CacheBash

Get tasks created for programs to work on

Check for pending messages from programs. Default call DRAINS the caller's own inbox (matching messages transition to delivered). Fleet-supervision reads of another program's inbox MUST pass markAsRead:false explicitly.

View messages that failed delivery. Admin only.

Update working status visible in the app

Send a message to a running program. Relay v0.2.

Create a new task for a program to work on

Claim a pending task to start working on it

Mark a task as complete (done) or failed

Send an alert notification to the user's mobile device

List all API keys for the authenticated user. Returns metadata, never raw keys.

Query the Gate audit log. Admin only.

Query sent messages from a program's outbox. Admin can query any source.

Get aggregated relay message metrics by period. Counts by status, avg delivery latency, per-program breakdown.

Get health status of all programs. Shows heartbeat age, pending messages/tasks per program.

Query full message history with bodies. Requires at least one of: threadId, source, target.

Get aggregated cost/token spend for completed tasks. Supports period filtering and grouping by program or type.

Get aggregated operational metrics from the telemetry event stream. Task success rates, latency, safety gate stats, delivery health. Admin only.

Query execution traces for debugging. Filters: sprintId, taskId, programId, tool, since/until.

Get a sprint's full state including definition, stories, and stats.

Merge an alternate Firebase UID into a canonical account. Admin only. Maps the alternate UID to the canonical tenant so all data access is unified.

Get ACK compliance report. Returns statistics on DIRECTIVE messages and their ACK status.

Ingest content for CLU analysis. Stores raw content in a session for later analysis.

Run CLU analysis on ingested content. Extracts patterns, opportunities, gaps, and blind spots using LLM.

Generate a formatted report from CLU analysis results.

Get a single task by ID with full details including completion status and result

Unclaim an active task, returning it to created status for re-claiming. Circuit breaker flags tasks with 3+ unclaims.

Claim multiple pending tasks in a single call. Each task claims independently (not all-or-nothing).

Record measured token/cost telemetry on a completed task (fill-if-null; never overwrites self-reported values). For host-side telemetry hooks. Completing program or admin only.

Complete multiple tasks in a single call. Each task completes independently (not all-or-nothing).

Get task claim contention metrics. Shows claims attempted, won, contention events, and mean time to claim.

Dispatch work to a target program with enforced pre-flight checks, durable obligation persistence, auto-wake, and uptake verification. Canonical dispatch API: returns success only when the target claims the task or ACKs the directive; otherwise returns a tracked pending handle.

Retry a failed or completed task. Resets the task to created status for re-claiming. Optionally updates target program and/or priority.

Abort a running or pending task. Marks the task as permanently cancelled (different from unclaim which requeues).

Reassign a task to a different program without losing context. Preserves original source and instructions.

Escalate a task's priority and/or route it up the chain. Default chain: builder → iso → vector → Flynn.

Quarantine a program to block all task dispatches. Used for programs experiencing repeated failures. Requires dispatch.write capability.

Unquarantine a program to restore task dispatch. Resets failure count. Requires dispatch.write capability.

Replay a completed task with optional modifications. Creates a new task cloned from the original with links preserved. Use this to re-execute tasks with modified instructions, different targets, or changed priorities.

Approve a task in supervised mode. Transitions task from completing → done. Only works on tasks with policy_mode=supervised that are awaiting approval.

Query the lineage chain of a task. Returns ancestors (tasks this was replayed/retried/reassigned/escalated from) and descendants (tasks derived from this one). Also returns the state transition log.

Export tasks with full details including lineage fields and state transitions. Supports filtering by status and date range.

Wave 16: Suggest optimal target programs based on historical success rates. Returns ranked list of programs with success rates for matching task characteristics. Requires dispatch.read capability.

PLAN-W1: fleet-internal only. Scans the caller's tasks collection for documents with no expiresAt field, classifies each via the reviewed classifyForBackfill() rule, and (only when execute:true) writes expiresAt onto them in batches of <=400 -- never overwrites a doc that already has the field, neve…

PLAN-W4: fleet-internal only. Scans the caller's tasks collection for documents where expiresAt EXISTS and is <= now, and (only when execute:true) deletes them in batches of <=400. NEVER deletes a document lacking expiresAt (W4-R1). Dry-run (execute:false) by default -- reports scanned/fieldLessCoun…

Check for pending dream sessions (lightweight check for shell hooks)

Atomically activate a dream session

Submit feedback (bug report, feature request, or general) which creates a GitHub Issue

Read GSP state entries by namespace/key/tier. Returns a single entry or scans a namespace.

Write an operational GSP state entry. Governance enforcement rejects constitutional/architectural writes with a redirect hint to gsp_propose. Uses Firestore transactions for atomic writes.

