Pi Herdr Agents

Agents: the tool list is in What's Included and the operating guide is skills/pi-herdr-agents/SKILL.md. Contributors: read AGENTS.md.

Asynchronous subagents for Pi, running exclusively in Herdr.

Delegate investigation, implementation, and review without blocking the parent session. Each child runs as a real Pi process in its own Herdr surface; results return automatically when the child finishes.

Features

• Non-blocking delegation — subagent acknowledges launch immediately while the parent keeps working.
• Parallel execution — run independent scouts, workers, and reviewers at the same time.
• Live supervision — track process and turn state in Pi's subagent widget; interrupt one child turn without destroying its session.
• Managed worktrees — isolate writing agents in retained Herdr workspaces with explicit Git ownership and recovery details.
• Conversation handoff — continue the active Pi conversation in a new worktree with /worktree while preserving the parent session.
• Pack-neutral roles — use project or global definitions and installable role packs; this package ships no default roles or workflows.
• Persistent specialists — retain one policy-bound Pi session for sequential, turn-based tasks.
• Optional sidebar focus — the companion Pi Herdr Agents Sidebar Herdr plugin hides marked delegated children from Herdr's Agents view while you run Focus. Markers are display-only and follow the pane, so see its limits.

Requirements

• Pi with package support
• Herdr and its CLI
• HERDR_ENV=1 — start Pi from inside Herdr

Other terminal multiplexers are not supported. Session startup skips worktree inventory to avoid blocking Pi initialization; use /worktree list or worktree_list to inspect managed worktrees. Outside Herdr, explicit inventory tools still report unavailable inspection as unknown. Worktrees isolate Git checkouts, not processes or permissions; child agents and installed Pi packages run with your user account's access.

Install

Install from npm:

pi install npm:pi-herdr-agents

Install project-locally or try it for one run:

pi install -l npm:pi-herdr-agents
pi -e npm:pi-herdr-agents

Then start Pi inside Herdr:

herdr
pi

Restart or /reload Pi after installation. Review package source before installing any Pi package.

Upgrading from 2.x? Bundled roles and several commands moved out. See 3.0.0 release notes.

Contents

• Requirements, Install, Safety and uninstall
• Quick start, Release notes, How it works
• What's Included: tools, commands, the operating skill, role packs
• Async Subagent Flow: status, configuration, model routing, sidebar markers
• Spawning Subagents, Persistent specialists, Interrupting, Cancelling
• The /worktree Workflow, Custom Agents, Tool Access Control
• Development, License

Safety and uninstall

• Child agents are real Pi processes running with your user account's permissions, inside Herdr panes. Worktrees isolate Git checkouts, not processes or permissions.
• The extension creates Herdr panes, tabs and managed worktrees only when a launch asks for them. It never pushes, merges, opens pull requests, deletes branches or removes worktrees on its own; cleanup is an explicit parent action (worktree cleanup).
• What it writes: $PI_CODING_AGENT_DIR/herdr-agents/config.json (only through /subagents-init or the writer tool, never on startup); per-launch artifacts under the parent session's artifacts/<session-id>/ directory beside Pi's session store, which hold the child's full task text and any systemPrompt as Markdown files, activity snapshots and worktree manifests; and the child's own Pi session file. A managed worktree that carries its own .pi/agent directory receives that child's session inside the checkout. Treat task text as potentially sensitive when you share or inspect those files.
• Managed worktree children record their PID, process start time, boot ID, and PID namespace in <session>.process.json beside their session. With sidebar.enabled, every other child records the same facts, and a resumed child records them under artifacts/<session-id>/process-identity/. The parent then reports one display-only pane token, piha_delegated_v1, on each child's pane through the Herdr socket. With the setting off, the extension writes no pane metadata.
• To uninstall: first list and remove any retained worktrees while the extension is still loaded (/worktree list, then /worktree remove <target>), because those commands leave with the package. If you installed the optional sidebar plugin, remove it with herdr plugin uninstall pi-herdr-agents.sidebar (or herdr plugin unlink for a linked copy). Then run pi remove npm:pi-herdr-agents, and delete $PI_CODING_AGENT_DIR/herdr-agents/ and the artifacts/ directories above if you no longer want the configuration and launch records.

Quick start

This package is an execution host and ships no agent roles. Named roles come from your project or global definitions or from an installed role pack. The scout examples below assume one of those provides scout; see Migrating from bundled roles. A bare launch without agent needs no role.

Ask Pi to delegate naturally:

Use two scouts in parallel to map the authentication flow, then summarize their findings.

Or launch a named role directly:

/subagent scout Analyze the authentication module and report relevant files and risks

For an isolated writing task:

/worktree auth-fix Implement the approved authentication fix and run the focused tests

Pi can also call the tool directly:

subagent({ name: "Auth scout", agent: "scout", model: "<provider>/<fast-tier-id>", thinking: "low", task: "Map the authentication flow" });
subagent({ name: "DB scout", agent: "scout", model: "<provider>/<fast-tier-id>", thinking: "low", task: "Map the session schema" });
// A bare launch needs no installed role.
subagent({ name: "Auth summary", model: "<provider>/<fast-tier-id>", thinking: "low", task: "Summarize the authentication flow" });
// All return immediately; each result comes back independently.

Use ordinary panes for read-only agents. A single or sequential writer can work in the parent checkout; give each parallel independent writing agent a unique managed worktree. The parent acts as coordinator: decompose work, give each child one bounded outcome with its goal, allowed files, verification, and commit instruction, and keep dependent writes sequential. Children are leaves by default; the parent owns integration and final verification. See Worktree subagents.

Release notes

CHANGELOG.md is generated from Git history on each release. Hand-written upgrade notes for breaking releases live here.

3.0.0

3.0.0 makes the host pack-neutral. It also adds conditional task-model writes (expectedConfigRevision, an advisory lock, and the reported configRevision), the pi-herdr-agents operating skill, and subagent_cancel.

Breaking changes. The package no longer ships the seven former bundled roles, /plan and its plan skill, or /skill:orchestrate; these moved to optional role packs. /iterate, /btw, and /btw-close were removed without replacement. roles.bundled is now a deprecated no-op. A named launch of a role that no definition or installed pack supplies now fails before Herdr creates a pane or worktree.

