# pi-persona technical reference This is the canonical technical guide for the runtime surfaces that users configure through personas, agents, teams, flows, contracts, presets, and environment variables. The design contract lives in [`ARCHITECTURE.md`](./ARCHITECTURE.md); strategy shapes and their complete parameter table live in [`STRATEGIES.md`](./STRATEGIES.md). User-oriented orientation and operating examples start in the root [`README.md`](../README.md) and [`EXPERIENCE.md`](./EXPERIENCE.md). ## Core API The runtime is composed from a small, fixed set of primitives. A strategy is a TypeScript file composing the **Strategy SDK**; a persona picks whether and how those run. `magi` is a `.md` persona plus a file that calls `parallel` and `reduce.vote`. **Strategy SDK** (`src/orchestration/sdk.ts`) — the primitives available to a strategy: | Primitive | Does | |---|---| | `agent(spec, { reserveChildren? }?)` | run one sub-agent and return a structured `AgentResult`; `spec` may carry `name`, model, tools, skills, or `outputContract`; optional reservations protect mandatory later legs from model recovery | | `parallel(thunks, {concurrency})` | run many at once, bounded by the run limits; this also provides the basis of `map` | | `reduce.aggregate(results)` | merge N results into one, used by fan-out | | `reduce.vote(candidates, opts)` | tally the candidates' own votes and preserve `winner`, `tie`, `no_consensus`, `invalid_outputs`, and dissent | | `reduce.judge(candidates, order?)` | anonymise and label candidates for an impartial judge, then map the verdict back with `pick(label)` | | `roster.team(name)` | resolve the agents of a named team | | `signal` · `limits` · `log` | cooperative abort, admission/runtime limits, and progress reporting | | series and loops | plain `await` and `for`; strategies are TypeScript, so pipelines and critic loops are ordinary code | The complete built-in strategy catalog and each strategy's declared parameters are maintained in [`STRATEGIES.md`](./STRATEGIES.md). Parameters are declared per strategy and looked up by `knownParams()`; `/doctor` exposes the same schema live. Unknown parameter keys warn but do not hard-fail. **Supervisor surface** — the tools and persona fields that drive those primitives: | Surface | Does | |---|---| | `delegate` tool | spawn one or more sub-agents; interactive calls are background-first, while `sync: true` blocks the turn | | `council` tool | select any installed strategy over a named roster or temporary inline `members`; voting strategies preserve tally and dissent. Use when the topology adds value, not for routine work | | `intercom` tool | observe and control internal runs with `peek`, `result`, `wait`, `steer`, and `stop`; `list`, `inbox`, `message`, `reply`, and `send` drive the coaching bus. `wait` snapshots without blocking by default with UI; `sync: true` opts into a bounded blocking join (headless defaults to blocking; use `sync: false` for a snapshot); queued steering does not preempt an explicit join | | `timer` tool | read the fresh clock with `now`, or schedule an idle wake with `arm`, `cancel`, and `list`; alarms are one-shot and in-memory per session | | `monitor` tool | run a bounded job (`mode: exit`) or event-producing program (`mode: output`) and wake the supervisor; see [`MONITORS.md`](./MONITORS.md) | | `flow` tool and `/flow` | run a journaled DAG of strategies; `gate: true` creates an approval checkpoint | | `models` tool | list or search authenticated model ids in `provider/id` form; the session provider is marked | | persona `mode:` | choose `solo`, `parallel`, `pipeline`, `strategy:`, or `flow:` | | persona `coaching:` | give children `contact_supervisor` for progress and blocking decisions | | persona `delegation:` | require briefs, set an output contract, schedule disjoint writes, and order fresh verification | | `isolation: worktree` | run in a clean throwaway Git worktree and return an exported diff; non-Git, dirty, or oversized cases fail closed | | `council: { preset }` | expand a reusable `presets/.preset.json`; authored fields override the preset | | `contracts/*.json` | define a hot-editable structured return requested through `outputContract` and pinned per run | The delegate tool's optional `name` is carried into `AgentRunSpec.name` and reaches both engine backends, keeping a caller-provided run label available in the live tree and runtime result. Every settled async run also emits an immediate UI banner with that name and its terminal outcome. The model-facing report remains independently idle-gated; a UI notification failure cannot discard it. An explicit `delegate` `async` flag overrides both `sync` and the session default: `async: true` forces background execution; `async: false` forces synchronous execution. ### Pi 1.0 codemode Pi's native `codemode` can call callable pi-persona tools through the