{"report":{"v":1,"package":"@agentistics/mcp","version":"2.217.1","ranAt":"2026-11-16T11:25:48.843Z","install":{"ok":true,"integrity":"sha512-RP8VluuBO42i8uy69yz8nEHA5sVOdfhrCJNJLRdB9cvmTGr1jpVT5CBisyKOPajvwPtQwgb6Az+PyVNPxjVcKg==","deps":95,"scripts":[],"bins":["agentistics-mcp"],"license":"FSL-1.1-ALv2","engines":"node >=18","repository":"git+https://github.com/agentistics/agentistics.git","bin":"agentistics-mcp"},"start":{"ok":true,"ms":233,"serverName":"agentistics","serverVersion":"1.0.0","protocol":"2025-06-18"},"tools":[{"name":"agentistics_tasks","description":"List every task on the ALM board with its rollup (cost, rounds, sessions used, tokens) plus a board-wide overview: tasks in flight, delivered, average cost per task and per task, top models and harnesses. START HERE to see what work exists and what the other assistants are on.","params":[]},{"name":"agentistics_task","description":"Open ONE task: its attempts (one per configuration tried), the sessions filed under it with their live state, its comments, subtasks, files, links and full metrics (models, harnesses, agent runs, tokens, delivery time). Comments come twice: `comments` (flat, oldest first, each with `subtaskId` when it was left on a subtask or group) and `commentThreads` — the same comments grouped by their OWN target (`target: {kind: 'task'|'group'|'subtask', id, title}`; the task first, empty targets omitted). `ref` is the task id or its exact title.","params":["ref"]},{"name":"agentistics_task_create","description":"Create a task. A title is required; a description is optional. Set `priority` (urgent | high | medium | low — default low, and nobody chose a default) and `type` (an id from agentistics_task_types) when you know them: a board of defaults orders nothing. Creating a title that already exists returns the existing task rather than a duplicate (its priority and type are left alone — use agentistics_task_edit), so this is safe to call before filing work under a name you are not sure exists.","params":["title","detail","priority","type"]},{"name":"agentistics_task_status","description":"Move a task to a status. The status VOCABULARY is an editable list now, not a fixed set — call agentistics_task_statuses first if you are unsure what this board currently has; a fresh board starts with exactly `todo`, `in_progress`, `blocked` and `done`. Only `done` stamps a task and closes rounds-to-delivery; `abandoned` (when the board has it) records that the work was given up on, which is a real outcome and not a failure to report — but it is NOT created by default, so use agentistics_task_status_edit to add it (or any other status) first if it is missing. **`blocked` REQUIRES a `reason` or a `blockedBy` list** and is refused (422) without one: it is the status that names a problem somebody has to go and solve, and a blocked card that does not say what it is waiting on cannot be unblocked by anyone but you. **`done` REQUIRES the task to have at least one session filed under it** and is refused (422, `done_needs_session`) without one. A `status` naming nothing this board currently has is refused (400, `unknown_status`). Pass `actor` (who you are) so the move is recorded against you in the activity log.","params":["ref","status","actor","allowOpenSubtasks","reason","blockedBy"]},{"name":"agentistics_task_statuses","description":"List every status this board currently has (id, label, color, and whether it is protected — todo/in_progress/blocked/done can never be deleted or have their id changed, but their label and color can). Call this before agentistics_task_status if you are not sure a status you want to move a task to actually exists.","params":[]},{"name":"agentistics_task_types","description":"List the task TYPE vocabulary (id, label, color, order, usageCount). A type is a second classification beside the status (e.g. CORE); a task may carry one or none. Call this before setting `type` through agentistics_task_edit.","params":[]},{"name":"agentistics_task_type_edit","description":"Add, rename/recolor, or delete a task type. Pass `label` (and optional `color`, a `#rrggbb` hex string) alone to CREATE a type — its id is derived from the label. Pass `id` with `label` and/or `color` to EDIT. Pass `id` with `remove: true` to DELETE: refused (422, `in_use`, with `usageCount`) while any task still carries it.","params":["id","label","color","remove"]},{"name":"agentistics_task_status_edit","description":"Add, rename/recolor, or delete a status. Pass `label` (and optional `color`, a `#rrggbb` hex string) alone to CREATE a new, non-protected status — its id is derived from the label and returned. Pass `id` with `label` and/or `color` to EDIT an existing status's label/color — works on a protected one too, only its id can never change. Pass `id` with `remove: true` to DELETE it: refused (422, `protected`) for todo/in_progress/blocked/done regardless of usage, and refused (422, `in_use`, with `usageCount`) for any other status still referenced by at least one task or subtask.","params":["id","label","color","remove"]},{"name":"agentistics_task_comment","description":"Leave a comment on a task — or on ONE of its subtasks or subtask GROUPS by passing `subtaskId` (a group's own id is allowed; omit it to comment on the task itself). `body` may be empty when `attachments` is given. Comments carry attachments as REFERENCES into the chat's own attachment store (`{name, path}`; agentistics_task returns them on each comment with a `url` that serves the file) — there is no second store. Use it to tell the person and the other assistants what you did, what you found, or what you are blocked on; comment where the work is (the subtask you are on) rather than on the whole task. An unknown, deleted or other task's `subtaskId` is REFUSED with a sentence (422, `no_such_subtask` / `wrong_delivery`) — never silently filed on the task. Threads read downward: a subtask's thread is its own comments, a group's thread is its own plus its members', the task's is everything. The activity log names the target. `author` is free text — say who you are (e.g. 