Add a display-only `activity` field to AgentRef plus `setActivity`, fed
from the subagent progress chokepoint with a short gist of the agent's
latest intent (or current tool). Render it in the `irc list` output, the
subagent peer roster, and the TUI peer card, beside the role-derived
display name. setActivity emits no event — the roster reads on demand —
so the per-tool-call rate stays off the registry listener path. Peers
with no activity render without a dangling clause.
Refs #2470
Op: extend
Add an optional `role` field to the task spawn contract, threaded end to
end through resolveSpawnItems/spawnParamsFor into the executor. A role
injects a specialization preamble into the subagent system prompt and
becomes the subagent's display name and telemetry identity (label
normalized, length capped), so delegated trees stop being clones of one
generic worker. Empty/absent roles fall back to the agent type name.
Refs #2467
Op: extend
- Added detached spawn metadata to lifecycle, progress, and executor session payloads so task and evaluator runs can mark background jobs.
- Updated subagent session tracking and HUD rendering to only show active detached spawns.
- Extended HUD tests to verify non-detached sync and eval spawns are excluded while detached flags propagate through events.
- Replaced task-simple-mode with a `task.batch` setting enabled by default.
- Updated task schema to use batch `{agent, context, tasks[]}` payloads.
- Migrated task execution to spawn one async job per task and merge outputs.
- Removed per-call schema passing while preserving legacy flat task calls.
- Removed `resume` from task params and schema, requiring agent and assignment inputs.
- Dropped resume continuation paths in task execution and call rendering, always spawning a new agent.
- Removed the `irc.enabled` setting and computed IRC availability by task-depth rules.
- Updated task follow-up guidance to use IRC messaging/history links instead of `task(resume:)`.
The task tool now takes a single { agent, assignment, description, ... } and always runs the subagent in the background — the batch tasks[] array and shared context parameter are gone. Fan-out is parallel task calls; shared background flows through a '/Users/can/.omp/agent/sessions/-Projects-.tree-pi-commit/2026-06-10T15-36-32-782Z_019eb22d-970e-7000-8964-72c98becf3e8/local' file referenced in each assignment.\n\nIntroduces a persistent subagent lifecycle: finished subagents stay live as idle, the lifecycle manager parks them to disk after task.agentIdleTtlMs (default 7 minutes; 0 keeps them live until exit), and they revive automatically when prompted from the Agent Hub, messaged on IRC, or resumed via task. New task(resume: "<id>") revives an idle or parked subagent and runs a follow-up assignment in its existing session.\n\nAdds soft request budgets (explore/quick_task 40, others 90, configurable via task.softRequestBudget, 0 disables): crossing the budget injects a one-time wrap-up steer into the child; crossing 1.5× aborts the run gracefully. Cancelled/aborted subagent salvage replaces the old (no output) with the child's last activity snippet plus request/token stats; SingleResult tracks a per-child requests counter (assistant message_end events) used to sort agent lists in runtime-ascending order in both the live progress view (finished agents above pending/running) and the finalized result view, so rows no longer reshuffle on finalize. Adds a task gallery fixture variant for the resume path (renderer key separated from fixture key).\n\nAll task tests are reshaped around the single-call contract; tests for the discarded shared-context flow are removed, and new task-guards/task-resume/task-schema tests pin the new contract surface.
Same shape of bug the reviewer flagged for custom tools: forwarding
`LoadExtensionsResult` from parent to subagent reused Extension instances
whose factories closed over the parent's `ExtensionAPI` — cwd, eventBus,
and runtime all pointed at the parent. Any tool/handler/command that
referenced `api.exec()`, `api.events`, or `api.runtime` still acted on the
parent session/worktree from inside an isolated subagent.
Forward only the path list; each session rebuilds extensions through
`loadExtensions` so factories see the right `ExtensionAPI`.
- `extensibility/extensions/loader.ts`: extract `discoverExtensionPaths`
(FS scan only) from `discoverAndLoadExtensions`. The combined helper now
composes the two. New export added to the package barrel.
- `sdk.ts`:
- Add `discoverSessionExtensionPaths()` (the `disableExtensionDiscovery`-aware
path-only counterpart of `loadSessionExtensions`).