normal nested-tool pipeline. Argument validation and persona permission gates still apply to every inner call; scripts do not turn a denied tool into an allowed one. Native `model-only` and `hidden` tools are not callable. Use scripts for independent batches or for filtering large results before returning them to the model, not as a wrapper that prints every result in full. `Promise.allSettled()` preserves successful siblings when a call rejects; a fulfilled `bash` call still requires checking `exit_code`. Only printed/returned script output reaches the model. Nested results therefore do not contribute raw bytes or invisible reminders to pi-persona's context-burn nudges; an actual successful nested `delegate`/`council` still resets the hand-off streak. Pi identifies calls made through codemode's `ctx.executeTool` with `parentToolCallId`. If a nested delegation report is relayed with `[BLOCKED]` or `FLAG: UNKNOWN`, only the outer model-visible result gets the persistence checkpoint and its expandable card. Discarding the report creates no reminder and never alters the programmatic inner result. See [the nudge contract](./ARCHITECTURE.md#supervision--the-waiting-model) for the differences between synchronous and background reports. The outer code/result card belongs to Pi, not pi-persona's delegate renderer. Keep bulky details expandable (`Ctrl+O` by default; `/hotkeys` shows active bindings). `Ctrl+T` toggles thinking blocks separately. A smaller script output also reduces model context; merely collapsing its card does not. No replacement sandbox or renderer is required for this integration. `outputContract` takes a contract **name**, not an inline shape — `/doctor` lists what's installed (the built-in `default` and bundled `finding` are available without custom files). Naming one that isn't installed is rejected before any worker spawns, checked across every task in a parallel batch: `delegate: unknown output contract(s) "x" — nothing was spawned. Installed contracts: default, finding` (the list is capped at 16 names; a `, …` marker is appended only when more than 16 are installed). ### Supervising running sub-agents — the `intercom` plane The supervisor's internal communication with children exposes these surfaces. Tree/F9/peek distinguish queued, running, waiting for supervisor, suspected stalled and terminal runs. Waiting is derived from a live pending ask on the exact child handle, never inferred from its last activity text: | Surface | Needs | What you get | |---|---|---| | **Observe and control** | any persona | `peek` inspects async sub-agents; `wait` is nonblocking by default with UI (`sync: true` joins; headless defaults to joining, with `sync: false` for snapshots); `steer` sends a soft correction; `stop` requests cancellation. The `f9` overlay provides the same controls (`s`, `x`). Queued steering does not preempt an active join. Cancelling a worker does not undo existing changes | | **Retrieve** | any persona | `result { to }` returns one retained run payload; `message { messageId }` returns one retained bus message. History is bounded to 256 messages and 256,000 body characters, and retrieval never starts a worker | | **Message bus** | `coaching: true` | children get `contact_supervisor`; progress appears in results and `intercom inbox`, while a blocking `decision` wakes the supervisor for `intercom reply` | | **Sibling peer comm** | strategy opt-in | `debate` and `pair` members always get one-way `contact_peer`; `map` and `synthesize` add it with `params: { peers: true }` | With coaching enabled, an async child's `decision` or `interview` waits for a supervisor reply. A synchronous child posts the question one-way and continues instead: the supervisor is holding that turn and cannot answer a blocking ask. Progress messages are always one-way. Delegated output is untrusted and is fenced before it reaches the supervisor. Interactive async work is background-first: continue independent work, and when only children remain, end the turn for automatic completion follow-ups rather than monitoring or joining all children. Use `wait` only when the next step depends on results. Async runs report failures to the supervisor, and a repeated identical failed delegation is stopped by the runtime ledger before another child is spawned. A run whose model's provider fails at call time (auth, outage, 5xx) may retry the same model id through a compatible fallback provider when its selection is unpinned; explicit delegate provider pins remain strict. Strategies and flows instead recover a model/provider failure once on the current main model, including unavailable saved or pinned choices. They never substitute a peer's model or search arbitrary providers. Each recovery is visible and both attempts consume usage/child budget; stop, timeout, contract and infrastructure failures do not retry. Any leg that