'claude:3f5f'). THREADS: a thread is a topic on the task; post into one with `threadId`, or open one with `threadTitle` (sessions: handbacks and blocks only). A thread is a RECORD, not a chat: a comment there reaches nobody by itself, so never wait for an answer in a thread. When you post in a thread you join it; the person may explicitly SEND a message from it, which arrives in your own chat starting `[Agentask · … · sent by the owner]` — answer there, in your chat, which is where conversations live. If you need something from the person, ask in your own chat as always.","params":["ref","body","author","subtaskId","threadId","threadTitle","threadKind","kind","attachments"]},{"name":"agentistics_task_subtask","description":"Joining a staged subtask to a group folds its queued delivery and brief into the group, keeping one queue card and an audit reason. Add a subtask, or change one. Pass `title` to add; pass `id` with `done`, `status`, `dueDate`, `startDate`, `blockedBy` or `parentGroupId` to edit one. `startedAt`/`deliveredAt` are system-stamped facts (when real work began / when this piece reached `done`) and are never set through this tool — they are read back on the subtask record. A subtask carries the same columns its parent does; its own cost, tokens, rounds and harness are rolled up from the sessions filed under it — the PARENT task's rollup is the sum of all its subtasks' sessions. A subtask CAN be blocked by another subtask of the SAME parent (pass `blockedBy` as the full list of sibling subtask ids that must be `done` first; a sibling outside this task, or the subtask itself, is dropped rather than accepted) — this is separate from agentistics_task_blocked_by, which blocks a whole TASK on other tasks. **`done` REQUIRES the subtask to have at least one session filed under it — or a VERIFIED report on it (agentistics_task_comment with its subtaskId) from a session of the same task, e.g. the leader that closed it; the report must carry `kind: \"handback\"` (a plain comment without a kind does not count)** — and is refused (422, `done_needs_session`) without one — EXCEPT for a group MEMBER (`parentGroupId` set), which can never hold a session of its own (see below) and therefore reaches `done` through this same edit with no session required; a GROUP itself (`isGroup: true`) and an ordinary loose subtask both still need one. **The moment every one of a task's top-level subtasks (loose subtasks and groups — a group's own members never count separately) reaches `done`, the parent task auto-delivers too.** To put a session ON a subtask use agentistics_task_session with `subtaskId` — a subtask holds ANY NUMBER of sessions, so the link lives on the session, not in a field here. **SUBTASK GROUPS are a real hierarchy level, not a label**: pass `isGroup: true` with `title` to create a GROUP — a peer row that CAN hold a session, exactly like a loose subtask. Pass `id` (a subtask) with `parentGroupId` (a group's own subtask id, from the SAME parent task) to make it a MEMBER of that group — join with the group's id, leave with `parentGroupId: ''`. A refused reference (the id names no subtask, names one that is not `isGroup: true`, names a different task, or the subtask being patched is itself a group) is refused as `invalid_group` (422). **Joining a group is ALSO refused (422, `subtask_has_sessions`) when the subtask already has a session filed on it** — a member gets no rollup bucket of its own, so that session's cost would silently drop out of every visible breakdown while the task's own total kept counting it; detach the session first. **A MEMBER can never hold a session of its own** — filing on one is refused by agentistics_task_session with `subtask_in_group` (422); o","params":["ref","title","id","done","status","dueDate","startDate","blockedBy","isGroup","parentGroupId"]},{"name":"agentistics_task_link","description":"Attach a link to a task — a pull request, an issue, a document. Pass `url` (http/https only) with an optional `label` and `kind` ('pr', 'issue', …). Pass `remove` with a link id to detach one.","params":["ref","url","label","kind","remove"]},{"name":"agentistics_task_blocked_by","description":"Set which tasks must finish before this one can proceed. Pass the full list of blocker task ids (it replaces, it does not append). A task cannot block itself.","params":["ref","blockedBy"]},{"name":"agentistics_task_next","description":"THE ORCHESTRATION CALL: what you can pick up right now. Returns the tasks that are open, not blocked by unfinished work, and not in another agent's hands — most urgent first, numbered from 1. It also returns `withheld`, every task that is NOT available with the reason (closed / blocked / claimed / status), because an agent told 'nothing' learns nothing; and `progress`, which separates 'it is all done' from 'everything is in somebody's hands right now'. Pass `actor` (who you are) so a task you already hold comes back as available to you rather than as taken.","params":["actor","limit"]},{"name":"agentistics_task_claim","description":"TAKE a task before working on it, so two agents cannot do the same one twice. This is a LEASE, not a lock: it expires (default 30 minutes) and the task returns to the board on its own, so an agent that dies does not hold work forever. Re-claiming as the same `by` REFRESHES the lease — do that while you are still working. Pass `release: true` to give it back when you finish or stop. A refusal names who holds it and until when; `takeover` is for a person overriding a live claim on purpose and an agent should not send it.","params":["ref","by","release","leaseMs","sessionId","note","takeover","force"]},{"name":"agentistics_task_activity","description":"What has been HAPPENING, newest first — status moves, claims, releases, priority changes, sessions filed. Pass `ref` for one task, or nothing for the whole board. On a board several agents drive this is how you find out what the others did without asking