- Add `preloadedExtensionPaths?: string[]` to `CreateAgentSessionOptions`.
Three loader branches: `preloadedExtensions` (CLI same-process reuse,
still shallow-cloned), `preloadedExtensionPaths` (subagent: skip scan,
reload locally), or full discovery.
- Document `preloadedExtensions` as same-process-only; subagent
forwarding MUST use `preloadedExtensionPaths`.
- `tools/index.ts`: `ToolSession.extensionsResult` → `extensionPaths:
string[]` for the same reason.
- `task/executor.ts` and `task/index.ts`: forward `extensionPaths`. Drop
the forward for the isolated `runSubprocess` branch — worktree cwd ≠
parent cwd, so the subagent re-discovers extensions against its own
tree.
- New `test/sdk-extensions-per-session-binding.test.ts` pins the contract:
two `loadExtensions` calls on the same path with different `cwd` and
different `EventBus` instances yield distinct Extension + runtime
objects whose factories close over the per-call bindings.
- Updated `executor-pass-through` and `sdk-preloaded-extensions-isolation`
tests for the new option name and comment context.
Refs PR review on #2193
Reviewer flagged that forwarding `LoadedCustomTool[]` from a parent session
to a subagent reused tool instances whose factories had closed over the
parent's `CustomToolAPI` — `cwd`, `exec`, `pushPendingAction`, and `ui` all
pointed at the parent. In isolated tasks the tool would `exec` against the
parent worktree and queue pending actions on the parent session.
Forward only the path list; let each session rebuild tools through
`loadCustomTools` so factories see the right `CustomToolAPI`.
- `extensibility/custom-tools/loader.ts`: extract `discoverCustomToolPaths`
(FS scan only) from `discoverAndLoadCustomTools`; export
`ToolPathWithSource`. The combined helper is now `discoverCustomToolPaths`
+ `loadCustomTools`.
- `sdk.ts`: replace `preloadedCustomTools` (`LoadedCustomTool[]`) with
`preloadedCustomToolPaths` (`ToolPathWithSource[]`). The custom-tools
block runs `loadCustomTools` unconditionally; only the path scan is
skipped when the caller pre-discovered it.
- `tools/index.ts`: `ToolSession.loadedCustomTools` →
`ToolSession.customToolPaths` for the same reason.
- `task/executor.ts` and `task/index.ts`: forward `customToolPaths`.
Drop the forward for isolated subagents — the worktree shifts `cwd`, so
the subagent re-discovers tools against its own working tree.
- New `test/sdk-custom-tools-per-session-binding.test.ts` pins the contract:
two `loadCustomTools` calls on the same path with different `cwd` and
different `pushPendingAction` callbacks yield distinct tool instances
whose factories see the per-call bindings.
- Updated `executor-pass-through` and `sdk-preloaded-extensions-isolation`
tests for the new option name and added a `ToolPathWithSource` fixture.
Refs PR review on #2193
Each `runSubprocess` call re-ran `loadCapability<Rule>()`,
`loadSessionExtensions()`, and `discoverAndLoadCustomTools()` because
`ExecutorOptions` and the `createAgentSession()` call inside the executor
omitted three pass-through fields the parent had already paid for. The
already-correct paths (skills, context files, workspace tree, MCP manager)
showed the intended pattern.
- Cache `rules`, `extensionsResult`, and `loadedCustomTools` on the
parent's `ToolSession`.
- Add `rules` / `preloadedExtensions` / `preloadedCustomTools` to
`ExecutorOptions`; forward them from both `runSubprocess` call sites
in `task/index.ts` and into the executor's `createAgentSession()`.
- Add `preloadedCustomTools` to `CreateAgentSessionOptions` and skip
`discoverAndLoadCustomTools()` when it is supplied.
- Shallow-clone `extensionsResult.extensions` when reusing
`preloadedExtensions`, so the per-session autoresearch + custom-tools
inline wrappers never leak back into the caller's array.
Fixes#2190
- Updated `convertToLlm` logic to emit `developer` role for custom, hook, and file-mention inputs.
