@arhen/pi-core-subagent

Install

Requires the pi coding agent — install it first: npm install -g @earendil-works/pi-coding-agent.

pi install npm:@arhen/pi-core-subagent
# or locally: pi install /path/to/pi-subagents

Minimalist pi extension: fast in-process subagents with single / parallel / graph modes, background runs, cancellation, intercom (child↔leader) and an agent↔agent mailbox.

Built for one job: delegate work to isolated subagents without bloating the parent context.

One rule underneath everything else:

A task is a graph, not a checklist. Nodes are workers, edges are data dependencies. The edge both gates the dependent and hands it the upstream output — so "the coordinator forgot to pass X" stops being a failure mode. Where there are no edges, there is no graph: flat fan-out stays flat.

That is Graph Protocol, applied to the runtime rather than to the prompt.

flowchart LR
    subgraph w1["wave 1 — runs in parallel"]
        api["api<br/><i>api-mapper</i>"]
        db["db<br/><i>db-mapper</i>"]
    end
    gate{{"gate"}}
    subgraph w2["wave 2"]
        doc["doc<br/><i>writer</i>"]
    end
    api -- "route map" --> gate
    db -- "schema map" --> gate
    gate -- "both outputs<br/>prepended to the prompt" --> doc

The subagent tool call plus the live above-editor widget: per-agent activity, tool counts, turns, token counters and timers.

Design principles

• The delegation is the graph. needs declares edges; the scheduler runs each wave of ready tasks in parallel and gates the rest. One code path for single, parallel, chain and graph — chain is just needs: [previous]. (why)
• Edges carry data, not just order. An upstream task's output is prepended to its dependents' prompts automatically. The coordinator cannot forget to pass it, because it never passes it.
• A bad graph fails before it spawns. Unknown ids, self-edges and cycles are rejected at call time — never halfway through a run with three children already burning tokens.
• Proof is an exit code, never a self-report. Tasks are asked for a runnable Verify: command; the leader checks git diff --stat. Agents auditing their own work score ~0. (why)
• No ceremony without edges. Six independent reviewers stay six independent reviewers — no waves, no gates, no graph vocabulary imposed on flat work.
• Agent files respected. A spawn goal (name + task) that matches a user agent file's description (.agents/agents, .claude/agents, .pi/agents — project then home) loads that file — body = system prompt, frontmatter model/tools apply, file model validated against the pi model registry. File wins over inline; no match → on-demand definition.
• Two toolsets only. Read-only (read, grep, find, ls — default) or write (read, grep, find, ls, bash, edit, write — write: true). No per-agent tool config surface.
• In-process — children are AgentSessions in the same runtime. No process spawn, no context bleed.
• Zero parent-context injection. No catalog, no context hook. 6 slim tools total.
• Throttled updates — widget/stream updates coalesce to ~6/s; no per-event deep clones.
• No silent hangs — watchdog aborts children that produce no events for 3 minutes.
• No default runtime cap — tasks run until done, stalled (watchdog), or aborted by the user. maxRuntimeMs is opt-in (default 0 = unlimited).

How it runs

Children are not subprocesses. They are separate AgentSessions inside the same pi process — which is why spawning is instant, and why a child's transcript never lands in your context:

flowchart TB
    subgraph proc["one OS process — no spawn, no IPC"]
        direction TB
        L["<b>leader</b><br/>your session, your context"]
        subgraph kids["isolated child sessions"]
            direction LR
            A["api-mapper"]
            B["db-mapper"]
        end
    end
    L -- "task text in" --> A
    L -- "task text in" --> B
    A -. "final answer only" .-> L
    B -. "final answer only" .-> L

The dotted arrows are the whole point: a child may burn 200k tokens reading files, and the leader receives only its final answer.

Usage — the leader invents the agents

Define agents inline per call, or reference a named agent file (see Agent files). Model resolution: explicit model → agent-file model (validated against the pi model registry) → the parent's current model → settings default.

