Continuum Console

continuum_record_checkpoint

Write an immutable state snapshot. Provide active/dormant/broken entries and a reason. Returns the persisted snapshot with hash. Use this at session end, after significant commits, or when state has materially changed.

continuum_get_state

Fetch the StateSnapshot in effect at the given ISO-8601 timestamp (or now if omitted). Answers "what was true on May 14?" — returns the most recent snapshot at or before the requested time. Returns null if no snapshots exist yet.

continuum_get_digest

Fetch a composed narrative for a time window. V0 returns template-based digests derived from recent checkpoints + observations. V0.5+ adds ruvllm/ruv-FANN local-AI narrative generation.

continuum_search_docs

Progressive Disclosure Layer-1: full-text keyword search across indexed observations. V0 uses SQLite FTS5 for high-precision exact/code-snippet matching. V0.5+ adds semantic vector fusion (RuVector). Returns compact hits — id + 1-line title + ~50-100 tokens per result. After narrowing here, drill via continuum_timeline (Layer 2) and continuum_get_observations (Layer 3); do NOT use this as the only step.

continuum_timeline

Progressive Disclosure Layer-2: observations in chronological order around a reference point. After Layer-1 (continuum_search_docs) narrows hits by keywords, use this to read causal context — what happened just before/after the moment of interest. Returns compact hits (~80-150 tokens each) with an `offsetSec` field so the AI sees relative timing without having to compute it. Anchor by observation ID (uses that observation's timestamp) OR by ISO-8601 timestamp; if neither given, anchors at now.

continuum_get_observations

Progressive Disclosure Layer-3: batch full-text fetch for specifically-narrowed Observation IDs. This is the EXPENSIVE step — ~500-2000 tokens per observation depending on content. You MUST have narrowed via Layer-1 (continuum_search_docs) or Layer-2 (continuum_timeline) first; do NOT paginate-via-this-tool. Caps at 50 IDs per call; extras silently dropped — batch into multiple calls if you genuinely need more. Returns full Observation records with content + metadata + refs.

continuum_delete_observation

**INCIDENT RESPONSE ONLY** — permanently delete an Observation by ID. Hard-delete: removes the row from the relational store, cleans the FTS5 index, and queues removal from any vector index (hybrid backend). There is NO undo. Use this ONLY for: secrets that leaked past the privacy filter, PII landed via adapter metadata, external operator requests for data removal, accidental ingest of confidential markdown. Pairs with the write-time privacy filter to close the privacy loop — the filter catches known patterns at write time; this catches the rest. Does NOT cascade through derived state (digests, briefings) — those regenerate on next read. Snapshots that referenced the deleted ID via refs[] keep the dangling reference as a historical artifact. Returns { deleted: true, id } on success, { deleted: false, id } if no row matched. Do not use this tool for normal lifecycle operations — observations are append-only by design.

continuum_get_todos

List todos in the live pipeline. Pass status="open" (or "in_progress" / "blocked" / "done") to filter, or omit to return all. Newest first. The continuum://todos/open resource is the cheap polling surface; this tool is for filtered lookups.

continuum_create_todo

Create a new todo in the pipeline. Use this when a discussion produces a commitment that should be tracked through to verification. refs[] links to observation IDs that motivated the todo; verifyCommand is a shell command that proves the todo is satisfied when it returns 0. IMPORTANT: for deployment, release, migration, or "ship" todos, you MUST populate verifyCommand — it is the gate that proves the change actually landed (e.g. health-check curl, smoke test, grep for the deployed artifact). Todos without verifyCommand cannot be auto-resolved by scheduled clients (e.g. cron-driven Hermes) and must be closed manually.

continuum_update_todo

Update mutable fields on a todo — status transitions, title edits, verifyCommand changes, blockedBy dependencies. Transitioning to status="done" stamps completedAt automatically.

continuum://todos/open

Live list of todos with status="open" or "in_progress". Cheap polling surface for scheduled clients (e.g. cron-driven Hermes runs) to check what work is queued.

continuum://state/current

Most recent StateSnapshot — what is active in production, what is dormant, and what is known broken right now. For historical state queries, use the continuum_get_state tool with an ISO timestamp.

continuum://digest/latest

Composed narrative for the last 24 hours. V0 returns a template-based summary of recent checkpoints; V0.5+ adds ruvllm/ruv-FANN local-AI narratives.

continuum://session/briefing

Layer-0 Progressive Disclosure — a pre-rendered markdown document combining current state, open todos, and recent activity. AI clients should read this FIRST at the start of every session: it is a single cheap read (~2–5 KB) that often answers a session's opening questions without any further tool calls. Pair with the continuum.session_start Prompt.

continuum.session_start

Canonical session warm-up. Instructs the AI to read continuum://session/briefing first (Layer-0), then use continuum_search_docs to filter by IDs (Layer-1) before fetching full content (Layer-3). Enforces the ~10x token-savings retrieval pattern.

continuum.cite

Citation discipline. When asserting any fact about the project, cite the Observation ID(s) that prove it. If no citation is possible, say so explicitly rather than asserting unverified claims.