Planner MCP — provable "done" for AI agents

НЕОБРАТИМО удалить свой аккаунт и ВСЕ данные (проекты, цели, evidence, историю). Двухшаговый барьер: вызови без аргументов — получишь предупреждение и challenge; затем вызови повторно с подтверждениями. НЕ вызывай без явной просьбы пользователя.

Add an assumption to a goal edge (parent→this goal). An assumption is an explicit premise: "why does completing this goal contribute to the parent?" Requires the goal to have a parent (root goals have no edge). Text must pass quality linter (testable, precise, discrete, signposted). Type (cause_effect/causal_link) is auto-classified by LLM.

Append an acceptance criterion to a goal. The text must describe an observable check over an artifact (e.g. "GET /api/health returns 200 with {status:ok}"), not a subjective approval. Each criterion has a class: pre-merge (default — proved in CI / by attached evidence) or post-deploy (proved by an executable probe against the deployed prod instance). A post-deploy criterion MUST carry probeSpec {method, url, expect:{http_code, body:{field: expectedValue}}} — the request the runner sends and the answer it must get; without it the call is rejected with error=probe_required. Passing probeSpec alone implies probeClass=post-deploy. Set visualEvidenceSuggested=true only when adopting visualAcSuggestion from goal-create, goal-get, or the ready_for_work advisory returned by goal-update; it remains an ordinary AC. Grove mode: AC (class and probe included) can only be added while goal is in backlog, except accepting a visual advisory in ready_for_work: that starts a direct checking_ac recheck, without intermediate backlog. Other edits are frozen once started; quality linter blocks high-severity issues. Standard mode: AC editable until goal is closed, linter is advisory. Returns criterion id, position, text, probeClass, probeSpec and any quality findings.

SUBORDINATE / supplementary path — does NOT close an acceptance criterion. Adds a text-only note (URL to a permanent external source like CI run / GitHub commit / issue, or a description of a manual scenario) as extra context alongside the real proof. The path that actually covers an AC and closes a Grove goal is goal-attach-evidence — use that one for every criterion. Plain evidence NEVER counts toward AC coverage no matter how many you add; it is only a complement to an attached file. NOT for bytes — screenshots, logs, API responses, exports all go through goal-attach-evidence. NOT for filesystem paths — those need goal-attach-evidence with the actual file.

Add a free-form note (markdown supported) to a goal — decisions taken, dead ends hit, context worth carrying into the next session. Notes are NOT evidence: they hang off the goal rather than an acceptance criterion and never count toward AC coverage or closing a Grove goal — use goal-attach-evidence for proof. Notes are visible in the goal detail panel and returned by goal-get under notes[].

Attach text evidence to an assumption and judge it. The judge evaluates whether the evidence supports or challenges the assumption and automatically transitions its status: supports → supported, first challenges → challenged, second (repeat) challenges → invalidated. Invalidated is terminal — further evidence no longer moves the status. Judge infra failures fail open (errored verdict, no status change).

PRIMARY path to close a Grove goal: this is the ONLY tool that covers an acceptance criterion. Attach binary evidence (screenshot, log dump, API response, export) to an AC — call it once per criterion to satisfy the close gate. The subordinate goal-add-evidence-text only adds context for proofs with NO bytes (URLs to permanent external sources, manual repro descriptions) and does NOT cover an AC. Caption is optional but strongly recommended: state what the file captures and the reproduction conditions (URL/commit/session/inputs) so a third reviewer can reproduce. ⚠ PICK THE RIGHT TRANSPORT BEFORE YOU CALL THIS TOOL ⚠ • BEST for ANY file > ~1 KB raw — and the ONLY no-token path, so use it in a claude.ai / hosted-agent session that has no raw X-Auth-Token → call the sibling MCP tool `goal-request-upload` with this same criterionId. It returns a one-time {uploadUrl, expiresAt}; then stream the raw bytes with a single PUT: `curl -sS --fail --upload-file "/abs/path/to/file.png" "<uploadUrl>"` (optionally add -H "X-Content-Sha256: <hex sha256>" so corruption fails fast). No base64, no token — the signed ?t= ticket in the URL is the only credential, single-use, criterion-scoped. The PUT response is the same evidence JSON this tool returns. • ALTERNATIVELY, if you DO have the raw X-Auth-Token in your shell → the `planner-attach.sh` helper (zero-install bash, binary-safe). The MCP base64 path below is unreliable for non-trivial files: long string arguments get truncated or whitespace-corrupted on the agent side BEFORE the JSON-RPC request is sent. Measured 2026-05-20 on prod: a 4 KB PNG arrived at the server as 1874 decoded bytes (file_hash_mismatch); a 2 KB payload arrived with stray whitespace (failed base64_decode). The server itself accepts up to 25 MiB raw — the bottleneck is the agent-side serialisation of contentBase64, NOT the server. planner-attach.sh COPY-PASTE RECIPE (replace 3 placeholders, run in your shell): curl -sS https://planner.monopoly-gold.com/api/c…

