--- name: tuicr description: Use tuicr's review CLI to read and add comments in active TUI review sessions, and launch tuicr in a Zellij split pane when the user needs an interactive review. --- # tuicr Review Workflow Use `tuicr review` as the default agent interface. The TUI is where the human reviews code; the CLI is how the agent discovers active sessions, reads user comments, and, only when appropriate, adds agent-authored comments. ## Core Rule First decide which workflow the user is asking for: 1. **User-led review of agent-generated changes** - The user wants to inspect the patch and write comments in tuicr. - Your job is to open or find the session, then retrieve the user's comments with `tuicr review comments` when they say comments are ready. If you are explicitly waiting while the user reviews, poll the same command periodically and look for new comment IDs. - Do not add your own review comments, do not preemptively review your own patch, and do not impersonate the user's comments. 2. **Agent review of an AI-generated patch** - The user wants you to understand, critique, or summarize a patch. - You may inspect the patch and propose findings. - If you can confidently identify this workflow and the target session, add findings directly with `tuicr review add` and an explicit `--username` identifying the agent. Ask first when the workflow or session is ambiguous. If the user's intent is ambiguous, ask which workflow they want. ## Attach To A Session 1. Determine the repository directory from the user's request, current working directory, or recent file operations. Ask if it is ambiguous. 2. List persisted sessions: ```bash tuicr review list --repo /path/to/repo # checkout + its repo's PR sessions tuicr review list --repo owner/repo # all sessions for a forge repo tuicr review list --all # every session across all repos ``` `--repo` is a selector: a checkout path also surfaces PR sessions for that checkout's `origin` repo, and a forge coordinate like `owner/repo` matches local and PR sessions by owner/repo. Each row carries a `kind` (`local` or `pr`) and a usable `slug`. Use `--all` when you don't know the repo. `[]` with exit 0 also means "not a repo root" — a subdirectory returns it too. Pass the root, then `--all`, before concluding nothing is open. 3. Choose the session: - If the CLI clearly reports exactly one relevant active session with `"active": true`, attach to it. - If multiple sessions are active, or the correct session is not clear, ask the user which slug to use. One repo can hold a worktree and a commit-range session at once, and adding to the wrong one exits 0. - If the user provided a slug or session JSON path, use it directly. - For a PR review, pass the PR slug from the listing (e.g. `gh:owner/repo/pr/N`) to `--session`; it is self-contained and needs no `--repo`. - If there is no active session, start or wait for one as described below. - Until active-session discovery is formalized as a stable protocol, treat `"active": true` as a convenience signal. If slug resolution fails, ask the user for the slug or repo path used by the session. The CLI works even if the agent is not running inside Zellij, so do not require a multiplexer just to connect to an existing active session. ## Start A Session When the user needs an interactive tuicr pane and no active session exists: | Environment | Action | | ---------------- | ------------------------------------------------------------------------------------------------------------------------------------ | | `$ZELLIJ` is set | Run `tuicr-wrapper-zellij /path/to/repo -- ` | | None is set | Tell the user you are waiting for them to start `tuicr` in the repo, then attach with `tuicr review list` after they say it is ready | `` is `-w` for uncommitted working-tree changes or `-r ` for a commit range — always pass one explicitly so the user is never left to pick staged/unstaged/commit-range manually in the TUI. tuicr supports both git and Jujutsu (jj) repositories, and jj workspaces may have no `.git` directory at all. Do not pre-check the directory with `git rev-parse` or refuse to launch because git does not recognize it; always run the wrapper and let it validate the repository. If your tool supports command timeouts, use a long timeout, such as 10 minutes, because the Zellij wrapper waits for the TUI to exit. ## Reconstruct The Diff To review a patch yourself, rebuild the diff the user sees. The slug's source segment says which: | Slug segment | Diff | | ----------------------------------------------- | -------------------------------------- | | `worktree/`, `staged-and-unstaged/` | `git diff HEAD` | | `staged/` | `git diff --cached` | | `unstaged/` | `git diff` | | `commits/..` | `git diff ~1..