pi-delegate setup

This is a guided walkthrough for the person setting up pi-delegate, not a script to run unattended. Work through the steps below in order. Report each check's result in plain language before moving on. Wherever a step says ask, stop and wait for the user's answer — do not barrel through to the next step on your own judgment.

1. Detect

Check whether pi is on PATH:

command -v pi

If this prints nothing, pi is not installed and nothing else in this plugin can work yet. Tell the user:

• Install it with npm install -g @earendil-works/pi-coding-agent (https://www.npmjs.com/package/@earendil-works/pi-coding-agent)
• pi-delegate also needs Node ≥ 22.

Then stop here — do not continue to step 2 until command -v pi finds it.

If pi is found, check its version and run the environment check:

pi --version
node "${CLAUDE_PLUGIN_ROOT}/bin/pi-doctor" --check

Read the effective block in the JSON output — that's who a dispatch will actually reach. Report it to the user plainly: which provider and model dispatches will use, and where that came from (source is either pi-delegate config.json or pi settings.json).

• This plugin emits no --provider / --model flags by default. A dispatch simply inherits whatever pi itself is already pointed at — anthropic, openai, a local OpenAI-compatible server. There is no requirement to run a local inference server; whatever the user already uses with pi interactively is what dispatches will use too.
• If effective.provider and effective.model are both empty, pi has no defaultProvider / defaultModel set in ~/.pi/agent/settings.json, and pi-delegate has no ~/.claude/pi-delegate/config.json pinning one either. In that case dispatches will land on whatever pi's own resolver picks (the first model with a usable API key) — tell the user this is usually not what they want, and that configuring a provider for pi itself is outside this plugin's scope: run pi interactively once and complete its own provider setup, or set defaultProvider / defaultModel directly in ~/.pi/agent/settings.json. Then re-run the check above before continuing.
• Note the problems array (if any) but don't act on it yet — that's step 3.

2. Discipline mode — ask which one

Explain the three modes, then ask which one the user wants for this project before doing anything:

• off — pi-delegate's tools are available, nothing is enforced. Right for a project where delegating to pi isn't wanted at all.
• soft (the default) — you get a reminder when you edit existing product source by hand, but nothing is blocked.
• strict — a PreToolUse hook actually denies those edits and tells you to delegate instead. Worth explaining why this exists: with a written rule in place telling the model to delegate, roughly 80% of the characters that got committed were still typed by the main model by hand. strict is what made that stop. Be honest about its limit too: the hook only matches the Write and Edit tools, so the same edit made through Bash (sed -i, a heredoc, python - <<EOF, …) is never intercepted — it's a discipline rail against your own habit, not a security boundary.

The mode is stored against the project root and applied by the edited file's project root, so it follows the project rather than your session's working directory.

If the user picks strict, do NOT just run the command — strict needs to know what to protect, and its built-in fallback only fits a src/ directory of TS/JS/Svelte/Python files. Follow /pi-delegate:mode's "Setting strict: survey the project first" section: survey the layout, propose protect / allow globs with reasons, wait for the user to approve them, and only then write the policy.

"Existing product code" means a file that already exists under the project's src/; tasks/, scripts/, docs/, markdown, config files, and brand-new files are always allowed through regardless of mode.

Once the user answers, apply it (this also confirms the choice was actually recorded, even if they picked the default):

node "${CLAUDE_PLUGIN_ROOT}/bin/pi-mode" <mode>

Replace <mode> with off, soft, or strict. This is exactly what /pi-delegate:mode <mode> does, and it can be changed later the same way.

3. Offer to fix what is fixable

Look at the problems array from step 1's --check output.

• If it's empty, say so and move on to step 4.
• Otherwise, explain each problem's real-world consequence in one line before asking anything:
  • reasoning-missing / compat-missing — this only shows up for a local OpenAI-compatible server (e.g. a self-hosted omlx, LM Studio, llama.cpp, or vLLM endpoint). The model is missing reasoning: true or the compat.chatTemplateKwargs.enable_thinking binding in ~/.pi/agent/models.json, so --thinking off silently has no effect — the model keeps "thinking" instead of calling tools, and a dispatch can burn its whole timeout without ever writing anything. This class can be fixed automatically.
  • drafter-selected — the model a dispatch would actually reach looks like a speculative-decoding draft/assistant model (a co-pilot, not a target). Calling it directly returns HTTP 500. This is never auto-fixed — which model to point at is the user's decision, not something to guess. Point them at switching defaultModel (or pi-delegate's own model override), or adjusting drafter_patterns in ~/.claude/pi-delegate/config.json if it's a false positive.

