pi-goal-xx

内容来源：README.md（说明文档） · 原始地址 · 查看安装指南

原始内容

pi-goal-xx

Owner: buihongduc132
Upstream: Fork of pi-goal-x (which forked @capyup/pi-goal)

pi-goal-xx is a long-running goal extension for pi. It gives the agent a durable objective, a visible lifecycle, and schema-gated tools for drafting, executing, pausing, resuming, and completing work.

The extension is designed around one rule: the user owns intent; the agent executes only after the goal is explicit and confirmed.

Table of Contents

Features

  • Two goal styles — Regular goals for open-ended research and implementation. Sisyphus goals for patient ordered execution, one step at a time.
  • Simple goal creation — Use /goals to discuss and confirm a draft. Use /goals-set to skip discussion and start immediately.
  • Full lifecycle — Pause blocked goals, resume when unblocked, abort obsolete work, complete when done. Auto-continue keeps the agent working across turns until completion, interruption, or the empty-turn guard.
  • Multiple open goals — Keep several goals in .pi/goals/. Each session focuses one at a time; switch with /goal-focus.
  • Above-editor status widget — See the current goal, status, file path, and progress at a glance while the agent works.
  • Structured task lists with subtasks — Break goals into trackable tasks. Agents can mark individual tasks or subtasks complete without stopping the turn. Subtask IDs are validated for uniqueness and depth.
  • Verification contracts — Attach plain-text requirements to a goal or task (e.g. "Run npm test, zero failures"). The agent must provide matching evidence before complete_goal or complete_task will succeed.
  • Independent completion auditor — When a goal is marked complete, a separate pi agent inspects the workspace, verifies every success criterion, and approves or rejects before the goal is archived. You can press Escape during an audit to abort it. Configure the auditor model via /goal-settings.
  • Schema-gated tools — Agents see only the tools relevant to the current lifecycle phase: drafting, active, paused, or tweaking. Lifecycle tools like pause_goal, complete_task, and abort_goal appear and disappear automatically.
  • Immutable objective — The agent cannot silently change your goal. Objective updates require a /goal-tweak drafting flow with explicit user confirmation.
  • Built-in questionnaire tools — During drafting, agents can ask structured questions through goal_question and goal_questionnaire without depending on external packages.
  • Disk-backed state — Active and archived goals persist in .pi/goals/. Goal state survives session compaction, workspace switches, and context churn.
  • Configurable settings — Tune the auditor model, disable the task system or contracts, and set subtask depth through /goal-settings or .pi/pi-goal-xx-settings.json.
  • Worker session isolation — When spawned as a pi-agent-teams worker (PI_TEAMS_WORKER=1), the extension skips goal focus inheritance from the leader's branch context. Workers start goal-unfocused and can still read goal files from disk without inheriting the leader's active goal.

Fork lineage: pi-goal-xx ← pi-goal-x ← @capyup/pi-goal

pi-goal-xx preserves all upstream features and adds: verification contracts (per-goal and per-task), unified goal+task acceptance in a single confirmation dialog, recursive task lists with subtasks, an immutable objective enforced by tools, deferred archival with cleaner lifecycle hooks, an improved completion auditor with configurable model and progress widget, drafting UX refinements, worker session isolation for pi-agent-teams, and lifecycle reliability fixes (including zombie process prevention via comprehensive timer cleanup).

Install

From npm:

pi install npm:pi-goal-xx

From a local checkout:

pi install .

Try once without installing:

pi -e .

Quick start

Regular goal

/goals add structured logging to the auth module

Flow:

  1. The agent clarifies, researches, or grills only when the goal contract needs it.
  2. The agent calls propose_goal_draft with a concrete objective once the contract is clear.
  3. pi shows a full plain-text confirmation report.
  4. If confirmed, the full finalized goal is printed into the conversation and written to .pi/goals/.
  5. The new goal becomes this session's focus. Existing open goals remain in .pi/goals/ and can be selected later with /goal-focus.
  6. The agent works only on the focused goal until it calls complete_goal, pauses, aborts, produces an empty/non-progress turn, or the user interrupts.

Sisyphus goal

/sisyphus Refactor the auth flow: 1) extract token validation. 2) wire it into login. 3) update tests.