Attach a small file (mockup, diagram, screenshot, PDF) to a GOAL as a description attachment — NOT evidence: it is not bound to any acceptance criterion, is never judged and never counts toward closing the goal. Returns {attachment.url, attachment.markdown}; paste `markdown` into the goal description (goal-update description) and the web UI renders images inline. contentBase64 is reliable only for payloads ≤ ~1 KB (long string arguments get truncated on the agent side before the request is sent). For anything bigger call goal-request-upload with goalId and stream the file with a single PUT — same result shape, no base64.

Add a blocker to a goal (blockers are additive — each call appends a new one, existing blockers are preserved). Sets status to blocked. Optionally create an inline resolver goal (resolverTitle) or link an existing one (linkedGoalId) — mutually exclusive. When the last active blocker is removed via goal-remove-blocker, the goal returns to its previous status.

Create a goal/task/milestone/habit in the planning tree. Two modes: grove (default) — enforces acceptance criteria gate on status transitions (AC required before ready_for_work, file-evidence per AC before done); standard — carries the same AC but without the evidence gate (advisory linting only). Returns the created goal with id, webUrl, mode, and nextStep hint. For an explicit visible UI result, the response may include an advisory visualAcSuggestion; adopt its criterion through goal-add-criterion with visualEvidenceSuggested=true, or ignore it. Every goal must belong to a project (pass projectId or inherit from parent).

IRREVERSIBLY delete a goal and all its descendants (children, evidence, blockers). Prefer goal-update status=cancelled to preserve history, or goal-block to mark an obstacle. Use delete only for erroneous/duplicate entries. Returns the deleted title and children_deleted count.

Dismiss the red-team finding that blocks a goal from starting, when you disagree with the verdict. Requires a reason of at least 80 characters explaining why the counterexample does not apply. The dismissal is recorded in goal history and surfaced in goal-get, so "agent overrode the gate" stays distinguishable from "no hole was found". Prefer fixing the acceptance criteria via goal-add-criterion — dismissing leaves the hole open.

Fetch full details of a single goal: title, description, status, priority, type, mode, children, acceptance criteria (with evidence coverage), blockers, tags, deadline, estimate, history log, and project. The primary drill-down tool after goal-list or goal-tree. For a Planner goal URL, do not browse the UI: extract the UUID after /goal/ and pass it as required goalId.

List goals with optional filters: projectId (UUID), status (backlog/checking_ac/checking_ac_failed/checking_ac_errored/ready_for_work/in_progress/blocked/done/cancelled), type (goal/milestone/task/habit), parentId (UUID — direct children only). Returns up to limit results (default 50, no offset pagination). Sorted by priority ASC then createdAt DESC. Each entry has id, title, webUrl, status, priority, type, progress, parent_id, project_id.

Move a goal to a different parent or project. Pass newParentId=UUID to reparent (inherits project from new parent), newParentId=null to make root. Optional projectId=UUID overrides the target project when making root (cross-project move to root). Cascades project change to all descendants. Prevents cycles.

Read attached evidence through MCP without changing its format. Pass evidenceId from goal-get. Text (text/plain, JSON, NDJSON) returns newline-preserving redacted lines; offset/limit are zero-based line pagination. Images return a native MCP image with original MIME and bytes. For a complete video, PDF, ZIP or other binary, binaryDelivery=link (default) returns a resource_link to the authenticated original download — no base64 and no conversion. If the client cannot open external links, binaryDelivery=base64 embeds the exact original bytes in a resource blob; this expands data by about 33% plus JSON overhead and can consume the client context window, so use it only when needed. byteOffset/byteLimit require binaryDelivery=base64. Text secrets are redacted; binary evidence is unmodified.

Pull-инбокс для «подхватить и довести» в неосновное окно. По одному проекту, за окно N дней, статусы для подхвата (очередь исполнителя — ready_for_work и checking_ac_errored: backlog-цели брать нельзя). Возвращает обогащённые записи (title/description-preview/parent/counts), чтобы выбрать без goal-get. Параметры: project (slug, required), withinDays (int, default 14), statuses (list, default ["backlog"]; допустимо {backlog,checking_ac,checking_ac_failed,checking_ac_errored,ready_for_work,in_progress,blocked}), limit (int, default 10).