For any reasoning-missing / compat-missing problems, ask before fixing. If the user agrees, run:

node "${CLAUDE_PLUGIN_ROOT}/bin/pi-doctor" --fix

This backs up ~/.pi/agent/models.json to ~/.pi/agent/models.json.pi-delegate.bak before writing, and only ever adds the thinking binding to a model that is already registered and confirmed to be a local chat-template endpoint — it never invents a provider or a model. Report the result (fixed / remaining in the output) back to the user.

Never run --fix without asking first, and never fix a drafter-selected problem — there is nothing to fix, only a model choice for the user to make.

4. Offer a verification dispatch

Ask first — a real dispatch calls the user's actual provider, which costs real time and (for a paid provider) real tokens. If they'd rather skip it, go straight to step 5.

If they want to try it, create a scratch task in a temp directory:

DIR=$(mktemp -d) && cat > "$DIR/TASK.md" <<'EOF'
Create a file named hello.txt in the current directory containing exactly the text:
pi-delegate setup verified
Do not create, read, or modify any other file.
EOF
echo "$DIR"

Then call the pi_dispatch tool yourself with task_file set to <DIR>/TASK.md, cwd set to <DIR>, and mode=sync — leave provider / model unset so it exercises the exact same resolution path a real dispatch would use. sync is deliberate here even though async is the normal default: this one call exists so the user watches it finish, and there is nothing else to get on with while it runs. Report the verdict plainly: status, files written, duration. To show the user it's real, not just claimed:

cat "$DIR/hello.txt"

If the dispatch times out or never writes, that is itself informative — it's usually the reasoning-missing symptom from step 3 (a local model stuck "thinking" instead of acting). Point back at step 3's fix rather than guessing at a new cause.

5. Offer the status line indicator

Ask first. This one writes to ~/.claude/settings.json, which is the user's own global config and probably already has a status line in it.

Explain what it is: one row per running pi dispatch at the bottom of the screen, each showing how long that dispatch has been going and which model it is on. It disappears entirely when nothing is running. The rows are this session's own; another Claude Code window's dispatches never appear in them.

Two things to say before they answer, because neither is visible from a screenshot:

• Claude Code allows exactly one statusLine, and a plugin cannot ship its own. If they already have one (claude-powerline, ccstatusline, a personal script), this composes with it rather than replacing it — their line runs untouched and the pi row goes underneath.
• It needs refreshInterval, which reruns their status line too, once a second for as long as Claude Code is open. On a status line that shells out to git, that is a real if small background cost — roughly 10% of one core for a status line taking ~100ms.

If they want it, invoke /pi-delegate:statusline and follow it — it probes their existing command by actually running it, shows before and after, and backs up settings.json before writing. Do not shortcut it from here — that skill assumes nothing about a setup it has not looked at.

If they decline, say it can be added any time with /pi-delegate:statusline, and move on.

6. Close with what to do next

Tell the user setup is done, and summarize where things stand: the mode they picked, and what a dispatch will resolve to. Then point at:

• The seven MCP tools this plugin adds: pi_dispatch, pi_status, pi_steer, pi_abort, pi_result, pi_transcript, pi_stats — see skills/delegating-to-pi/SKILL.md (loaded automatically when relevant) for how and when to use each one.
• /pi-delegate:mode <mode> — change the discipline mode later, any time.
• /pi-delegate:probe — get a one-time bypass in strict mode to make a single hand-edit (a probe), before writing the verified recipe into a task book for pi.
• /pi-delegate:statusline — add (or later remove) the live dispatch indicator in the status line.
• /pi-delegate:doctor — re-run this environment check any time something seems off (after a pi upgrade, after editing ~/.pi/agent/models.json by hand, after switching providers).