{
  "agent": "api-reviewer",
  "prompt": "You are a strict API reviewer. Check auth, rate limiting, and error handling. Cite file:line.",
  "task": "Review src/api/upload.ts"
}

Parallel — mixed toolsets, siblings can talk via mailbox:

{
  "allowIntercom": true,
  "tasks": [
    { "agent": "researcher", "prompt": "You find facts. Cite paths.", "task": "Map the auth flow", "write": false },
    { "agent": "implementer", "prompt": "You make minimal changes.", "task": "Implement POST /api/upload", "write": true }
  ]
}

Chain — {previous} is replaced with the prior agent's output:

{
  "chain": [
    { "agent": "planner", "prompt": "You write a step list.", "task": "Plan the change", "write": false },
    { "agent": "doer", "prompt": "You follow the plan exactly.", "task": "Execute: {previous}", "write": true }
  ]
}

Agent files

A user agent file in an agents directory is matched by its description frontmatter against the spawn goal (agent name + task) — not by name. When matched, the file is authoritative: body = system prompt, frontmatter model/tools apply, inline prompt/model/tools are ignored. No match → the inline on-demand definition stands. The model stays in control: it names the agent and states the goal; user files that describe that goal take over.

---
name: api-reviewer
description: reviews APIs for auth, rate limiting, and error handling
model: claude-opus-4-6
tools: read, grep, find, ls
---
You are a strict API reviewer. Check auth, rate limiting, and error handling. Cite file:line.

Lookup order (first directory with a match wins):

1. .agents/agents/ then .claude/agents/ then .pi/agents/ in each directory from the task cwd up to the filesystem root (nearest ancestor wins).
2. Home: ~/.agents/agents/ (single source) → ~/.claude/agents/ → ~/.pi/agents/.

Within a directory the file with the highest description-overlap score wins (≥2 shared meaningful tokens). A file model is validated against the pi model registry (unknown model fails the task with a catalog message). Files without a description frontmatter never match.

Worktree isolation (write agents)

In a git repo, a write: true subagent runs in an isolated git worktree at <repo>/.git/subagents/<run>/<task> on branch subagents/<run>/<task> — the child's cwd is the worktree, so project context (AGENTS.md chain) still loads, node_modules is symlinked, and the main tree stays clean while the child works. Parallel write agents can't collide on files.

On completion the extension commits the child's changes (the child is told not to touch branches) and reports branch + diffstat + changed files in the result. The leader reviews, then merges:

git merge --no-ff subagents/<run>/<task>

Isolation follows the toolset the child actually receives: explicit tools: ["bash", "edit", "write"] earns a worktree even without write: true, and an agent file that narrows the child to read-only gets no branch at all.

Cleanup, in order of trust:

| When | What | |---|---| | Session start | Registered worktrees can't be live yet — an interrupted child's uncommitted work is committed, the branch kept, the dir dropped. Then merged branches are reaped and dirs git no longer tracks are removed. | | After a run | Merged branches + their dirs, never touching a branch a live run owns (checkout state re-checked per branch). | | Task failed/canceled | Partial work is committed first, so the branch keeps it; the dir is dropped. |

Non-git repos fall back to in-place edits.

Graph mode — needs

parallel runs everything at once; chain runs everything one at a time. Most real work is neither. Give a task an id and list the ids it needs:

{
  "tasks": [
    { "id": "api", "agent": "api-mapper", "task": "Map every route in src/api/" },
    { "id": "db",  "agent": "db-mapper",  "task": "Map the schema in src/db/" },
    { "id": "doc", "agent": "writer", "needs": ["api", "db"], "write": true,
      "task": "Write ARCHITECTURE.md from the maps above. Verify: test -s ARCHITECTURE.md" }
  ]
}

The call line renders the graph in §2 notation as the model types it:

subagent graph 3
  6 at a time
  wave1[api ∥ db] → gate → wave2[doc]
  api api-mapper Map every route in src/api/
  db  db-mapper Map the schema in src/db/
  doc writer ✎ ← api, db Write ARCHITECTURE.md from the maps above. Verify: test -s ARCHI…

✎ marks a write-toolset task; ← lists its edges. With no needs anywhere the wave line is omitted entirely.