Sisyphus mode is for patient ordered execution. It uses the same lifecycle and tools as a regular goal; the difference is the prompt style and completion standard: preserve the user's order, do not rush, do not invent preflight/reconnaissance steps, and stop to ask when blocked.

If the objective is already final and should start immediately, use:

/goals-set add structured logging to the auth module
/sisyphus-set Refactor auth flow exactly as ordered: 1) extract token validation. 2) wire it into login. 3) update tests.

User commands

/goals <topic>          Discuss/research/grill a regular goal, then confirm a draft
/sisyphus <topic>       Discuss/grill a Sisyphus-style goal, then confirm a draft
/goals-set <objective>  Immediately create and start a regular goal
/sisyphus-set <objective> Immediately create and start a Sisyphus-style goal
/goal-status            Show focused goal state
/goal-list              List all open goals in .pi/goals/
/goal-focus             Choose this session's focused goal
/goal-tweak <change>    Draft a revision to the focused active/paused goal
/goal-pause             Pause the focused active goal
/goal-resume            Resume a paused goal
/goal-settings          Configure pi-goal settings, including auditor model settings
/goal-abort             Abort/archive the focused goal or cancel drafting
/goal-clear             Archive the focused goal or cancel drafting

Pressing Esc or aborting an active run pauses the goal so it does not remain falsely active.

Multiple open goals and focus

pi-goal separates durable goals from session focus:

  • Goal pool: every open goal is an active_goal_*.md file under .pi/goals/.
  • Focused goal: the current pi session has one focused goal id stored in a pi-goal-focus custom session entry.
  • No focus in markdown: goal files describe the goal itself; they do not record which session is focused on them.
  • Branch-local focus: because focus is reconstructed from the current session branch, /tree navigation can restore a different focus for a different branch.
  • One continuation chain: auto-continue only schedules work for the focused goal in the current session.

Creating a goal with /goals, /sisyphus, /goals-set, or /sisyphus-set no longer clears other open goals. It creates a new active goal file and focuses it. Use /goal-list to inspect open goals and /goal-focus to switch the session focus. If the latest focus entry explicitly clears focus, or points at a missing/stale goal, a remaining single open goal is not auto-focused; single-open auto-focus only happens when no focus entry exists at all. If multiple open goals exist and the session has no valid focus, /goal-resume, /goal-clear, /goal-abort, /goal-pause, and /goal-tweak ask the user to choose a goal instead of acting on all of them.

Multi-session goal focus

When multiple pi sessions run in the same cwd (worktree-per-feature, parallel verification, an ad-hoc second session), a fresh session previously auto-focused the only open goal and started running it — stealing the goal from the session that was actively working on it. pi-goal now coordinates concurrent sessions with a lease-based advisory lock.

  • Per-goal lock sidecar: when a session focuses an active goal, it writes <cwd>/.pi/goals/.locks/<goalId>.lock (JSON). The file records owner.sessionId, owner.pid, acquiredAt, expiresAt, and heartbeatAt. The .locks/ subdir keeps these out of the active-goal pool scan.
  • Two-signal liveness: a lock is HELD iff BOTH the owning PID is alive AND the lease has not lapsed. Either signal going stale makes the lock reapable. PID-alive correctly treats EPERM (cross-user process) as alive — only ESRCH (no such process) counts as dead. This catches crashes near-instantly (PID dead) and hangs within the lease window (lease lapses while the PID still looks alive).
  • Auto-run chokepoint: queueContinuation (the auto-run trigger) only fires when THIS session holds the focused goal's lock. No lock, no lock file, or a lock held by another live session → no auto-run.
  • Auto-focus restricted to reason: "resume" by default (LD3): a brand-new session, a hot-reload, a fork, or /tree navigation does NOT auto-focus the only open goal. Only a resumed session (the user coming back to their own session) auto-focuses. This eliminates the "I opened pi for an unrelated task and it stole my goal" case.
  • /goal-focus override is advisory: running /goal-focus <id> on a goal locked by another LIVE session prompts "Session (pid ) looks alive — take over anyway?". On a STALE lock (PID dead or lease lapsed), the lock is silently reaped and acquired with no prompt. In headless contexts (!ctx.hasUI), override is refused with a warning.
  • Heartbeat refresh: the lock owner extends its lease via a single 60-second backstop setInterval timer while focused and active. The timer refreshes the 3-minute lease ~3× within its window, covering both idle presence and long tool executions. No event-driven refresh on turn_end or tool_execution_end (timer-only — the least-resistant path).
  • Fail-open on fs errors: a permissions misconfig on .locks/ does not crash the session, and manual/explicit goal work proceeds. Auto-run is NOT fail-open — if the session cannot prove it holds the lock, the chokepoint still blocks queueContinuation.

