Summary: pi-delegate-mcp lets you delegate coding/repo tasks to background pi agents running on any of ~38 (often cheap) models, then monitor, steer, question, and follow up on them while their context stays out of your main conversation.
Initialize once (init) — discovers reachable models, permitted tools, and default delegate config; every other tool refuses until it runs.
Delegate in the background (spawn) — starts a pi agent, returns a sessionId immediately; poll later. Also run for a blocking quick question.
Fan out (spawn_batch) — launch up to 10 delegates in one validated call (batch-level model, cwd, tools, extensions, overridable per task); nothing starts if any task is bad.
Track sessions (status, sessions) — state, turn count, tools used, latest text, pending questions, and an ordered toolCalls trace (verbose adds call ids/results); sessions filters by state.
Steer mid-task (steer) — inject a redirect that lands after the current tool call, before the next model call, without aborting.
Answer agent questions (answer) — unblock a delegate waiting on a question surfaced by status (only possible with extensions: true).
Continue finished work (follow_up) — give a completed delegate another turn with everything it already read still in context.
Stop work (abort) — halt a session, with partial output still readable.
Manage history (forget) — drop a finished session to free its id; newest 50 are kept.
Choose models per call (models, model arg) — pi's scoped set plus custom providers; overrides PI_DELEGATE_MODEL, and unresolvable names are hard errors, never silent fallbacks.
Safe by default — tools locked to read, grep, find, ls at session construction; widen deliberately via PI_DELEGATE_ALLOW_TOOLS/ALLOW_WRITE, or add network reach via extensions (e.g. web_search) instead of bash.
Traceability — pass your own id and label so delegates are identifiable in traces and the status line.
Status line integration — a wrapper binary shows this workspace's delegates (running, waiting on questions, finished) in Claude Code, attributed by process lineage.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@pi-delegate-mcp Delegate a repo-wide search for all usages of the deprecated API to pi."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
██████╗ ██╗ ██████╗ ███████╗██╗ ███████╗ ██████╗ █████╗ ████████╗███████╗
██╔══██╗██║ ██╔══██╗██╔════╝██║ ██╔════╝██╔════╝ ██╔══██╗╚══██╔══╝██╔════╝
██████╔╝██║ ██║ ██║█████╗ ██║ █████╗ ██║ ███╗███████║ ██║ █████╗
██╔═══╝ ██║ ██║ ██║██╔══╝ ██║ ██╔══╝ ██║ ██║██╔══██║ ██║ ██╔══╝
██║ ██║ ██████╔╝███████╗███████╗███████╗╚██████╔╝██║ ██║ ██║ ███████╗
╚═╝ ╚═╝ ╚═════╝ ╚══════╝╚══════╝╚══════╝ ╚═════╝ ╚═╝ ╚═╝ ╚═╝ ╚══════╝
███╗ ███╗ ██████╗██████╗
████╗ ████║██╔════╝██╔══██╗
██╔████╔██║██║ ██████╔╝
██║╚██╔╝██║██║ ██╔═══╝
██║ ╚═╝ ██║╚██████╗██║
╚═╝ ╚═╝ ╚═════╝╚═╝
MCP server that exposes the pi coding agent as a delegable, steerable worker.
Point Claude Code (or any MCP host) at it and delegate work to any of pi's ~38 providers (DeepSeek, Grok, GLM, Kimi, Qwen, Codex, OpenRouter, local llama.cpp) with the sub-agent's context staying out of your main conversation.
Your main harness runs on an expensive model, with a context window you care about. A lot of what it does doesn't need that model, and actively damages that context: grepping a repo for every call site, reading a 2000-line file to answer one question, auditing what a refactor left behind.
Hand that work to a delegate instead:
Cost. The grunt work runs on DeepSeek, GLM, Kimi, Qwen, or a local llama.cpp. You pay frontier prices only for the reasoning that actually needs them.
Context. The delegate reads the files on its own budget and returns a result. The 200 KB it read never enters your conversation.
Blast radius. Delegates are read-only by default (read, grep, find, ls), enforced at
session construction. A cheap model doing exploratory work cannot touch your tree unless you
opt it in.
The delegate is always the pi agent. Codex, Grok, DeepSeek and the rest supply the model behind it; this is not a wrapper around their CLIs.
Related MCP server: SubagentBridge
A delegate is only steerable if two channels stay open: you must be able to redirect it mid-task, and it must be able to ask you something and block until you answer. Most ways of driving a coding agent from another program close both.
| opencode SDK | this server | |
Runs in-process | no (subprocess) | no (HTTP client to | yes ( |
Redirect a running turn | no |
|
|
Agent can ask you something | no ( | not in the session API |
|
Model per call | no | yes |
|
pi -p and --mode json set ctx.hasUI = false. A delegate started that way is fire-and-forget
by construction: it cannot raise a question, and you cannot redirect it.
opencode's SDK is a typed client for a separate server process: createOpencode() boots
opencode serve and talks HTTP to it. Clean design, but it means a second process to supervise,
and the session surface it exposes (prompt, abort, revert, messages) has no mid-turn
steering and no path for the agent to ask the caller anything.
pi ships createAgentSession as an embeddable library. This server holds the session object
in-process, so session.steer() can land a message after the current tool call and before the
next model call, and a synthetic uiContext catches the agent's questions and parks them for
answer. Nothing is shelled out; nothing has to be supervised.