What one edge does

An edge is not just ordering. It is a delivery:

sequenceDiagram
    participant S as scheduler
    participant A as api
    participant D as db
    participant W as doc

    Note over S,D: wave 1 — both start together
    S->>A: "Map every route in src/api/"
    S->>D: "Map the schema in src/db/"
    A-->>S: route map
    Note right of W: doc is queued,<br/>waiting at the gate
    D-->>S: schema map
    Note over S: gate opens: every need settled
    S->>W: ## Output of api<br/>&lt;route map&gt;<br/><br/>## Output of db<br/>&lt;schema map&gt;<br/>---<br/>"Write ARCHITECTURE.md…"

The leader never copies those outputs into the prompt — so it cannot forget to.

One scheduler, four shapes

Single, parallel, chain and graph are not four code paths. They are four shapes of the same wave loop:

flowchart LR
    subgraph one["single"]
        direction TB
        s1(("a"))
    end
    subgraph par["parallel — no needs"]
        direction TB
        p1(("a")) ~~~ p2(("b")) ~~~ p3(("c"))
    end
    subgraph ch["chain — needs: [previous]"]
        direction TB
        c1(("a")) --> c2(("b")) --> c3(("c"))
    end
    subgraph gr["graph — needs"]
        direction TB
        g1(("a")) --> g2(("b"))
        g1 --> g3(("c"))
        g2 --> g4(("d"))
        g3 --> g4
    end

The loop

flowchart TD
    start(["subagent call"]) --> validate{"graph valid?<br/><small>unknown id · self-edge · cycle</small>"}
    validate -- no --> reject["reject the call<br/><b>zero children spawned</b>"]
    validate -- yes --> loop{"tasks left?"}
    loop -- no --> done(["run finished"])
    loop -- yes --> ready["frontier =<br/>tasks whose needs are all settled"]
    ready --> spawn["run that wave in parallel<br/><small>throttled by concurrency</small>"]
    spawn --> collect["record each output<br/>mark tasks settled"]
    collect --> loop

Two consequences worth stating plainly:

• A bad graph costs nothing. Validation happens before the first spawn, never halfway through with three children already burning tokens.
• A broken upstream stops its branch. If a need fails or is aborted, its dependents are marked aborted rather than run against a prompt with a hole in it:

flowchart LR
    api["api ✓"] --> doc
    db["db ✗ failed"] --> doc["doc ⏹ skipped<br/><small>never spawned</small>"]

And the rule that keeps this from becoming ceremony: zero needs anywhere = plain parallel. No waves, no gates, no graph vocabulary imposed on flat work.

Background (default) + intercom — the run returns a runId immediately; you stay steerable while it works:

{
  "agent": "auditor",
  "prompt": "You audit dependencies.",
  "task": "Audit package.json for outdated deps",
  "allowIntercom": true
}

Steering a running child: while a background run is active the leader stays responsive, and you can push a message into a live child's session mid-run with steer_subagent — e.g. steer_subagent({ runId, taskId, message: "Ignore tests/, only audit runtime deps" }). The message queues as a steer if the child is mid-turn and lands at its next model boundary. Omit taskId to steer every still-running task in the run. Combined with notifyPerTask, this makes a background run feel like a live team you can redirect, not a fire-and-forget blob.

