Plannotator Reference

Plannotator CLI Reference

Plannotator is a local, browser-based review layer for agent workflows: it opens plans, diffs, and documents in an annotation UI, the human marks them up, and the structured feedback comes back to you on stdout. It installs as a single plannotator binary plus per-host hooks, so plan review fires automatically when you exit plan mode; every other surface is launched explicitly, with the plannotator tool when you have one and with the CLI otherwise. A session runs on a random localhost port (fixed port 19432 in remote mode) and blocks until the reviewer submits feedback, approves, or closes the tab.

This skill is the knowledge layer. The plannotator-review, plannotator-annotate, and plannotator-last skills are thin launchers for the three most common actions; use this reference when you need to pick the right command or flags yourself.

If you have a plannotator tool, always use it

Check your tools before you run any command below. If you have a tool named plannotator (it can be listed with a prefix, for example mcp__plannotator__plannotator in Claude Code, and can need to be loaded through tool search before you call it), always call it instead of running plannotator annotate, plannotator review, or plannotator last. This includes approvals: call it with { "action": "annotate", "target": "<file>", "gate": true }, not --gate --json.

The tool returns at once. End your turn after you call it and wait: the reviewer's decision arrives later as a message. Do not also run the CLI, poll, or reopen the session.

Use the CLI only when you have no such tool, or for what the tool does not do: archive, guide, sessions, review flags other than --base, annotate flags other than --gate and --markdown, and strict gates that a script checks by exit code (--require-approval, --result-file).

Choose the command

Session model

Every review or annotate command starts a local web server, opens the browser, and blocks until the human decides. That can take minutes, or more than an hour for a large pull request. Launch it with a long (or no) command timeout, or in the background, then read stdout when the process exits. In Claude Code, a background command is stopped after 30 minutes unless you pass run_in_background with a longer timeout (up to 7200000 ms). Do not kill the process to "finish" a review; a session that ends without a decision reads as no feedback. If a session was stopped by a time limit, run the same command again: annotation drafts are restored.

Stdout is the interface, but its contract is command-specific. For annotate and its last-message variants:

• Plaintext (default): empty output on close, The user approved. on approve (an approval with a note prints an "Approved with Notes" message instead; treat the notes as guidance, not a change request), otherwise the feedback text. Address returned feedback in the same conversation.
• --json: one JSON record with decision (approved, dismissed, or annotated) and optional raw feedback. An approval may still carry notes in feedback; treat those as guidance, not a change request.
• --hook: hook-native output for real PostToolUse/Stop hook contexts only. Approve/close emits nothing (hook passes); annotations emit {"decision":"block","reason":"..."}. --hook implies the gate UI. Never use it for a normal interactive invocation.

plannotator <command> --help prints usage without launching anything. Bare plannotator is the hook entry point and expects hook JSON on stdin.

plannotator review

plannotator review [--git | --gitbutler] [--base <ref>] [--diff-type <type>] [--local | --no-local] [--patch-file <path | ->] [--no-git-remote-check] [--tailscale] [--json] [DIRECTORY | PR_URL]

Reviews local VCS changes, or a pull request when a URL is given. Default stdout stays plaintext: the existing close message, approval prompt, or feedback.

With --json, direct review emits one record: { decision: 'approved' | 'annotated' | 'dismissed', message: string }. message is the CLI-rendered text exactly as default plaintext would print it, without the final console newline. It includes customized prompts and non-blocking approval-with-notes framing; a denial suffix is included only when annotations.length > 0, including in PR mode, not for zero-annotation platform status.

Classify the outcome only by decision, never by message text. Notes on an approved review are guidance, not a blocking change request. This rendered message contract is separate from the raw feedback JSON used by annotate and the unchanged opencode-review integration. --hook is annotate-only.