* Questions come from pi extensions, so that channel is open only for delegates spawned with
extensions: true. See Web search and other extension tools.
(The table compares the delegation channel, not sandboxing; opencode has its own permission config. See Read-only by default for what this server does and does not enforce.)
Tool | Purpose |
| Call first. Reports reachable models, permitted tools, and how to drive a delegate. Every other tool refuses until it has run once. |
| Delegate in the background. Returns |
| Fan out up to 10 delegates in one call. Validated as a batch, so nothing starts if one task is bad. |
| Delegate and block until done. For quick questions only. |
| State, turns, tools used, latest text, and pending questions. |
| Redirect a running agent. Lands after its current tool call. |
| Give a finished delegate another turn. It keeps everything it read, so you do not re-explain the task. |
| Answer a question surfaced by |
| Stop a session; partial output stays readable. |
| List models this delegate may use. |
| List sessions, running and finished. Filter by |
| Drop a finished session from history, freeing its id. |
Requires Node.js 22.19+ and a working pi install that has been logged in once
(pi, then /login).
claude mcp add pi -e PI_DELEGATE_MODEL=openrouter/stealth/ox-alpha -- npx -y pi-delegate-mcp.mcp.json{
"mcpServers": {
"pi": {
"command": "npx",
"args": ["-y", "pi-delegate-mcp"],
"env": { "PI_DELEGATE_MODEL": "openrouter/stealth/ox-alpha" },
"timeout": 1800000
}
}
}npx resolves the package on every launch. To pin it, install globally and call the binary
directly:
npm install -g pi-delegate-mcp{ "mcpServers": { "pi": { "command": "pi-delegate-mcp", "timeout": 1800000 } } }Keep the server key short, since it prefixes every tool name (mcp__pi__spawn).
git clone https://github.com/howznguyen/pi-delegate-mcp && cd pi-delegate-mcp
npm install && npm run build && npm linkAsk your agent to delegate something. It calls init once to learn what this server can reach,
then spawn:
{ "id": "audit-01", "label": "who still imports onnxruntime",
"prompt": "Search this repo for anything still importing onnxruntime and list the files.",
"cwd": "/path/to/repo" }{ "sessionId": "audit-01", "state": "running", "model": "opencode-go/deepseek-v4-flash",
"activeTools": ["read", "grep", "find", "ls"] }spawn returns immediately. Poll with status for the ordered tool trace and the answer, or
sessions when several are in flight. If init fails, it says exactly what is missing: pi not
installed, no provider logged in, or a model scope that matches nothing.
Model names in the examples below are illustrative. Run models to see what your own pi install
can actually reach.
spawn and run both accept your own id and a free-text label:
{
"id": "search-audit-01",
"label": "what ONNX removal left behind",
"prompt": "...",
"model": "opencode-go/deepseek-v4-flash"
}Ids are [A-Za-z0-9._:-], 1-64 chars, must start alphanumeric, and must be unique among live
sessions. Omit for a UUID.
Finished sessions stay readable via status and sessions instead of vanishing, so you can go
back and check what a delegate actually did. The newest PI_DELEGATE_HISTORY (default 50) are
kept; forget drops one early.
status returns an ordered toolCalls trace: every tool the delegate ran, with arguments and
timing. Add verbose: true for call ids and results:
{
"seq": 1,
"id": "call_467b4bb4…",
"name": "bash",
"state": "ok",
"ms": 10,
"args": "{\"command\":\"echo hello-trace\"}",
"result": "hello-trace\n"
}Arguments and results are clipped (PI_DELEGATE_TRACE_ARGS, PI_DELEGATE_TRACE_RESULT) with the
dropped length recorded, so one read of a large file cannot flood your context.
A finished delegate is not spent. pi keeps its session in memory, so follow_up re-prompts
the same agent with everything it already read still in context:
{ "sessionId": "search-audit-01", "prompt": "Now check whether the build files reference it too" }{ "sessionId": "search-audit-01", "state": "running", "turnsSoFar": 1 }The delegate picks up where it left off. It still holds the files it read on the first turn, so the second question costs one model call rather than a fresh session re-reading the repository.
This is the cheap way to have a conversation with a delegate. Spawning a fresh one means re-explaining the task and paying for it to re-read the same files, and its answer arrives with none of the reasoning that led there.
follow_up refuses a delegate that is still working, because redirecting one mid-task is
what steer is for. The two are not interchangeable: steer lands between tool calls on a
running agent, follow_up starts a new turn on a finished one.
spawn_batch starts a whole batch in one call. Tasks inherit the batch-level model, cwd,
tools and extensions, and override them individually where they need to:
{
"idPrefix": "audit",
"model": "opencode-go/deepseek-v4-flash",
"cwd": "/repo",
"tools": ["ls"],
"tasks": [
{ "prompt": "What still imports onnxruntime?", "label": "imports" },
{ "prompt": "Which build files still reference ONNX?", "label": "build" },
{
"prompt": "Any ONNX model files left on disk?",
"label": "artifacts",
"model": "opencode-go/ox-alpha-free"
}
]
}That names them audit-01, audit-02, audit-03 and returns in a few milliseconds, since
launching a delegate does not wait for it to think.