Tools

| Tool | Purpose | |---|---| | subagent | single / tasks (parallel or graph via needs) / chain ({previous}); every run is background — returns a runId, completion notifies you; autoAwait:true parks the call until the run finishes and returns the final result inline; allowIntercom:true enables child talk tools; notifyPerTask (default true) wakes you as each task completes | | subagent_status | live per-task snapshot (non-blocking), including each child's session file path | | subagent_result | full output of a run or one task | | await_subagent | block until a run finishes (optional timeoutMs) | | reply_subagent | answer a child's ask_parent question | | steer_subagent | inject a steering message into a running child's session (queues as steer if mid-turn; lands at its next model boundary) | | subagent_cancel | abort a running/queued run |

Per-task fields

agent (name you invent — required), task (required), prompt (system prompt, optional — minimal default used), write (toolset, default read-only), plus optional model (provider/model-id), thinking (validated enum: off|minimal|low|medium|high|xhigh|max), tools (explicit allowlist), cwd, maxRuntimeMs, id, needs (dependency edges — see Graph mode). Top-level only: autoAwait, notifyPerTask, allowIntercom, concurrency.

Child talk tools (when allowIntercom: true)

| Tool | Meaning | |---|---| | ask_parent | blocking question to the leader; parent answers via reply_subagent | | notify_parent | one-way message to the leader | | send_agent_message | message to a sibling subagent's mailbox (to = its task id, or "leader") | | poll_agent_messages | drain this subagent's mailbox |

Intercom anti-deadlock: children are told to never block indefinitely on intercom replies — ask_parent keeps the stall watchdog fed while awaiting a parent reply, and sibling polls are capped (~5 tries) with a proceed-with-best-judgment fallback. Gated siblings (later waves) may not be running yet — waiting on them is the top stall cause, so children are instructed not to.

Commands

• /subagents — list runs; /subagents peek (or ctrl+shift+a) — browsable pane
• /subagents auto-limit on|off — toggle leader-imposed maxRuntimeMs caps (persists to ~/.pi/agent/subagents-config.json; default on). off strips ALL task timeouts: tasks run unlimited until done, stalled, or aborted — only for runs where a hard bound is genuinely required is a cap kept (none, when off). Bare /subagents auto-limit shows the current state.

Peek — /subagents peek or ctrl+shift+a

Read-only pane over the session's subagents:

• shift+↑/shift+↓ (or j/k) — move between agents; bare arrows work too where the terminal doesn't reserve them
• enter — live tail of that child's session file (esc goes back)
• x then y — abort ONE subagent (only mutation; n/any other key cancels)
• esc — close

Watching a child from outside

A child has no terminal of its own — but it does write a real transcript file, and that file is the seam every external viewer can use:

flowchart LR
    child["child session<br/><small>no TTY</small>"] -- writes --> file[("session.jsonl")]
    file -- "peek · enter" --> pane["in-pi tail"]
    file -- "tail -f" --> term["any terminal pane<br/><small>herdr · tmux · zellij</small>"]

subagent_status returns that path for every running child:

tail -f /path/from/subagent_status.jsonl

In a terminal multiplexer, that is a pane per agent — e.g. with Herdr:

herdr pane split --current --direction right
herdr pane run w1:p2 "tail -f /path/from/subagent_status.jsonl"

The extension has no multiplexer integration and does not want one: it exposes the path, your agent already knows how to drive its own terminal. For an in-pi view of the same stream, use /subagents peek.

Context budget

• Parent tools: 6 schemas with short descriptions. No catalog, no context hook — nothing injected per request.
• Background completion: 3-line notice. Full text only via subagent_result.
• Children: isolated sessions; talk tools injected only when allowIntercom; each child's prompt states its own task id and its siblings' so mailbox addressing works. Model resolution: explicit provider/model-id or bare id via the pi model registry → the parent's current model → settings default. Thinking levels validated against the resolved model's thinkingLevelMap.