Migration.

• Install pi-herdr-roles for the six generic roles (scout, planner, worker, reviewer, adversarial-reviewer, visual-tester), /plan, and /skill:orchestrate.
• Install pi-herdr-pstack for the poteto-mode methodology skill and command. It ships no named roles; its delegates are bare.
• Remove roles.bundled from $PI_CODING_AGENT_DIR/herdr-agents/config.json.
• Alternatively, copy a former role's definition into .pi/agents/ or the global agents directory to keep it without a pack.

See Migrating from bundled roles for the full mapping and pack-compatibility notes.

How it works

A subagent call selects the target checkout, reuses its Herdr workspace, and gives the child a pane in an extension-owned Agents tab. Four panes fit in each tab by default; overflow opens another tab in the same workspace. A worktree is created only when explicitly requested for checkout isolation. The call launches a child Pi session and returns started. The parent watcher combines Herdr process state with child activity details and projects the result into a live widget:

╭─ Subagents ──────────────────── 1 active · 1 open ─╮
│ 00:23  Scout: Auth (scout)        active · read 7m │
│ 00:45  Reviewer (reviewer)              waiting 2m │
╰────────────────────────────────────────────────────╯

When the child completes, the parent receives one bounded subagent_result message and starts a new turn with that result in context. Disposable ordinary panes close after result delivery; Herdr removes a tab when its last pane closes. Persistent specialists keep their pane between tasks, and managed worktree roots return to retained interactive shells. Callers never need to poll, tail session files, or wait in a shell loop.

Troubleshooting completion delivery

If a child finishes but the parent returns an empty or unrelated response, first verify that the result reached the parent session:

jq -c 'select(.type == "custom_message" and .customType == "subagent_result")' "$PI_SESSION_FILE" | tail -1

If the entry exists, spawning and result extraction worked; investigate parent wake-up and model-facing delivery rather than the child process. Completion wake-ups must contain the bounded result directly—do not send a separate message that merely tells the parent to look at an adjacent custom message.

Git package refs are pinned. To move an installed development copy back to the current main, install that ref explicitly and reload the active Pi session:

pi install git:github.com/giuseppecrj/pi-herdr-agents@main
# Then run /reload inside Pi.

Smoke-test delivery with an autonomous subagent instructed to return one exact marker. Success means the marker itself—not only a generic wake-up notice—automatically appears in the parent turn.

Subagent tabs, panes, and worktree workspaces are created without stealing keyboard focus. Launch commands target child panes by explicit ID, so focus and command delivery are independent. If a fresh or resumed launch fails, the extension closes the ordinary pane that it created and preserves the original launch error. It does not close a caller-supplied surface, and managed worktree workspaces remain retained on failure. Note: the interactive option controls parent status notifications, not terminal focus.

What's Included

Extensions

Subagents — 10 parent-session tools + 3 commands, plus 2 child-only tools:

| Tool | Description | | -------------------- | ------------------------------------------------------------------------------------------- | | subagent | Spawn a sub-agent in a dedicated herdr pane (async — returns immediately) | | subagent_interrupt | Interrupt a running Pi-backed subagent's current turn | | subagent_cancel | Cancel a running ordinary subagent: no fallback, one cancelled result after confirmed termination | | subagent_send | Deliver a follow-up task to an idle persistent specialist | | subagent_stop | Gracefully stop a persistent specialist after its active task settles | | subagents_list | List available agent definitions | | worktree_list | Parent-only inspect-only inventory of managed worktrees and cleanup blockers | | worktree_remove | Parent-only explicit removal by target path, branch, or workspace ID; optional preserve: true commits dirty state first | | subagent_resume | Resume a previous Pi-backed sub-agent session in a new ordinary pane (async) | | subagents_write_task_models | Parent-only internal tool that validates and atomically writes models.tasks preferences, optionally conditional on expectedConfigRevision and carrying an unsaved ranking basis |

| Skill | Description | | ----- | ----------- | | pi-herdr-agents | Operating guide for this host: launching, supervising, interrupting, cancelling and resuming children, worktrees, persistent specialists, model routing and configuration. Loaded by agents on demand; see skills/pi-herdr-agents/SKILL.md |

| Pi child-only tool | Description | | ---------------- | ------------------------------------------------------------------------- | | caller_ping | Ask the parent for help; ordinary children exit, persistent specialists stay alive | | subagent_done | Mark an interactive child complete and exit; autonomous agents auto-exit |

| Command | Description | | -------------------------- | ------------------------------------ | | /worktree <name> [task] | Continue this session in a new managed worktree (/worktree list lists them) | | /subagent <agent> <task> | Spawn a named agent directly (/subagent list lists available agents) | | /subagents-init [preferences] | Draft task-category model preferences from the live authenticated registry, with optional ranking preferences; a loaded approval extension can own the write |

Taxonomy and discovery

This package distinguishes directly runnable agent roles, Pi-native skills, and authenticated Pi runtimes. A multi-stage user outcome may be a command or skill that composes roles; it is not itself an agent role.

The host's own orchestration surfaces are:

| Surface | Entry point | Behavior | | --- | --- | --- | | Delegation | subagent, /subagent <agent> <task> | Launches one bare or named child; results return automatically. | | Worktree handoff | /worktree <name> [task], /worktree list | Forks the active conversation into a managed worktree. |

Planning, review, and other multi-stage workflows belong to role packs and skills, not to this package. See ADR-0002, ADR-0009, and ADR-0013.

Roles and role packs

This package is a pack-neutral execution host: it ships no agent roles, no /plan command, and no planning or review skills. Its one skill, pi-herdr-agents, is a general operating guide for the host. A named launch resolves agent from project definitions, global definitions, and roles registered by installed role packs, in that precedence order (see below). An empty catalog is valid. A bare launch without agent always works. An explicitly named role that is missing or invalid fails before Herdr creates a pane or worktree; it is never silently replaced by a bare agent.

All subagents execute through Pi. Claude models remain available through normal Pi provider/model routing. Legacy role definitions that contain cli fail before Herdr creates a pane or worktree; remove cli and cli-model, then select an authenticated Pi provider/model-id.

Role packs own their roles' prerequisites, such as a required skill, and must document them. This package does not install role packs or prerequisites.

