knowledge-engineering

Validate and append a DurableTask envelope to its domain JSONL queue.

Read current task state from a domain JSONL queue (latest-line-wins per id).

Append a state transition for a task. Validates the transition is legal.

Send a typed message from one agent to another (or broadcast). Appended to recipient's JSONL.

Read unacked messages for an agent from its mailbox JSONL.

Ack a message — appends an ack row to the mailbox JSONL.

Summary counts (pending/in_progress/blocked/completed/failed) for all domain queues.

Lint a .github/workflows/claude*.yml file for OAUTO13 security baseline. Returns an array of findings, each with severity (ERROR|WARNING|INFO), rule, line, message. Empty array = clean.

List all devices in the Tailscale network with their hostnames, IPs, OS, and last-seen time.

Look up a Tailscale device by hostname fragment and return its full record (IPs, OS, last seen). Does not send ICMP — uses API lookup.

List subnet routes advertised on this tailnet, including their enabled/advertised state.

Fetch the current ACL policy for the tailnet. Returns the Tailscale HuJSON ACL document.

Create a durable task via pg_durable

Retrieve a task by id — L1 → L2 → pg_durable

Mark a pg_durable task as completed and evict from cache

List pending tasks from a pg_durable queue

List the support collections in the local mirror (id-slug + title).

List article .md files in a collection (by its id-slug).

Read one article's markdown by collection id-slug + filename.

The 15-agent registry (5 functions x 3 tiers) + the named teams.

Fan one Envelope out to every member of a named team (the cowork coordination primitive).

Reply on Discord. Pass chat_id from the inbound message. Optionally pass reply_to (message_id) for threading, and files (absolute paths) to attach images or other files.

Add an emoji reaction to a Discord message. Unicode emoji work directly; custom emoji need the <:name:id> form.

Edit a message the bot previously sent. Useful for interim progress updates. Edits don't trigger push notifications — send a new reply when a long task completes so the user's device pings.

Download attachments from a specific Discord message to the local inbox. Use after fetch_messages shows a message has attachments (marked with +Natt). Returns file paths ready to Read.

Fetch recent messages from a Discord channel. Returns oldest-first with message IDs. Discord's search API isn't exposed to bots, so this is the only way to look back.

Fetch recent iMessage history as readable conversation threads. Each thread is labelled DM or Group with its participant list, followed by timestamped messages. Omit chat_guid to see all allowlisted chats at once; pass a specific chat_guid to drill into one thread. Reads chat.db directly — full nati…

Progressive-disclosure tool discovery for the codemode runtime. Searches across servers/<name>/README.md catalogs and returns matching tools with their qualified names + (optional) descriptions. Use this to find tools by capability instead of loading every tool definition into context up front. Cite…

Download counts for a package over a window (last-day | last-week | last-month). Source: https://api.npmjs.org/downloads/point/{period}/{package}.

List all packages owned by an npm organization. Returns name + role per package. Source: https://www.npmjs.com/org/{org}/package.

Fetch a package document from the npm registry: versions, dist-tags, maintainers, latest manifest. Source: https://registry.npmjs.org/{package}.

Full-text search the npm registry. Returns up to `size` results with name, description, and score. Source: https://registry.npmjs.org/-/v1/search.

List posts on https://www.anthropic.com/engineering. Returns title + url for each post discovered on the index page.

Fetch a specific engineering post. Accepts either a full https://www.anthropic.com/engineering/<slug> URL or the slug alone. Mirror-first: returns local body when present (source:'mirror'), else falls back to live HTTP (source:'http').

Text-match the engineering index titles. Case-insensitive substring search; returns the matching entries.

Tiered semantic-cache health: per-tier (L1/L2/L3) hit ratios and hottest keys from dw.fact_cache_hits, live Redis L2 key count, and L3 semantic_cache row count. Sources that are unreachable report null instead of failing.

Search the mirrored citation corpus (CSL-JSON items extracted from vendor/). BM25-ranked (title boosted) with substring fallback for id-shaped queries. Returns up to `limit` items plus ranking meta.

Fetch one CSL-JSON citation item by its id (e.g. anthropic-sitemap:research:clio). The item loads unchanged into citeproc/Zotero and maps onto Claude citations blocks.

Publication-year histogram over the citation corpus (only items with a parsed issued date).

Read a citation memory (managed-agents path+content shape). Reads dw.dim_memory's current row when postgres is configured (logging the access); otherwise derives the memory from the mirrored corpus. Accepts a memory path (/citations/<slug>.md) or a csl id.

Write a citation memory version (SCD II: closes the current row, inserts curation_source='agent'). Requires postgres (dim_memory allowed_operations: scd2_rewrite). Body is markdown; preserve the csl-json block when editing an existing memory.

Citations attributed to an Anthropic research team page (economic-research, alignment, interpretability, societal-impacts).

List posts on https://www.claude.com/blog. Returns title + url for each post discovered on the index page.