The batch is validated before anything starts: id format, ids duplicated inside the batch, ids already live, blocked tools, and every model name. One bad task fails the call and launches nothing. Half a fan-out is the worst outcome, because you pay for the delegates that did start and still have to work out which ones did not.
Poll the whole batch with one sessions call rather than one status per delegate. Drop to
status only for the delegate you actually want to read. steer and abort stay per session.
model on any call overrides PI_DELEGATE_MODEL. An unresolvable name is a hard error, never a
silent fallback to the default model, because a silent fallback is how you end up billing a model
you never asked for.
Which names resolve is decided by pi's own enabledModels scope, which this server enforces
rather than merely displays:
opencode-go/deepseek-v4-flash -> ok (listed in enabledModels)
opencode-go/glm-5.3 -> refused (out of scope)
knowns-hub/claude-opus -> ok (custom provider, see below)Custom providers bypass the scope. Any model served by a provider declared in
~/.pi/agent/models.json is offered even when enabledModels does not name it, on the grounds
that declaring a provider by hand is already an intent to use it. This is why the list can be
much longer than enabledModels: three entries in the scope plus two custom providers can easily
mean fifteen offered models. init says so explicitly in models.scopeNote when it applies.
Two switches change that:
Effect | |
| Honour |
| Drop scoping altogether. Every authenticated model is usable. |
Call models to see what is actually reachable under whichever setting is in force.
Claude Code allows exactly one statusLine command, so pi-delegate-statusline wraps whatever
you already run and appends a segment showing this workspace's delegates:
{
"statusLine": {
"type": "command",
"command": "PI_DELEGATE_STATUSLINE_WRAP=ccstatusline pi-delegate-statusline",
"refreshInterval": 10
}
}Drop PI_DELEGATE_STATUSLINE_WRAP to print the pi segment alone.
π ▸ audit engine·t1·12s audit index·t2·8s running, with turn counts and elapsed time
π ▸ migrate·t7·3m04s ?1 waiting one delegate is blocked on a question
π ✓2 finished, nothing runningFiltering by directory is not enough: two Claude Code sessions open on the same repository would show each other's delegates. Attribution uses process lineage instead.
The MCP host spawns one server per session, so the server records process.ppid, the host's
pid. The status line, spawned by that same host, walks its own ancestry and keeps only the
state files whose hostPid it finds there. Same repo, two sessions, no crosstalk. The
directory filter remains as a fallback for state files written before this existed.
State lives in $XDG_STATE_HOME/pi-delegate-mcp/<pid>.json (PI_DELEGATE_STATE_DIR to
relocate). Files are pruned when their process is gone, ESRCH only, since EPERM means the
process is alive under another user. Servers also exit on their own when stdin closes or the
host pid disappears, so a host that dies without closing the transport leaves nothing behind.
Tools are locked to read, grep, find, ls at session construction. Anything else is refused
before a session is even created.
Omit tools to use the defaults. Pass tools: [] to disable all tools.
To widen that, name the extra tools on the server:
"env": { "PI_DELEGATE_ALLOW_TOOLS": "bash" }or PI_DELEGATE_ALLOW_WRITE=1 to permit everything.
bash is not a middle ground. pi ships no permission system, so a delegate holding bash
can write files, delete them, and reach the network regardless of whether write and edit are
on its list. Refusing those two while allowing bash records your intent; it does not enforce
anything. Claude Code's permission prompts and hooks never see what pi does. If you need a real
boundary, run this server inside a container.
pi's own tools are read, grep, find, ls, bash, powershell, write, edit. There is no
search and no fetch among them. Those come from pi extensions, which register their own tools, and a
delegate can use them.
Set extensions: true on the call and permit the tool names on the server:
"env": { "PI_DELEGATE_ALLOW_TOOLS": "web_search,fetch_content" }{ "prompt": "Find the current Node LTS version and tell me just the number",
"extensions": true, "tools": ["read", "grep", "find", "ls", "web_search"] }{ "seq": 1, "name": "web_search", "state": "ok", "ms": 2568,
"args": "{\"query\":\"latest stable Node.js LTS version\",\"numResults\":5}" }This is how you give a delegate network reach without handing it bash. web_search can search
and nothing else, and it passes through the same allowlist as every other tool, so the read-only
default is unchanged for calls that do not ask for it.
Which tools exist depends on what the user running the server has installed. pi-web-access provides
web_search, fetch_content, source_check and get_search_content. pi-mcp-adapter bridges the
MCP servers in ~/.pi/agent/mcp.json and exposes them as mcp. pi has no MCP client of its own, so
that extension is the only route to one.
extensions: true trusts every installed extension, not just the one you wanted. They load as a
set, they run with the full privileges of this server's process, and some open sockets and timers
that outlive the session. Turn it on per call, for the delegates that need it, rather than leaving it
on by default. It also costs real startup time, which is why it is off unless asked for.