Recovering a stuck lock

Locks are reaped lazily by the next acquirer, so a crashed session's lock clears automatically the next time any session tries to focus that goal. To clear manually:

  • Delete the sidecar: rm <cwd>/.pi/goals/.locks/<goalId>.lock, or
  • Wait for lease expiry (~3 minutes from the last heartbeat) — the lock becomes stale and the next focus reaps it.

The .locks/ directory is a cache; rm -rf <cwd>/.pi/goals/.locks is safe.

Migration note

Default behavior changed: fresh sessions (reason: "new"/"startup"/"fork"/"reload") and /tree navigation no longer auto-focus the only open goal. Users who relied on the prior auto-focus-anywhere behavior set PI_GOAL_AUTO_FOCUS=all.

Agent tools

The extension exposes tools only when they make sense for the current lifecycle phase.

Tool Visible when Purpose
goal_question drafting / tweak drafting Ask one focused user question
goal_questionnaire drafting / tweak drafting Ask multiple structured questions
get_goal always Read the focused goal state; mentions other open goals when present
propose_goal_draft drafting only (goal creation) Submit a concrete draft for user confirmation
propose_goal_tweak tweak drafting only Submit a revision to an existing goal (shows Confirm / Continue Chatting dialog)
complete_goal focused active or paused goal Mark the focused goal complete — supply a verificationSummary covering all contract items. When the auditor is disabled, supply confirmBypassAuditor: true after user confirmation to bypass the audit
pause_goal focused active goal Pause the focused goal because of a real blocker
abort_goal focused active or paused goal Abort/archive an obsolete, impossible, unsafe, or user-cancelled focused goal
propose_task_list active or paused goal Propose a structured task list for user confirmation (stops the turn)
complete_task active or paused goal Mark a task complete with optional verificationSummary. If the task has a verificationContract, the summary is required (does not stop turn)
skip_task active or paused goal Mark a task skipped with a required reason (does not stop turn)
propose_goal_tweak tweak drafting only Submit a revision to the focused goal (shows Confirm / Continue Chatting dialog)
step_complete hidden / legacy Compatibility no-op; Sisyphus no longer requires a step counter
create_goal hidden Direct calls are rejected; normal creation goes through propose_goal_draft

Tools that interrupt

These tools stop the current turn and block subsequent work tool calls (except read-only tools like read, bash with safe commands):

Tool Effect When to use
pause_goal Pauses goal, stops turn, blocks subsequent work tools Agent encounters a real blocker (missing info, dependency failure, unclear requirement)
abort_goal Archives goal, stops turn Goal is obsolete, impossible, unsafe, or user cancels
complete_goal Marks complete, runs auditor, stops turn All success criteria met, ready for independent verification
propose_goal_tweak Starts tweak drafting, stops turn User wants to revise the objective or task list
propose_goal_draft Confirmation dialog, stops turn Agent has clarified intent, ready to create goal
propose_task_list Task confirmation dialog, stops turn Agent wants to break goal into trackable tasks

Turn-stopping mechanism:

When these tools execute, they call setTurnStopped() which:

  1. Sets a turn-scoped marker (turnStoppedFor with turnSeq)
  2. Blocks all subsequent tool calls in the same turn (except POST_STOP_ALLOWED_TOOLS)
  3. Forces the agent to yield the turn instead of continuing work

Allowed tools after stop:

  • Read-only tools: read, bash (with safe commands like ls, grep, find)
  • Inspection tools: get_goal, goal_status
  • No write/edit/bash with side effects

Stale checkpoint guard:

If a continuation is queued for a goal that becomes inactive (paused, cleared, replaced) before the turn starts, the stale checkpoint guard blocks work tools with:

Cannot call <tool>: the goal checkpoint that triggered this turn is no longer active.
Goal <id> has been paused, cleared, or replaced.