does NOT run in `isolation: worktree` shares the real checkout, so its result gets a "what changed" block appended: `--- FILES CHANGED DURING THIS LEG (shared checkout; parallel legs may overlap) ---` followed by one `- ()` line per path (capped at 40, `+N more` beyond that); a changed path outside the leg's own declared `writeSet` is flagged `⚠ outside declared writeSet`. This is a `git status` before/after snapshot, not a private diff — a parallel sibling touching the same file can appear here too — and it is skipped when there is nothing to report, or when the leg IS worktree-isolated (that path already returns a real unified diff instead). The cross-process broker is on by default for child-engine, worktree, and MCP legs. It provides the same supervisor communication surface over a session-scoped POSIX socket or Windows named pipe. `PI_PERSONA_BROKER=off` restores the pre-broker child spawn behavior. See the comm-plane contract in [`ARCHITECTURE.md`](./ARCHITECTURE.md#the-comm-plane-in-practice). > **MCP in sub-agents.** A normal in-process leg has no MCP connection. Set `mcp: true` in an agent > definition or delegation task to route it through the child engine, which initializes MCP from > `~/.pi/agent/mcp.json`. A fresh UI-less child cannot complete an interactive OAuth flow, so > authenticate the server in a normal session first. A backend that keys state by a session id must > receive that id in the task. A local registration, stdio process, or loopback endpoint does not > prove local target egress; verify source routing with the executor and canary required by the > engagement. This is a behavioral boundary, not a network sandbox. ### Identity and Exocom An unnamed standalone session receives a provisional process/session-derived handle. On its first turn where `agent_name` is callable, the runtime context asks the model to invent a short handle and call `agent_name` before other work. The chosen handle is separate from the active persona and role, is persisted in the session identity entry, and remains across later turns and persona changes. Once chosen, it is not re-injected into every provider request; the runtime and session entry remain authoritative. `agent_name` is unavailable while Exocom is active; its call-time capability check also rejects empty, unsafe, generic, or reserved persona/agent names. Naming needs one tool action, not a recurring prose confirmation. A stable system-level runtime rule distinguishes skill-prefetch bookkeeping from new user requests; it neither hides updates nor rewrites conversation history. Periodic async status is a durable operator-only entry, not a model wake or context message. After updating extension code, use Pi's `/reload` once delegated work has settled (or restart/resume); `/persona reload` reloads definitions only. With Exocom active, the corresponding tool is `exocom_name`. The first incoming peer wake can be the first model turn in an RPC session; the context hook supplies the naming bootstrap for that custom wake as well as for an ordinary user turn. A chosen Exocom handle updates the live registry while its session-qualified routing target remains stable across renames. Naming metadata remains callable when a pending ask permits `exocom_name`, but it does not settle that work ask. Exocom is a flat, external plane between independent top-level Pi sessions. It provides presence, fenced postcards, and a durable claim/ask ledger. `--exocom` joins the current workspace; an exact `--exocom=Ab0T` joins an existing workspace scope from another cwd. The four-character code is a same-host join reference, not authentication. Foreign members advertise the workspace they can actually inspect and cannot claim repository-relative paths in another workspace's ledger. `exocom_status` (a tool, taking no arguments) and `/exocom` (a command, which additionally lists live peers) both render the same ownership view: your claims, peers' claims, asks waiting on you, and your own open asks. A claim's `write_set` is advisory, not enforced: a supervisor's own `write`/`edit` call whose path overlaps a PEER's open claim is blocked once with a reason to coordinate (`exocom_ask`) or claim the path itself; calling the same tool again on the same path is then allowed through (one warning per claim + normalized path, not a standing lock). ## Configuration and data Personas and agents are Markdown files with YAML frontmatter. Teams, flows, contracts, and presets are data files. Discovery precedence is builtin < user `~/.pi/agent/persona` < project `.pi/`. Personas and agents are loaded from the user and project agent folders; bundled persona and agent files are seed sources. A fresh install has no installed personas until `/persona seed`, `/persona restore`, or opt-in first-run seeding installs them. | Definition | Location | Purpose | |---|---|---| | Persona | `personas/*.md` or