• VCS is auto-detected (JJ, GitButler, Git, and P4 where supported). --git forces plain Git; --gitbutler forces GitButler (requires the but CLI 0.21.0+). Running from a non-VCS parent folder that contains nested repos produces a combined workspace diff.
• The default diff is "everything a PR would show now": merge-base of the trunk vs the working tree plus untracked files. --base <ref> opens the session against a different compare target (branch, origin/<branch>, tag, or commit) and --diff-type <type> opens it in a different mode (since-base, merge-base, branch, uncommitted, staged, unstaged, last-commit, local-vs-remote, all). Both are session-only: the reviewer can change either in the UI, and neither writes the user's saved defaults.
• Reviewing one layer of a stacked branch? Pass --base <the branch below yours> — plannotator review --base feature/part-1 shows only what this layer adds, instead of everything since main.
• Both flags are git-only: they error on jj, GitButler, Perforce, multi-repo workspace reviews, and PR URLs (a PR's base comes from the pull request). A --base ref that does not resolve is a startup error naming near-match branches, never a silently wrong diff.
• --patch-file <path> reviews a static caller-supplied unified diff with no repository at all (use - to read it from stdin): the session serves the patch as-is with no file-system affordances that need a worktree. It cannot be combined with a PR/MR URL, --base, --diff-type, --git/--gitbutler, or --local/--no-local. Every working-tree affordance is off in that session (staging, hunk-context expansion, open-in-editor, code navigation, diff-type/base switching), and the endpoints behind them answer 400.
• Pass a directory to review another repo or worktree: plannotator review ../feature-worktree or plannotator review ./backend --diff-type last-commit. Paths resolve relative to the invoking directory; quote paths containing spaces. A directory selects the review workspace, not a file filter. Accepts one directory or PR URL; a directory cannot be combined with --patch-file. Invalid targets fail instead of falling back to the current repo.
• PR review (plannotator review https://github.com/owner/repo/pull/123, GitLab MR and Bitbucket Cloud PR URLs too) needs an authenticated gh or glab CLI; Bitbucket Cloud (https://bitbucket.org/<workspace>/<repo>/pull-requests/<id>) instead needs an Atlassian API token in PLANNOTATOR_BITBUCKET_TOKEN (plus PLANNOTATOR_BITBUCKET_EMAIL). --local (the default) builds a local checkout of the PR head in the background for full file access; --no-local skips it and reviews the platform diff only.
• --no-git-remote-check stops the session contacting the git remote at all: no git ls-remote for the default branch or the "behind GitHub" baseline check. The compare target then comes from local refs only and the staleness banner never shows, which also takes away its one-click Fetch button (the /api/fetch-base endpoint stays available); fetching from a terminal is unaffected. Use it when a remote probe is expensive or intrusive — most sharply when SSH authentication is backed by a hardware token, where each probe is a physical touch prompt. The same opt-out is available session-wide as PLANNOTATOR_GIT_REMOTE_CHECK=0 or { "gitRemoteCheck": false } in ~/.plannotator/config.json (the flag beats the env var, which beats the config key). Without it the remote is queried when the review opens, on diff load, on a diff-type/base switch, and on Fetch — never on a timer.
• --tailscale publishes the loopback session over the user's tailnet via tailscale serve (HTTPS, never public) and prints the URL with a QR code. A publish failure exits nonzero instead of leaving the server hanging.

plannotator annotate

plannotator annotate <target> [--markdown] [--no-jina] [--app | --static] [--render-html] [--tailscale] [--gate] [--json] [--hook]

Opens one document, page, or app in the annotation UI and returns the human's annotations on stdout.

Plain annotate is feedback-only: it shows Close but no Approve button. When the user asks to review, approve, accept, or gate a generated plan/spec/document saved as a file, call the plannotator tool with "gate": true if you have it; otherwise always add --gate --json. Do not tell the user they can approve a plain annotate session. If the plan is being handed off through the host agent's native plan flow, do not launch annotate; let the plan-exit hook open the approval UI automatically.

Targets:

• Markdown and text files: .md, .mdx, .txt.
• Plain-text config and data files, rendered as text: .yaml, .yml, .json, .jsonc, .json5, .toml, .ini, .cfg, .conf, .properties, .csv, .tsv, .log, .xml, .env.example. .env itself is deliberately refused (it commonly holds secrets, and annotate history copies file contents). Source-code files belong to plannotator review, not annotate.
• Diagram sources, opened in the full diagram viewer (zoom, pan, popout, click a node/edge/cluster to comment): .mmd, .mermaid (Mermaid) and .dot, .gv (Graphviz). The file is the whole diagram — no fence needed — and comments carry the part's id plus its real file line.
• HTML files (.html, .htm): rendered as the raw page by default; --markdown converts to markdown instead. --render-html is accepted for compatibility; raw rendering is already the default.
• URLs (https://...): fetched and converted via Jina Reader by default; --no-jina uses plain fetch plus Turndown instead.
• Running local apps: a loopback http://localhost:PORT/ URL whose probe returns HTML opens in live-app mode (annotate the real running page). --app forces live mode and fails loudly when it cannot apply; --static forces the classic conversion pipeline. Non-loopback URLs always use the conversion pipeline.
• Folders: plannotator annotate docs/ opens a file browser over the folder's supported files.
• Several files: plannotator annotate spec.md mock.html notes.md opens them as one review, in the order given, with one decision. Every argument must be an existing file named by its path (no prose, URLs or folders among them). With the plannotator tool, pass the list as target.

Single files are capped at 2MB. Files are read from disk at stable project paths; keep the reviewed source where it lives.

Argument tolerance: extra words are fine (plannotator annotate look at notes.md please opens notes.md), but two resolvable targets mixed with prose, a URL or a folder is an error naming them, and an unrecognized dashed token disables the tolerance so flag typos fail loudly. When nothing resolves in a plain multi-word invocation, the CLI prints an agent-addressed handoff on stdout and exits 0: read it, work out the concrete target, and re-run with that exact path or URL.