Record one completed LeanSolver iteration against the current tracked contract. Result artifact refs must resolve to existing evidence or goal_attachment rows owned by this goal. Rework iterations accept only registered finding IDs and require a separate judged evidence row as repeatCheckEvidenceId; Planner derives the repeat verdict server-side. Only certificates admitted here may affect the numeric implementation rework rate.

Register an immutable LeanSolver finding against the currently tracked goal contract. subjectRefs must be goal-scoped refs of kind goal, criterion, or assumption. Returns the server-generated finding id to pass later as addressesFindingId; rejects stale contract tracking.

Remove an assumption from a goal edge.

Remove a blocker from a goal by blocker UUID. If it was the last active blocker, the goal automatically returns to its previous status (in_progress or backlog). Does not delete the linked resolver goal if one exists.

Remove an acceptance criterion from a goal. Grove mode: only while goal is in backlog (frozen once started). Standard mode: until goal is done. Cascades to all evidence on the criterion. Returns confirmation with removed criterion details.

Delete an evidence record by UUID. Forbidden if the owning goal is already done (evidence is frozen after close). Removes both the database record and the attached file (if any).

Remove a note from a goal by note UUID. Note ids come from goal-get (notes[].id). Deletes only the note — never touches acceptance criteria, evidence or blockers.

Set display order of sibling goals within the same parent and priority band. Pass an array of goal UUIDs in the desired order — each is assigned position = its array index. All goals must share the same parent and priority.

Request a change to an acceptance criterion. grounds=unreachable (default): the AC cannot be proven — requires ≥3 failed evidence attempts (weak/mismatch) with 0 matches. grounds=wrong_criterion: the AC measures the wrong thing (owner never approved this wording, or a revealed fact shows it tracks something else) — no failed attempts required, but you MUST pass quotedFragment: a verbatim slice of the criterion text you claim is wrong (checked against the stored text), and the reason must name the unapproved wording or the finding; difficulty alone is rejected. An LLM judge screens the reason either way. On pass, creates an escalation for the goal owner to resolve (edit AC text, split to sub-goal, drop AC without creating a child, or reject). Grove mode only, goal must be in_progress.

PREFERRED path to attach a LARGE binary evidence file (screenshot, log dump, PDF, session transcript — anything > ~1 KB) to an acceptance criterion. Returns a one-time {uploadUrl, expiresAt} scoped to this criterion. Then STREAM the raw file to it with a single PUT — no base64, no token: curl -sS --fail --upload-file "/abs/path/to/file.png" "<uploadUrl>" Optionally pass the hex SHA-256 of the file so the server fast-fails on any in-flight corruption: curl -sS --fail -H "X-Content-Sha256: <sha256>" --upload-file "/abs/path/to/file.png" "<uploadUrl>" The PUT response is the same evidence JSON that goal-attach-evidence returns (evidence id, serverSha256, judge verdict, criterion evidenceCount). A non-2xx PUT means the upload was rejected (expired/already-used/wrong-criterion/hash-mismatch) and NO evidence was created — request a fresh URL and retry. Use this instead of goal-attach-evidence for any non-trivial file. Use goal-add-evidence-text only for byte-less context (external URLs, manual repro notes) — it does NOT cover an AC. GOAL ATTACHMENTS (description illustrations, NOT evidence): pass goalId INSTEAD of criterionId. The PUT then creates a goal attachment and returns {attachment.url, attachment.markdown} — paste `markdown` into the goal description and the web UI renders images inline. Such a file is not bound to any AC and never counts toward closing the goal.

Resolve a pending AC escalation. Owner decides: edit (provide new AC text), split (move AC to a child goal), drop (remove the AC outright, optionally with a linkedGoalId audit reference), or reject (agent must find another way). All prior evidence on the AC is deleted for edit/split/drop so the agent must submit fresh proof.

Explicitly start LeanSolver iteration tracking for the goal contract currently stored by Planner. Planner computes the contract fingerprint server-side from title, description, acceptance criteria, and assumptions. Calling again without a contract change is idempotent; after a contract change it starts a new version while retaining old certificates as superseded history.

LLM generates suggested assumptions for a goal edge (parent→child). Returns 2-4 assumptions with signposts and type classification. Author should review, edit, and accept relevant ones via goal-add-assumption.

Aggregate statistics across all goals (or scoped to a projectId): total count, breakdown by status, root count, blocked count, and overdue count. Useful for daily standups and dashboard views.