Auditor subscriptions

Auditor subscriptions allow you to forward lifecycle events asynchronously to the auditor channel. These are non-blocking notifications — they log events and show UI notifications, but do NOT intercept or block tool execution.

Configuration

In .pi/pi-goal-xx-settings.json:

{
  "auditorSubscriptions": [
    {"event": "pause", "mode": "async"},
    {"event": "abort", "mode": "async"},
    {"event": "audit_started", "mode": "async"},
    {"event": "task_skip", "mode": "async"},
    {"event": "contract_violation", "mode": "async"}
  ]
}

Available events

Event When emitted Payload
pause After pause_goal succeeds {goalId, details: {reason, suggestedAction}}
abort After abort_goal succeeds {goalId, details: {reason, archivePath}}
audit_started When completion auditor begins {goalId, details: {provider, model}}
task_skip After skip_task succeeds {goalId, taskId, details: {reason}}
contract_violation When complete_task lacks required evidence {goalId, taskId, details: {contract, message}}

Behavior

  • Non-blocking: Events are forwarded on the microtask queue; failures are swallowed and logged
  • UI notifications: Each event shows a UI notification (e.g., "Auditor subscription: pause (goal=abc123)")
  • Ledger entries: Each event appends an audit_subscription_emitted entry to the goal ledger
  • Unknown events: Silently skipped (no error, no notification)
  • Mode restriction: Only "async" is supported; other modes are silently dropped

Example: Track all pauses and aborts

{
  "auditorSubscriptions": [
    {"event": "pause", "mode": "async"},
    {"event": "abort", "mode": "async"}
  ]
}

When the agent pauses or aborts a goal, you'll see:

Auditor subscription: pause (goal=abc123)
Auditor subscription: abort (goal=abc123)

Future: Question gating

Not yet implemented. To intercept goal_question → forward to auditor first → auditor decides if question should go to user, you would need:

  1. New config: gateQuestions: true
  2. goal_question tool intercepts → calls auditor agent → auditor approves/rejects/rewrites question → then shows to user

This feature is planned but not in the current release. Current goal_question goes directly to user.

Drafting behavior

/goals and /sisyphus start a lightweight intent discussion, not a heavy runtime sub-state. The agent clarifies, researches, and grills only when needed, may proceed directly for fully specified requests, and then calls propose_goal_draft to show the user a Confirm / Continue Chatting dialog. goal_question and goal_questionnaire are available when structured input helps, but plain conversation is acceptable.

/goals-set and /sisyphus-set skip the discussion and confirmation dialog. They directly create and focus an active goal from the supplied objective so execution can begin immediately.

The agent may do minimal read-only reconnaissance when it directly improves the goal contract, but should not begin substantive implementation before confirmation. The strict runtime starts after the user confirms the draft and an active goal is created.

When a draft is proposed, the confirmation UI shows a full plain-text report with draft details, the original topic, and the proposed goal. If the confirmation UI throws in interactive mode, creation fails closed and confirmation remains active; it never auto-creates a goal. When a draft is confirmed, the tool result includes the full final objective, not a one-line summary, and normal work tools (write, read, bash, edit) are available for execution. This makes the confirmed contract visible in the conversation as well as on disk.

While goal confirmation or tweak drafting is active, old goal execution is suspended: active-goal prompts, accounting, and auto-continue checkpoints do not run for the previously focused goal.

Completion behavior

Completion is also explicit and is checked by an independent pi auditor agent. The executor calls complete_goal with its completion claim:

{
  "status": "complete",
  "completionSummary": "What was completed and what evidence proves it."
}

Before archiving the goal, complete_goal starts a separate pi agent in an isolated in-memory session. The auditor receives the objective, the executor's completion claim, and current goal metadata, then can inspect the workspace with read-only-oriented tools (read, grep, find, ls, and bash). It must end its report with exactly one marker:

  • <approved/> archives the goal as complete.
  • <disapproved/>, no marker, an error, or an abort rejects completion and leaves the goal open.