Fetch a specific Claude blog post. Accepts a full https://www.claude.com/blog/<slug> URL or the slug alone. Mirror-first; falls back to live HTTP.

Text-match the Claude blog index titles. Case-insensitive substring search.

Execute multiple KG operations in a single round-trip. Steps run in order; each result is appended to the output array. Supported tools: kg_read_graph, kg_search_nodes, kg_open_nodes.

Post a mailbox_outcome summary (achieved or blocked) to the configured Slack channel.

Post a blocked status with evidence to the configured Slack channel.

Post a generic message to the configured Slack channel.

Read recent messages from the configured Slack channel (operator feedback injection).

Execute a GraphQL-style query against the KE knowledge graph. Supported: { entities }, { searchNodes(query: "X") }, { node(name: "X") }, { graph }.

Create one or more typed knowledge-graph entities. Idempotent: duplicate names are skipped.

Create directed relations between named entities. Idempotent: duplicate from|relation|to triples are skipped.

Append observations to an existing entity.

Delete entities by name. Also removes all relations that touch the deleted entities.

Delete specific directed relations.

Return the full knowledge graph (all entities and relations).

Full-text search across entity names, types, and observations. Returns matching entities.

Retrieve specific entities by name plus all relations that touch them.

List the curated set of llms.txt namespaces this bridge knows about.

Fetch one llms.txt namespace and return its raw text. `id` is one of the values returned by `llms_namespaces` (e.g. 'code.claude.com/docs').

Case-insensitive line-grep across every known llms.txt namespace. Mirror-first per namespace (no HTTP if the body is in vendor/<name>/llms.txt mirror) — Phase 4 speed win. Returns each hit with its source URL.

List pending messages without consuming. Returns up to limit messages, does not mutate state.

Return all messages in a thread_id, across all agents. Scans all JSONL under MAILBOX_DIR.

Shorthand: send an OutcomePayload (type=outcome, ack_required=false). Designed for agent→orchestrator reporting.

Sync a TaskPayload to the shared task ledger (_task_ledger.jsonl). Upserts by task_id. Notifies assignee if set.

Send version-controlled mail to another agent (or 'broadcast'). Writes mail/<to>/inbox/<id>.json in the repository; commit it with your current todo so the recipient's next session sees it. Returns the delivered mail including its id.

Read your unread repo mail (own inbox + broadcast). By default marks own-inbox mail read by moving it to mail/<agent>/read/ — commit that move so other sessions see the state. Broadcast mail is never moved.

Full version-controlled mail thread (inbox + read + broadcast) for an agent, oldest first.

Report git working-tree status as structured JSON. Replaces Bash(`git status --porcelain`) at roughly half the token cost. Use to detect drift before committing or pushing.

Walk the project tree (skipping node_modules, .git, dist, .terraform, coverage, .cache, .next) and return files matching an optional substring pattern + optional size constraints. Replaces Bash(`find ...`) + `stat` chains at no shell-spawn cost.

List recent commits as structured JSON (sha, subject, author, timestamp). Replaces Bash(`git log --pretty=...`) at roughly a third the token cost. `since` accepts any git revspec ('HEAD~10', 'origin/main..HEAD', '2026-05-01').

List the curated set of support.claude.com collection slugs we recognize. Each entry is the slug used by the bridge tools.

Fetch a collection page and return the article links it lists. `slug` is e.g. '14445694-claude-code'.

Fetch a specific support article. Accepts a full https://support.claude.com/en/articles/<id-slug> URL or the bare '<id>-<slug>' fragment. Mirror-first; allowlist-enforced HTTP fallback (rejects URLs outside support.claude.com/en/articles/).

Text-match support.claude.com article titles against the local mirror manifest. Returns URL + slug-derived title for each hit, no HTTP required. Sourced from vendor/claude-sitemap/urls.md (341 EN articles under support/en/articles/).

List agent session cost records from the COSTS_JSONL file. Returns sessions with token counts, cache efficiency, and cost.

Query Prometheus for claude_code.cost.usage metrics. Returns time-series data scoped to the org.

Compute prompt-cache efficiency metrics for one or all sessions. Parity with Console /usage/cache hit rate view.

Append an agent session cost record to COSTS_JSONL and emit OTel metrics. Use after each agent SDK session.

List the vendors mirrored under vendor/. Each entry includes the discovered llms.txt URL, last crawl timestamp, and the count of URLs in the runtime allowlist.

Fetch a vendor doc by URL. Returns local mirror body when available (source:'mirror'); otherwise falls back to live HTTP (source:'http'). The URL must be in some vendor's allowlist (see vendor_list).

Case-insensitive line-grep across the local vendor mirror. Optionally restrict to one vendor. Returns each hit with the vendor, original URL, mirror-relative path, line number, and the matching line.

Fetch a single knowledge-graph node by its entity_id with full content and all observations. Returns null if not found.

Full-text search over crawled vendor documentation pages. Returns vendor, path, and the first 500 chars of content ranked by relevance. Use vendor to scope to a specific namespace (e.g. 'cloudflare', 'anthropics').