exact anchored line(s) and the model is certain of a grounded mechanical fix (a cited rule or an obvious best practice). It NEVER guesses an unverifiable value (a URL, path, route, ID, version, config key, API signature); such concerns become a verification-question in the rationale instead. --patch-repair (requires --suggest) runs one focused 2nd LLM pass to recover a repairable near-miss patch. --approval clean|threshold submits APPROVE only when the policy and safety gates pass; the first approval body is short (LGTM) and links to the code review summary unless --approval-note none is set. Approval is head-SHA scoped: a later clean push can be approved again bodyless, threshold re-approvals carry the threshold note, and the same commit is not approved twice. Permission/self-approval failures warn and degrade rather than failing the review.
• -o pretty is the human-readable local format; -o json is for agents; -o sarif for editors/CI.
• Multi-provider profiles. Add a named provider (e.g. z.ai/glm) with kind, base_url, model, auth, and either auth_env or auth_command; select with --provider <name>. Built-in kinds: anthropic, openai (ChatGPT-plan OAuth via miucr login). Transient GitHub/network errors auto-retry with backoff. Optionally cap a provider instance's usage with [providers.<name>.quota] (dimension = tokens|requests, limit, window = <Go duration like 5h/24h>|monthly); uncapped by default, fail-closed, over-quota → typed quota.exceeded.
• Thinking on by default; deterministic fallback. Capable models (Claude, gpt-5/o-series, codex, z.ai GLM 4.5+) review with extended thinking/reasoning (deeper analysis; temperature is omitted because thinking forces temp 1). Models without thinking (gpt-4o, plain glm-4 chat) sample at temperature 0 for stable, reproducible findings. Both are config-exposed: [review].thinking (auto|off|low|medium|high, default auto) and [review].temperature (0-2, default 0). Set MIUCR_TRACE_REASONING=true to capture that reasoning into the review trace (miucr trace <id>).

Output contract: miucr.cli/v1 envelope (parse this)

Every command prints one JSON object on stdout (default -o json). Field order:

{
  "ok": true,
  "api_version": "miucr.cli/v1",
  "kind": "review.result",
  "command": "review",
  "request_id": "req_...",
  "summary": { "findings": 2, "gate": "high" },
  "data": { "...": "command-specific" },
  "artifacts": [],
  "warnings": []
}

• ok (bool) is the branch point. artifacts and warnings are always present ([] when empty).
• summary, data, page, stats are present only when relevant (omitempty).
• Errors use the same envelope with ok:false, kind:"error", and an error object:

{ "ok": false, "api_version": "miucr.cli/v1", "kind": "error", "command": "review",
  "error": { "code": "review.gate_failed", "message": "<redacted>", "hint": "...",
             "retryable": false, "safe_to_retry": false } }

kind per command: version, review.result, config.show, rules.check, rules.init, init.result, login.result, whoami, logout, history.list, history.record, history.prune, eval.result, trace.show, error (REST: review.accepted / review.result). Secrets never appear in the envelope, logs, or on disk (credential-named fields are scrubbed; finding rationale/suggested_patch prose is exempt). The --instruction/--conversation flags and the /miucr review <prompt> comment trigger only add input (a USER-turn context block); they never change the envelope or the finding JSON (still miucr.cli/v1); parse the same shape.

Exit codes (gate → exit mapping)

Severities low to high: info < low < medium < high < critical. A review.gate_failed error is emitted after the normal review.result envelope (the findings still print), then exit 2.

Typed error codes (branch on error.code)

The day-1 provider/auth/timeout failures classify into a stable taxonomy (same code across all backends: anthropic/openai/codex), each with an actionable hint and a correct retryable:

Unknown failures stay internal.error (never mislabeled as retryable). Classified messages are redacted: no token fragment ever appears.

The codex backend retries 429/502/503/504 (and a response.failed stream event) with bounded, jittered exponential backoff (≤3 attempts) like the SDK backends, honoring Retry-After/resets_in_seconds and aborting on cancel/timeout. A persistent rate limit returns provider.rate_limited with the usage-cap reset window in error.details.resets_in_seconds (or retry_after_seconds); branch on that to decide wait-vs-switch-provider.