Quick-capture a task or improvement idea into a project. Creates a backlog task with priority 4 (low) and auto-adds the "suggestion" tag. Resolves project by slug (falls back to the default project). Lightweight alternative to goal-create when you need minimal ceremony.

Fetch the full goal hierarchy as a nested tree. Optional filters: projectId (UUID) to scope to one project, rootId (UUID) to get a subtree. Each node includes id, title, status, priority, type, progress, and nested children array.

Partial update of a goal — only the fields you pass are changed; omitted fields are untouched. Updatable: title, status, priority (1-5), description, type, tags, deadline, estimate, mode, attemptCommit (git SHA of the attempt commit; once a deployed revision contains it, post-deploy probes of this goal run automatically and record machine verdicts — pass "" to clear), dismissVisualAcSuggestion. Status transitions follow a matrix: backlog→ready_for_work (author; grove: requires ≥1 AC, red-team check when the project requires it — the goal parks in checking_ac and becomes ready_for_work, checking_ac_failed (hole in AC set) or checking_ac_errored (check did not run)); ready_for_work|checking_ac_errored→in_progress (executor, no checks); in_progress→done (file-evidence on every AC + session_history) | blocked | cancelled. A visible-UI goal with an outstanding visualAcSuggestion returns an advisory before ready_for_work freezes the AC contract: accept it through goal-add-criterion(visualEvidenceSuggested=true), or retry this same update with dismissVisualAcSuggestion=true. The first MCP dismissal persists the choice and notifies Product Events with the goal title and link, full proposed AC, and why screenshot proof was useful; UI, repeat, and no-active dismissals are silent, and notification failures do not undo the dismissal. backlog→in_progress and backlog→done are rejected. Returns the updated goal with all fields.

Update text or signpost of an assumption. Status cannot be changed manually — transitions happen only via evidence judge verdicts.

Update an acceptance criterion: its text and/or its class + probe. probeClass=post-deploy requires an executable probeSpec {method, url, expect:{http_code, body}} (kept from the criterion when omitted and one already exists; otherwise error=probe_required); probeSpec alone implies post-deploy; probeClass=pre-merge drops the probe. Grove mode: only while goal is in backlog (frozen once started), quality linter blocks high-severity issues. Standard mode: until goal is done, linter is advisory. Returns updated criterion details (text, probeClass, probeSpec) and any quality findings.

Declare that one project depends on another (depends-on relationship). Accepts UUID or slug for both sides. Idempotent: if the dependency already exists, returns it with already_existed=true. Self-references are rejected. No cycle detection — the caller is responsible for avoiding circular chains.

Create a new project container for goals. Requires title and a unique slug (lowercase a-z, 0-9, hyphens). Optional: description, status (active/archived/paused, default active), tags, icon (emoji), repositoryPath, repositoryUrl, sourceRef ({type:"git",url,ref} — remote-research source), isDefault. Returns the created project with id, slug, and webUrl. Slug must be unique — duplicates are rejected with existing slugs list.

Delete a project. If the project has goals, pass force=true to cascade-delete them all; without force, the call is rejected with the goals count. Returns deleted title, slug, and goals_deleted count.

Fetch full project details: title, slug, status, description, icon, tags, repository info, sourceRef ({type,url,ref} for remote-research or null), goals breakdown by status (backlog/in_progress/blocked/done/cancelled counts), and depends_on list of project dependencies.

List all projects with optional status filter (active/archived/paused). Returns id, title, slug, webUrl, icon, and goals_count for each project. Use project-get for full details with goals breakdown by status.

Remove a project dependency. Two lookup modes: pass dependency_id (UUID of the link itself), or pass both projectId and dependsOnProjectId (UUID or slug). Returns removed=true on success, removed=false if the dependency was not found.

Partial update of a project — only the fields you pass are changed; omitted fields are untouched. Updatable: title, slug, description, status (active/archived/paused), tags, icon (emoji), repositoryPath, repositoryUrl, sourceRef ({type:"git",url,ref} — remote-research source), clearSourceRef (true → сбрасывает Project.sourceRef в null; несовместим с sourceRef), isDefault, redTeamMode (optional/advisory/required), evidenceJudgeMode (off/optional/required), allowedGoalMode (any/grove_only), fmeaGenerationMode (off/on), escalationResolveTimeout (off/1h/4h/24h), escalationAutoResolveBy (none/planner-agent), escalationAutoResolveStrategy (suggest-edit/reject-default). Returns the updated project with all fields.