Env var | Default | Meaning |
| pi's own default | Model used when a call omits |
| unset | Comma list of extra tools to permit, e.g. |
| unset |
|
|
| Finished sessions kept for review |
|
| Max chars of tool arguments kept in the trace |
|
| Max chars of tool results kept in the trace |
|
| Ceiling on tasks per |
|
| Above this, |
| XDG state dir | Where status-line state is published |
| unset | Status line command to wrap and append to |
| unset | File to append a timestamp to on every status line render, for debugging |
|
| Progress notification interval during |
| unset |
|
| unset |
|
|
| Where pi's |
The MCP TypeScript SDK defaults to a 60 second request timeout, which a real task will blow through. Three defences, in order of preference:
Use spawn + status. Nothing blocks, so no timeout applies.
run emits periodic progress notifications, which reset the host's timeout.
Raise the ceiling with "timeout" in .mcp.json or MCP_TOOL_TIMEOUT in the environment.
CLAUDE_AUTO_BACKGROUND_TASKS=1 makes Claude Code background long MCP calls after ~2 minutes.
Note that progress notifications are discarded once a call is backgrounded, so pick (1) or (3),
not both.
The server does not handle credentials. pi authenticates itself from ~/.pi/agent/auth.json,
then environment variables. MCP hosts often launch servers with a stripped environment, so
prefer auth.json (run pi once and /login) over exporting keys in a shell profile.
npm install
npm run build # tsc, src/*.ts -> dist/
npm run typecheck # tsc --noEmit, strict
npm run test:ci # offline: boots the server over stdio and lists its tools
npm test # full suite: needs a logged-in pi, makes real model callstest:ci is what CI runs and what prepublishOnly gates on, because it needs no credentials and
no network. npm test drives real delegates against real providers, so it costs money and only
works where pi has been logged in.
Path | What lives there |
| Every environment variable, read in one place |
| The tool allowlist and the gate that enforces it |
| Session map, id claiming, history eviction |
| One module per group of MCP tools |
| Everything that touches the pi SDK |
| State file publishing and the status line binary |
Releases are tag-driven. npm version patch && git push --follow-tags runs the build and tests,
then publishes over OIDC trusted publishing, so no npm token is stored anywhere in the repository.
Issues and pull requests are welcome. If you are reporting a delegate that misbehaved, the
toolCalls trace from status with verbose: true is the useful thing to attach.
abatilo/pi-mcp-bridge takes the simpler route:
spawn pi --mode json -p --session-id <uuid> and let pi persist sessions on disk, so the bridge
holds no state at all. Elegant, and worth reading. It trades away steering, questions, and tool
control to get there.
MIT
Stop a running pi session. Partial output stays readable via status.
| Name | Required | Description | Default |
|---|---|---|---|
| sessionId | Yes |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
No annotations are provided, so the description carries the full burden. It discloses that partial output remains readable after abort, which is a useful behavioral detail beyond the simple action. It does not detail other side effects, but for a simple abort tool this is adequate.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is one clear sentence with an additional clause. It is front-loaded with the main action and adds a relevant detail. No wasted words, though it could be slightly more structured.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given the simplicity of the tool (one param, no output schema), the description covers the core action and a key consequence. However, it lacks details on error conditions, idempotency, or what 'running' means precisely. Acceptable but not comprehensive.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 0%, and the description does not explain the sessionId parameter at all. The agent must infer that sessionId refers to the ID of the session to abort. With low coverage, the description should compensate, but it does not.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
States the specific action 'Stop a running pi session' with a clear resource (pi session). It distinguishes from siblings like run, status, and spawn by focusing on termination. The verb 'abort' is reinforced by the description.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
Implies when to use (to stop a running session) but does not explicitly contrast with alternatives. However, it does mention that 'Partial output stays readable via status', which hints at using status afterwards. No explicit when-not to use, but the mention of status provides some context.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Answer a question raised by a pi agent. Get requestId from status. Only pi extensions can ask, so questions appear only for delegates spawned with extensions: true; the MCP adapter's tool-approval and elicitation prompts are the usual source. A delegate waiting on one is blocked until you answer it.
| Name | Required | Description | Default |
|---|---|---|---|
| value | Yes | Chosen option, text, or boolean for a confirm | |
| requestId | Yes | ||
| sessionId | Yes |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations, the description carries the full burden, and it does disclose a critical behavioral trait: 'A delegate waiting on one is blocked until you answer it,' i.e. this call has a real unblocking side effect. It also explains the origin of requests (MCP adapter tool-approval and elicitation prompts), which helps an agent recognize legitimate questions. It omits permission/auth requirements for the caller and what a successful answer returns.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Three tight sentences, front-loaded with the core purpose and immediately followed by the operational detail (where to get requestId). No filler or restatement of the tool name. Slightly dense mid-sentence clause, but nothing wasted.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a 3-required-param, no-output-schema tool with no annotations, the description covers purpose, prerequisite lookup, source of requests, and the blocking consequence. The main remaining gap is that `sessionId` is never explained and there is no statement of what a valid answer looks like beyond the schema's `value` types.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema coverage is only 33% (only `value` is described), so the description must compensate. It does add genuine meaning for `requestId` by pointing to `status` as its source. However, `sessionId` is left entirely unexplained in both schema and description, so the compensation is partial.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description opens with a specific verb+resource ('Answer a question raised by a pi agent') that is clearly distinct from siblings like status, spawn, or steer. It scopes the tool to a narrow case (questions from pi extensions), so an agent can tell it apart from generic session tools. It stops just short of 5 because it never characterizes what an 'answer' does downstream beyond unblocking the caller.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
It gives a clear prerequisite chain: get `requestId` from `status`, and questions only exist for delegates spawned with `extensions: true`. That effectively tells the agent the condition under which this tool is callable. It doesn't state a when-not-to-use or name a fallback tool if no question is pending, so it falls short of explicit exclusions.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Send another prompt to a delegate that has already finished, keeping everything it read and said. Use this instead of spawning a fresh delegate and re-explaining the task: the session still holds its own context, which yours never had to absorb. Returns immediately; poll with status as usual. For a delegate that is still working, use steer instead.