them.","params":["ref","limit"]},{"name":"agentistics_live_delivery","description":"What is being DELIVERED right now — the Live Delivery line, the same source as the screen. Use it to answer \"what is running\", \"when does X ship\" and \"is anything stalled\". Returns a `summary` (active deliveries, agents working, stalled, the next delivery with an estimate), the active `deliveries` with their stage (queue, running, qa, adjust, leader, merge, published), the agents on each (harness, model, account, live state), time in the stage and in total, the step inside the stage (`step`, with `stepSource` = `reported` or `signals` and `stepSignal` naming the observed fact — null `step` means nothing is observable, say so; the screen then draws no percent), the ETA with its source and confidence (`kind: none` means there is NO estimate — say so, never invent a time), the last action, and the `leaders` band (each leader's current action, its workers and what waits on a decision — leaders are not deliveries while the task has an open piece of its own; a leader working alone on a task with none gets ONE piece, `leaderPiece: true`, on the leader session's own clock). Each delivery carries its Agentask `priority` (urgent | high | medium | low; a subtask inherits its task's). Filters: `task` (id or part of the title), `stage`, `stalledOnly`, and the screen's own tabs — `tab`: `active` (Em andamento: every stage except queue, published, completed and old — a blocked or adjusting piece with no agent on it is in), `queue` (Na fila), `published` (Publicado) or `all` (Tudo: every piece in its real column, old legacy reviews marked `old`; the default here, while the screen opens on `active`) — and `period`, the Publicado period: `24h` (default), `7d`, `30d` or `all`. The answer's `view` carries `{tab, period, counts, defaulted, capped}`: `tab`/`period` are the EFFECTIVE values (a missing or invalid one falls back to `all` / `24h`, and `defaulted` names which did), `counts` is the count of EVERY tab under the same `task` filter and period, the very numbers the screen's tabs print, and `capped` (`{cap, omitted}`, else null) says the list was cut at 400 pieces per kind and that many older ones are NOT in the answer — then a count is not the total. Plus `sort` (how each stage lists its cards; default `priority`, highest first, ties by longest in the stage). Stage names on screen: queue = Na fila, running = Em desenvolvimento, qa = QA. There is no Histórico column any more: a piece published earlier is in `deliveries` whenever `period` covers it.","params":["task","stage","stalledOnly","tab","period","sort"]},{"name":"agentistics_task_edit","description":"Set a task's fields: `title`, `detail`, `priority` (urgent | high | medium | low), `type` (an id from agentistics_task_types; empty string clears it), `dueDate` / `startDate` (yyyy-mm-dd), `labels`. An absent field is left alone; an EMPTY STRING clears it. `priority` defaults to `low`; a legacy `none` is stored as `low`. `startedAt`/`deliveredAt` are system-stamped facts and are never set through this tool. Pass `actor` so the change is recorded against you in the activity log.","params":["ref","title","detail","priority","type","dueDate","startDate","labels","actor"]},{"name":"agentistics_task_session","description":"File a SESSION under a task, which is what makes the task measurable: its cost, tokens, rounds and harness all come from the sessions filed under it. Pass `subtaskId` to file it under one of the task's subtasks (the unit of WORK, for a task broken into steps) — this INCLUDES a subtask GROUP (see agentistics_task_subtask's `isGroup`), which holds a session exactly like any other subtask; omit `subtaskId` to file the session directly on the task itself (the unit of DELIVERY) — **only LEADER sessions may be filed on the task itself; a worker is refused (422, `leader_only`: only leader sessions are filed on the task itself; file this session on a subtask)**, so pass a `subtaskId` for workers. Both can be used on the same task — some sessions direct, some under subtasks — and each subtask's own rollup plus the directly-filed sessions' own rollup are both readable via agentistics_task's `subtaskRollups` (keyed by subtask id, with `id: null` for the direct ones), so the breakdown never disappears. **Filing on a subtask that is a MEMBER of a group (has a `parentGroupId`) is refused (422, `subtask_in_group`)** — a member can never hold a session, only the group itself can; file on the group's own id instead. Use agentistics_task to read the subtasks, or agentistics_task_subtask to create one, then pass `ref` + `sessionId` (a managed session id) + optionally `subtaskId`. Pass `detach` with a session id to unfile one. The task inherits the session's repository when it has none. **HISTORICAL conversations — pass `conversationId` INSTEAD of `sessionId`** to file a conversation that has NO live/registry row: typically a session that was killed and whose row agentop then purged, leaving its subtask `done` with 0 sessions and its cost missing from every rollup. The conversation must still be in the metrics store (its tokens/cost are what get counted); the link is kept on the board, never in the fleet, so nothing appears in the Sessions workspace and there is nothing to open — the row is marked `historical: true` in agentistics_task's `sessions`. Filing is a MOVE: one link per conversation, so filing it elsewhere replaces the old link and the answer reports `movedFrom: {taskId, subtaskId?