installed `agents/*.md` with `persona: true` | supervisor prompt, capabilities, delegation, council, orchestration, and coaching | | Agent | `agents/*.md` or project `.pi/agents/*.md` | worker prompt, optional tools/model/isolation/MCP | | Team | `teams.yaml` | named roster of bare agent names or inline `{ agent, role, model, skills }` members | | Flow | `flows/*.flow.json` | strategy phases connected by `needs`, optionally gated | | Contract | `contracts/*.contract.json` | structured output fields requested by name and pinned per run | | Preset | `presets/*.preset.json` | reusable council strategy, roster, and parameter bundle | `tools` accepts an allowlist such as `[read, grep]`, an `allow`/`deny` object, or `[]` for no tools. Omit it to inherit the session tool set. `model`, `isolation: worktree`, `mcp: true`, and `purpose` are optional agent metadata. The body is the agent prompt. Inline roster members specialise one agent with a role, model, or skills without requiring another agent file. ### Multi-agent quick start Follow the [installation instructions](../README.md#install), then start a coding supervisor: ```bash pi --persona dev ``` Delegate a bounded task with the complete brief that Dev requires: ```js delegate({ agent: "scout", name: "Login-Trace", task: "Inspect the login flow and report the root cause.", brief: { objective: "Explain the login failure with code evidence", scopeRoe: "Read this repository only; do not contact external systems", position: "Start from the current checkout and its existing tests", constraints: ["Do not modify files"], requiredArtifacts: ["File and line references, with a suggested minimal fix"], stopConditions: ["Stop after finding the cause or report the missing evidence"] } }) intercom({ action: "peek" }) ``` For independent Pi processes, open the same workspace in two terminals with `pi --exocom` and use the qualified target returned by `exocom_list`: ```js exocom_name({ name: "" }) exocom_list({}) exocom_send({ target: "", message: "Review the API boundary and send back concrete risks." }) exocom_claim({ work_key: "api-review", write_set: ["src/api"], slice: "Review and fix the API boundary" }) exocom_ask({ target: "", work_key: "api-review", question: "Is the transport contract safe to change?" }) exocom_wait({ work_key: "api-review", ask_id: "" }) exocom_answer({ work_key: "api-review", ask_id: "", ok: true, evidence: "Tests and file references" }) exocom_release({ work_key: "api-review" }) ``` ### Recipes Everything in these examples is data. Drop files into the discovery locations above and switch personas with `f8`. A brand-new strategy shape is the only case that requires a source file. **Opportunistic delegation** — a simple persona with no orchestration block: ```markdown --- name: my-researcher persona: true --- You research thoroughly. For independent sub-questions, fan out `scout` sub-agents in ONE `delegate` call (`tasks: [...]`), each with a disjoint scope, then synthesise their findings. ``` **A review council and a per-call override** — `council:` is the persona default; the tool can override it for one call without editing the file: ```yaml # teams.yaml — one reviewer agent, three lens roles review: - { agent: reviewer, role: "Focus ONLY on the SECURITY lens" } - { agent: reviewer, role: "Focus ONLY on the PERFORMANCE lens" } - { agent: reviewer, role: "Focus ONLY on the TESTS lens" } ``` ```markdown --- name: myaudit persona: true council: { strategy: synthesize, roster: review, params: { synthesizer: reviewer, peers: true } } --- Convene the council before sign-off, then apply its merged findings yourself. ``` ```js // one-off: run this decision as a debate over the same roster — no file edit council({ question: "cache this or recompute?", strategy: "debate", roster: "review", params: { bestOf: 2 } }) ``` A one-off group can be composed without editing a team file. The external arbiter is also an agent specification, independent of the panel: ```js council({ question: "Choose between these designs; prioritize correctness and operational cost.", strategy: "judge", members: [ { agent: "reviewer", role: "Evaluate correctness and failure boundaries." }, { agent: "reviewer", role: "Evaluate operating cost and maintainability." } ], params: { judge: { agent: "reviewer", role: "Choose impartially from the anonymised options." } } }) ``` `members` overrides the named roster for this invocation only. Roles/skills do not grant tools or bypass capability gates. Model selection covers roster members and declared auxiliary actors such as `judge` and `synthesizer`; supported agent-valued params accept inline member specs too. Model precedence is inline `model`, saved role assignment, saved agent assignment, agent frontmatter, then the session model (`src/persona/model-participants.ts`). With an active persona, UI, and at least two configured