io.github.sionjang101-create/taskattest
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
| root | No | Absolute path the MCP server is allowed to read. Any path resolving outside this directory is refused, including paths reached through the spec's own references. Recommended; without it, the server can read whatever the running process can. |
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
Verify that a task is actually finished by checking the files that were produced. Write the definition of done into a spec file first, or call draft_spec to have one proposed from the files and then review it, then call verify_task. Judging is deterministic: no model is consulted, so the same inputs always give the same verdict and the same evidence digest. Call describe_checks to learn the spec format before writing one.
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| verify_taskA | Judge whether a task is finished by checking the files it produced against a spec file that states the definition of done. Returns pass, partial or fail, plus a replayable evidence bundle. Deterministic: no model is consulted. |
| replay_evidenceA | Re-run the verification recorded in an evidence manifest and report whether the digest still matches. Use this to test whether a claimed verdict holds up, rather than taking the manifest's word for it. |
| draft_specA | Read the input, output and trace files and propose a spec: structure checks for outputs, seals on the inputs, checks relating output ids to input ids, and for traces the ordering and counts it can infer. Also returns a profile of every field so you can write your own conditions without opening the files. Deterministic, no model calls. It never fits a threshold to an output; the spec is marked draft and every verdict against it carries a warning until a person removes the flag. |
| describe_checksA | List the check types this server can run, what each one verifies, and two spec templates: a record reconciliation and a JSON-plus-report deliverable. Call this before writing a task definition. |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
No prompts | |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
No resources | |
TDQS
Scored across 4 tools
Each tool has a distinct role: describe_checks is reference/help, draft_spec authors a spec, verify_task runs the judgement, and replay_evidence audits an existing manifest. The only mild overlap is verify_task vs replay_evidence, since both 'verify', but the descriptions clearly separate live verification from re-checking a recorded digest. describe_checks and draft_spec both touch spec content, but one documents check types and the other generates a draft, which the descriptions make explicit.
All four names follow a clean verb_noun snake_case pattern: describe_checks, verify_task, replay_evidence, draft_spec. No mixed conventions or vague stand-alone verbs.
Four tools is squarely within the well-scoped range and each maps to a necessary step in the workflow: learn the checks, draft a spec, verify, and replay. Nothing appears redundant or padding.
The discover-draft-verify-replay loop covers the core lifecycle of deterministic spec-based task verification. Minor gaps remain: there is no tool to finalize or edit a spec (removing the draft flag is left to a human out-of-band) and no way to list or discover existing evidence manifests, so callers must already know paths.