| Name | Required | Description | Default |
|---|---|---|---|
| prompt | Yes | The next turn for this delegate | |
| sessionId | Yes |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
Given no annotations, the description carries the full burden. It discloses context preservation, immediate return (non-blocking), and the need to poll with `status`. It also states the pre‑condition (delegate finished). It doesn't cover error cases or side effects explicitly, but for this tool the key behaviors are well communicated.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Two sentences pack purpose, usage, alternatives, and behavior with zero filler. The primary instruction (use for finished delegates) is front-loaded, and each clause serves a purpose.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a simple 2-parameter tool without an output schema, the description covers the essential workflow: when to use, what it does, and how to obtain results (poll via `status`). Minor gaps like error handling or repeated follow‑ups are not critical for an agent to call it correctly.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
The schema documents the `prompt` parameter, but `sessionId` is left undocumented. The description does not add explicit meaning for `sessionId` beyond the name and context. With 50% schema coverage, this is borderline; the description could have compensated by explaining that `sessionId` identifies the finished delegate session.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
Clear action (send another prompt) targeted at a specific resource (a finished delegate) while preserving context. Explicitly distinguishes from spawn and steer, making the tool's purpose unambiguous even among many siblings.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
Provides direct guidance to use this instead of spawning a fresh delegate, and explicitly directs to `steer` for a working delegate. The when-to-use and when-not-to-use conditions are spelled out, leaving no ambiguity.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Drop a finished session from the review history, freeing its id for reuse.
| Name | Required | Description | Default |
|---|---|---|---|
| sessionId | Yes |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description carries the full burden. It clearly discloses the destructive effect ('Drop'), the target ('review history'), and the side effect ('freeing its id for reuse'). It does not mention irreversibility or error behavior, but the core behavior is transparent and specific.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
A single, front-loaded sentence with no filler. Every word contributes meaning, and the most important action ('Drop') comes first.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a one-parameter tool with no output schema, the description states the operation, the constraint (finished session), and the outcome (id reuse). It could have pointed to sessions for finding sessionId or to abort for running sessions, but nothing critical is missing for selecting and invoking the tool.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
The schema gives no description for sessionId, so the description adds meaning by tying it to a finished session's id and explaining the reuse consequence. It does not explicitly name the parameter or explain where to find sessionId, but the parameter name plus the description make the intent clear.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description states a specific verb ('Drop') and resource ('finished session from the review history'), and clarifies the consequence ('freeing its id for reuse'). It also distinguishes itself from abort by specifying the session must be finished.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description implies this is for finished sessions only, which separates it from running-session tools like abort. However, it never explicitly names an alternative or states when not to use this tool, leaving some routing inferred rather than stated.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
READ THIS FIRST. Reports what this server can reach and how to drive it: permitted tools, the default model, models available per provider, and the recipes for delegating. Every other tool refuses until this has been called once.
| Name | Required | Description | Default |
|---|---|---|---|
| cwd | No | Repository you intend to delegate in; picks up its project-local pi model scope | |
| models | No | Substring to filter the model list, e.g. "deepseek" |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations, the description carries the full burden, and it discloses the most important behavior: all other tools refuse until init runs. It also makes clear this is a reporting/discovery call, implying read-only use, though it does not explicitly address repeated-call safety or side effects.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Two short sentences front-load the critical directive and include the refusal behavior without any filler. Every sentence earns its place.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a discovery tool with two optional parameters and no output schema, the description sufficiently lists what the agent will learn (reachable tools, default model, provider models, delegation recipes) and the prerequisite behavior, making it complete enough to call correctly.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema coverage is 100%, so the baseline is 3. The description does not add meaning beyond the schema's existing explanations for cwd and models, though its mention of model listing aligns with the models parameter.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description uses a specific verb ('Reports') and names concrete resources: permitted tools, default model, provider models, and delegation recipes. It clearly positions init as the mandatory discovery/entry-point tool, distinguishing it from siblings like models, which only list models.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description gives explicit ordering guidance: 'READ THIS FIRST' and 'Every other tool refuses until this has been called once.' This tells an agent exactly when to call init and that no sibling can substitute for it.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
List models this delegate may use: pi's own scoped set plus any custom provider. Use to pick a model value.