}` (a plain `sessionId` filing reports `movedFrom` too when it displaced another filing). It counts everywhere a live session does (rollups, `subtaskRollups`, the board headline, `done_needs_session`, the group-join rule). Refusals: `no_such_conversation` (422 — the store has no record of that id; nothing to price, so nothing is filed), `conversation_in_fleet` (422 — a registry row already exists; the answer carries its `sessionId`, so call again with that as `sessionId`), plus the same target refusals as a session (`no_such_task`, `no_such_subtask`, `wrong_delivery`, `blocked`, `subtask_in_group`, `leader_only`). Optional `harness` disambiguates and `note` is kept on the link. To remove a historical link pass `detach` with the conversation id (or the row's `hist:<conversationId>","params":["ref","sessionId","conversationId","harness","note","subtaskId","detach"]},{"name":"agentistics_task_delete","description":"Delete a task and its comments, subtasks, files and links. The SESSIONS filed under it are kept — deleting a board entry never deletes work.","params":["ref"]},{"name":"agentistics_delivery_calendar","description":"BETA. The DELIVERY CALENDAR — when each piece of work started, when it is expected to ship, which release version it rides and when a version really went out; the same source and rules as the \"Calendário de entregas\" screen (`agentop flow calendario` prints the same). Returns `pieces` (task, subtask, `groupId`/`groupTitle` = the group subtask it sits under when it is a member of one (the task → group → subtask tree of the Cronograma), stage, `version`, harness/model, priority, `startedAtMs`, `stageEta` and `publishEta` — `kind: none` means there is NO estimate: say so, never invent a time — `publishedAtMs`, `dueDate`, `confirmed`, `late`), `versions` (each with `state` published|planned, `atMs` = the real publication or the LATEST piece estimate, `basis` published|estimate|partial|none, the pieces and the count per stage) and `marks` (an account's usage window renewing, a task's due day). Times are epoch ms; day arguments are BRT days (America/Sao_Paulo). By default only pieces of tasks being worked on now; pass `all` for the history too. Filters: `task` (id or part of the title), `subtask`, `version` (one or a list), `stage`, `harness`, `priority`, `from`/`to` (yyyy-MM-dd or an ISO instant — keeps what touches the period), `confirmedOnly` (published or a medium/high-confidence estimate). `format: \"ics\"` returns the filtered view as an .ics file instead of JSON.","params":["task","subtask","version","stage","harness","priority","from","to","all","confirmedOnly","format"]},{"name":"agentistics_task_phase","description":"FLUXO.CICLO-COMPLETO — the minute `tick` also moves a piece in revisao_lider or pr_merge to ci BY ITSELF once its PR is merged (found from the piece's PR number, branch, the PR containing the sha QA approved, or its author worktree's branch; the merge gates `delecoes` + `mergeado` still run, and a refusal leaves the piece where it is with the reason), publishes ci pieces from the FIRST tag that contains the merge, and reminds the LIVE leader of the task (never the owner) of a piece waiting on them for 30 min, then every 2 h, at most 3 times. `get` with a `subtaskId` returns `tickReason`: why the tick has not moved it. `reconcile` (no `subtaskId`; `apply` true to write) is the one-shot BACKFILL: every piece in revisao_lider/pr_merge/ci matched to its merged PR and the first release tag containing it — a report of what moves and what cannot be proven, with the reason. `tick` also enables the mandatory owner-request/placement flows through the official engine switches and reads canonical pedidos; after 10 minutes without any attended link it wakes the current live task leader at most once per pedido, across handoffs and restarts, and never asks the owner. Complete predispatch, a running/closed delivery or a deliberate hold/dequeue counts as attended; failed sends retry after 30 minutes. Queue actions: `dequeue` withdraws a queued piece with a required free-text `reason` (task leader or owner only), keeps the draft/history, and refuses dispatch until explicitly restaged; `tick` reconciles stale queued pieces against board review/done/abandoned and records a comment. Vacancy counts live backend processes; lost/exited/closed sessions reserve no slots. FLUXO.ENTREGA-DSL — read or move a DELIVERY (one subtask) through its fixed phases: fila → executando → qa → (ajuste ⇄ qa) → revisao_lider → pr_merge → ci → publicado, plus bloqueado (from any phase, back to the stored origin only). Any transition outside that table is REFUSED (422 `fase_pulada`) with a sentence naming what is missing — executando → revisao_lider, ajuste → revisao_lider and every other skip included. The gates each edge needs are run BY THE SERVER from evidence (predispatch, vaga, salvo = sha on origin per ls-remote, qa-modelo = QA model ≠ author model from the registry and never Sonnet over Sonnet, cabeca = the pushed branch is still at the sha QA tested, delecoes = no file deleted outright by the PR, mergeado = gh says the PR is MERGED, the merge commit is on origin/main and the PR head is the sha QA approved — the merge sha is READ FROM THE PR, never from you, release = the batch's version tag exists AND the release.yml run that produced it is green, read in the repository the release recorded); salvo needs every check run on the pushed head to have concluded green (pending or absent refuse); you never pass proof. `action`: `get` (default; omit `subtaskId` to list every delivery) | `phase` (`to` + what that edge reads: `verdict` APROVADO/REPROVADO + `findings` from qa, `reviewedSha` = ","params":["action","antigasAction","preview","all","expect","apply","subtaskId","version","mergeSha","repo","subtaskIds","to","verdict","findings","reviewedSha","head","detail","name","phase","decision","note","reason","text","step","patch"]},{"name":"agentistics_task_stage","description":"Placement is deterministic: a queued piece of the same task sharing a file, area tag or UI surface is returned as an `encaixe` warning so you can extend that brief or group the pieces. Withdrawn pieces are excluded. FLUXO.ENTREGA-DSL pre-dispatch — write the STAGED SESSION on a subtask (prompt/brief, harness, model, effort, account, cwd) and open its delivery in `fila` (\"Na fila\"). A leader receiving a request from the owner stages it in the SAME turn: the demand flow reminds the leader (never the owner) after 10 minutes of an owner request with no staged subtask. The fields merge over what