The auditor is semantic, not a paperwork checklist: it should reject scaffold-only, alpha, generated-template, proxy-metric, build-only, or weakly verified completions when the real user outcome is not satisfied.

By default the auditor uses the current/default pi model. Configure it via .pi/pi-goal-xx-settings.json, or interactively with /goal-settings (see Configuration).

The completion result prints a full report into the conversation:

  • Goal complete.
  • optional completion summary / evidence supplied by the executor
  • the auditor's approval report
  • full current goal details, including objective, status, usage, mode, and file path

Sisyphus goals use the same completion tool as regular goals. The stricter part is the prompt/criteria standard: the agent should only call completion after the whole ordered objective is actually satisfied and likely to survive independent auditing. A paused goal can also be completed directly when the agent already has enough evidence that every requirement is satisfied; it does not need a resume just to call complete_goal.

Schema gates

The shipped gates are intentionally small and mechanical.

Gate Prevents
Focus consistency /goals accidentally becoming Sisyphus, or /sisyphus becoming regular mode
Confirm-before-commit The agent silently creating or replacing a discussion-based goal
Direct set intent /goals-set and /sisyphus-set are explicit user shortcuts that bypass draft confirmation
Completion auditor gate Archiving completion unless an independent pi auditor agent returns <approved/>
Abort gate Aborting missing, stale, completed, or reasonless goals
Direct-create rejection Hidden create_goal calls creating goals without the confirmation flow
Post-stop block Continuing to call tools after pause_goal, abort_goal, complete_goal, or propose_goal_tweak stops the turn
Empty-turn guard Pure chat loops that would keep auto-continuing without meaningful goal work
Abort pause Active goals staying active after user abort / Ctrl-C
Disk reconciliation External pause/archive/delete/status changes being ignored or overwritten by stale memory
Post-compaction reminder Losing the active objective after session compaction

Files

.pi/goals/active_goal_<timestamp>_<id>.md
.pi/goals/archived/goal_<timestamp>_<id>.md

Multiple active_goal_*.md files may exist simultaneously. This is the project-level open goal pool. The selected/focused goal is intentionally not stored in these files; focus lives in session custom state.

Each file contains:

  1. extension-owned JSON metadata;
  2. a user-editable # Goal Prompt section;
  3. progress/status information.

Before commands, tools, and lifecycle hooks act on a focused goal, the runtime reconciles the focused record against the active goal file on disk. External archive/delete/status changes therefore win over stale in-memory state and cannot resurrect deleted active files. Prompt-body edits are still picked up from the # Goal Prompt section; focus is never stored in goal markdown.

Goal paths are constrained to .pi/goals/ and .pi/goals/archived/; absolute paths, traversal, NUL bytes, symlinks, and unsafe metadata paths are rejected.

Configuration

All settings live in a single file: .pi/pi-goal-xx-settings.json

Configured interactively via /goal-settings, or edited directly:

{
  "disableTasks": false,
  "disableContracts": false,
  "subtaskDepth": 1,
  "provider": "fireworks",
  "model": "accounts/fireworks/models/deepseek-v4-flash",
  "thinkingLevel": "high",
  "disabled": false,
  "disabledTools": ["goal_question"],
  "auditorSubscriptions": [{"event": "pause", "mode": "async"}]
}
Field Default Purpose
disableTasks false Suppress task list features entirely when true
disableContracts false Suppress verification contract enforcement when true
subtaskDepth 1 Maximum nesting depth for subtasks
provider system default Provider name for the auditor agent
model system default Model name for the auditor agent
thinkingLevel system default Thinking level: off, minimal, low, medium, high, xhigh
disabled false When true, skip the completion audit entirely
disabledTools [] Tool names to hide entirely (never registered, agent never sees them). All tool names are eligible including lifecycle tools (complete_goal/pause_goal/abort_goal); you accept breakage if you disable a lifecycle tool. Unknown tool names are silently skipped.
auditorSubscriptions [] Events to forward asynchronously to the auditor channel (non-blocking). Each entry: {event: string, mode: "async"}. Arbitrary event strings allowed (lifecycle: pause, abort, complete, audit_started; task: task_skip, contract_violation; any custom string). Unmatched event names are silently skipped.
auditorMode "inherit" Auditor resource mode: "inherit" (start with all main-session resources, opt out via auditorExclude) or "minimal" (start with baseline read-only tools, opt in via auditorInclude).
auditorExclude {} Resources to exclude in inherit mode. Object with tools, mcp, skills, extensions arrays; glob patterns allowed (*, ?).
auditorInclude {} Resources to add in minimal mode. Same shape as auditorExclude; matched against the main session's resources.
auditorPromptMode "global-local" Prompt resolution: "global-local" (local overrides global), "local" (local only, global ignored), "global-local-merge" (global + \n\n + local).
auditorPrompt unset Inline auditor prompt string; takes precedence over all file-based prompts and modes.
goalPromptMode "global-local" Goal custom-prompt resolution (injected into runtime goal/continuation prompts): "global-local" (local overrides global), "local" (local only, global ignored), "global-local-merge" (global + \n\n + local). Mirrors auditorPromptMode.
goalPrompt unset Inline goal custom-prompt string; takes precedence over all file-based prompts and modes.
leaseMs 180000 Lease window (ms) for the multi-session focus lock. A lock is stale once now > expiresAt. See Multi-session goal focus.
heartbeatMs 60000 Interval (ms) for the timer that refreshes the focused goal's lease while active.

Env var overrides:

  • PI_GOAL_DISABLE_TASKS=1 — disable task features (takes precedence over file)
  • PI_GOAL_DISABLE_CONTRACTS=1 — disable contract enforcement (takes precedence over file)
  • PI_GOAL_DISABLED_TOOLS=tool_a,tool_b — comma/whitespace-separated tool names to hide (takes precedence over file)
  • PI_GOAL_SETTINGS_FILE=custom-path.json — alternative settings file path (relative to cwd or absolute)

Environment variables

Variable Default Purpose
PI_GOAL_AUTO_CONFIRM unset When 1, auto-confirms drafts in headless/test contexts
PI_GOAL_DISABLE_TASKS — When 1, disable task features (overrides settings file)
PI_GOAL_DISABLE_CONTRACTS — When 1, disable contract enforcement (overrides settings file)
PI_GOAL_DISABLED_TOOLS — Comma/whitespace-separated tool names to hide (overrides settings file)
PI_GOAL_SETTINGS_FILE .pi/pi-goal-xx-settings.json Alternative settings file path (relative to cwd or absolute)
PI_TEAMS_WORKER unset When 1, worker session mode: skips goal focus inheritance from leader's branch context. Workers start goal-unfocused but can still read goal files from disk. Set automatically by pi-agent-teams when spawning worker sessions.
PI_GOAL_AUTO_FOCUS resume Multi-session auto-focus policy. resume (default): only a session with reason: "resume" auto-focuses the single open goal. all: restore legacy behavior — auto-focus on any session reason (including new/startup/fork/reload and /tree navigation, which is routed as null). See Multi-session goal focus.

Configurable auditor

The completion auditor verifies goal completion before archiving. By default it inherits the main session's tool list (filtered through auditorExclude) and uses a hardcoded prompt. You can make it stricter, looser, or project-specific.

Two modes

  • inherit (default): the auditor starts with all the main session's tools/MCP/skills/extensions, then removes anything matching auditorExclude. Use this when you trust the auditor to verify anything the executor could.
  • minimal: the auditor starts with the baseline read-only toolset (read, grep, find, ls, bash, report_auditor_progress) and adds anything matching auditorInclude from the main session. Use this for strict, predictable verification.

Wildcard patterns

auditorExclude / auditorInclude accept glob patterns in every array:

  • * — any run of characters (incl. empty)
  • ? — exactly one character
  • no wildcard — exact match (case-sensitive)

Prompt modes

Auditor prompts resolve in this order (first non-empty wins):

  1. Inline settings.auditorPrompt (always wins)
  2. File-based, combined per auditorPromptMode:
    • global-local (default): .pi/auditor-prompt.md overrides ~/.pi/auditor-prompt.md
    • local: only .pi/auditor-prompt.md (global never checked)
    • global-local-merge: ~/.pi/auditor-prompt.md + \n\n + .pi/auditor-prompt.md
  3. Hardcoded default prompt (built into the extension)