Migrating from bundled roles

Earlier versions bundled seven roles and two workflows. They have moved to separately installed packs or have been removed. Nothing is installed automatically, and your global and project role files and config.json are not modified.

| Former bundled surface | Status | Destination | | --- | --- | --- | | scout, planner, worker, reviewer, adversarial-reviewer, visual-tester roles | Moved | pi-herdr-roles role pack | | /plan command and its plan skill | Moved | pi-herdr-roles | | /skill:orchestrate and its adversarial-review resources | Moved | pi-herdr-roles | | poteto role | Removed; replaced by the poteto-mode skill and /poteto-mode command | pi-herdr-pstack | | /iterate | Removed, not relocated | Call subagent({ name, task, fork: true, interactive: true }) directly | | /btw, /btw-close | Removed, not relocated | None | | roles.bundled setting | Deprecated no-op | Remove it from config.json |

Both destination packs are experimental, unpublished candidates; their install sources and compatible host versions are not final. Install a role pack with pi install like any Pi package and enable it beside this extension; a role pack stays inert without it. Older versions of this package that still bundle these roles, /plan, or orchestrate collide with the replacement packs: a bundled role rejects a same-named pack role, and Pi resolves duplicate commands and skills by discovery order. Setting roles.bundled: false on an older host only removes its roles, not its command or skill. Upgrade the host instead of combining an older host with the replacement packs.

Existing models.agents preferences keyed by role name keep applying when a pack or definition supplies that name. To keep a former role without installing a pack, copy its definition into .pi/agents/ or the global agents directory.

Roles use model defaults from config.json when configured; otherwise they inherit the parent model. Thinking defaults still come from agent frontmatter or the parent level. This resolution chain remains available as a fallback, but orchestrators should explicitly set each child's exact authenticated provider/model-id and supported thinking level. Select the model tier first: fast for bounded mechanical work and recon, mid for ordinary implementation or review, and frontier for architecture, security, hard diagnosis, or adversarial review. Then select thinking within that model's supported range. Cross-family independent review requires a reviewer from a different model family than the author. For ordinary review, prefer a different authenticated model family. When no other authenticated model family is available, ordinary review may use a same-family reviewer in a fresh standalone session. Disclose that this review is context-isolated, not cross-family independent. Cross-family verification must not use this fallback. A stronger model in the same family is a quality escalation, not cross-family independent review. Family is the independence boundary; project policy may separately require a different provider.

Discovery loads definitions in package → global → project order, so effective priority remains project (.pi/agents/) > global ($PI_CODING_AGENT_DIR/agents/, defaulting to ~/.pi/agent/agents/) > package. Package definitions are the roles contributed by installed Pi role packs; there is no bundled layer. Both subagents_list and /subagent list show each visible definition's source; contributed roles include their package identity, for example (package:@acme/security-roles). A hidden higher-priority definition still suppresses a visible lower-priority definition.

Custom roles and installable role packs are the package's main extension points. See Custom Agents for the complete create, package, verify, and launch workflow.

Async Subagent Flow

1. Agent calls subagent()          → returns immediately ("started")
2. Sub-agent runs in herdr pane    → widget shows live status
3. User keeps chatting             → main session fully interactive
4. Sub-agent finishes              → result steered back as a normal completion/failure
5. Main agent processes result     → continues with new context

Multiple subagents run concurrently — each steers its result back independently as it finishes. Active watchers survive parent /reload, /new, /resume, and /fork transitions, so completion is delivered into the replacement session. Quitting Pi still stops parent-side delivery. The live widget above the input tracks every agent still in flight:

╭─ Subagents ──────────────────── 1 active · 2 open ─╮
│ 01:23  Scout: Auth (scout)             active · read 7m │
│ 00:45  Reviewer (reviewer)                   stalled 4m │
│ 00:12  Scout: DB (scout)                      starting… │
╰─────────────────────────────────────────────────────────╯

Completion messages render with a colored background and are expandable with Ctrl+O. Results larger than 16,000 characters are abbreviated in the parent context while preserving their beginning, conclusion, and session path; the complete result remains in the child session. The extension includes that bounded result and a continuation instruction directly in the single custom subagent_result message that triggers or steers Pi, avoiding empty turns caused by a separate context-free wake-up. The renderer uses the unadorned bounded result from structured details. Completed rows are removed from the widget as soon as their result is delivered or suppressed.

In-progress status updates

The widget projects each sub-agent from a process + turn lifecycle:

• Herdr pane inspection is the coarse authority for whether the child process is present and whether Herdr reports it as idle, working, blocked, or done.
• Child activity snapshots enrich the label with Pi-only detail (tool name, streaming, etc.) when available.
• Session JSONL is still used for transcript, resume, lineage, and result extraction — not for liveness.

Projected labels include:

• starting — launched; pane/activity confirmation is still settling
• active — processing work (agent turn, provider request, streaming, or tool execution)
• blocked — Herdr reports the child as blocked
• waiting — turn finished; the process is intentionally open for more input or another stage
• interrupted — the current turn was cancelled (Escape / subagent_interrupt); the process stays open and is not treated as active processing
• stalled — pane inspection is unhealthy long enough that the parent can no longer trust the run
• running — fallback when only coarse process presence is known (e.g. non-Pi backends)
• finalizing — completion was observed and delivery is in progress; the process elapsed timer freezes here
• cancelling… / cancel unconfirmed — a subagent_cancel is terminating the run, or its termination could not be confirmed and the run stays live

The widget header counts active vs open:

• active — active, starting, running, or blocked
• open — everything else still tracked (waiting, interrupted, stalled, finalizing, …)

When activeCount === 0 (every tracked row is open), the border uses an amber accent. Process elapsed time (MM:SS on the left) freezes when the process reaches finalizing/completed/failed. Interrupt does not freeze that process clock; the interrupted state shows its own duration on the right while the process remains open.

A fixed internal watchdog marks a run as stalled when pane inspection fails or the pane disappears without a completion sidecar; valid long-running active or waiting states do not become stalled just because time passes. When a run enters stalled or recovers from it, the parent agent receives a steer message so it can react. All other status transitions stay in the widget only.