| Name | Required | Description | Default |
|---|---|---|---|
| cwd | No | Picks up a project-local pi model scope | |
| filter | No |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
No annotations are provided, so the description carries the full burden. It discloses the list scope (pi's scoped set plus custom providers) but does not mention side effects (none expected), output format, or ordering. For a read-only listing tool this is adequate but not rich; it does not contradict any structured data.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Two succinct sentences. The main purpose and usage are front-loaded, with no filler or redundant wording. Every phrase earns its place.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a simple listing tool with two parameters and no output schema, this description covers the core purpose and usage. However, the `filter` parameter is left undocumented, and the description does not mention the return format or any caveats, leaving some operational details unspecified.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is only 50%: `cwd` has a description but `filter` does not. The tool description adds no meaning for `filter` or `cwd` beyond the schema, so the undocumented parameter remains unexplained. It fails to compensate for the schema gap.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states 'List models this delegate may use' with a specific verb and resource. It also specifies the scope ('pi's own scoped set plus any custom provider'), which distinguishes it from the sibling command-like tools (init, run, etc.) that perform actions rather than listings.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description explicitly says 'Use to pick a `model` value', giving clear context for when to invoke it. It does not mention when not to use it or name alternatives, but given the sibling set are mostly action-oriented commands, the guidance is sufficient for selecting it correctly.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Delegate a task to a pi agent and wait for the final answer. Blocks until done. Prefer spawn for long work; this is for quick questions.
| Name | Required | Description | Default |
|---|---|---|---|
| id | No | Your own session id for traceability, e.g. "search-audit-01". 1-64 chars of [A-Za-z0-9._:-], must start alphanumeric, must not already be in use. Defaults to a UUID. | |
| cwd | No | Working directory for the agent | |
| label | No | Free-text note shown in `sessions`, e.g. what this delegate is for | |
| model | No | Model as "provider/modelId", e.g. "openrouter/stealth/ox-alpha" | |
| tools | No | Tool allowlist for this delegate. [] disables all tools. Omit for defaults: read, grep, find, ls. Permitted on this server: read, grep, find, ls. | |
| prompt | Yes | The task for the pi agent | |
| extensions | No | Load pi extensions for this delegate. Off by default; they add startup cost and can misbehave. |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations, the description carries the full burden, and it discloses the most consequential trait: the call blocks until the delegate finishes, and it returns the final answer. It does not cover failure/timeout behavior, cost, or isolation of the delegate, which leaves some gaps for an unattended blocking call.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Three short sentences, zero waste, and the blocking constraint plus the routing guidance are front-loaded. Every sentence conveys a decision-relevant fact.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a 7-parameter delegation tool with full schema coverage and no output schema, the description supplies the missing return semantics ('final answer') and the blocking contract, which the schema cannot express. Minor gap: no mention of what happens on delegate failure or timeout.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 100%, so all seven parameters (including id, tools allowlist, extensions, model) are already documented in the schema. The description adds no syntax or format meaning beyond that, so the baseline of 3 applies.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
States a specific verb+resource ('Delegate a task to a pi agent') and clarifies the interaction model ('wait for the final answer'), then explicitly differentiates itself from the `spawn` sibling. An agent can distinguish this from spawn, spawn_batch, and follow_up without opening any schema.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
Names the alternative (`spawn`) and the exact condition that selects it ('long work') versus the condition for this tool ('quick questions'). This is an explicit when/when-not/alternative statement, which is the top of the scale.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
List pi sessions held by this server, running and finished. Finished ones stay readable for review until evicted (keeps the newest 50).
| Name | Required | Description | Default |
|---|---|---|---|
| state | No | Filter by state: starting, running, done, aborted, error | |
| verbose | No | Include full text and tool calls |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
No annotations are provided, so the description carries the full burden. It does disclose that finished sessions remain readable until evicted and keeps the newest 50, which is useful behavioral context. However, it does not mention any authentication requirements, rate limits, or the exact return format (e.g., whether it returns summaries or raw session objects). For a non-destructive list operation, this is partial transparency but not exhaustive.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is two tight sentences. The first states the core purpose, and the second adds a relevant retention policy. No fluff or redundancy; every word earns its place. It is front-loaded with the primary action and scope, making it easy for an agent to quickly grasp the tool's function.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given the low complexity (2 optional params, no output schema, no annotations), the description covers the essential aspects: what it lists, the state scope, and retention behavior. It does not explicitly describe the output format, but for a list tool this is often inferable. The lack of an output schema is partially offset by the clear description of the tool's behavior, making it adequately complete for correct invocation.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 100%, with both 'state' and 'verbose' already documented. The description adds no new parameter-specific detail beyond what the schema provides—it references 'running and finished' which maps to a subset of state values, but this is already inferred from the schema's enum-like description. Since the schema carries the parameter semantics, the description's contribution is minimal, aligning with the baseline of 3.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description begins with a specific verb and resource: 'List pi sessions held by this server, running and finished.' It clearly identifies the tool as a listing operation for sessions, and the mention of 'running and finished' distinguishes it from siblings like 'status' which likely targets a single session. The purpose is unambiguous.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description implies usage for listing sessions but does not explicitly contrast it with alternatives such as 'status' or 'forget'. No when-to-use/when-not-to-use guidance is provided, relying on the agent to infer when this tool is appropriate based on the action of listing. The retention note hints at review use cases, but no explicit routing to siblings.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Delegate a task to a pi agent running in the background. Returns a sessionId immediately, so nothing blocks. Poll with status, redirect with steer, answer its questions with answer. Use this for anything that might take more than a minute.