Install

curl -fsSL https://cr.miu.sh/install.sh | sh   # asset-aware latest
curl -fsSL https://cr.miu.sh/install.sh | sh -s -- v0.65.0   # pin
brew install vanducng/tap/miucr                                                     # Homebrew
go install github.com/vanducng/miu-cr/cmd/miucr@latest                              # Go 1.25+

Verify: miucr version → {"ok":true,...,"data":{"version":"v0.65.0"}}. Config (optional) at ~/.config/miu/cr/config.toml; state DB at ~/.config/miu/cr/state.db. Repo rules at .miu/cr/rules/*.md (never a flat .miucr/).

Onboarding (miucr init)

miucr init is the fastest path to a working config. It walks a clean, sectioned wizard (provider → provider-aware auth → project rules), then writes ~/.config/miu/cr/config.toml (dir 0700, file 0600, deltas only: the chosen provider block, never the full built-in defaults) and ends on the literal miucr review --staged.

miucr init                                  # interactive wizard (idempotent: Overwrite? y/N)
miucr init --non-interactive --provider anthropic --auth-env ANTHROPIC_API_KEY --yes

• Provider-aware auth menu: openai offers browser login (OAuth, default) (review on your ChatGPT/Codex plan, no API key) plus env-var or paste; it runs the same PKCE loopback flow as miucr login and caches the token in oauth.json (config records just default_provider = "openai", no secret). anthropic offers env-var (default) or paste, no OAuth (Anthropic ToS). custom asks kind + base URL, then env-var or paste.
• Default writes no secret: only the env-var name (auth_env). A literal auth_token lands only on explicit paste + confirm (after a plaintext-on-disk warning).
• Flags: --provider anthropic|openai|custom, --auth oauth|env|paste (non-interactive selector), --auth-env <NAME>, --base-url <gateway>, --no-rules, --force, --yes, --non-interactive. --auth oauth is interactive-only (needs a browser); non-interactive errors init.aborted toward miucr login. Envelope kind: init.result (data.auth_method = oauth|env|paste, data.next = miucr review --staged); errors init.aborted / config.write_failed.
• init is optional: zero-config still works when a provider key is on the env. With no config and no key, review prints a soft one-line nudge to run init.

Examples (copy-paste starters)

The repo ships an examples/ tree: rules/{go-api,typescript-node,python-data}.md, github-action/code-review.yml (fork-safe pull_request_target), workflows/miucr-review.yml (the dual-trigger default: pull_request + a /miucr review <prompt> comment trigger), mcp-setup/{claude-code,cursor,codex} + README-mcp.md, review-local/{pre-commit,Makefile,agent-review.sh} (review-your-own-changes recipes), and review-host/{Dockerfile,docker-compose.yml,config.example.yaml} (pure-Go CGO_ENABLED=0 Alpine image + Postgres host stack for miucr serve). Onboarding walkthrough lives at the docs Getting started page.

Comment-triggered review (/miucr review <prompt>). With the dual-trigger workflow installed, a write or admin collaborator posts /miucr review <prompt> as a top-level PR comment (e.g. /miucr review focus on the auth changes) to steer a re-review with free text. The issue_comment event runs the trusted base-branch workflow with full secrets even on fork PRs, so the workflow self-gates: the commenter must have write|admin (an actions/github-script repos.getCollaboratorPermissionLevel check; author_association alone is insufficient), the comment body is read via an env var (never inline-interpolated into run:), the released binary is used (fork head code is never built with secrets), and miucr adds a 👀 reaction to acknowledge an accepted command. Known limits (v1): only top-level issue_comment is caught (not inline review comments or review summaries), and an unchanged head SHA short-circuits (there is no --force on the comment path yet).

