MCP server for orchestrating agent-based workflows with project management, knowledge base, plot (protocol) management, and agent lifecycle tracking
Continuum Orchestrator demonstrates solid tool design with 17 well-named, verb-prefixed tools (registry_list, agent_get, knowledge_search, plot_update, etc.). All tools have descriptions (avg ~180 chars, within 10-1024 baseline). Input schemas are present with typed parameters and enums (status filters, kind filters). However, output schemas are NOT documented, descriptions state what fields are returned but JSON Schema output is absent. Error handling is basic (toolError/toolSuccess wrappers return text, not structured recovery guidance). Tool composition is strong: registry_list→agent_get, knowledge_list→knowledge_get, and knowledge_search all chain properly. Parameter descriptions are adequate but could be more prescriptive about constraints (e.g., slug regex /^[a-z][a-z0-9-]*$/ is mentioned in knowledge_create but not enforced in schema). No tool annotations (readOnlyHint, destructiveHint) despite clear risk levels (READ_ONLY, WRITE, DESTRUCTIVE marked in metadata).
Create a new agent in draft state under `project`. Returns full record.
Get full record for an agent by `project` + `slug`. Returns every field including `request`, `plan`, `implPrompt`, `coordinationBrief`, `postMergeNotes`. Use this after `registry_list` to drill into any agent whose body you need.
Update an agent in `project`. Allowed transitions: draft→active, active→merged|abandoned. merged requires mergedCommit (len≥7). abandoned requires abandonedReason. plan/implPrompt/coordinationBrief patchable while status ∈ {draft, active}; frozen on merged/abandoned.
Create a new knowledge lesson under `project` attributed to `agentSlug`. `slug` matches /^[a-z][a-z0-9-]*$/ and is unique within the project. `content` is the full lesson body (markdown OK). Use this tool when recording a reusable lesson at post-merge file-back. Optional `kind` is `"fundamental"` (binding rule applied to every dispatch) or `"situational"` (context-specific). Defaults to `"situational"`.
Hard-delete a knowledge lesson by `project` + `slug`. Returns `{deleted: true}` on success. Use this tool to retire a lesson that no longer applies.
Output schemas not documented. Descriptions state return fields (e.g., 'Returns {data, page, limit, hasNextPage}') but no JSON Schema output definition exists. LLMs cannot plan downstream tool calls without knowing response structure.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels marked in metadata (READ_ONLY, WRITE, DESTRUCTIVE). Agents cannot distinguish safe reads from destructive writes without explicit hints.
Error handling returns plain text (toolError/toolSuccess wrappers). No recovery guidance, error classification (retryable vs fatal), or actionable next steps. LLMs receive 'User not found' with no hint to call search_users().
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
Get one knowledge lesson by `project` + `slug`. Returns the full row including `content`. Use this tool when you already know which lesson to read.
List knowledge lesson **metadata** for `project` (newest first). Returns array of `{slug, kind, agentSlug, createdAt, updatedAt}` — no `content`. `project` is the canonical absolute path (your `pwd`). Optional `kind` filters to `"fundamental"` or `"situational"`. Use this tool to enumerate the inventory (e.g., load all fundamentals at Phase-1 Intake by passing `kind: "fundamental"`), then call `knowledge_get({project, slug})` for each lesson whose body you need to read.
Semantic vector search (cosine-equivalent ranking on L2-normalized embeddings) across knowledge lessons in `project`. Returns ranked **metadata only** — array of `{slug, kind, agentSlug, createdAt, updatedAt}`, no `content`. Call `knowledge_get({project, slug})` per slug whose body you want to read. Embedder is required and runs in-process by default; semantic only. At least one of `q` or `kind` is required. `q` is a free-text semantic query (no SQL wildcards). `kind: "fundamental"` filters to binding lessons that must be followed on every dispatch; `kind: "situational"` filters to context-specific lessons. Optional `limit` (default 10, max 50). Use this tool at Phase-1 Intake with `q` for topical relevance; for the full fundamental set use `knowledge_list({project, kind: "fundamental"})` instead.
Whole-content replace of a knowledge lesson by `project` + `slug`. Pass new `content` (full body) and/or new `agentSlug`. Slug rename is not supported — delete+create instead. Empty patch throws `no_change`. Optional `kind` patches the lesson kind ("fundamental" or "situational").
Returns the orchestrator protocol text (PLOT.md) for the given project. Call this first when entering orchestrator mode. `project` MUST be the canonical absolute path of the working tree (your `pwd`).
Apply a unified-diff (git-format) to the project PLOT.md. Diff MUST include `--- a/PLOT.md` and `+++ b/PLOT.md` headers. Hunks must match current content with zero fuzz — get-edit-diff loop on mismatch.
Create a new project keyed by canonical absolute path. `name` defaults to the basename of `path`. Initialises default PLOT. Knowledge starts empty (no rows).
Delete a project and CASCADE all of its plot/knowledge/agents. Returns {deleted: true, cascadedAgents: N}.
Get a project by canonical absolute path. Use your `pwd` as `path`. Returns null if not found (no error).
List all projects (orchestrator project records). Returns array of {id, path, name, createdAt, updatedAt, deletedAt}.
Rename an existing project. `path` identifies the project; `name` is the new display name.
List agent dispatch metadata for `project` (newest first), paginated. Returns `{data, page, limit, hasNextPage}` where each `data` entry carries `slug, status, branch, worktree, reservedPaths, createdAt, dispatchedAt, updatedAt, mergedAt, mergedCommit, abandonedReason` — no `request`, `plan`, `implPrompt`, `coordinationBrief`, `postMergeNotes`. Call `agent_get({project, slug})` for the full record (artifacts + notes). Optional `status` filters to `draft|active|merged|abandoned` (use `status: "active"` for the collision matrix). Optional `page` (default 1) and `limit` (default 10, max 50). When `hasNextPage` is true, increment `page` and call again.
Parameter constraints (e.g., slug regex /^[a-z][a-z0-9-]*$/, mergedCommit len≥7) documented in descriptions but not enforced in JSON Schema. LLMs cannot validate inputs before calling; invalid calls fail at runtime.
plot_update requires unified-diff input (git-format with headers/hunks). No schema validation, no example, no error guidance for malformed diffs. LLMs will struggle to construct valid diffs without trial-and-error.