- Simplified OpenAI compact output filtering to retain only `user` and `assistant` messages, removing legacy `system-reminder` pattern checks.
- Adjusted compaction and session tests to match the new developer-role mapping and expected compacted content.
- Propagated `schemaOverridden` from `YieldTool` into executor `YieldItem` metadata.
- Bypassed schema validation on override or schema-builder errors and kept payload output with success exit.
- Emitted `SUBAGENT_WARNING_SCHEMA_OVERRIDDEN` so accepted override results no longer surface as `schema_violation`.
`finalizeSubprocessOutput` always spliced collected `report_finding`
entries onto a top-level `findings` array regardless of the active output
schema. A caller-supplied schema with `additionalProperties: false` and
no `findings` property would accept the raw payload in-tool (via the
`yield` validator, which only sees the pre-injection data) but then fail
post-mortem validation — emitting `schema_violation: findings: must not
be present` and propagating as a fatal `RuntimeError` through
`agent-bridge.ts` and the eval Python/JS preludes, collapsing the entire
workflow cell along with any prior successful subagent work.
`normalizeCompleteData` now takes the resolved validator and only
performs the injection when the augmented candidate validates. When the
schema rejects it, the raw payload is returned instead — which the in-
tool yield validator already accepted, so the lockstep guarantee
documented at the top of `output-schema-validator.ts` is honored.
Findings remain visible via the agent progress stream and JSONL
artifact, so no information is dropped when injection is suppressed.
Both finalize call paths (yield-success and no-yield fallback) now share
the single validator build instead of constructing it twice, and the
yield-path schema_violation branch is now reached only via the
explicit malformed-schema check, never via spurious findings rejection.
Fixes#2070
- Added `/tan` slash command registration and interactive handling.
- Added TanCommandController validation and async task scheduling for `/tan` dispatch.
- Added session cloning that suppresses breadcrumbs, copies artifacts, and handles abort cleanup.
- Added `promptCacheKey` support in Agent and inherited `providerPromptCacheKey` in session creation.
- Used `||` so empty stderr falls through to abortReason in agent bridge.
- Preferred assistant errorMessage over "Cancelled by caller" on internal aborts.
- Forced `maxRuntimeMs: 0` for eval subagents via ExecutorOptions override.
- Skipped count/concurrency normalization when --bench is set.
- Errored when no OAuth accounts resolve for the provider.
- Updated flag docs to run one request per OAuth account.
- Added `loadOverallPlanReference` to resolve a session plan reference from local storage and skip empty or missing files.
- Updated task execution to read the active plan reference (except in plan mode) and pass it into each spawned subagent.
- Extended the subagent system prompt and session SDK/tools plumbing so subagents receive and render the approved plan path and contents.
- Renamed `TodoWriteTool` to `TodoTool` and its source/prompt files.
- Updated tool registration, schema, renderers, and gating to `todo`.
- Adjusted cursor provider native tool names and tests to match.
- Renamed strike-animation constants and `todo-error-reminder` type.
The catch around the subagent yield-reminder prompt previously logged
every exception at ERROR. User cancel (^C) and compaction-driven aborts
both surface as ToolAbortError through awaitAbortable, so benign control
flow generated 9 spurious 'Subagent prompt failed' errors in 2 days on
the reporter's instance.
Gate the ERROR branch on '!abortSignal.aborted && !(err instanceof
ToolAbortError)' and route the abort path to logger.debug. The outer
catch + finally still mark the run aborted, so observable behaviour is
unchanged.
Fixes#1623
- Converted [SECTION]...[/SECTION] markers to "SECTION\n===" format in system prompt templates.
- Updated system conventions doc to reference the new marker style.
- Updated tests to match against the new header pattern.
- Added parentMnemosyneSessionState propagation from session state through SDK, executor, and task options into nested sessions.
- Added getMnemosyneSessionState() and rekeying logic to refresh Mnemosyne IDs during session sync, switch, and restore.
- Added Mnemosyne reset and teardown cleanup on unaliasing or restoration to avoid stale state.
- Added `mnemosyne.scoping` setting: `global`, `per-project`, and `per-project-tagged`.