Global prompt file: ~/.pi/auditor-prompt.md • Local prompt file: <cwd>/.pi/auditor-prompt.md

Configurable goal prompt

The runtime goal/continuation system prompts that drive the active goal agent (not the auditor) can also be customized. This is the channel for project-specific execution rules (delegation policy, TDD discipline, blocker handling, verifier-loop requirement). Resolution mirrors the auditor prompt exactly:

  1. Inline settings.goalPrompt (always wins)
  2. File-based, combined per goalPromptMode:
    • global-local (default): .pi/goal-prompt.md overrides ~/.pi/goal-prompt.md
    • local: only .pi/goal-prompt.md (global never checked)
    • global-local-merge: ~/.pi/goal-prompt.md + \n\n + .pi/goal-prompt.md
  3. Nothing injected when unset (fully additive — zero behavior change by default)

Global prompt file: ~/.pi/goal-prompt.md • Local prompt file: <cwd>/.pi/goal-prompt.md

The resolved block is appended to both goalPrompt() (agent start) and continuationPrompt() (checkpoint resume), after the Sisyphus discipline block when present. The /goal and /sisyphus drafting instructions are also customizable via prompts.goal-drafting (see unified prompt configuration).

Examples

Read-only auditor with everything else stripped (default is already broad; tighten it):

{
  "auditorMode": "inherit",
  "auditorExclude": {
    "tools": ["write", "edit", "bash"],
    "extensions": ["cc-safety-net*"]
  }
}

Minimal auditor that can also query GitNexus:

{
  "auditorMode": "minimal",
  "auditorInclude": {
    "tools": ["gitnexus*"],
    "mcp": ["gitnexus"]
  }
}

Project-specific auditor prompt (create <project>/.pi/auditor-prompt.md):

You are auditing a financial trading repo. Verify no magic numbers,
all numeric literals are named constants, and every order path has
an idempotency guard.

{ "auditorPromptMode": "local" }

Inline override (no files needed):

{ "auditorPrompt": "Reject unless all tests are green and the diff is < 500 lines." }

Unified prompt configuration

The auditorPrompt/goalPrompt channels above are the legacy surface for two specific prompts. unified-prompt-config generalizes the same resolution pattern to all 7 runtime prompts + auditor + every tool prompt field, with two new modes (override, off) on top of the original three, plus per-command hooks and {{snippet}} contract templating.

See the dedicated docs for full detail:

Quick reference

{
  "prompts": {
    "goal-running":      { "mode": "append",   "inline": "Delegate implementation to a team." },
    "goal-continuation": { "mode": "append" },
    "auditor":           { "mode": "override", "inline": "Reject unless verifier-loop hash present." }
  },
  "promptsDir": ".pi/pi-goal-xx/prompts/",
  "commandHooks": { "enabled": true, "goals": { "mode": "append" } },
  "hooksDir": ".pi/pi-goal-xx/hooks/",
  "contractTemplates": true,
  "contractsDir": ".pi/pi-goal-xx/contracts/"
}

Backward compatibility: legacy auditorPrompt/auditorPromptMode and goalPrompt/goalPromptMode continue to work — they are silently mapped to prompts.auditor / prompts.goal-running when no explicit prompts.<key> block is present. Explicit prompts.<key> always wins.

Tool prompts require /reload: tool promptSnippet/promptGuidelines resolve at extension load (per design D3). Editing a tool-<name>.md file requires reloading the session to pick up the change. Runtime prompts (goal-running etc.) and command hooks are similarly load-time resolved.

Worker session isolation

When pi-agent-teams spawns a worker session with contextMode: "branch", the worker inherits the leader's session entries including goal focus state. This can cause workers to accidentally work on the leader's goal instead of their assigned task.

Setting PI_TEAMS_WORKER=1 (done automatically by pi-agent-teams) triggers worker isolation mode:

  • No focus inheritance: The worker skips reading pi-goal-focus and pi-goal-state entries from the branch context
  • Starts unfocused: Worker sessions begin with focusedGoalId=null and no active goal
  • Can still read goals: Workers can read goal files from .pi/goals/ via disk, but don't auto-focus any goal
  • Leader unchanged: Leader sessions continue to inherit focus normally (backward compatible)

This prevents the bug where 22 zombie test processes accumulated at 90% CPU for 2+ days due to workers inheriting and executing the leader's goal.