models, unassigned participants can be picked and saved for the active persona, even when borrowing another persona's council profile. Dismissing a picker suppresses another prompt for that participant during the session, not for later teams with different participants. Headless runs never prompt. An unavailable choice is not overwritten: a provider or unknown-model failure may recover once on the main model, with the switch reported and both attempts accounted. This does not relax ordinary delegate provider pins. The MAGI persona retains its canonical core team; the Judge persona retains impartial arbitration. Other personas choose whichever strategy fits the task. **A delegation policy on any persona** — fields are generic and are never keyed to a persona name. The runtime rejects incomplete briefs before a model call, supplies a missing contract, rejects overlapping or missing ownership for parallel writers, and refuses a declared verifier that would start before or beside a material writer in the same call or while a mutating background run is in flight. Only `read`, `grep`, `find`, and `ls` count as read-only; shell-capable, MCP-enabled, and unknown toolsets fail closed as writers. `writeSet` is a scheduling contract, not an OS sandbox. ```markdown --- name: my-project-lead persona: true delegation: requireBrief: true outputContract: finding requireDisjointWrites: true requireFreshVerification: true verificationAgents: [my-checker] council: { strategy: critic-loop, roster: my-repair-team, params: { rounds: 3 } } --- Delegate bounded work, then accept completion only after fresh evidence. ``` The six required `brief` fields are `objective`, `scopeRoe`, `position`, `constraints`, `requiredArtifacts`, and `stopConditions`. `verificationAgents` names ordinary installed agents; there is no built-in verifier identity. When `requireBrief` is enabled, it also applies to read-only scouts. Single-worker calls put the brief alongside `agent` and `task`; parallel calls put a complete brief inside **each** `tasks[]` entry. A top-level brief is not inherited by the batch, and prose in `task` does not replace these fields. A missing-brief error means the batch was rejected before any worker started: correct all affected entries before retrying. **A mandatory-orchestration persona** — `orchestration:` runs automatically every turn, while `council:` runs only when requested. Only `ok: true` is presented as a ruling; failures, cancellation, and unresolved output remain labelled evidence to repair and verify. ```markdown --- name: myguard persona: true orchestration: { mode: strategy, strategy: magi, roster: magi, params: { reflect: false } } --- Every turn is decided by the MAGI triarchy first (no reflection round); present and act on it. ``` **Coaching** — add `coaching: true` to give children `contact_supervisor`, then read and answer progress or decisions with `intercom inbox` and `intercom reply`. **A custom agent** — the body is the worker prompt and the frontmatter sets its capabilities: ```markdown --- name: hardener description: Locks down a change's security posture — authz, input validation, secrets handling. tools: [read, grep, edit, bash] model: opus isolation: worktree --- You are the Hardener. Given a diff or area, find and FIX authz/access-control gaps, injection and unsafe-sink risks, secrets/token mishandling, and missing input validation. Cite `file:line`. ``` **A flow with a human checkpoint** — phases are journaled and dependents wait for a `gate: true` approval. Save as `.pi/flows/gated-build.flow.json`: ```json { "name": "gated-build", "phases": [ { "id": "plan", "strategy": "magi", "roster": "magi", "gate": true }, { "id": "build", "strategy": "critic-loop", "roster": "repair", "needs": ["plan"] }, { "id": "verify", "strategy": "fanout", "roster": "review", "needs": ["build"] } ] } ``` **A structured-return contract** — save as `.pi/contracts/ship-verdict.contract.json` so votes and judges can tally fields mechanically: ```json { "name": "ship-verdict", "fields": { "vote": { "type": "string", "required": true }, "severity": { "type": "enum", "values": ["low", "medium", "high", "critical"] }, "confidence": { "type": "number", "min": 0, "max": 1 } } } ``` **A preset keeps persona files to one line** — save as `.pi/presets/magi-rounds.preset.json`: ```json { "strategy": "council-rounds", "roster": "magi", "params": { "rounds": 3, "bestOf": 3 } } ``` ```yaml # any persona's frontmatter council: { preset: magi-rounds } ``` For one-line strategy calls and the canonical parameter defaults, use the table in [`STRATEGIES.md`](./STRATEGIES.md). A new vote rule or custom loop adds `src/orchestration/strategies/.ts`, registers it in `strategy.ts`, and gets a stub-engine unit test; the data-driven definitions above require no core change. ## Commands, flags, and environment **Keys** — `f8` cycles