- `per-project-tagged` writes to a project-local bank while merging global memories on recall.
- Refactored `MnemosyneSessionState` to manage scoped recall/retain targets and deduplication.
- Updated hindsight tools to route recall/retain through scoped methods.
- In `resolveApproval`, yolo mode now returns the user policy directly (`allow`/`prompt`/`deny`) and ignores tool `override` prompts.
- Updated approval-mode and approval unit tests to match the new behavior for critical bash patterns under yolo and auto-approve.
- Updated docs and settings metadata to describe yolo as user-policy-driven rather than override-driven.
- Added `ToolTier`, `ToolApproval`, and `ToolApprovalDecision` types and exported approval APIs.
- Updated approval-mode options from `auto|prompt|custom` to `always-ask|write|yolo` and defaulted mode to `yolo`.
- Changed approval resolution to apply per-tool decisions first, then mode-tier limits, with legacy-mode migration.
- Assigned read/write/exec `approval` and approval-detail prompts across built-in, custom, extension, and MCP tools.
- approval: user 'tool: deny' now wins over critical-pattern override
(the override only tightens allow->prompt; it must never re-arm a denied tool).
- approval: rename hindsight policy keys to match registered tool names
(recall/retain/reflect, not hindsight_recall/hindsight_retain).
- approval: head+tail truncation for bash/ssh command prompts so a
destructive suffix buried after a long benign preamble stays visible.
- task/executor: force tools.approvalMode='auto' in createSubagentSettings
so subagents (which have no UI) cannot deadlock on per-tool prompts;
the parent's approval of the task call is the authorization.
- docs/approval-mode: rewrite so every example surfaces tools.approvalMode
and explains that tools.approval is ignored outside 'custom' mode.
- Removed per-session run queues from JS and Python backends, allowing async cells on the same session id to interleave.
- Introduced `getEvalSessionId` on ToolSession so subagents spawned via `task` inherit the parent's executor id and share JS VM and Python kernel state.
- Switched JS runtime state from module-level fields to AsyncLocalStorage so concurrent runs route output and tool calls to their own context.
- Changed Python runner to an asyncio event loop with per-request tasks and ContextVar-based run id tracking for concurrent execution.
- Added mtime-based module cache eviction to preserve singleton state across re-imports of unchanged local files.
- Unified output schema construction and validation by adding buildOutputValidator and using it in YieldTool and task executor.
- Added MAX_SCHEMA_RETRIES so YieldTool now retries schema failures three times with hints before overriding.
- Updated failure handling to use shared summarizeValidationFailure and formatters for required-field reporting.
- Added tests for output-schema-validator and YieldTool covering malformed schemas and nested-array retry edge cases.
Queued extension-delivered user messages when deliverAs is set and waited for session_start extension message sends before prompting subagents.
Fixes#1343
- Captured `tool_execution_update` snapshots for `task` calls into in-flight progress state for live nested rendering.
- Cleared in-flight task snapshots at task start and completion to prevent stale nested progress from persisting.
- Updated progress rendering to combine completed and in-flight task details through a dedicated nested task tree view.
The report_finding tool's priority is exposed as a string enum
("P0"-"P3") for ergonomics, but the reviewer agent and every
custom review agent declare priority as `type: number` in their
JTD output schema. The cast at executor.ts:1473 lied about the
runtime shape, so the auto-injected `findings[].priority` flowed
through as strings and every yield with at least one finding was
rejected with `findings.0.priority: expected number, received string`,
forcing the run into the schema_violation exit path.
Added `toReviewFinding(details)` in tools/review.ts that maps the
priority enum to its numeric ordinal via the existing PRIORITY_INFO
table and use it at the boundary in executor.ts. Render paths still
see the original `ReportFindingDetails` shape (string priority)
through normalizeReportFindings, so display formatting is unaffected.
Fixes#1350
- Added `retry.maxDelayMs` to the settings schema and interfaces, with a default cap for provider backoff delays.
- Updated session auto-retry logic to fail fast when a requested wait exceeds the cap without fallback, emitting terminal auto-retry failure state.
- Propagated retry state and failure data into task progress and rendering so children show retry/wait details and reminder prompts stop after terminal errors.
