Gossipcat

Inspect durable Codex bridge jobs and recover results after MCP reconnect. These jobs belong to the local separate Codex session, not the currently open Desktop chat.

Collect results from dispatched tasks. Waits for completion by default. Use consensus: true for cross-review round.

Manage runtime feature-gate flags stored in .gossip/runtime-flags.json. Actions: list (show all flags + source), get (one flag), set (write flag), unset (remove flag), reload (discard cache after hand-edit).

Dispatch tasks to agents. mode:"single" (default) sends to one agent. mode:"parallel" fans out to multiple agents. mode:"consensus" dispatches with cross-review instructions. Returns task IDs for collecting results. Image attachments: pass `images` (mode:"single") or per-task `images` (parallel/consensus) as local absolute PNG/JPEG paths to send screenshots to vision-capable relay providers (openai/google/anthropic/grok/local) — max 4 images, ≤4 MB each, magic-byte-sniffed. When omitted, absolute PNG/JPEG paths in the task text are auto-attached (up to 4). Text-only providers and native agents…

Return the canonical CONSENSUS_OUTPUT_FORMAT block. Use this when you need to write an ad-hoc Agent() prompt for a native subagent that should produce findings the consensus engine can parse. Paste the returned string into your prompt verbatim — the format trains the agent to emit <agent_finding> tags instead of prose.

Show the gossipcat handbook for humans — architectural invariants, operator playbook, caveats, hallucination patterns, glossary. Call this when you want to READ the docs, not when you need LLM context. The LLM-facing auto-load lives in gossip_status().

Send a short message from the orchestrator to a configured agent. It is delivered (fenced as untrusted data) into that agent's prompt on its next dispatch; the agent can answer with an <agent_message in_reply_to="..."> tag. Not for tasks, results or findings — use gossip_run / gossip_dispatch for work. To answer the human use `gossip_chat_reply`. Pass in_reply_to with an agent message id (from its inbox fence) to close it. Returns the new inbox id.

Plan a task with write-mode suggestions. Decomposes into sub-tasks, assigns agents, and classifies each as read or write with suggested write mode. Returns dispatch-ready JSON for approval before execution. Use this before gossip_dispatch(mode:"parallel") for implementation tasks.

Show progress of active tasks and consensus rounds. Call during long-running operations to see what agents are doing.

Feed a native agent result back into the gossipcat relay. Call this after a Claude Code Agent() completes a task dispatched via gossip_dispatch or gossip_run. Enables consensus cross-review and gossip for native agents.

Feed native agent cross-review result back into a pending consensus round. Called after dispatching Agent() for cross-review.

Terminate the MCP server process so Claude Code respawns it with a fresh bundle. Use after 'npm run build:mcp' to hot-reload without quitting Claude Code. Returns a notice, then exits 100ms later.

Search an agent's archived knowledge files, or manage the memory sidecar index.

Walk open findings in .gossip/implementation-findings.jsonl and auto-resolve any whose cited code has been fixed (cited identifier absent from the touched file after comment stripping). Conservative — multi-cite AND, file-scoped not range-scoped, structural insight-tag exclusion. Writes a hash-chained audit entry per resolution to .gossip/finding-resolutions.jsonl and advances the watermark. Pass full:true to bypass the watermark and rescan everything.

Run a task on a single agent and return the result. For relay agents (Gemini), this is a single call — dispatches, waits, returns. For native agents (Sonnet/Haiku), returns dispatch instructions with gossip_relay callback. Use agent_id:"auto" to let the orchestrator decompose and assign agents automatically.

View agent performance scores from consensus signals. Shows accuracy, uniqueness, reliability, and dispatch weight for each agent. Use to understand which agents are performing well and which need improvement.

Save a cognitive session summary to project memory. The next session will load this context automatically on MCP connect. Call before ending your session to preserve what was learned.

Description is built at runtime.

Record or retract consensus performance signals. Use action "record" (default) to record signals after cross-referencing agent findings — call IMMEDIATELY when you verify. Use action "retract" to undo a previously recorded signal.

Fetch the full markdown of ONE of your bound skills that was not injected into this dispatch. The prompt lists fetchable names under "Skills available on demand". Read-only — never binds or changes your skill set.

Manage agent skills. Actions: list (show skill index), get (return a skill's raw markdown by name), bind (attach skill to agent), unbind (remove skill from agent), build (create skills from gap suggestions), develop (generate skill from ATI competency data), prune (drop skill-index entries for agents no longer in config.json).

Check Gossip Mesh system status, host environment, available agents, dashboard URL/key, and agent list with provider/model/skills. Pass slim:true to skip the handbook inline (~21KB savings) on reconnect refreshes.

List all available gossipcat MCP tools with descriptions. Call after /mcp reconnect to discover new tools.

Check for or apply gossipcat updates. Detects install method (global npm, local dep, git clone) and fetches the latest version from the npm registry.

On-demand staleness check for a memory file claim. Reads the memory file, dispatches a native haiku-researcher utility to verify the claim against current code, and returns a structured verdict (FRESH | STALE | CONTRADICTED | INCONCLUSIVE) with file:line evidence. Call before acting on any backlog item from memory. Spec: docs/specs/2026-04-08-gossip-verify-memory.md

Pull signals recorded since a cursor timestamp. Returns {events, next_cursor, count, truncated}. Stateless; pass next_cursor back on subsequent calls. Max 24h lookback, max 500 events per call.

Search agent memory by keyword

Return your own identity: agent_id, runtime (native or relay), provider, and model. Use when you need to know who you are — for example, to pick the right tool variant (gossip_remember vs memory_query) or to cite your own past findings. Identity is also injected into your system prompt at dispatch time, so you usually do not need to call this.

Execute a shell command (60s timeout). Use for: npm install, npm run build, npx tsc --noEmit, etc. NEVER run dev servers (npm run dev, npm start) — they run forever and will timeout.

Fetch the full markdown of ONE of YOUR OWN bound skills that was not injected into this cross-review. Read-only — it never binds, unbinds, or changes your skill set. Budget: 2 calls per cross-review round.

Suggest a skill that would help with the current task. Non-blocking — logs the suggestion and you keep working.

Run tests and get a peer review of your changes. Call this after writing files to verify correctness. Returns test results + reviewer feedback.