the subtask already holds, and the predispatch gate is checked right away (the reply says what is still missing). Nothing starts here — `agentistics_task_phase action=stage-run` starts it once the vaga gate passes. `pedidoId` links the owner request this answers.","params":["subtaskId","prompt","harness","model","effort","account","cwd","versionTarget","pedidoId"]},{"name":"agentistics_task_evidence","description":"ENTREGA.GARANTIDA — attach, list or remove the EVIDENCE a piece needs before it can move to review (the server decides what is enough; you never supply a hash, a time or a commit — it computes them). `ref` = the TASK id; `subtaskId` = the piece (omit for a task-level piece). `action`: `add` (default) | `list` | `remove`. `add` needs `kind` = screenshot | api-capture | test-log | report | probe and ONE payload: `filePath` (a file in your worktree, uploaded — 8 MB cap), `fileId` (a TaskFile of this task), `repoRef` {repo, sha, path} (a `report` pushed to a remote branch) or `probe` (http | bundle | cli | structural; closed set, no extra fields). Validators: screenshot = real PNG/JPEG/WebP header ≥ 320×200 and ≥ 4 KB; api-capture = UTF-8 ≤ 1 MB with a request/status line and a body; test-log = runner summary with 0 failures and ≥ 1 pass (or an `agentop run` footer `exit 0`); report = markdown ≥ 300 B with a `Verdict:`/`Veredito:`/`Result:`/`Resultado:` line. The same bytes can back evidence on ONE piece and ONE revision, forever (`evidence_reused`). Evidence is bound to the piece's verified HEAD: a new commit or an uncommitted change makes it stale — recapture. `list` returns the items and the binding (HEAD, tree clean) they are judged against. `remove` (`evidenceId`) is refused (409 `evidence_in_use`) when a checkpoint relies on the item. Only the piece's own session, the task leader or the owner may attach.","params":["ref","action","subtaskId","kind","filePath","name","fileId","repoRef","probe","evidenceId"]},{"name":"agentistics_task_queue","description":"FILA.PRIORIDADE — read the delivery QUEUE (\"Na fila\") exactly as the task page's Fila tab shows it, from the same source, in the order pieces LEAVE: the task's priority (urgent → low), then arrival. Each item: `position` (in its task) and `globalPosition`, `priority`, the subtask/group (with `group.members`), the full PRE-DISPATCH (`predispatch`: harness, model, effort, account and its name, cwd, prompt), the slot gate's last verdict (`slot`; ok=false carries the raw reason such as `no slot: 4 workers running (machine ceiling 4)`), `hold` (the leader said `nao <motivo>` or `agendar <ISO> <motivo>`: kind, reason, until, who, `active`), `ask` (the system question to the leader: when, delivered, approved, `expiresAt` = 10 min, `onTimeout` = `sim` with autonomous mode on, `esperar` with it off), the leader, and the subtask's notes (`notes.last` the newest 3, `notes.total`). `recent` lists pieces the queue opened in the last 24 h with the session they opened. Pass `taskId` for one task's queue. To change the order, change the TASK's priority with `agentistics_task_edit`; to edit a piece's pre-dispatch use `agentistics_task_stage`; to start now / schedule / answer use `agentistics_task_queue_answer`; to withdraw use `agentistics_task_phase action=dequeue`.","params":["taskId"]},{"name":"agentistics_task_queue_answer","description":"FILA.PRIORIDADE — answer the queue's SYSTEM QUESTION (the message headed `[from flow fila · confirmar]` that arrives in a leader's composer when a slot opens and its piece is next), or act on a queued piece from the Fila tab. Leader of the task or owner only. `answer` is one of: `sim` (open the session now — if no slot is free right now the approval is kept and the next tick starts it; never opens twice), `nao <motivo>` (the piece stays queued with the reason on its card; the flow moves on to the next one), `agendar <ISO> <motivo>` (stays queued, becomes a candidate again only from that FUTURE instant, shown on the card as `começa às HH:MM`). A `nao` or `agendar` without a reason, and an `agendar` in the past, are refused with a sentence. Every decision is recorded as a comment on the subtask (who, why, when). With no answer in 10 minutes: autonomous mode ON → the flow carries on as `sim` and records it; OFF → it keeps waiting and reminds you. Set `runNow: true` instead of `answer` to run the queue flow once right now (what happens automatically when a worker ends and every minute).","params":["subtaskId","answer","runNow"]},{"name":"agentistics_session_groups","description":"List the user's SESSION GROUPS — the named folders in the Sessions sidebar (\"Saved to later\", \"Pelvie\", …) — with the sessions in each. A group is how a person keeps a fleet of assistants organised; every session belongs to at most one. Each member carries its `id` (the managed session id you can pass to the other tools), `title`, `state` and `harness` when the session is on this machine right now, or only its `key` when it is gone. Call this BEFORE creating a group, so you file a session under an existing one instead of making a near-duplicate.","params":[]},{"name":"agentistics_session_group_create","description":"Create a session group, optionally filing sessions into it in the same call. Use it to keep the assistants you start organised: start a session (for example `agentop session start …`, which prints its id), then create or reuse a group and file it. `sessions` takes session REFERENCES — a managed id (the `agentop-…` id `agentistics_session_groups` and `agentop session ls` print), a conversation id, an exact title, or a unique id prefix; any reference that matches nothing or more than one refuses the WHOLE call (404 / 409) and creates nothing, so a group is never half-filled. Filing moves a session out of any other group and unpins it if it was pinned. Duplicate group names are allowed, but they make later calls ambiguous — check `agentistics_session_groups` first and reuse a group when one fits.","params":["name","sessions"]},{"name":"agentistics_session_group_edit","description":"Change