Interactive subagents stay silent. Long-running user-driven subagents (for example, an interactive planning role or a bare interactive: true fork) do not wake the parent session on stalled/recovered transitions — the user is working directly in the subagent's pane, and a steer message there would just burn an orchestrator turn on a no-op "still waiting" ping. The widget still updates normally, and activity snapshots are still recorded/classified regardless of the interactive setting. By default, agents with auto-exit: true are treated as autonomous and get stall pings; agents without it are treated as interactive and stay quiet. Override per-agent with interactive: true|false in frontmatter, or per-spawn with interactive: true|false on the tool call.

Configuration

The durable user configuration is $PI_CODING_AGENT_DIR/herdr-agents/config.json, defaulting to ~/.pi/agent/herdr-agents/config.json. It is not read from the installed package root, so npm and git package upgrades do not overwrite it. Create it by copying the installed package's config.json.example, or run /subagents-init to seed and draft model task preferences. This is a breaking migration: manually move an existing package-local config.json to this path, or re-run /subagents-init.

{
  "status": {
    "enabled": true
  },
  "models": {
    "agents": {}
  },
  "persistent": {
    "maxAgents": 3
  },
  "supervision": {
    "forcePolling": false,
    "hangWarningMinutes": 15
  },
  "panes": {
    "mode": "grouped",
    "direction": "right",
    "maxPerTab": 4
  }
}

If config.json is absent, status, role, pane, and persistent-specialist settings fall back to config.json.example. Model routing does not read the example: no model overrides apply until a real config.json exists.

The copyable example is model-neutral, so it works without requiring credentials for a specific provider. To configure models, replace the empty section with exact IDs from your authenticated model catalog:

{
  "models": {
    "default": "your-provider/your-default-model",
    "agents": {
      "scout": "your-provider/your-fast-model",
      "reviewer": "your-provider/your-review-model"
    },
    "tasks": {
      "coding": ["your-provider/your-coding-model"],
      "review": ["your-provider/your-review-model"],
      "recon": ["your-provider/your-fast-model"],
      "qa": ["your-provider/your-qa-model"],
      "architecture": ["your-provider/your-architecture-model"],
      "docs": ["your-provider/your-docs-model"]
    },
    "tasksMeta": {
      "generatedAt": "2026-09-17T00:00:00Z",
      "method": "research"
    }
  }
}

models.tasks candidates are ordered exact authenticated IDs. Use task:<category> only in the subagent tool's model argument; it is not valid in frontmatter or model defaults. Cross-family independent review requires a reviewer from a different model family than the author. For ordinary review, prefer a different authenticated model family. When no other authenticated model family is available, ordinary review may use a same-family reviewer in a fresh standalone session. Disclose that this review is context-isolated, not cross-family independent. Cross-family verification must not use this fallback. Use an exact authenticated shortlist provider/model-id when the authoring family is known; task:review does not establish independence. Family is the independence boundary; project policy may separately require a different provider. This is guidance, not extension enforcement.

Run /subagents-init [preferences] to draft task-model preferences. For example:

/subagents-init Prefer capability over price for implementation; keep recon inexpensive

The command supplies a sanitized snapshot of all available models from the active session registry, including extension-registered providers, exact IDs, display names, reported base token costs, context/output limits, input modalities, reasoning, and supported thinking levels. Safe extension-registration and auth-source metadata is included when Pi exposes it; credentials, endpoints, and raw auth labels are not. Configured authentication does not prove account access or a successful request. Missing costs remain unknown; reported zero does not mean free, and OAuth does not establish subscription billing. The brief uses compact JSON without truncating models and reports its model count and JSON character count (not a token estimate); large catalogs still consume context. This is the current synchronous snapshot: a dynamic provider whose initial catalog refresh has not completed might be absent. Init does not refresh providers or probe the network for availability. The brief also includes configRevision, the revision of the exact config bytes that supplied the saved preferences. Init does not start while a turn runs or messages are queued, and an empty registry ends with a notice instead of a prompt. Neither case writes anything.

The draft considers current saved task, default, and per-agent preferences. Optional command arguments set ranking preferences. Otherwise it favors capability for substantive work and efficiency for bounded reconnaissance and test execution. Categories describe work, not complexity tiers:

| Category | Work | | --- | --- | | coding | Implementation workers | | review | Code reviewers | | recon | Reconnaissance scouts | | qa | Software and test runners | | architecture | Planning and diagnosis | | docs | Documentation workers |

Init asks the agent to research major candidates across providers using primary sources, disclose uncertainty and notable exclusions, and avoid duplicate upstream models across routes unless deliberate redundancy is explained. Display names help identify candidates but, like aliases, do not prove upstream equivalence; research is still required. Price or context size alone is not quality evidence. No live model probes run.

The proposal carries a ranking basis. It is {"kind":"registry-only"} when search is unavailable or yields no usable evidence. It is {"kind":"research"} only with the http(s) sources consulted in this run, the influence each had on the ranking, and the remaining uncertainty. Code checks the basis's shape, not whether a source was read, so the basis is labeled as submitted and unverified.

While init runs, it emits pi-herdr-subagents:task-models:init:approval:v1. With no offer, the prompt directs the model to subagents_write_task_models with expectedConfigRevision. When exactly one loaded extension offers, init opens that extension's flow, and the extension's instructions replace the writer instruction; the extension then owns approval and the write. Two or more offers, or an invalid one, stop init before any extension opens or a prompt is sent. Once an offer is recorded, a refusal or failure also stops init: it never falls back to the direct writer. See Task-model init events.

The writer validates and atomically replaces models.tasks and tasksMeta, preserving unrelated settings. Its tool schema accepts partial nonempty categories (omitted categories are removed), rejects empty tasks: {} input, and rejects exact duplicate refs within a category after trimming; IDs remain case-sensitive. Its result includes normalized saved tasks, tasksMeta, configPath, missingCategories, and configRevision. Init requests all six categories and a before/after table based on that saved result, not the unsaved draft, and must explain missing categories or changed choices.

Optional basis must have the same kind as tasksMeta.method. A research basis needs at least one source with an http(s) URL that has a host and a nonblank influence, plus a nonblank uncertainty. The writer validates it before writing, returns it in the result, and never saves it; only tasksMeta.method persists. Calls without basis, including research calls, keep the earlier contract.