Adds optional autoloadSkills field to agent frontmatter that automatically loads listed skills when a sub-agent is spawned. Uses the same buildSkillPromptMessage + sendCustomMessage mechanism as interactive skill loading, queued via sendCustomMessage({ triggerTurn: false }) before the first session.prompt(task). No extra agent turns, no new injection path. Skills stay in listing for sub-resource access. Compaction behavior matches manual loading. Unknown skill names silently skipped.
Lore-id: f85fdbdc
Constraint: autoload must use buildSkillPromptMessage + sendCustomMessage, never modify systemPrompt or use contextFiles
Constraint: triggerTurn must be false to avoid extra agent turns
Rejected: append to systemPrompt | agent cannot distinguish skill content from own instructions
Rejected: contextFiles injection | agent sees opaque file blob, cannot discover sub-resources
Rejected: promptCustomMessage per skill | N extra agent turns with model inference
Directive: autoload skill names are resolved against parent session skill list at spawn time in task/index.ts
Tested: TypeScript compiles clean with tsc --noEmit
Tested: parseAgentFields parses array and CSV string frontmatter
Tested: parseAgentFields returns undefined for absent and empty fields
Not-tested: bun test cannot run locally due to missing pi_natives native addon (requires Rust toolchain)
Confidence: high
Scope-risk: moderate
Reversibility: clean
buildOutputValidator already exists in task/executor.ts but only ran on
the fallback JSON parse path. Subagents that called yield directly with
a data payload skipped validation entirely, so a schema-conforming-
empty-object (e.g. `{}` against a schema requiring `findings`) was
returned as a successful task.
finalizeSubprocessOutput now invokes the validator on every yield path
and on the fallback completion path. On failure the result carries
error="schema_violation", a typed message, the missing required field
list, and a truncated preview of the offending data. exitCode=1 and
isError=true so existing consumers in task/index.ts surface it as a
failed task without code changes.
- Runtime-limit timeout is now tracked with a sticky runtimeLimitExceeded flag, so later caller aborts during teardown cannot downgrade the timeout state and let a late yield report success.
- Awaited setup operations before the first prompt are now raced against the subagent abort signal, including auth discovery, model refresh, model resolution, session open, session creation, extension session_start, prompt, and waitForIdle.
- Added defensive abort re-check after registering the abortSignal listener plus a checkAbort() immediately before await session.prompt(...), so a wall-clock timer that fires during pre-prompt setup is no longer lost between listener registration and the prompt call.
- Late yield events arriving after a wall-clock timeout no longer flip the result to success: a runtimeLimitExceeded flag derived from the internal abortReason forces wasAborted=true and exitCode=1 regardless of hasYield, while yield payloads remain captured in extractedToolData.
- Async task progress now copies contextTokens and contextWindow from the completed SingleResult onto AgentProgress, so backgrounded tasks still surface their context gauge to the UI.
- Replaced fromTypeBox conversion with a JSON-schema validator flow in ai tool handling and execution paths.
- Added recursive schema validation and expanded TypeBox checks for refs, enums, uniqueItems, and constraint keywords.
- Sanitized Azure/CCA tool schemas by dropping unsupported fields and rewriting oneOf tool branches as anyOf.
- Tightened argument and model-config validation, preserving unknown tool fields and adding apiKey plus compatibility flags.
- Added canonical `pi.zod` schema API exports and removed TypeBox package exports/imports.
- Migrated Tool schema typing from TypeBox to shared `TSchema`/Zod flow with legacy TypeBox compatibility.
- Updated AI provider adapters and MCP/agent builders to convert tool params through `toolWireSchema()`.
- Reworked schema validation from AJV to Zod-safe parsing with `fromTypeBox`, `toolWireSchema`, and meta schema checks.
- Added a task.maxRuntimeMs setting with a default disabled state for per-subagent runtime limits.
- Updated subprocess execution to enforce the configured wall-clock timeout, mark timeouts as aborts, and include timeout-specific abort messaging.
- Tracked per-turn context size and context window through task progress/results and updated UI renderers to display current context against window with cumulative tokens rendered separately.
metaphorics