Commands & exact flags

Global flags (all commands): -o, --output json|pretty|sarif (default json), --timeout <dur> (default 30s; review auto-bumps to 900s and eval to 30m unless --timeout is set explicitly). sarif is review-only: it emits a SARIF 2.1.0 document (NOT the envelope) for code-scanning/IDEs (ruleId=category, level from severity, repo-relative paths); upload it with github/codeql-action/upload-sarif. pretty is a local reporter (jumpable file:line, excerpt, patch; color on a TTY). review --pr also takes --filter-mode added|diff_context|file|nofilter (default diff_context) controlling which findings are inline-eligible; file/nofilter route off-diff findings to the summary/SARIF/local output, never inline.

review: needs exactly one mode

miucr review --staged                     # staged changes vs the index
miucr review --from main --to HEAD        # ref range (--from and --to required together)
miucr review --commit HEAD~1              # one commit vs its parent
miucr review --pr owner/repo#123          # a GitHub PR (dry-run by default)
miucr review --staged --instruction "focus on the auth changes"   # steer this one review
miucr review --pr owner/repo#123 --conversation                   # also read the prior PR thread

review.result data (local and --pr):

"data": {
  "findings": [
    { "file": "internal/foo/bar.go", "line": 42, "end_line": 42,
      "title": "…optional short scannable summary…",   // omitted when the model emits none
      "rule": "go",                                     // optional: stem of the project rule that motivated this finding (omitted when none)
      "severity": "high", "category": "bug",
      "rationale": "…why this is a problem (may cite a convention the model can see, e.g. \"differs from mapWriteError\")…",
      "suggested_patch": "…optional minimal fix…",
      "quoted_code": "…verbatim source the finding anchors to…" }
  ],
  "stats": { "files_changed": 3, "files_reviewed": 2, "findings_total": 2,
             "findings_dropped": 1, "max_severity": "high", "gate": "high",
             "truncation_level": "full",            // full | hunks_only | filenames_only
             "rules_applied": 5, "rules_truncated": false },
  "review_id": "rev_...",  // additive: the saved history record id ("" only with --no-save; on an incremental skip it is the prior review id)
  // additive, only on a --pr run when an unchanged head SHA short-circuited (no
  // LLM pass). Dry-runs skip from the local history store; --post can skip from
  // the completed-publish marker in the summary comment. Both fields are omitted
  // on a normal review. On the skip path findings is [] and stats is {} (never null):
  "skipped_unchanged": true, "prior_review_id": "rev_prior",
  "pr": {  // only on --pr
    "owner": "owner", "repo": "repo", "number": 123, "head_sha": "deadbeef",
    "is_fork": false, "posted": false, "posted_inline": 0,
    "summary_action": "none",         // fate of the ONE upserted summary issue comment: none | created | edited | fork_fallback | failed
    "approve_action": "commented", "approve_reason": "not_requested", "suggestions_posted": 0,
    "patches_repaired": 0,            // additive, omitempty: findings whose rejected SuggestedPatch the --patch-repair 2nd pass recovered into a now-clean one-click suggestion (0/absent when --patch-repair is OFF)
    // additive, omitted when empty:
    "mode": "review",                 // review (default) | checks
    "check_run_id": 0, "check_conclusion": "",  // --mode checks only (success|failure)
    "fallback_annotations": 0 }       // >0 when a fork-PR 403 under Actions fell back to ::error:: workflow annotations (review did NOT hard-fail); summary_action then "fork_fallback"
}

findings_dropped = findings rejected by line-anchor drift (their quote no longer matches the reviewed revision, kills position drift). --post keeps the summary and the inline findings in separate homes: inline comments post as a PR review (body left empty, never a 422 on a no-inline run), and the summary is ONE issue comment that is UPSERTED: miu-cr lists the PR's issue comments, finds the one carrying the <!-- miu-cr-review --> marker, and edits it in place; if none exists it creates it. So a re-run updates the single summary instead of stacking a review per commit. summary_action is created (new summary issue comment), edited (upserted in place), fork_fallback (a fork PR lacked comment-write scope, degraded, no hard fail), failed (summary upsert failed after inline posting), or none (--no-post, --mode checks, or a skipped summary path). A same-commit --post re-run short-circuits after a matching completed-publish marker; pass --force to review anyway. per-comment <!-- miucr:fp=... --> line-free fingerprints prevent inline dupes across commits. The summary body itself carries a finding lifecycle ledger - Open + Resolved tables with per-finding status/origin+resolved commit/priority-before→after/timestamps - persisted storelessly in a hidden <!-- miu-cr-ledger:<base64> --> marker so the open/resolved history survives across pushes (and ephemeral CI) without a DB. The miucr.cli/v1 JSON envelope is unchanged: the ledger is a rendering/comment concern, not an envelope field. A public-PR dry-run needs no GitHub PAT (LLM key still required); --post and private repos need a PAT with repo scope.