Development

npm install
npm test
npm run check
npm pack --dry-run

The fast unit suite uses Node's built-in test runner and covers core parsing, drafting gates, lifecycle policy, abort policy, questionnaire formatting, centralized tool names, Sisyphus prompt-style behavior, completion reporting, and display helpers.

The experiment harness under experiments/ runs full pi sessions against real model calls and mechanical rubrics.

cd experiments
bash harness/run.sh C1-vague-goal-set --count 3 --grade --no-smoke

Package contents

The npm package ships only the runtime extension, docs, and package metadata. The extension is split into small modules:

extensions/goal.ts                 orchestration, commands, tools, events, timers
extensions/goal-record.ts          goal record types, normalization, creation helpers
extensions/goal-pool.ts            open-goal pool, focus resolution, list/selector text helpers
extensions/goal-core.ts            display helpers
extensions/goal-draft.ts           lightweight confirmation prompt, proposal validation, drafting tool gate
extensions/goal-policy.ts          lifecycle, pause/resume/complete, and Sisyphus policy
extensions/goal-auditor.ts         independent pi auditor agent for completion approval, config, and progress tracking
extensions/goal-ledger.ts         event append, read, validation, sanitization, and reconstruction
extensions/goal-questionnaire.ts   built-in question UI and question tool registration
extensions/goal-tool-names.ts      centralized published tool names and allowlists
extensions/prompts/goal-prompts.ts active, continuation, tweak, and stale prompts
extensions/storage/goal-files.ts   goal file paths, serialization, parsing, archive IO
extensions/widgets/goal-widget.ts  above-editor goal beacon component
extensions/widgets/goal-notifications.ts widget-style notification text

Design principles

  • User owns intent: only the user starts, replaces, resumes, clears, or confirms goals; the agent may only pause, complete, or abort through schema-gated lifecycle tools with evidence/reason.
  • One commit path: normal goal creation goes through drafting and confirmation.
  • Schema beats prompt walls: recurring failure modes are handled by validators and tool-call interceptors.
  • Visible contracts: confirmed goals and completion reports are printed fully into the conversation.
  • Lifecycle-shaped tool surface: the agent sees only tools appropriate to the current phase.
  • Disk-backed continuity: goal state survives context churn and can be audited from .pi/goals/.
  • Human-owned focus: the agent may work on the focused goal, but only user commands/UI selection switch focus.

Upstream

This repository is a downstream fork of pi-goal-x (which forked @capyup/pi-goal). To sync with upstream changes:

git fetch upstream
git merge upstream/main
# resolve conflicts, test, commit

The upstream remote should point to https://github.com/tmonk/pi-goal-x.git.

Release policy

This repository can be validated locally with tests and packaging checks. Publishing a new npm version, pushing tags, and running pi update are explicit release steps and are not part of ordinary implementation goals unless requested.

Recent changes

v0.1.0 (2026-07-03)

  • Worker session isolation — Workers spawned by pi-agent-teams no longer inherit the leader's goal focus. Prevents workers from accidentally executing the leader's goal instead of their assigned task. Fixes zombie process accumulation (22 processes at 90% CPU for 2+ days).
  • Comprehensive timer cleanup — All internal timers (statusRefreshTimer, auditAnimationTimer, debugMockAuditTimer, continuationTimer) are now cleared in session_shutdown handler, preventing test runner hangs and zombie processes.
  • Test reliability — Re-enabled afterEach cleanup in test suite with proper timeout guards. All 654 tests pass and exit cleanly within 1.5s.
  • disabledTools config — Hide specific tools entirely from the agent via settings file or PI_GOAL_DISABLED_TOOLS env var.
  • auditorSubscriptions config — Forward events asynchronously to the auditor channel for non-blocking audit tracking.
  • Widget timing fix — Convert milliseconds to seconds before passing to formatDuration for tool timing display.

License

MIT

本页内容由上游仓库自动镜像，如有疑问请访问原始仓库。