| Name | Required | Description | Default |
|---|---|---|---|
| id | No | Your own session id for traceability, e.g. "search-audit-01". 1-64 chars of [A-Za-z0-9._:-], must start alphanumeric, must not already be in use. Defaults to a UUID. | |
| cwd | No | Working directory for the agent | |
| label | No | Free-text note shown in `sessions`, e.g. what this delegate is for | |
| model | No | Model as "provider/modelId", e.g. "openrouter/stealth/ox-alpha" | |
| tools | No | Tool allowlist for this delegate. [] disables all tools. Omit for defaults: read, grep, find, ls. Permitted on this server: read, grep, find, ls. | |
| prompt | Yes | The task for the pi agent | |
| extensions | No | Load pi extensions for this delegate. Off by default; they add startup cost and can misbehave. |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations present, the description carries the burden and delivers the single most important trait: it returns a sessionId immediately and nothing blocks, so the agent knows not to expect an inline result. It omits other behavioral context such as permissions, session lifetime/cleanup (handled by `forget`/`abort`), and failure behavior.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Three short sentences, front-loaded with the core action and followed by the async contract and lifecycle routing. No filler; every sentence carries actionable information.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a 7-parameter, no-output-schema tool this covers what an agent most needs: the async contract and the follow-up tools for polling, steering and answering. Residual gaps are minor – return payload shape beyond sessionId and failure/timeout behavior are unstated.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 100%, so the schema already explains id, cwd, label, model, tools and extensions in detail. The description adds no parameter-level meaning beyond that (e.g. it never mentions the tools allowlist or model override), so the baseline 3 applies.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
States a specific verb and resource: "Delegate a task to a pi agent running in the background," which is unambiguous. It implicitly separates itself from `run` via the timing heuristic ("more than a minute"), but never names `run` or `spawn_batch` explicitly, so sibling differentiation is inferred rather than stated.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
Gives an explicit when-to-use rule ("anything that might take more than a minute") and then routes the agent through the full lifecycle: poll with `status`, redirect with `steer`, answer questions with `answer`. This is exactly the alternative-naming guidance an agent needs to pick and follow this tool correctly.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Fan out several delegates in one call. Each task inherits the batch-level model, cwd, tools and extensions unless it overrides them. The whole batch is validated before any delegate starts, so a bad model name or a duplicate id fails everything instead of leaving half a fan-out running. Poll the result with sessions, which reports all of them at once, rather than one status per delegate.
| Name | Required | Description | Default |
|---|---|---|---|
| cwd | No | Default working directory for every task in this batch | |
| model | No | Default model for every task in this batch | |
| tasks | Yes | 1 to 10 delegates to start | |
| tools | No | Default tool allowlist for every task in this batch | |
| idPrefix | No | Names the tasks `<prefix>-01`, `<prefix>-02`, ... e.g. "audit" gives "audit-01" | |
| extensions | No | Default extensions setting for every task in this batch |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations, the description fully covers behavioral traits: whole-batch validation before any start, failure semantics for bad model or duplicate id, and inheritance/override rules. It discloses key outcomes and edge-case behavior, though it does not mention async execution or session cleanup.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Two sentences with no filler. The core purpose and key behavior are front-loaded; the polling guidance is a natural close. Every clause serves a purpose.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given no output schema and no annotations, the description covers purpose, validation, inheritance, and polling—enough for correct invocation. Minor gaps like precise return format are covered by the `sessions` reference, so it is nearly complete for the complexity level.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema coverage is 100% with descriptions for all 6 parameters, so baseline is 3. The description adds meaningful semantics beyond the schema: it explains how batch-level defaults are overridden per task, clarifies the idPrefix format, and states the validation impact on parameters. This elevates it above the baseline.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the action: 'Fan out several delegates in one call.' It specifies the resource (delegates) and the batch nature, distinguishing it from the single-delegate spawn. The inheritance and validation details further clarify its exact role.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description explicitly advises polling with `sessions` instead of per-delegate `status`, giving concrete guidance on expected follow-up. It implies batch use case but does not explicitly contrast with `spawn` for single-delegate scenarios, leaving some room for inference.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Check a background pi session. Returns state, turn count, tools used, latest text, and any pending questions the agent is waiting on. A non-empty questions array means it is blocked until you call answer. toolCalls traces every tool the delegate ran, in order.