serve: webhook daemon (default) + opt-in poll

WEBHOOK_SECRET=… GITHUB_TOKEN=… ANTHROPIC_API_KEY=… \
  miucr serve --addr :8080 --repos owner/repo,owner/other --gate high

Env: WEBHOOK_SECRET (required unless poll-only), GITHUB_TOKEN/GH_TOKEN (required unless [github] mode=app), ANTHROPIC_API_KEY (or compatible). MIUCR_LOG_LEVEL=debug enables progress/tool-turn logs; MIUCR_TRACE_LOG=true adds bounded debug trace payloads (MIUCR_TRACE_LOG_MAX_BYTES, default 4096) that are redacted/truncated but may include prompt/diff context. MIUCR_TRACE_REASONING=true captures the model's reasoning as a reasoning trace step (Claude/GLM thinking verbatim; OpenAI a token count + [hidden by provider]) - off by default, redacted in storage, needs [review].thinking on. Endpoints: POST /webhook (HMAC), GET /healthz. Each new head SHA = one full LLM review; allowlist + per-head dedup are the only spend guards. serve inherits --suggest OFF and review.approval.mode: off.

In serve --host YAML, thread_resolution_sync is an object and defaults off. Enable it per repo when manual GitHub conversation resolution should update the summary table between commits:

review:
  thread_resolution_sync:
    mode: poll
    interval: 5m

Opt-in REST API: set MIUCR_API_TOKEN (env-only, no flag) to register /v1:

MIUCR_API_TOKEN=$(openssl rand -hex 32) WEBHOOK_SECRET=… GITHUB_TOKEN=… ANTHROPIC_API_KEY=… \
  miucr serve --addr :8080 --repos owner/repo
# Queue (202 + server-generated crypto/rand id):
curl -sS -X POST https://host/v1/reviews -H "Authorization: Bearer $MIUCR_API_TOKEN" \
  -H 'Content-Type: application/json' -d '{"owner":"acme","repo":"widgets","number":42}'
# Read back (whitelist: id,status,created_at,findings,stats, never the clone path):
curl -sS https://host/v1/reviews/<id> -H "Authorization: Bearer $MIUCR_API_TOKEN"

Status lifecycle: pending → done/failed. HTTP map: 400 bad body, 401 missing/wrong bearer (empty token can never auth), 403 off-allowlist, 404 unknown id, 405 wrong method, 413 body > 64 KB. Single-operator: one shared bearer = one trust boundary (not multi-tenant).

GitHub App auth (opt-in alternative to PAT): [github] mode=app in config (see below).

Host mode: miucr serve --host loads YAML from MIUCR_CONFIG or ~/.config/miu/cr/host.yaml, requires Postgres (store.backend: postgres; MIUCR_PG_DSN preferred), and watches multiple repos from one daemon. Validate without opening DB/secrets:

MIUCR_CONFIG=examples/review-host/config.example.yaml miucr serve --host --dry-run-config -o json

Dry-run emits envelope kind serve.host_config with a redacted config plus summary counts (repos, accounts). Host YAML uses providers, store, github.accounts, top-level review/agent, host.review, and repos[]; layering is defaults -> review/agent -> host.review -> repo overrides. PAT accounts support auth_env/auth_file/auth_command; App accounts support App/installation ids from literal/env plus private keys from path/env/command. Provider credentials also support auth_env and auth_command.

Repo prompts override the global prompt (system_prompt or system_prompt_file, mutually exclusive). repos[].rules can point to markdown files or a non-recursive directory of *.md; these are trusted host context and cannot change the protected finding schema. Host review write policy lives at the effective repo level: post, force, suggest, patch_repair, thread_resolution_sync, and approval. First host mode does not push code.