a session group. `action` is one of: `add` — file `session` under `group` (moves it out of any other group, unpins it if pinned); `remove` — take `session` out of whichever group holds it (the session itself is untouched); `rename` — give `group` a new `name`; `delete` — delete `group` (its sessions are NOT deleted, they only leave it). `group` is a group id OR its exact name (case-insensitive); a name shared by two groups answers 409 with the ids, so use the id. `session` is a managed id, a conversation id, an exact title or a unique id prefix. Refusals name what was wrong: 404 for a group or session that matches nothing, 409 for an ambiguous one, 400 for a blank name or a missing argument.","params":["action","group","session","name"]},{"name":"agentistics_session_notify","description":"Switch a session's notifications on or off, or read the switch. `ref` is a managed id, a conversation id, an exact title or a unique id prefix. With `notify` ('on' | 'off') it sets the switch; with `ref` alone it reads it. Muting removes the INTERRUPTION only (bell, Nay card, sound, desktop toast, peer message): the session still shows as waiting and its events are still recorded in the inbox. The mute follows the conversation, so it survives a reopen. 404 for a ref that matches nothing, 409 for an ambiguous one.","params":["ref","notify"]},{"name":"agentistics_session_message","description":"Send a message to ANOTHER agentistics session — the standard channel between a session and the one that started it (or the ones it started). `to` is a session reference (managed id, conversation id, exact title or unique id prefix); `kind` is handback (work finished/committed), block (you cannot continue) or question (you need a decision); `body` is the text. It is typed into the target's own conversation through the same path its composer uses, headed `[from session <id> · <kind>]`, and — when YOUR session is filed on a task or subtask — also recorded there as a comment of that kind. Use it to report to the session that started you instead of asking the user; a parent may message its children the same way. Refused for a session that does not exist, for yourself, from a process that is not a verified agentop session, and more than once per 2 seconds to the same session. INTERMEDIATE handback: a handback is read by the runtime as \"I finished everything\" (the product may close you once you are stopped, clean and saved). If you stop on purpose between phases, waiting for a signal, pass `awaiting` (kind=handback only, one line ≤200 chars, e.g. \"engine merged <sha>\"): the runtime then refuses to close you while that is your newest handback. A later handback WITHOUT `awaiting` counts as final again.","params":["to","kind","body","awaiting"]},{"name":"agentistics_wait","description":"EXPERIMENTAL (AGENTISTICS_FLOWS) — a DURABLE wait the SERVER keeps for YOU: instead of polling or running a background command, register what you wait for and a deadline; when the condition is met (or the deadline passes, once) the server types a wake message into your own session, headed `[from flow wait · nudge]`, naming the wait id and generation. Conditions are a closed set: `{kind:'handback', from:<session id>, since?:ISO, match?:text}` (a handback FROM that session; `match` = case-insensitive substring its text must contain, so another subject does not wake you), `{kind:'memory', minAvailableMB}` (free memory, two stable readings), `{kind:'time', at:ISO}`, `{kind:'gate', gate:'sessao-parada'|'arvore-limpa'|'trabalho-salvo', sessionId}`. action=register needs `condition` + `deadlineAt` (ISO UTC, ≤30 days; pass `requestId` to make a retry the same wait); action=ack (`waitId`, `generation`) confirms you RECEIVED the wake — without it the wake is sent again (at most 3 times), so ack every wake, and treat a repeated wake for the same id+generation as already handled; action=cancel (`waitId`) before it fires; action=list your open waits; action=handoff (`from`: the previous leader on the same task) moves its waits to you. A wait is never 'working' nor an approval.","params":["action","condition","deadlineAt","label","requestId","reopen","waitId","generation","from","all"]},{"name":"agentistics_session_role","description":"Get or set a session's ROLE (leader | worker | auto). With no `ref` it is YOUR OWN session. Setting `leader` stores the role, renames the session to \"[ LEADER ] - <title>\" (\"[ LÍDER ] - …\" in a Portuguese UI) keeping the user's title, and the briefing then tells the model it is a leader; `worker`/`auto` strip the prefix. Use it to declare yourself leader when the work grows to several fronts — do not rename the session by hand. Setting ANOTHER session's role works only for sessions you started yourself. Idempotent. With no `role` it just reports the current one and the role line.","params":["ref","role","retirePredecessor"]},{"name":"agentistics_session_switch_account","description":"Continue a session on ANOTHER account — of its own harness or of another harness (multiple accounts must be on). Use it when a session hit its account's usage limit and another account has room. With no `accountId` it lists every account the session can move to, each with its `harness` (its own harness's other accounts first, then every other harness that can start a session here; the recommended one per harness from the plan router) and whether its last turn stopped on a usage limit. With `accountId` and no `harness` (or its own harness) it ends the session, copies its conversation to that account (never moves it) and resumes the SAME conversation there — Claude Code and Codex only. With `harness` set to ANOTHER harness it starts a NEW session there in the same folder, task, subtask and role, briefed with a summary of the conversation, and ends the old one only after the new one is running (a failure leaves the old one untouched). The new id is returned. `resendPending: true` re-sends, once, the message the limit refused. An account that is off for