Optional expectedConfigRevision makes a write conditional on the config the proposal was read from. A revision is sha256: followed by 64 lowercase hex digits of the SHA-256 of the exact config.json bytes (not normalized JSON or only models.tasks), or the literal missing when the file is absent. Any byte change, including whitespace or unrelated settings, makes the revision stale; a missing revision rejects a file that now exists, and an existing revision rejects a file that was removed. A stale revision fails with Stale task model config revision without replacing configuration: re-read, re-propose, and re-approve rather than retrying. Malformed revisions, including null and empty strings, fail closed. Omitting the field keeps the unconditional write. A write to a missing file still seeds from the packaged config.json.example. configRevision is the revision of the exact bytes written and can serve as the next precondition.

Every writer call, conditional or not, holds an exclusive config.json.lock sibling while it reads one snapshot, checks the revision, and atomically renames a private temporary file into place. A held lock fails immediately with Task model config writer busy; there are no waits or retries. The lock is never broken automatically: after a crash, the error names the recorded owner and reports when that process is no longer running, and you remove the lock only after confirming that no writer is active. Each call removes only its own lock and temporary file. This lock is advisory: it serializes cooperating writers but cannot constrain a text editor or another process that ignores it. The revision check detects changes made before the snapshot is read; it is not a filesystem transaction against arbitrary external writers.

task:<category> values select subagent models; they are not slash commands and do not change the parent model. Ordered authenticated candidate plans resolve before launch. Ordinary nonpersistent runs can retry later candidates after launch failure or after a running child settles with a provider/agent error, not after a completed negative task result. Persistent specialists do not advance after a running-child error. This is not per-step routing; worktrees use the first authenticated candidate only, without fallback retries. Shortlists do not enforce reviewer independence. Cross-family independent review requires a reviewer from a different model family than the author. For ordinary review, prefer a different authenticated model family. When no other authenticated model family is available, ordinary review may use a same-family reviewer in a fresh standalone session. Disclose that this review is context-isolated, not cross-family independent. Cross-family verification must not use this fallback. Another route to the same family is not independent review. Family is the independence boundary; project policy may separately require a different provider. Run /reload (or start a new session) after writing preferences.

Set persistent.maxAgents to the maximum concurrently retained persistent specialists. It defaults to 3; a persistent spawn at the cap is rejected before Herdr creates a pane or workspace, and no specialist is evicted.

roles.bundled is deprecated and has no effect because this package ships no roles. Existing true and false values are accepted so that current configuration keeps loading; a parent session reports one warning per extension load naming the file and asking you to remove the key. The extension never rewrites the file. Other values remain configuration errors, as do unknown keys under roles. Registered role packs are the entire package layer, and global and project definitions keep their precedence over them.

Task-model init events

Two versioned pi.events channels let another extension take part in /subagents-init without importing this package. Both carry the invoking command's live ExtensionCommandContext, so they are for trusted extensions in the same process. The context is current only while the event is emitted.

The host emits pi-herdr-subagents:task-models:init:approval:v1 while init runs:

type ApprovalRequest = {
  apiVersion: 1;
  brief: TaskModelBrief; // frozen: operatorPreferences, categories, configRevision, current, models (never empty)
  context: ExtensionCommandContext;
  offer(offer: { owner: string; open(): OpenResult }): "recorded" | "closed";
};
type OpenResult =
  | {
      kind: "ready";
      destination: { toolName: string; instructions: string };
      cancel(): void; // closes the flow this open created
    }
  | { kind: "blocked"; reason: string };

• Call offer synchronously, before the listener's first await. After emit returns, offer answers "closed" and records nothing.
• Offer whenever your extension owns task-model approval, even when it is busy, and refuse from open. owner is one printable token without spaces.
• The host opens nothing until collection ends. Two offers (even with the same owner) or an invalid offer stop init. With one offer, the host calls open() once, and it must return synchronously.
• ready must leave toolName active. Its instructions replace the host's writer instruction in the prompt the host sends.
• ready must include cancel. When the host does not hand the prompt to Pi (the tool is inactive, or pi.sendUserMessage throws), it calls cancel() once before reporting not-started. cancel must close only the flow that this open created, never a later one. A cancel that throws is reported with the outcome.
• A throw, a promise, an invalid result (including a missing cancel), or blocked stops init with a notice and no prompt. The host cannot undo what open changed before failing.
• Pi accepts the prompt asynchronously. It can still reject it after start reports started, for example when the selected model has no configured auth; the host refuses up front only when no model is selected. Pi delivers the prompt to input handlers (source "extension") before that check and emits before_agent_start and agent_start only after it passes. An extension whose flow outlives one prompt should close it when other input or a run arrives before those events, so a later request cannot inherit it.
• Pi's event bus logs and swallows listener exceptions. A listener that throws before offering looks the same as no listener, so init uses the direct writer. An extension that guards the writer still refuses that write.

Another extension's command emits pi-herdr-subagents:task-models:init:start:v1 to start the same init:

type StartRequest = {
  apiVersion: 1;
  context: ExtensionCommandContext;
  preferences: string; // ranking preferences, passed through as typed
  offer(offer: { owner: string; start(): InitOutcome }): "recorded" | "closed";
};
type InitOutcome =
  | { kind: "started"; destination: string }
  | { kind: "not-started"; reason: string };

The host offers synchronously as pi-herdr-agents and starts nothing until the emitter calls start(). Call it only when exactly one host offered. start runs init with the request's context and returns synchronously; the emitter shows a not-started reason. The host does not listen in subagent sessions and unsubscribes at session_shutdown. Unsubscribe an approval listener there too, as the role-pack bridge does.

Sidebar markers

Set sidebar.enabled to true to mark delegated children for the optional Pi Herdr Agents Sidebar Herdr plugin, whose Focus action hides marked children from Herdr's Agents view. The setting defaults to false.

{
  "sidebar": {
    "enabled": true
  }
}

Add the key to your existing config.json. The extension rejects a config.json without its status section, so copy config.json.example first if the file does not exist. Run /reload after changing the setting. Children launched before the change stay unmarked.