Host poller state lives in Postgres: repos, PR sessions, queued jobs, attempts, workspaces, poll cursors, and retention metadata. Startup applies versioned schema migrations under an advisory lock. Claims use row locks for concurrency, and host.retention prunes old jobs/attempts/sessions/workspace records/cursors. The review-host compose example defaults MIUCR_LOG_LEVEL=debug for local dogfood while leaving MIUCR_TRACE_LOG=false.

review.pr_filter layers top-level -> host.review -> repos[].review. Draft PRs are skipped unless include_drafts: true. default_action defaults to include; set default_action: exclude for allowlist repos. Rules are evaluated in order and the last matching rule wins. Matchers inside one rule are ANDed. Supported matchers: authors, author_types (Bot, User, Organization), author_associations, title_regexes, labels, requested_reviewers, base_branches, and head_branches. Use title regexes for generated PRs that may be opened by a human token, such as ^chore\(deps\):. comment_trigger_regexes are used by the GitHub Action comment-trigger workflow, not by serve --host poll_source: pulls.

review:
  pr_filter:
    default_action: include
    include_drafts: false
    comment_trigger_regexes:
      - '(^|\s)(/miucr review\b|@vanducng\b)'
    rules:
      - action: exclude
        title_regexes: ['^chore\(deps\):']

repos:
  - name: critical-dbt
    slug: example-org/critical-dbt
    review:
      pr_filter:
        default_action: exclude
        rules:
          - action: include
            authors: ["vanducng"]
          - action: include
            requested_reviewers: ["vanducng"]

rules: project review context

miucr rules init             # scaffold annotated .miu/cr/rules/example.md
miucr rules init --force     # overwrite
miucr rules check internal/foo/bar.go   # which loaded rules apply to a path (kind: rules.check)

Rule files are markdown with YAML frontmatter, then prose injected as context only (never gate):

---
description: Project-specific review context for changes under cmd/.
globs:
  - "cmd/**/*.go"
  - "internal/**/*.go"
alwaysApply: false
context_files:
  - "AGENTS.md"
---
# Prose below the fence is injected as CONTEXT for the reviewer.

Three layers merged by file stem (later overrides earlier): built-in defaults (Trusted, lowest) → user ~/.config/miu/cr/rules/*.md (Trusted) → repo .miu/cr/rules/*.md (Untrusted, highest). A file with no --- fence is skipped (never always-applied). Untrusted repo rules are fenced context-only, byte-capped, and dropped on fork PRs; the finding-JSON schema stays in the cached system prompt so injected prose can't redefine it. rules check data lists each applicable rule with provenance, stem, globs/always_apply, trusted, plus skipped body_only files.

Rule grounding. A finding may carry the rule stem of the project rule that motivated it. The wire layer validates the stem against the rules actually loaded this review (a hallucinated stem is dropped) and renders it as (per <stem>) on the inline comment + summary overflow. A repo rule (.miu/cr/rules/*.md) additionally links to its file, repo-relative at the head SHA; user and built-in rules are cited as text only (no link, a user-rule home path never leaks).

history: browse saved reviews

Every review auto-saves a full record (findings + stats + per-turn transcript + raw prompt/response) to the local store; --no-save opts out per run. Records are local only (~/.config/miu/cr/state.db, gitignored); no tokens are ever stored.

miucr history                                 # list recent reviews, newest first (kind: history.list)
miucr history --repo owner/repo               # filter by repo (PR) or repo dir (local)
miucr history --pr owner/repo#7               # filter to one PR
miucr history --since 7d                       # 7d / 24h / 2026-06-01
miucr history --limit 50                        # cap rows (default 20; 0 = no limit)
miucr history show <id>                         # one full record (kind: history.record; 404 → history.not_found)
miucr history show <id> -o pretty --raw         # pretty, with raw prompt/response inline
miucr history prune --keep 200 --yes            # keep newest N (kind: history.prune; reports deleted count)
miucr history prune --older-than 30d --yes      # delete records older than a span

prune needs at least one of --keep/--older-than plus --yes (destructive). An optional [history] max_records = N auto-prunes oldest on save. List rows carry {id, created_at, target, mode, findings, max_severity, status}; show data adds provider, model, head_sha, findings, stats, transcript, raw_prompt, raw_response. Errors: history.unavailable, history.not_found, history.prune_policy_required, history.prune_confirm_required, history.bad_pr, history.bad_time.

eval: compare reviewer quality against expected findings

miucr eval runs one or more reviewer commands over a JSON suite and scores them by file+line overlap against expected findings. Use synthetic/public fixtures only. For unlabeled public-PR smoke suites, read unmatched findings as review output to ins

…(truncated)