agents, removed or signed out is refused in words.","params":["ref","accountId","harness","resendPending"]},{"name":"agentistics_session_away","description":"Read or switch AUTONOMOUS MODE, or send the owner a notice while autonomous — the fleet works on its own, without asking permission or questions (the owner may be asleep or travelling). While it is on, every new session is born in the no-questions approval mode unless its caller names a mode. With no `on` it reports the state (`active`, `until`, what is pre-authorised). With `on: true` it switches it on: `until` is an ISO instant (omit = until the owner says they are back), `authorised` lists what the owner pre-approved (releases = publish releases the leader approved; workers = open worker sessions within the plan limits; merges = merge PRs the leader reviewed), `note` is free text. `schedule` makes it SCHEDULED instead of open-ended: `{enabled:true, days:{mon:{windows:[{from:\"09:00\",to:\"18:00\"}]}, tue:{windows:[],allDay:true}}}` (local clock, weekdays mon…sun, hour windows per day, `allDay` = whole day) — it is on only inside those windows. Omitted `authorised` = all three. `on: false` switches it off. Every session started while it is active is told in its briefing. With `target` (a session's managed id or conversation id) `on` switches THAT session's own autonomous mode instead — it alone is told the owner is out, even with the global switch off, and a running one is told the change; sessions it starts inherit it. Ceilings are per PROVIDER: `topCaps` maps a harness (claude, codex, gemini, copilot, antigravity, kimi, opencode) to how many workers may run on its top model tier at once. Only a LEADER session may switch it; use it when the owner tells you they are leaving or are back.","params":["on","until","authorised","note","schedule","maxWorkers","topCaps","target","notice"]},{"name":"agentistics_checkin_report","description":"Report the leadership check-in round. The owner reads it and is NOT technical: every field is a short plain-language sentence naming things by what they do — no session ids, commit hashes, PR numbers, ports, limit percentages or internal codenames unless the owner must act on one; a health row only when something needs attention, said in words. Include a concise headline (<=140 chars), delivered items, in-progress items with stage and expected release, next steps, and health rows (sessions/memory, limits, releases, MCP, handoff). Keep summary as the fallback for older/short rounds. The server records the round, updates the agenda, and adds a checkin activity event to the filed task.","params":["state","summary","headline","delivered","inProgress","next","health","actions","needs"]},{"name":"agentistics_summary","description":"Get an overview of AI coding usage metrics (across all tracked harnesses — Claude Code, Codex, Gemini, Copilot, Antigravity — or scoped to one): total tokens, estimated cost, sessions, streak, most used model, and top project. Good starting point for any metrics question.","params":["harness"]},{"name":"agentistics_harnesses","description":"Compare the tracked AI coding harnesses side by side. Lists every harness present in the data with its sessions, messages, token usage, estimated cost, and last-active date. Use this to answer 'which harness do I use most / costs most' questions. Rows tagged source \"projections\" (read from the journal) carry personTurns (the person's turns); there `messages` is a deprecated alias of personTurns, and subagent spend is reported apart (subagentTokens, subagentCostUSD).","params":[]},{"name":"agentistics_plan_limits","description":"Read this BEFORE opening sessions; follow the recommendation; a plan on hold gets no new worker. Returns subscription budgets, pacing, and a routing recommendation.","params":["kind","authorHarness"]},{"name":"agentistics_benchmark_observed","description":"The Benchmark's observed history, read-only and identical to the Benchmark page: cost per delivery, time to publish, active time, deliveries per week, coordination range, QA rejection and rounds, escaped defects and wasted cost, each with n, a 95% interval, the named gaps that kept records out and its origin. Also `history`: per task (with subtasks) and per sessions-per-piece bucket — cost (tokens × price), wall time, active time, sessions, QA rounds, rework, size by lines — with coverage. Observed and biased, never a verdict; a null value is not measured, not zero. Optional from/to (epoch ms or ISO date) restrict the period by publication/reference date.","params":["from","to"]},{"name":"agentistics_projects","description":"List projects with session counts, message counts, token usage, estimated cost, and last active date. Optionally scope to a single harness. Rows tagged source \"projections\" (read from the journal) carry personTurns (the person's turns); there `messages` is a deprecated alias of personTurns, and subagent spend is reported apart (subagentTokens, subagentCostUSD).","params":["harness"]},{"name":"agentistics_sessions","description":"Get the most recent sessions with harness, project path, duration, message count, token usage, and model. Optionally scope to a single harness. With AGENTISTICS_FLOWS on, a worker that handed back and is still open carries `deliveredStillOpen` ({managedId, handbackAt, lastRun?