` | | `pr/` | `gh pr diff ` | | `pristine` | none; every tracked file shown in full | Range endpoints are inclusive and printed oldest-first, so `` without `~1` drops the first commit. Check the file count against the listing row's `file_count`; a mismatch means every line number you derive will be wrong. ## Read User Comments This is the main review loop for user-led review. There is no push stream from tuicr to the agent. Read comments by running the CLI on demand. After the user says comments are ready, or after the TUI exits, run: ```bash tuicr review comments --repo /path/to/repo --session ``` The command emits JSON. Each comment includes fields like: - `id` - `location` - `path` - `start_line` - `end_line` - `side` - `comment_type` - `author` - `lifecycle_state` - `content` Treat these comments as the user's review feedback: - `issue`: blocking problem to fix first - `suggestion`: consider implementing or explain why not - `note`: answer or acknowledge - `praise`: no action required If you are waiting during an active review, poll this command about every 30 seconds and compare comment IDs with the previous result. Read immediately when the user says comments are ready. Stop polling once the user says the review is done or your tooling would block other work. An empty result does not by itself mean the review didn't happen. On exit, tuicr always prints a line like `tuicr-summary: reviewed 3/3 files, 0 comments added` to stderr (visible in the pane's scrollback), and `tuicr review list` reports the same `reviewed_count`/`file_count` for the session. If `reviewed_count` equals `file_count`, zero comments is a legitimate "nothing to flag" outcome — treat the review as complete, don't ask the user to confirm. Only ask whether the user saved comments in the intended session, or whether another active session should be selected, when `reviewed_count` is less than `file_count` (the user quit before reviewing everything) or you can't find a `tuicr-summary:` line at all. If the review may have continued while you were working, rerun `tuicr review comments` before claiming completion. ## Add Agent Comments Only add comments when the workflow allows it and, for agent-authored review, after the user approves writing them into tuicr. Defaults: - Prefer line comments when a specific file and line are known. - Use file comments for file-scoped feedback. - Use review-level comments only for whole-review summaries. - Use `--type issue` for problems by default. - Use `suggestion`, `note`, or `praise` when that better matches the intent. - Pass `--username` so agent comments are visually distinguishable. Examples: ```bash tuicr review add --repo /path/to/repo --session \ --target-file src/main.rs \ --line 42 \ --side new \ --type issue \ --username "LongCat" \ "Handle the empty case here." ``` ```bash tuicr review add --repo /path/to/repo --session \ --target-file src/main.rs \ --type suggestion \ --username "LongCat" \ "Consider splitting this file-level concern into a helper." ``` Omit `--target-file` for a review-level comment. Add `--end-line` for a range comment. Use `--side old` for removed lines and `--side new` for added or unchanged lines in the new file. For structured input, use `--input` with literal JSON, `@path/to/file.json`, or `-` for stdin. Supported target types are `review`, `file`, `line`, and `line_range`. One object per call — an array is a parse error. The file key is `file`, not `path`. `target.type` is inferred from the fields present: ```bash tuicr review add --session --username "LongCat" --input \ '{"file":"src/main.rs","line":42,"side":"new","comment_type":"issue","content":"Handle the empty case."}' ``` Then verify. A line outside the diff stores, prints back, and exits 0, but never renders — invisible to the user, successful-looking to you. Re-read `tuicr review comments` and check each `start_line` exists on the side you gave (`new` for added or unchanged, `old` for removed). Check `author` to distinguish your comments from the user's, and keep the returned `id`s to identify the exact comments in later reads. ## Legacy Export Output Older wrapper-driven flows may emit: ```text === TUICR INSTRUCTIONS === ... === END TUICR INSTRUCTIONS === ``` If present, process those instructions. Otherwise prefer `tuicr review comments`; it is the primary source of review feedback. If the wrapper mentions clipboard export, ask the user to paste it only when the CLI comments are unavailable. ## Zellij Tips - Switch panes: `Alt` + arrow keys - Close tuicr: press `q` - Resize panes: `Ctrl-n`, then arrow keys - Toggle fullscreen: `Alt-f` - Cycle stacked panes: `Alt` + `[` / `]` ## Error Handling | Situation | Action | | -------------------------------------------------------- | ----------------------------------------------------------------- | | Multiple plausible active sessions | Ask which session slug to use | | No active session, Zellij available | Start a new tuicr pane with the zellij wrapper | | No active session, no Zellij | Tell the user you are waiting for them to start `tuicr` | | `tuicr` not installed | Tell the user to install tuicr | | Not a repository | Ask for the correct repo directory | | Comments are empty, but `reviewed_count` == `file_count` | Treat as a completed review with nothing to flag — don't ask | | Comments are empty and `reviewed_count` < `file_count` | Confirm the selected session or ask the user to save/add comments | ## When Not To Use - The user only wants raw `git diff` output. - The user explicitly asks for a non-tuicr review workflow. - The task is remote PR review and no tuicr PR session is involved.