perseus_auto_skill
Instruct the agent to load a specific skill before starting work. Use at the top of context documents to enforce critical hygiene skills (e.g., memory-hygiene, agent-safety). Renders as a mandatory instruction block. Read-only.
How to use it
perseus_auto_skill is exposed by the Perseus MCP server. Add the server to your MCP client (Claude Desktop, Cursor, Windsurf and others), and the perseus_auto_skill tool becomes available to the model automatically. See the full listing for setup details and every tool this server provides.
Install Perseus
docker run --rm -v /path/to/workspace:/workspace perseus mcp serveOther tools in Perseus (38)
Execute local agent subprocess
Compile a bounded, task-scoped sanitized agent projection. Shows the exact agent view separately from provenance and selection reasons; receipts contain hashes/references only.
Release a previously previewed sanitized projection after matching per-scope consent. Durable release metadata excludes prompts, private bodies, secrets, and tool arguments.
List tasks from the project task board (tasks/.md files). Use to see what is open, in progress, or completed. Filter by status. Read-only; returns task array with id, title, status, scope.
Declare a token budget for the rendered context (renders as empty text). Enforced by perseus prompt-size: an over-budget render warns — or fails with strict — with a per-directive byte/token breakdown (#606). Declarations are read from source text before conditionals are evaluated; top-level only —
Write recent session checkpoints to Perseus Vault as durable memories (#713) — the write side of the memory loop, symmetric to @memory recall. Idempotent per checkpoint (re-render upserts, never duplicates). Use at session boundaries so lessons persist immediately instead of waiting for a scheduled
Answer one narrow question from at most 64 scoped records with evidence-linked validity/confidence, or an explicit insufficient-evidence/review/degraded/unavailable outcome.
Render a compact 'Since last session' delta (#714): git branch/commits, Agora task-board changes, new inbox messages, new checkpoints, and new vault session memories since the last recorded snapshot. Use at the top of a context document so the assistant spends zero turns re-orienting on unchanged st
Deterministically rank at most 64 caller-supplied candidates for one task/scope, preserving identity and provenance commitments without exporting raw private memory.
Current date/time
Detect drift between predicted and actual tool usage patterns via the Guide oracle. Use when tool behavior seems off or after config changes. For workspace hygiene checks, prefer perseushealth. Read-only; returns a markdown drift report.
Embed environment variable
The global-workspace tier: a small, capacity-bounded (default 32), salience-ranked set of items Perseus broadcasts into context — the shared 'what I'm working on now' set for the agent and its subagents. With no args, renders the current working set. add=/pin= admit items; the lowest-salience non-pi
Return the full rendered Perseus context for the workspace.
Run Daedalus context-maintenance heuristics — cache health, directive resolution stats, memory integrity check. mode=basic (default) returns the @health maintenance report; mode=doctor returns the same structured payload as perseus doctor --json (per-check status + summary), the MCP equivalent of th
Audit workspace context health: stale skills, duplicate tasks, oversized output. Use before starting work to catch drift. For deep Daedalus heuristics (cache, directive stats), use perseusgethealth. Read-only; returns status enum and metric counts.
Read agent-to-agent messages from the workspace inbox. Use to check for coordination messages from other agents. Filter to unread only. Read-only; returns message array with read/unread status.
Include and render another Perseus source file, recursively resolving its directives. Use to compose context from multiple files or share common sections across workspaces. Bound a growing file with last=N (final N lines) or since=14d/2w/24h (recent dated sections only). Use mode=reference (or rende
List directory contents or structured data. Use to discover files before reading with perseusread. Supports sorting by name, modified time, or size. Read-only; for hierarchical view, prefer perseustree.
Query the Mason code architecture concept map to find which files implement a feature. Use before editing code to understand where changes should go. Read-only; returns concept map and mapped file list.
Search LOCAL project memory (FTS5, zero-network) for past decisions and architecture notes. Use for in-workspace recall. For cross-session persistent facts, use perseusvault instead. Read-only; returns results array with mode and count.
Fetch rendered context from a remote Perseus instance by URL. Use to pull live workspace state from another machine or container. Read-only; caches results — re-fetch when remote state may have changed.
Select the per-model context profile for this document (#608): sets the context target and memory posture (ondemand/relevant/always) used by the automatic memory injection layer. Use at the top of a context document, e.g. @profile claude-sonnet-4-6. Unknown names fall back to the default profile. Fi
Define a system prompt block that instructs the AI assistant about how to use the rendered context. Use to set behavioral rules, memory hygiene gates, or context interpretation guidelines. Read-only; rendered as-is into the output.
Run a shell command and return stdout
Read and embed file contents into the rendered context. Use to inject config values, environment files, or any text file. Can extract specific keys from structured files. Read-only; use perseuslist or perseustree to browse before reading.
Search an EXTERNAL paper-search MCP server (BGPT by default) for scientific literature and inject per-paper Methods/Results blocks. Use to ground claims in published studies. Self-gates on research.enabled; degrades gracefully when the provider is unreachable. Read-only; speaks JSON-RPC over stdio (
Health-check all services listed in the workspace context (HTTP endpoints, Docker containers, shell commands). Use to verify the environment is healthy before starting work. May make network calls and execute shell commands per service definition — side effects depend on configured checks.
List recent session digests with task summaries and outcomes. Use to understand what was done recently across sessions. For the single most recent checkpoint, prefer perseuswaypoint. Read-only; returns session array with count.
List mined procedural-skill candidates (from session transcripts) pending operator review: trigger, steps, pitfalls, evidence sessions, token cost. Candidates are staged, never active — perseus skills approve <name> promotes one to the live skills dir. Opt-in surfacing: place the directive in your c
List available skills with descriptions and freshness status. Use to discover what capabilities are installed. Filter by category for smaller output. Read-only; stale skills flagged automatically.
Embed token budget for rendered context
Run an external tool that has been allowlisted in the Perseus configuration. Use for approved integrations only. Requires the tool name to be present in the allowlist. Destructive — executes the tool with the user's permissions.
Return filtered toolset metadata and usage statistics. Use to understand what tools are available and how they are being used. For full tool metadata, set full=true. Read-only; stats mode returns aggregated counts.
Display a directory tree with configurable depth. Use to understand project structure at a glance. For flat file listings with metadata, use perseuslist instead. Read-only; depth limits control output size.
Validate a rendered block against a JSON Schema. Use to enforce structure on configuration blocks, task definitions, or any schema-constrained section. Read-only; returns pass/fail with error messages.
Query Perseus Vault for scoped, durable context. Read-only; falls back to the local Vault FTS5 index when the service is unavailable.
Return the most recent session checkpoint: what was being worked on, status, and next steps. Use at session start to resume where you left off. Stale after TTL (default 24h). Read-only; lightweight — call freely.