Diff GSP state entries since a version or timestamp. Returns changed entries for reconciliation.

Get full context payload for an agent boot. Single call replaces 4+ boot API calls. Returns identity, constitutional state, operational state, agent memory, and pending context.

Seed constitutional or architectural state into GSP. Admin/orchestrator only. Bypasses gsp_write governance enforcement for authorized seeding.

Propose a change to constitutional or architectural state. Creates a governance proposal for review. Operational state should use gsp_write instead.

Subscribe to GSP state change notifications. Supports message-based and webhook callbacks. Use unsubscribe: true to deactivate.

Resolve a pending governance proposal. Approve, reject, or withdraw. Approved proposals apply state changes atomically.

Search GSP state entries by text query. Searches across keys, values, and descriptions. Returns scored results ranked by relevance.

Create a new per-program API key. Returns the raw key (only shown once).

Revoke an API key by its hash. Soft revoke — key stays in DB for audit.

Rotate the calling API key. Atomically creates a new key and grace-expires the old one (30s window).

Log a rate limit/throttle event from a session. Written to rate_limit_events collection with 7-day TTL.

Query rate limit events with optional period and session filtering. Returns events ordered by timestamp desc.

Project monthly spend based on current burn rate. Returns forecasted costs, daily burn rate, and top spending programs.

Track whether tasks meet expected completion time based on priority and action type. Returns compliance rate and breach breakdown.

Automated health scoring per program based on success rate, latency, error diversity, heartbeat freshness, and cost efficiency. Returns 0-100 health score with recommendations.

Scan all program states for learned patterns and auto-promote patterns when N+ agents have learned the same pattern. Promoted patterns are written to GSP knowledge store with tier 'architectural' and key format 'pattern/{domain}/{slug}'. Tracks provenance (contributors, confidence, evidence count).

Retrieve all consolidated/promoted patterns from the GSP knowledge store. Returns patterns grouped by domain with aggregate stats (confidence, contributor count, evidence count).

Create a new policy. Operational-tier policies can be created directly. Constitutional-tier requires gsp_seed, architectural-tier requires gsp_propose.

Update an existing operational-tier policy. Cannot update constitutional or architectural policies directly.

Delete an operational-tier policy. Cannot delete constitutional or architectural policies.

Get a single policy by ID.

List all policies with optional filters by tier, enforcement, or enabled status.

Check if a dispatch would violate any policies without creating a task. Useful for dry-run validation.

Read a program's persistent operational state. Programs can read their own state; admin/auditor can read any.

Write a program's persistent operational state. Programs can only write their own state. Partial updates merge with existing.

Query timestamped context snapshots (shadow journal). Returns entries newest-first. Same access rules as get_program_state.

Store a learned pattern into agent memory. Upserts by pattern ID — existing patterns with the same ID are replaced.

Recall learned patterns from agent memory. Supports optional domain filter and text search (grep-style, case-insensitive substring match across pattern, evidence, and domain fields).

Get memory health summary for a program. Returns pattern counts (total, active, stale, promoted), domains list, last update timestamp, and decay configuration.

Delete a learned pattern from agent memory by ID. Hard delete — pattern is removed, not just marked stale.

Reinforce an existing memory pattern. Bumps lastReinforced timestamp, optionally updates confidence and evidence. Resets stale flag.

List all registered programs for the tenant. Filterable by role, group, active status.

Update a program's metadata. Programs can update their own entry; admin can update any.

Create a new session to track work progress

List active sessions for the authenticated user

Query historical fleet snapshots with configurable resolution. Returns time-series data for fleet health visualization.

Write a fleet health snapshot for time-series tracking. Called by the Grid Dispatcher daemon.

Query context window utilization time-series. Returns contextHistory from session docs. If sessionId provided, returns that session; otherwise aggregates across active sessions.

Pause a program's task intake. Paused programs still send heartbeats but won't receive new dispatched tasks until resumed.

Resume a paused program's task intake. Re-enables dispatch of tasks to the program.

List available multicast groups and their members. Use group names as targets in send_message for multicast.

Send a directive to another program. Convenience wrapper for send_message that auto-sets message_type=DIRECTIVE and action=interrupt. Use for orchestrator→worker commands.

Ask the support team for help. Use when you're stuck or need assistance — they'll be notified and in touch shortly.

Create a recurring schedule: {name, target, cron, taskTemplate, budgetCap, enabled}

List all schedules (filterable by target program, enabled status)

Get schedule by ID with next/last run times

Update cron, budget cap, enable/disable a schedule

Remove a schedule

Send a question to the user's mobile device and wait for a response

Check if the user has responded to a question

Create a new sprint to track parallel story execution