Setting sidebar.enabled to false and running /reload retires every existing marker and clears the markers of running children. A clear can fail to reach Herdr, so a token still expires at most 15 seconds after the last write Herdr accepted. Setting it back to true marks only later launches; a child that was already running when you turned it off is never marked again.

With markers on, every fresh, resumed, and persistent child records its own process identity, as managed worktree children always do. After the parent verifies that identity, it reports the pane token piha_delegated_v1 with the value live, a 15-second time-to-live, and an increasing sequence number through HERDR_SOCKET_PATH. It renews the token on supervision checks while that exact process is alive. It clears the token when the run finalizes or is suppressed, or when the process exits or can no longer be verified. A /worktree handoff session is never marked.

Marker writes never fail a launch or keep Pi running. Each write has a 2-second timeout on an unreferenced socket, failed writes are not retried, and three failures in a row stop that child's marker. Process identity reads Linux /proc, so on other systems, and outside Herdr, nothing is marked. After a parent crash, a marker can remain for up to 15 seconds after the last write Herdr applied. A marker belongs to the pane, so a different process started by hand in a delegated pane stays hidden by Focus while the original process is alive and supervised, even if it is suspended. Run All first or use a fresh pane. See what a marker proves and reusing a delegated pane.

Supervision transport

On supported local filesystems, supervision uses file wake-ups plus one shared 4.8-second pane reconciliation. A wake-up only prompts fresh evidence collection; it never establishes a result by itself. While any child's wait for that check is parked, the coordinator's file watches stay referenced, so the sidecar is still observed when the parent has no other event-loop work. The reference is released when the wait settles, is aborted, or its child is unregistered. With no parked wait, the watches stay unreferenced and do not keep the process running. If the watcher or shared pane inspection becomes unavailable, supervision quietly returns to the legacy one-second polling cadence. No caller action is required.

Set supervision.forcePolling to true in the durable user config.json to disable wake-ups and use that legacy cadence deliberately. The setting is read when the coordinator is created, so run /reload after changing it. subagents_list reports the active transport mode (wake+batch, polling(forced), or polling(fallback)) and watcher count.

supervision.hangWarningMinutes defaults to 15; set it to 0 to disable no-progress advisories. For example, this keeps the default transport and sets a 30-minute advisory budget:

{
  "supervision": {
    "forcePolling": false,
    "hangWarningMinutes": 30
  }
}

While a child projects active or blocked, the parent compares durable session JSONL and activity-snapshot updates against this budget. An advisory is warning-only, fires once per no-progress episode, and never interrupts, kills, retries, or restarts a child. It identifies blocked-tool (an outstanding tool call may still complete), truncated-turn (an observed toolUse stop with no tool call; its cause is unknown), or generic-no-progress when neither condition is established, then includes the session path and manual recovery options. Ordinary children can be interrupted, cancelled with subagent_cancel, or, after termination, resumed or newly spawned. Persistent ordinary-pane specialists can be interrupted or stopped with subagent_stop and replaced; they cannot be resumed. Managed-worktree children, including persistent ones, retain their workspace and continue there only after the previous process has exited; do not use subagent_resume or start a concurrent writer. Interactive children stay quiet just as they do for stalled/recovered notices; their widget state still updates. A later durable update clears the episode and sends the corresponding recovered notice for non-interactive children. polling(fallback) means at least one tracked child is using per-child polling; other children can still use wake+batch.

A Linux manual benchmark on 2026-09-06 used isolated Herdr panes held pending, 20-second windows, and the extension's completion/supervision seams. At 10 children across three rotated rounds, wake+batch averaged 2.20 CLI launches/s versus 14.20 for forced polling (84.5% fewer); mean evidence-to-resolver latency was 3.2 ms versus 449.0 ms, and the largest reconciliation probe gap was 4.82 s. The benchmark measures /proc CPU ticks for the supervisor and isolated Herdr tree, not parent-model latency; raw samples are written to /tmp/issue29-bench/ by test/bench/supervision-bench.mjs.

panes.mode defaults to "grouped" when omitted. Ordinary public subagent and subagent_resume launches, including bare forks, fill extension-owned Agents, Agents 2, etc. tabs in the target checkout's existing workspace. panes.maxPerTab is a positive safe integer, defaults to 4, and counts all live panes in each owned tab, including user-added panes and retained shells. Overlapping launches in one parent respect this cap. It is independent of persistent.maxAgents.

Checkout matching uses Herdr's canonical worktree.checkout_path and includes descendant directories. Shell working directories do not establish workspace ownership. If no checkout matches (including non-Git directories), placement uses the caller's workspace; overflow never creates a workspace. A reviewer with cwd set to a managed checkout joins that workspace without creating another worktree. Resume placement uses the saved session's cwd.

Explicit panes.mode: "tab" preserves one new tab per ordinary child in the caller's workspace. Explicit "split" preserves splits of the stable parent pane. panes.direction is "right" (default) or "down" and applies to grouped and legacy splits. maxPerTab does not affect these legacy modes. Managed worktrees retain their separate workspaces.

Ownership is tracked by returned pane/tab/workspace IDs, never labels. Separate parent processes own separate groups; /reload preserves a parent's in-memory ownership, but a full restart does not adopt old tabs. Placement never moves existing panes or renames user tabs. Background launches preserve focus; Herdr may resize sibling panes when splitting or closing. User-added panes are never closed by automatic tab cleanup. An owned tab remains reusable while user panes remain, even after all child panes close.

Run /reload after changing role, model, or pane settings.

models.default sets the model for subagents that do not specify a model. models.agents sets per-agent defaults, keyed by the agent name passed to subagent({ agent: ... }). Explicit model tool arguments take precedence, followed by agent frontmatter, per-agent config, the global default, and finally the parent model. Model values must be exact authenticated provider/model-id references. A value can contain an ordered comma-separated fallback list, for example provider/preferred, provider/fallback. The tool argument also accepts task:<category> as its complete value (not in a list), for configured coding, review, recon, qa, architecture, or docs preferences. The extension validates every candidate before launch, then launches later candidates only after the selected child settles with a provider/agent error. Pi owns any automatic transient retrying inside that child; the extension does not infer retry counts or permanence from the error text. A later candidate that launches after a parent /reload, /new, /resume, or /fork uses the live parent session for its artifacts and lineage, as completion delivery does. Its candidate list and thinking level stay those of the original call. The original parent directory and process directory remain the bases for directory resolution. Role files are read again for each attempt, so a changed role cwd can redirect a fallback when the tool call did not specify cwd. If no live parent context is available, that candidate fails without launching. A completed child result, including a negative task result, never switches models. Completion metadata reports the requested candidate, every attempted candidate, the model actually used, and each raw model failure in attempt order when fallbacks are tried.