}): close it with agentistics_flow_run name=fechar-entregue {sessionId: managedId}, which only closes it when stopped, clean and saved.","params":["limit","harness"]},{"name":"agentistics_costs","description":"Get cost breakdown by model: token counts (input/output/cache read/write) and estimated USD cost per model. Scope to a single harness, or 'all' (default) for the unified breakdown.","params":["harness"]},{"name":"agentistics_component_catalog","description":"List all components available for the custom layout page, with IDs, categories, descriptions, and default grid sizes. Call this before building or modifying a layout.","params":["category"]},{"name":"agentistics_get_layouts","description":"Get all custom layouts: names, components in each, which is active, and each component's instance ID for removal.","params":[]},{"name":"agentistics_get_preferences","description":"Read the Nay persona preferences (chatResponder, chatWorkingDir, chatLabel and chatMcpEnabled).","params":[]},{"name":"agentistics_set_preferences","description":"Set Nay persona preferences. chatWorkingDir must be an existing absolute directory; the server canonicalises and validates it.","params":["chatResponder","chatWorkingDir","chatLabel","chatMcpEnabled"]},{"name":"agentistics_create_layout","description":"Create a new empty named layout on the custom page.","params":["name"]},{"name":"agentistics_add_component","description":"Add a single component to a layout. Use agentistics_component_catalog to get valid componentId values. Positioned automatically at the bottom of the grid unless x/y are given.","params":["layoutName","componentId","x","y","w","h"]},{"name":"agentistics_remove_component","description":"Remove a component from a layout by its instance ID (the 'instanceId' field from agentistics_get_layouts).","params":["layoutName","itemId"]},{"name":"agentistics_set_active_layout","description":"Switch the active layout shown on the /custom page.","params":["name"]},{"name":"agentistics_delete_layout","description":"Delete a named layout. Cannot delete the last remaining layout.","params":["name"]},{"name":"agentistics_export_pdf","description":"Generate a PDF report of Claude Code usage. Returns a URL that, when opened in a browser, automatically opens the PDF export modal pre-configured with the requested options. The user can review settings and click Download to save the PDF.","params":["range"]},{"name":"agentistics_build_layout","description":"Build a complete layout in one call: creates it, adds all requested components with auto-positioning, and optionally activates it. Ideal for setting up a themed dashboard from scratch.","params":["name","componentIds","activate"]},{"name":"agentistics_tags","description":"List all tags — saved cross-cutting filters over repos/projects/machines/teams/accounts — with their aggregate sessions, cost, and tokens. Use for 'what tags exist' or 'how much did tag X cost' questions.","params":[]},{"name":"agentistics_vault_status","description":"How this machine's vault is protected — never what it holds. Returns the state (open/locked/uninitialized…), the ways to confirm it is you already in use (touchid = Touch ID or the Mac password, hello = Windows Hello, password = the vault's own password, fido2 = a security key), the ways this machine offers (a security key only when its tools are present, marked advanced), whether an old app-code record is still on disk (it is no longer asked; it is deleted once a way to confirm is chosen), whether the 24-word recovery key exists, the auto-lock minutes, and the vault-password rules (the same table the page and the server enforce). Setting up, unlocking and confirming are the person's: on the /vault page or with `agentop vault` — never pass a vault password or recovery words through a tool.","params":[]},{"name":"agentistics_tag_detail","description":"Get full detail for one tag by name or id: aggregate totals plus per-source breakdown and distributions (top projects, models, harnesses, repos), daily activity series, and the active date window. Call agentistics_tags first to find the id/name.","params":["tag"]},{"name":"agentistics_repos","description":"List repositories, grouped by normalized git remote (independent of local path or which machine produced the session), with session/message/token/cost totals and last-active date. Sessions with no linked repository are grouped under 'unlinked'. Optionally scope to a single harness. Rows tagged source \"projections\" (read from the journal) carry personTurns (the person's turns); there `messages` is a deprecated alias of personTurns, and subagent spend is reported apart (subagentTokens, subagentCostUSD).","params":["harness"]},{"name":"agentistics_book_takeover","description":"Confirm a real successor's takeover of a task through the hand-off book. The MCP server attaches this session's proof automatically; never provide a token.","params":["taskId","bookmarkHash","lastDeltaHash"]}],"toolsHash":"ce18ecd1b9413ab8170e9e1dbff6ee1bc96a58452a32f30cddb8da1304837d0c","wrote":[],"callsAt":"2026-11-16T11:25:55.546Z","calls":[{"tool":"agentistics_summary","ok":false,"ms":9,"bytes":0,"error":"tool reported an error: Error: fetch failed"},{"tool":"agentistics_get_layouts","ok":false,"ms":2,"bytes":0,"error":"tool reported an error: Error: fetch failed"},{"tool":"agentistics_get_preferences","ok":false,"ms":1,"bytes":0,"error":"tool reported an error: Error: fetch failed"},{"tool":"agentistics_vault_status","ok":false,"ms":1,"bytes":0,"error":"tool reported an error: Error: fetch failed"}],"host":{"commit":"eaaffffaf52255d1b77dc4aacffdfaf353f0c12c","run":"https://github.com/itsHaddad/bragproof/actions/runs/43857070216"},"network":{"names":["registry.npmjs.org"],"conns":[],"callNames":[],"callConns":[]}},"summary":{"outcome":"started","needs":[],"contacts":[],"flags":[{"kind":"very-long","tool":"agentistics_task_subtask","quote":"3000 characters or more"},{"kind":"very-long","tool":"agentistics_task_session","quote":"3000 characters or more"},{"kind":"very-long","tool":"agentistics_task_phase","quote":"3000 characters or more"}],"tools":58,"wrote":0,"callsTried":4,"callsAnswered":0,"callContacts":[]},"history":[{"package":"@agentistics/mcp","version":"2.217.1","ranAt":"2026-11-16T11:25:48.843Z","outcome":"started","tools":58,"contacts":0,"flags":3,"serious":0,"toolsHash":"ce18ecd1b9413ab8170e9e1dbff6ee1bc96a58452a32f30cddb8da1304837d0c"},{"package":"@agentistics/mcp","version":"2.210.0","ranAt":"2026-11-12T17:46:23.047Z","outcome":"started","tools":56,"contacts":0,"flags":3,"serious":0,"toolsHash":"de56e0c2b83311ef049fb1f2f5df9180a0779988cf2750ed74adcf6519971e59"}],"page":"https://bragproof.smaverk.com/r/%40agentistics/mcp"}