@atrib/mcp

Slow test tool

Fake write primitive

Fake read primitive

Abort-aware read

Fast test tool

Verify counterparty handoff evidence before linking follow-up work through informed_by.

Return signed atrib records from the local mirror. The agent's own past, with each record's

Walk the local derived graph from a starting record_hash. Returns records reachable via the requested edge types up to the given hop depth, ordered by ascending weighted distance. Layer 1 covers four edge types: CHAIN_PRECEDES (weight 1), INFORMED_BY (weight 1), ANNOTATES (weight 2), REVISES (weight 2). SESSION_PRECEDES, SESSION_PARALLEL, CONVERGES_ON, CROSS_SESSION, and PROVENANCE_OF are deferred to subsequent releases. Useful for tracing the local causal neighborhood of a record before re-attempting a similar action. Legacy alias: new callers should prefer the `recall` tool with shape='walk' (no direction).

Return the aggregated annotation summary for a record: maximum annotation importance across all D058 annotation records pointing at it, the union of their topic tags, and the most recent summary string. Useful for surfacing the agent's prior critique on a record before re-attempting a similar action. Returns null annotations field when no annotation points at the record. Legacy alias: new callers should prefer the `recall` tool with shape='annotations'.

Return the D059 revision chain for a record, with per-entry content + sibling-fan-out awareness. Walks revises edges forward from the given record_hash, surfacing each revision in turn. Each entry carries the revision's record_hash, timestamp, and content (`new_position`, `reason`, `importance`) so the agent can read the chain inline without follow-up recall calls per entry. When more than one revision targets the same record, the chain follows the first-by-timestamp branch and lists the other branch heads as `sibling_hashes` on that entry so the agent learns about parallel revision branches (common in multi-agent flows). Useful for checking whether a position the agent previously held has been revised before acting on it. Returns an empty chain when no revision points at the record. Legacy alias: new callers should prefer the `recall` tool with shape='revisions'.

Free-form text search over the agent's signed past. Returns top-k records by hybrid retrieval: BM25 over each record's per-event_type indexable text (observation `what + why_noted + intent + rationale + topics`; tool_call `tool_name + args + result`; annotation `summary + topics`; revision `prior_position + new_position + reason + topics`; transaction counterparty + memo; directory_anchor tree_root; extension URIs via generic recursive string-walk per D086/D118) plus the annotation summary + topics when present as a curator-quality lift. Reranked by Park et al. weighted-sum scoring with annotation-derived importance and recency signals. Raw unannotated tool_call records are score-demoted so operation logs do not dominate substantive memory. BM25 contribution is clamped to [0, 1] before coverage scaling so the documented Park-component bound is honored. Responses include runtime metadata plus coverage.index, so callers can detect stale MCP processes and whether the durable content-token sidecar was hit, rebuilt, disabled, or bypassed. Useful when the agent has no specific filter and needs to ask 'what do I know about X?'. Legacy alias: new callers should prefer the `recall` tool with shape='content'.

Return all records in a context_id, ordered chronologically (oldest-first). The natural traversal of the CHAIN_PRECEDES topology for a single session/trace. Each entry carries record_hash, event_type, timestamp, display_summary (per-event_type human-readable text from D086), and producer label. Useful for 'what happened in this session?' without manually sorting filter results. Sibling tool to recall_my_attribution_history with built-in context_id scoping + chain-ascending order. Legacy alias: new callers should prefer the `recall` tool with shape='chain'.

Return records that are NOT cited by any other record via informed_by (loose ends: decisions or observations the agent made but never followed up on). Surfaces records whose record_hash does NOT appear in any other record's informed_by[] array within the local mirror. Optionally scoped to one context_id, one event_type, or one creator_key. Useful for the agent to discover dropped balls (e.g. 'I noted X but never built on it'). Legacy alias: new callers should prefer the `recall` tool with shape='orphans'.

Aggregate the local mirror by creator_key. Returns the distinct creators present + per-creator record count + latest record timestamp. Useful when the mirror is multi-signer (multi-agent flows, transactions with counterparty signers) and the agent wants to discover who else's records are in scope before deciding whether to filter with creator_key. Pure aggregation; no records returned directly. Use recall_my_attribution_history with the creator_key filter to drill into one creator's records. Legacy alias: new callers should prefer the `recall` tool with shape='by_signer'.

Walk a record's informed_by chain backward to surface the reasoning

Walk forward from a record_hash, surfacing the records that cited

Supersede a prior position with a stated reason. Produces a signed

Mark a past record\'s importance and meaning without superseding it.

Make a signed statement now: atrib's write verb. Signs an observation by default;

Read a set of records (by context_id and/or record_hashes) and ask

mock modern recall tool

Get the current temperature for a location

Store an action outcome with diagnostic evidence.

Read prior diagnostic outcomes from this Durable Object.

Return signed atrib records captured by the Durable Object.

Flush pending atrib log submissions.

Return the signed approval context for a runnable Cloudflare action.

Preview the approved Cloudflare Workers checkout patch before mutation.

Apply the approved Cloudflare Workers checkout patch to Durable Object SQLite.

Complete a checkout and return an ACP-style completion shape.