A catalog-listed model and configured authentication do not prove that the active provider account can use that model. Providers may reject an account / model combination only when the request is made. The completion preserves each raw provider reason with its model and suggests checking account access, spawning a new subagent with a supported model, or choosing an appropriate configured fallback. subagent_resume does not select a model and should be used only after the session's stored model is usable. Persistent session sidecars fail closed: v1 does not resume or revive a stopped or crashed specialist; retain its evidence and spawn a new specialist. The completion does not claim a permanent failure or a retry count that Pi has not exposed. Reliable structured permanence and retry counts require an upstream Pi/ExtensionAPI diagnostics seam for final provider errors and retry outcomes.

config.json is durable user state under the Pi agent directory and is loaded when the extension starts. Run /reload after changing it. Package-root config.json files are ignored; move them manually or re-run /subagents-init.

Spawning Subagents

Examples that set agent assume a role pack or a project/global definition supplies that role; this package ships none.

// Explicit fast-tier runtime for bounded reconnaissance
subagent({ name: "Scout", agent: "scout", model: "<provider>/<fast-tier-id>", thinking: "low", task: "Analyze the codebase..." });

// Force a full-context fork for this spawn
subagent({ name: "Fix", fork: true, model: "<provider>/<mid-tier-id>", thinking: "medium", task: "Fix the bug where..." });

// Explicit frontier-tier runtime for architecture work
subagent({ name: "Planner", agent: "planner", model: "<provider>/<frontier-tier-id>", thinking: "high", task: "Work through the design with me" });

// Explicit mid-tier runtime with a custom working directory
subagent({ name: "Designer", agent: "game-designer", model: "<provider>/<mid-tier-id>", thinking: "medium", cwd: "agents/game-designer", task: "..." });

// Isolated ticket branch in a Herdr-managed Git worktree
subagent({
  name: "Ticket 123",
  agent: "worker",
  model: "<provider>/<mid-tier-id>",
  thinking: "medium",
  worktree: { branch: "ticket/123", base: "main" },
  task: "Implement ticket 123, test it, and commit the result",
});

Parameters

| Parameter | Type | Default | Description | | ---------------------- | ------- | -------------- | ------------------------------------------------------------------------------------------------- | | name | string | required | Short stable child label; coordinated groups use <task>-<role>[-n] (widget and pane title) | | task | string | required | Task prompt for the sub-agent | | agent | string | — | Load defaults from agent definition | | fork | boolean | — | Override the child session mode: true forces fork, false forces standalone. Omit to inherit the agent session-mode frontmatter | | persistent | boolean | false | Keep one specialist session alive for sequential tasks; follow-ups use subagent_send only | | interactive | boolean | derived | Mark this spawn as interactive (don't wake the parent on stall/recovery). Defaults to the agent's interactive frontmatter, otherwise the inverse of auto-exit. | | model | string | configured or parent | Exact authenticated provider/model-id, ordered fallback list, or whole-value task:<category> (coding, review, recon, qa, architecture, docs). Task routing is tool-only; worktrees use its first authenticated candidate. Resolution is tool argument → agent frontmatter → per-agent config → global config → parent | | thinking | string | parent level | Pick the model tier first, then set thinking within that model's range: minimal/low for bounded mechanical work, medium for ordinary implementation or review, high+ for architecture, security, or hard diagnosis. Omitting still inherits the parent level; this is a discouraged fallback for orchestrated children. | | systemPrompt | string | — | Role text for a bare spawn, delivered as a role block at the top of the child's first message (not the system prompt); dropped for fork: true children. Named agents keep their definition body | | skills | string | — | Comma-separated skill names | | tools | string | — | Comma-separated tool names | | cwd | string | — | Working directory, or source repository when worktree is set (see Role Folders) | | worktree | object | null | — | Isolated Herdr-managed Git worktree; requires branch, with optional base (committed HEAD by default). Omit or pass null to use an ordinary pane in cwd when a client requires the property. |

A bare spawn's systemPrompt is not passed to Pi as a system prompt. The host prepends it as a role block to the child's first message, which is delivered through a task artifact file referenced with @path. A full-context fork (fork: true) receives only the raw task, so its systemPrompt is dropped. Set fork: false when a bare child must receive reference or role text through systemPrompt.

Naming coordinated children

Before launching a new group, choose a short task slug and label each new child <task>-<role>[-n], such as login-api or login-test2. Roles are plan, research, ui, api, build, test, review, browser, security, perf, and merge. Leave existing labels unchanged. After the final launch, print name | agent kind | role | model | worktree and use each name in prompts, handoffs, and results.

Isolated worktree runs

Use one worktree per parallel independent writing task; a single or sequential writer can work in the parent checkout, and read-only agents use ordinary panes. Omit worktree for an ordinary pane; clients whose generated tool schema requires every property may send worktree: null with the same effect. cwd selects the source Git repository, branch must be unique, and base is resolved to an exact commit before creation. If cwd is a linked checkout, Herdr provisioning uses the principal checkout while the requested checkout supplies the base SHA and manifest provenance. A successful launch from that linked checkout does not itself authorize cleanup there: cleanup checks the canonical principal/source repository under the invoking parent session's cwd, not manifest.sourceCwd or shared Git identity. If cleanup is needed, start the parent Pi session rooted at the principal checkout or an ancestor containing it, then use the normal explicit cleanup flow; changing directories inside an existing Pi session does not change its session cwd. If base is omitted, the source checkout's committed HEAD is used. Parent-checkout changes that have not been committed are not copied.

Choose a worktree from the task, not from a role name: the extension emits no role-specific worktree warnings. Read-only scouting and review normally use an ordinary pane; to inspect or review an existing worker result, start an ordinary child in that retained worktree path. Do not infer that a role cannot write because its tools omit write or edit: a read,bash allowlist is not an enforced read-only boundary because shell commands can mutate files. Report-only roles must restrict Bash to safe inspection and avoid artifact-generating verification in the reviewed checkout. Herdr worktree workspaces persist until explicitly removed.