| Name | Required | Description | Default |
|---|---|---|---|
| verbose | No | Include tool results and call ids in the trace | |
| sessionId | Yes |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description carries the full burden. It discloses the return contents (state, turn count, tools, latest text, pending questions) and explains the blocking semantics of the 'questions' array. It also notes that 'toolCalls' traces every tool run in order, adding behavioral depth beyond the tool name. It doesn't mention side effects, but 'check' implies read-only behavior, and the description sufficiently covers operational behavior.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is concise and well-structured: two sentences plus a note. It front-loads the purpose, then details the return fields and the blocking condition. Every sentence carries useful information with no fluff.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given the absence of an output schema, the description adequately explains what the call returns and how to interpret the 'questions' array. It also clarifies the nature of 'toolCalls'. Minor omissions like error behavior for invalid sessions are not critical for a status-checking tool, so it feels complete enough for an agent to call it correctly.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema coverage is only 50%: only 'verbose' has a description; 'sessionId' is undocumented. The tool description does not mention either parameter, failing to compensate for the uncovered 'sessionId'. It doesn't explain what constitutes a valid sessionId or how 'verbose' changes the output beyond the schema's minimal hint. With low coverage and no description support, this is a significant gap.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description states a clear verb ('Check') and a specific resource ('background pi session'), and enumerates exactly what it returns (state, turn count, tools used, latest text, pending questions). This distinguishes it from siblings like 'sessions' (which likely lists sessions) and 'answer' (which handles questions). The purpose is unambiguous.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description gives actionable guidance on how to interpret the result: a non-empty 'questions' array means the agent is blocked and must call 'answer'. However, it does not explicitly state when to use this tool versus alternatives (e.g., 'sessions' for listing all sessions, or 'steer' for modifying a session). The when-to-use context is implied but not contrasted with siblings.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Redirect a running pi agent mid-task. The message lands after its current tool call finishes, before the next model call. Use this instead of aborting when the agent is going the wrong way.
| Name | Required | Description | Default |
|---|---|---|---|
| message | Yes | ||
| sessionId | Yes |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations, the description bears the full burden of behavioral disclosure. It transparently explains the timing of when the message lands, which is a key behavioral trait. It does not mention error cases or lack of side effects, but for a simple steering action this is reasonably transparent.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is two sentences, concise, and front-loaded with the core action and timing. Every word adds value without unnecessary detail.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a simple tool with no annotations and no output schema, the description covers the primary use case and timing behavior. It lacks details on error handling or edge cases, but given the low complexity, it is fairly complete.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema coverage is 0% and the description does not explain the parameters at all. While 'sessionId' and 'message' are somewhat self-explanatory by name, the description fails to indicate which parameter identifies the agent or what the message content should be. The description should have compensated for the lack of schema documentation.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the action ('redirect') on a specific resource ('running pi agent') and adds timing context (after current tool call, before next model call). It explicitly contrasts with the sibling 'abort', making the purpose unambiguous.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
It gives explicit guidance: use this instead of aborting when the agent is going the wrong way. This clearly tells the agent when to choose this tool over a direct alternative, leaving no ambiguity.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
v0.1.1answer2 fields changedInput schema / properties / value / anyOfRemoved value: -[
- {
- "type": "string"
- },
- {
- "type": "boolean"
- }
-]Input schema / properties / value / typeAdded value: +[
+ "string",
+ "boolean"
+]run1 field changedInput schema / properties / tools / descriptionPrevious value: -"Tool allowlist for this delegate. Default: read, grep, find, ls. Permitted on this server: read, grep, find, ls."New value: +"Tool allowlist for this delegate. [] disables all tools. Omit for defaults: read, grep, find, ls. Permitted on this server: read, grep, find, ls."spawn1 field changedInput schema / properties / tools / descriptionPrevious value: -"Tool allowlist for this delegate. Default: read, grep, find, ls. Permitted on this server: read, grep, find, ls."New value: +"Tool allowlist for this delegate. [] disables all tools. Omit for defaults: read, grep, find, ls. Permitted on this server: read, grep, find, ls."v0.1.0abortanswerfollow_upforgetinitmodelsrunsessionsspawnspawn_batchstatussteerScored across 12 tools
Each tool targets a distinct state or action, and the three delegation entry points (spawn/run/spawn_batch) are explicitly differentiated by blocking behavior and fan-out. The message-sending trio (answer/steer/follow_up) is clearly disambiguated by delegate state, and descriptions even cross-reference each other (spawn vs steer vs follow_up).
Most names are concise lowercase verbs (spawn, run, steer, abort, forget, init) with a consistent compound pattern for spawn_batch/spawn variants. The deviation is noun-style list tools (sessions, models) and the short 'init', which break a strict verb convention but remain readable and predictable.
12 tools map cleanly onto the delegation lifecycle: launch (spawn/run/spawn_batch), monitor (status/sessions), interact (steer/follow_up/answer), stop (abort), and housekeeping (forget/models/init). Each earns its place with no redundant or filler tools.
The surface covers the full delegate lifecycle including batch fan-out, mid-task redirection, post-completion follow-up, Q&A unblocking, and cleanup. Minor gaps exist (e.g. no dedicated blocking wait-for-completion on a spawned session, no explicit per-provider model-filtering view), but agents can work around these via status/run.
Agent-native task management: your AI agent is the interface. Delegate to anyone by email.
Hand tasks, bugs and finished work to AI coding agents, and get back a write-up with evidence.
Coordination hub for AI coding agents: message teammates, ask humans, audit every event.
Build agents to automate any background task. Works with your ChatGPT/Claude subscription.
