pi-playbook

Drive the pi CLI as a benchmark target or dispatched worker. Use when running pi probes, comparing pi postures, launching pi-zero, or handing work to a pi agent. Covers the flag surface that maps onto

SKILL.md

GitHub

pi Playbook

pi is a coding agent CLI with a flag surface that maps unusually cleanly onto Skill Heaven's mechanisms — it can evict and readmit skills natively, so postures compose without any config-directory tricks.

Verified against pi 0.83.0. Re-check pi --version before trusting any dose measurement; a flag's behaviour has changed across versions before (see §5).

Rule 0 — run pi in a herdr pane, never through your Bash tool. The operator must be able to see which model you invoked. A probe the operator could not see is not evidence. Use pane send-text with a trailing newline inside the quotes — not pane run, which splits argv on whitespace and destroys quoted prompts:

herdr pane send-text "$PROBE_PANE" 'pi --model openai-codex/gpt-5.6-luna:low -p --no-session "probe text"
'
herdr pane read "$PROBE_PANE"

Building the door? Read packages/core/skills/harness-door-pattern/SKILL.md first — it carries the suppression-mechanism taxonomy, the probe methodology, and the traps. This playbook is the harness-specific detail underneath it.


1. Model selection

pi --list-models                 # all models
pi --list-models luna            # fuzzy search

Models are addressed as provider/id, with an optional :thinking suffix:

pi --model openai-codex/gpt-5.6-luna:low   "..."
pi --model openai-codex/gpt-5.6-luna:max   "..."

Session guinea pig: openai-codex/gpt-5.6-luna:low ("Luna Light") — use it for probes and high-volume runs. :max ("Luna Max") is the heavier tier for coding work.

Thinking levels: off, minimal, low, medium, high, xhigh, max — settable inline as :level or via --thinking <level>.


2. Non-interactive invocation

pi -p --no-session "your prompt"
  • -p / --print — process the prompt and exit.
  • --no-session — do not persist a session file. Always use this for probes, otherwise probe runs accumulate as session state and pollute later comparisons.
  • --mode json — structured output for machine consumption.

Read-only run (safe for evaluating a repo without edits):

pi --tools read,grep,find,ls -p "Review the code in src/"

3. Flags that map onto Skill Heaven mechanisms

pi flag Mechanism
--no-skills evict all skills
--skill <path> (repeatable) readmit exactly these — the curated posture
--no-extensions / --extension <path> the extension (door) surface
--no-context-files suppress AGENTS.md / CLAUDE.md discovery
--no-prompt-templates suppress prompt templates
--no-tools / --tools <list> / --exclude-tools <list> tool surface

--thinking is NOT a Skill Heaven mechanism. It sets how hard the model reasons. The Skill Heaven ladder is skill- and context-scoped — what is admitted into the session — and is a different axis entirely. The two share level names (off, low, medium, high, xhigh, max) purely by coincidence of vocabulary, and conflating them will produce meaningless benchmark arms. Hold --thinking constant across arms; it is a control variable, never the thing under test.

The curated posture is --no-skills followed by one --skill <dir> per admitted skill. This is a native primitive, not a workaround — it is why pi is the cleanest door to build after Claude Code.


4. Measuring standing dose

Ask the session what it can see, with and without suppression:

PROBE="List every skill you have available, by name. If none, say NONE."

pi --model openai-codex/gpt-5.6-luna:low -p --no-session "$PROBE"
pi --model openai-codex/gpt-5.6-luna:low -p --no-session --no-skills "$PROBE"

A correct suppression returns NONE. Anything else is a leak — record exactly what leaked.

Run every arm on the same model at the same thinking level, or the comparison is meaningless.


5. Known version-sensitive behaviour

packages/core/src/compile.ts carries a note pinned to pi 0.80.10: --no-skills immediately followed by -p silently lost the suppression, so the launcher emits tail args first. Whether this still reproduces on 0.83.0 is recorded in packages/pi-zero/PROBE.md — read that file rather than assuming either way.

The general rule: argv ordering is load-bearing on pi. When a probe gives an unexpected result, try reordering before concluding a flag does not work.


6. pi as a dispatched worker

Via herdr (see herdr-dispatch):

herdr agent start pi-worker --kind pi --pane <PANE_ID> --timeout 120000 \
  -- --model openai-codex/gpt-5.6-luna:low

herdr agent prompt pi-worker "Read /abs/path/brief.md in full and execute it. Begin now."

Direct, non-interactive:

pi --model openai-codex/gpt-5.6-luna:low -p "$(cat brief.md)"

pi has read, bash, edit, and write tools by default. Constrain with --tools when the worker should not be able to modify files.


7. Extensions

pi install <source>     # install an extension and add to settings
pi list                 # list installed
pi remove <source>
pi config               # TUI to enable/disable package resources
pi --extension <path>   # load an extension for one run, without installing

--extension is the right flag for testing a door surface without mutating the user's settings — it keeps the P3 "never mutate shared state" guarantee intact.


8. Anti-patterns

  • Omitting --no-session on probes. Session state leaks into later runs.
  • Comparing arms across different models or thinking levels. Not a comparison.
  • Assuming a flag works because it is in --help. Probe it, on the version in front of you.
  • Recording a dose without recording pi --version. Inadmissible.