What this is built on

Graph Protocol

needs is an implementation of Graph Protocol — a delegation discipline that treats a task as a graph (Delegation<A, E, R>) rather than a checklist. Its ten sections map onto this extension as follows:

| § | Protocol | Here | |---|---|---| | §1 | nodes, domains, edges | one agent owns one task; needs are the edges | | §2 | happy path as execution graph, waves + gates | wave scheduler: ready = tasks whose needs are settled | | §3 | one worker or many | single vs tasks | | §4 | break points: wrong context, missing input, misinterpretation | missing input is structurally impossible — the edge carries the output | | §5 | R: subgraph, method, verification command, WHY | prompt guidelines require a runnable Verify: line per task | | §6 | structured at the boundary | in: upstream outputs prepended as named blocks. out: prose (see below) | | §7 | observe without changing the graph | the widget and /subagents peek are read-only | | §8 | worker attention acquired and released | spawn → terminal status; aborted upstream releases dependents immediately | | §9 | prove it: delegated vs implemented | deliberately not self-reported — see below | | §10 | prompt = subgraph, return = implemented graph | prompt yes; return kept as prose |

Why §9 is a verification command, not a self-report

The protocol asks the coordinator to compare the delegated subgraph against the graph the worker says it implemented. We implement the comparison against the filesystem and the exit code, not against the worker's account of itself, because self-reports carry close to zero signal about exactly the failure §9 exists to catch:

• Asked to audit its own work against 34 real violations, an agent reported 0 — at 90–100 confidence. A fresh instance of the same model shown the same output caught 7 (p = 0.0156). A deterministic checker caught all 34. (Armalo Labs, 2026)
• Across 9,876 τ2-bench and 1,879 AppWorld trajectories, "false success" reached 75.8% of self-assessing coding-agent failures; adding an LLM judge scored 0.54–0.65 AUROC (0.5 = coin flip). (arXiv:2606.09863)
• LLM judges reading agent traces can be flipped by rewriting the trace — the exact surface a self-reported graph exposes. (arXiv:2601.14691)

The shape of the problem:

flowchart TD
    W["worker finishes"] --> Q{"who says it's correct?"}
    Q -- "the worker itself" --> S["self-report<br/><b>0 of 34 caught</b><br/><small>at 90–100 confidence</small>"]
    Q -- "another model reading the trace" --> J["LLM judge<br/><b>0.54–0.65 AUROC</b><br/><small>0.5 = coin flip</small>"]
    Q -- "the machine" --> D["exit code + git diff<br/><b>34 of 34 caught</b>"]

So §9 in practice is two things you already have:

# in the task text — the worker must prove it, not claim it
Verify: npx tsc --noEmit && bun test

# in the leader, after the run — ground truth, not narrative
git diff --stat

If files outside a worker's subgraph were touched, the diff says so. A structured return schema would only add a second, less trustworthy witness.

Why waves instead of "more agents"

Flat fan-out is not free — orchestration cost is critical path + α × cross-agent communication, and ignoring the second term is what makes added agents lose to a single one:

• Dependency-graph partitioning vs flat file-parallel spawning across 28 real repos: +14.0% pass rate, 2.10× wall-clock, −35% API cost, with the largest gains on the most dependency-dense projects. Flat parallel inflated cost 60% for a 1.56× speedup; an agent-team baseline was fastest but scored below sequential on code quality. (arXiv:2606.00953)
• Dynamic task graphs across 300 trials: 47.5% of baseline token cost, 79.7% accuracy vs 57.6% for a static graph — and, notably, static tied dynamic when the structure was genuinely known up front, which is the case needs targets. The "frontier" (ready set) in this scheduler is theirs. (arXiv:2605.06320)

The corollary is in the design principles: when there are no edges, don't draw a graph. Six independent reviewers stay six independent reviewers.

Development

bun install        # dev deps (typecheck/test only; runtime uses pi's bundled SDK)
npx tsc --noEmit
bun test           # pure-logic tests (wave scheduling, edge payload, mailbox, failure classification, watchdog)

Runtime state: runs persist to <parent-session>.subagents.json sidecar; restored (non-terminal → aborted) on session start.

License

MIT.