The child starts at the returned worktree root. Tell writing agents to test and commit when you want a commit-based handoff, and tell them not to push, merge, switch branches, or remove the worktree. The parent owns review and integration.

Successful, failed, and help-requesting worktree runs retain their workspace and root shell. A reviewer's disposable pane can close without closing that root, tab, or checkout. Completion includes the worktree path, Herdr workspace, branch, base/head SHAs, commits ahead, changed and untracked files, and clean/dirty/conflicted state. Here, clean means no uncommitted files; the branch may still contain commits. If Git inspection fails, state is reported as unknown rather than guessed.

An ownership manifest is written under the parent session's artifacts/<session-id>/worktree-runs/ directory before Herdr creates resources. V1 does not automatically recover watchers after a full process restart, and subagent_resume does not reattach the managed worktree lifecycle.

The extension does not push, create a PR, merge, cherry-pick, or remove the worktree or branch automatically. For task selection, lifecycle states, review commands, failure recovery, and safe cleanup, read Worktree subagents. The research report records the rationale and deferred roadmap.

Persistent specialists

Set persistent: true on a subagent launch to create one logical specialist with one v1 session generation. Its resolved tools, denied tools, model, thinking level, and optional worktree binding are snapshotted at launch and do not change when work is sent later. subagents_list shows each live specialist's logical ID, generation ID, state, completed-task count, and effective policy.

The initial task and each subagent_send({ id|name, message }) task are delivered exactly once with a task ID. A specialist accepts one task at a time. Sends while it is working are recorded as rejected-busy; no queue is retained. After a task result arrives, it is idle and accepts the next task. A persistent child's caller_ping records a help request but keeps the session alive; answer with subagent_send.

Use subagent_stop({ id|name }) to request graceful shutdown. If a task is active, stop becomes stop-pending and the task reaches its terminal outcome first. The parent reports stopped only after process-exit evidence is confirmed, then closes an ordinary pane it created and releases the name. If confirmation times out, the specialist is stalled in an unconfirmed-stop state: subagent_send rejects follow-up work while retaining evidence. Request subagent_stop again to make another bounded exit check, or spawn a new specialist. A pane or process disappearance without a stop directive produces one facts-only crash notice; persistent sessions cannot be resumed in v1, so spawn a new specialist. There is no automatic restart, replay, or revival.

A persistent specialist with a worktree holds that lease for its entire lifetime. It cannot be re-bound to another checkout. Otherwise it runs in an ordinary pane.

Interrupting a running subagent

Use subagent_interrupt to cancel the active turn of a running Pi-backed subagent:

subagent_interrupt({ id: "abcd1234" });
// or
subagent_interrupt({ name: "Scout" });

This sends Escape to the child pane, cancelling the in-progress model turn. The subagent session stays alive — the pane, session file, and background polling all remain intact. After the interrupt, the widget immediately labels the child as interrupted (counted as open, not active processing). Stale pre-interrupt activity snapshots are ignored so a lagging Herdr/active reading cannot overwrite the interrupt. The process elapsed timer keeps running because the pane is still open; only the interrupted-state duration freezes relative to the interrupt request. If the child starts work later, newer observations return it to active; completion, failure, and caller_ping still flow through normally.

id and name are each optional, but execution requires one usable target: an exact running ID or an exact, unambiguous display name. When both are supplied, id is used. Duplicate names are rejected.

This is a turn-level interrupt, not a method for forcibly terminating a subagent session. To end the run, use subagent_cancel.

Cancelling a running subagent

Use subagent_cancel to end one ordinary (non-persistent) managed run, including an interrupted one:

subagent_cancel({ id: "abcd1234" });
// or
subagent_cancel({ name: "Scout" });

Target resolution matches subagent_interrupt: an exact running ID or an exact, unambiguous display name. Persistent specialists are rejected; use subagent_stop, whose graceful v1 semantics are unchanged.

The cancel intent is recorded before anything is aborted or killed. From then on the run never advances its model shortlist, retries, or recovers, even when terminating the pane makes the watcher observe a lost pane, and even when a fallback launch was already in flight. The result reports one status:

| Status | Meaning | | --- | --- | | confirmed | Termination is confirmed. One cancelled result is delivered automatically. | | requested | A launch or fallback acquisition is still in flight. Its owner is terminated as soon as it is acquired; no later model is tried. | | unconfirmed | Termination failed (the error is reported). The run stays live, owned, and supervised, and nothing is delivered or cleaned up. Call subagent_cancel again to retry. | | already-terminal | The run already took a natural result or was retired; nothing was cancelled. |

Repeated cancels join an in-flight termination and keep the first request time. The parent receives exactly one subagent_result whose details carry error: "cancelled" and a cancellation record (requestedAt, termination, confirmedAt). Its message says the run was cancelled and lists any models already attempted. It is never presented as a provider failure. A child's natural result taken before the cancel stays authoritative. A cancel while a provider error is still eligible for fallback wins and stops that fallback.

Termination follows surface ownership:

• Ordinary pane: the pane is closed. Confirmation is Herdr reporting it absent; closing a Herdr pane terminates its terminal session, but this is not a separate OS process check. Other panes, tabs, and user panes are never closed.
• Managed worktree: the retained root pane, workspace, checkout, branch, commits, and manifest are kept. At launch, the child records its own process identity (PID, kernel start time, boot ID, and PID namespace) beside its session; the parent accepts it only while that process is alive in the parent's PID namespace as the Herdr pane shell or a descendant of it. SIGTERM goes only to that identity, re-verified immediately before the signal. Confirmation requires that identity to no longer exist, or the pane to be gone while it is not known alive. Command-line text and Herdr's foreground list are never evidence, because Pi rewrites its process title. A live identity (for example a suspended Pi), or one whose SIGTERM fails while it is still alive, leaves termination unconfirmed even if the pane is gone. An identity that was not captured (including on non-Linux hosts), is unreadable, or whose PID now names another process is never signalled and leaves termination unconfirmed unless pane absence confirms it. E