The local pfm chat fleet — cross-chat communication between independent running chats, never parent/child agent communication. Provides tools for chat management, message injection, transcript reading, and fleet operations.
18 tools with mixed quality. Naming is verb-forward (chat_*, issue_*) and mostly clear, but descriptions vary widely in depth and actionability. Most tools have input schemas with types and descriptions, but output schemas are undocumented, LLMs cannot plan downstream calls or extract structured data. Error handling descriptions exist but lack recovery guidance (e.g., 'not_found' status with no suggestion for next steps). Tool composition is reasonable (single responsibility per tool), but critical chaining IDs are missing from responses. STDIO transport caps protocol readiness at 50, limiting remote accessibility.
Captures a live chat's screen text — "what is on chat X's screen", "show me its scrollback". Call chat_capture{target:"my-chat"} or chat_capture{target:"my-chat", tail_lines:200}. Returns status ok with text (truncated flags a cut, most recent kept); not_found = no such chat; ambiguous = several match; a tool error = the capture itself failed. For the last answer only, chat_last; for a killed or old chat, chat_read.
Finds chats by transcript excerpt — "find / read an old transcript". Call chat_find{excerpt:"search term"}. Returns candidates with match confidence and confirmation status.
Types and submits a message into another live chat — every "send / tell / message / reply to / inject into chat X" ask. Call chat_inject{target:"my-chat", message:"…"}; follow-up steers go in then. Cross-chat only — one independent chat addressing another; a sub-agent reports to its parent by returning its result and never calls this. Returns status delivered with proof; queued = target mid-turn, submits after; refused or undelivered = not sent, message says why; not_found = no such chat; a tool error = delivery itself broke.
Presses tmux keys in a live chat — "press Escape / Enter in chat X", accept a modal, interrupt a turn. Call chat_keys{target:"my-chat", keys:["Escape"]}; raw text is keys:["y"] with literal:true. For a whole message use chat_inject; a sub-agent never drives its parent's pane. Returns status ok with count sent; not_found = no such chat; dead = the pane vanished mid-sequence, count says how many landed; a tool error = an unknown key name, the valid ones listed.
Output schemas completely undocumented. No tool describes what fields it returns, forcing LLMs to infer structure and plan blind. chat_ls returns 'rows plus matched, truncated, and the filter echoed back' but no schema. chat_resolve returns 'socket_path and pane' but no type info. Breaks downstream tool chaining.
Error handling lacks recovery guidance. Descriptions state error codes (not_found, ambiguous, refused) but do not tell LLMs what to do next. E.g., chat_resolve returns 'not_found (1)' with no suggestion to call chat_ls or chat_find. Agents cannot self-correct.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2025-06-18+ | v2 |
Terminates a chat session — "kill / end chat X". Call chat_kill{target:"my-chat"}. Closes the chat and saves its state.
Returns the last answer from a chat — "what did X answer last". Call chat_last{target:"my-chat"}. Returns the most recent assistant response.
Lists live and resumable chats as rows — "what chats are running", "is there a chat named X". Call chat_ls{} or chat_ls{project:"substring", all:true}. Returns rows plus matched, truncated, and the filter echoed back; rows empty with matched 0 = nothing matched; a tool error = the fleet could not be read. Address a row with chat_inject by its session or name.
Renames a chat — "rename chat X to Y". Call chat_name{target:"my-chat", new_name:"new-name"}. Changes the chat's display name.
Spawns a new chat — "spawn / start a new chat". Call chat_new{project:"my-project", name:"my-chat"}. Returns the new chat's session and socket.
Opens a chat in the terminal — "open chat X". Call chat_open{target:"my-chat"}. Brings the chat window to focus.
Reads transcript from a chat — "find / read an old transcript". Call chat_read{source:"chat-id", last_n:10, max_bytes:1000}. Returns turns with role and text, count, and truncation flag.
Resolves one exact name to its tmux socket and pane — "where does chat X live", checking a target before chat_inject or chat_keys. Call chat_resolve{kind:"label", name:"my-chat"}. Returns status ok (code 0) with socket_path and pane; not_found (1) = no such chat; ambiguous (2) = several match, listed in candidates; a tool error = the resolver itself failed.
Dumps a chat's transcript to a file — "dump my transcript to a file". Call chat_save{source:"my-chat", path:"/path/to/file"}. Saves the full transcript.
Compacts THIS chat in place after its turn settles and KEEPS the session (crons, sub-agents, pane) — the only answer to "compact yourself" / "self-compact at this milestone", this tool and nothing else, never a hand-typed /compact. Call chat_self_compact{focus:"one line", then:"one steer"} — exactly ONE post-compact steer, a string never a list. Only focus and then cross the boundary — write durable state to disk FIRST. END THE TURN IMMEDIATELY after it returns, run no further tool; more work lands the steer beside the compaction. Main chat only — a sub-agent has no pane.
Reports a chat's status — "is chat X idle, what is it doing". Call chat_status{target:"my-chat"}. Returns idle/busy state and current activity.
Resumes a killed chat — "bring back a killed chat". Call chat_unkill{target:"my-chat"}. Restores a previously terminated chat session.
Reports THIS chat's own identity — "who am I", "what is my address" — the session name another chat targets with chat_inject. Call chat_whoami{}. Returns status ok with session, socket_path, pane, engine, id; not_found = this transport carries no caller identity, message names the pfm chat command to run from the chat's own shell instead. A sub-agent has no chat identity — it reports to its parent by returning.
Files an issue with the servicedesk — "report a problem / file a bug". Call issue_servicedesk{title:"...", description:"..."}. Creates a support ticket.
Destructive tools (chat_kill, chat_unkill, chat_self_compact) lack confirmation or dry-run support. No pattern to prevent accidental termination. Descriptions warn 'END THE TURN IMMEDIATELY' but do not offer a confirmation step.
Parameter descriptions lack format constraints. chat_new accepts 'project' and 'name' but does not specify allowed characters, length limits, or naming conventions. chat_inject 'message' parameter unbounded. Invites invalid input from LLMs.
STDIO transport only. Not remotely accessible; cannot be used by hosted MCP clients. Limits deployment to local/containerized environments. Blocks cloud-based agent platforms.