persona; `f9` or `/agents` opens the agent overlay: ↑↓ navigate, ⏎ open, `x` stop, `s` steer, and Esc close. **Commands** — `/persona [name|off|list|reload|seed|restore]` · `/models [query]` · `/orchestrate ` · `/flow ` · `/peek [id]` · `/exocom` · `/doctor`. `/doctor` reports the cross-process broker's lifecycle on one line: `broker: on — endpoint …` once the host is listening, `broker: on — (starting…)` mid-attempt, or `broker: on — (not started — no child-engine build yet)` before any child-engine build has run. A bind failure reports on its own, without the `on —` prefix or a connected-children count (there is neither while down): `broker: failed — ` — children built while it is down spawn without a bus endpoint, so they never burn connect backoff against a dead socket, and the next child-engine build retries. A connected broker child's status shows `⇄ `; if that connection drops, the status flips to `⇄ offline` and a fresh ask on it fails fast instead of hanging. `/doctor` also reports the installed host's version against this extension's floor: `pi: (requires ≥ 1.0.0)`, with a trailing `— BELOW FLOOR` marker once an installed version fails the check (`src/core/pi-compat.ts`). `pi: unknown` means the host's own `package.json` could not be resolved (a non-standard install layout) — not a failed check. **CLI flags** — `--persona ` starts with that installed persona and errors if it is missing; `--exocom` joins this workspace's Exocom plane; `--exocom=Ab0T` joins another workspace's scope from a different cwd. Pi owns `--model `, `--thinking `, and `--list-models`. | Variable | Default | Effect | |---|---|---| | `PI_PERSONA_ENGINE` | in-process | `child` spawns each sub-agent as an isolated `pi` process | | `PI_PERSONA_CHILD_THINKING` | inherited | reasoning effort for child-engine sub-agents | | `PI_PERSONA_SEED` | off | `on` auto-installs bundled personas on first run | | `PI_PERSONA_BROKER` | on | `off` disables the cross-process comm plane for child, worktree, and MCP legs | | `PI_PERSONA_PEEK_MS` | 30000 | fast async stall/message wakeup tick; `0` disables it | | `PI_PERSONA_CHECKIN_MS` | 300000 | durable expandable operator status entry while async children run; does not start a model turn; `0` disables it | | `PI_PERSONA_AGENT_MAX_MS` | off | opt-in per-agent hard wall-clock cap; unset leaves healthy work running | | `PI_PERSONA_AGENT_STARTUP_MS` | 300000 | startup deadline for a child that makes no progress; `0` disables it. Like the idle window, a leg waiting on a supervisor reply is not idle: the watchdog re-arms while its ask is pending, instead of killing on a blind ceiling | | `PI_PERSONA_NUDGE` | on | `off` silences delegation and persistence nudges; synchronous checkpoints stay in model-visible tool results and are mirrored to durable, expandable TUI-only cards | | `PI_PERSONA_EXOCOM` | off | `1` joins the current workspace Exocom scope; cross-workspace selection is CLI-only | | `PI_PERSONA_SPINE` | off | `on` or a path enables the shared behavioral layer; see [`SPINE.md`](./SPINE.md) | | `PI_PERSONA_SPINE_LEGS` | follows `PI_PERSONA_SPINE` | selector for delegated legs, allowing all four spine measurement arms | | `PI_PERSONA_DISABLE` | unset | any non-empty value disables the whole extension for this process (also set internally as the fork-bomb guard on delegated legs) | | `PI_PERSONA_DIRS` | none | extra discovery directories (`;`/`,`-separated), searched alongside the user and project agent folders | | `PI_PERSONA_DEFAULT` | none | persona activated automatically when no persona is otherwise selected | | `PI_PERSONA_KEY` | `f8` | keybinding that cycles installed personas | | `PI_PERSONA_PERSIST` | on | `off` stops remembering the last-selected persona across sessions | | `PI_PERSONA_DELEGATE_DEFAULT` | allow | `deny` (or an off-word) makes an unlisted delegate target default-deny instead of default-allow | | `PI_PERSONA_STATE_FILE` | a global default path | overrides where the last-selected-persona state is stored | | `PI_PERSONA_ASYNC_RETAIN` | 25 | how many settled-but-uncollected async runs are retained before FIFO-evicting the oldest | | `PI_PERSONA_LEDGER_V2` | off | `on` widens the delegation-ledger retry-veto key to also include role/tools/isolation, not just agent+model+task | | `PI_PERSONA_LEG_CHANGE_REPORT` | on | `off` disables the non-worktree files-changed report (two `git status` spawns per leg) | ## Development ```bash npm install npm run typecheck # strict tsc --noEmit npm test # node --test ``` The architecture contract is [`ARCHITECTURE.md`](./ARCHITECTURE.md), orchestration behavior is [`STRATEGIES.md`](./STRATEGIES.md), the shared prompt layer is [`SPINE.md`](./SPINE.md), and the observer/export contract is [`TELEMETRY.md`](./TELEMETRY.md). After updating Pi packages, run `npm run typecheck`: it catches changes at the in-process session and `ThinkingLevel` seams.