mcp

@coretexa/mcp — reference MCP server

Gives any MCP-speaking agent (Claude Code, Cursor, Windsurf, custom runtimes) an accountability queue: work it can't close itself, verification by a different actor, corrections and regressions filed with lineage.

Setup

cd mcp && npm install

Apply ../schema.sql to any Postgres (Supabase free tier works — paste it into the SQL editor).

Wiring it up (Claude Code example)

Two entries, two identities — that's how no-self-close is enforced:

{
  "mcpServers": {
    "coretexa-doer": {
      "command": "node",
      "args": ["/path/to/coretexa/mcp/src/index.js"],
      "env": {
        "DATABASE_URL": "postgres://...",
        "CORETEXA_ACTOR": "doer-claude-code"
      }
    },
    "coretexa-verifier": {
      "command": "node",
      "args": ["/path/to/coretexa/mcp/src/index.js"],
      "env": {
        "DATABASE_URL": "postgres://...",
        "CORETEXA_ACTOR": "verifier-claude-chat"
      }
    }
  }
}

Typical split: your working session uses the doer server; a separate session/agent (or your chat Claude) uses the verifier server. If the same actor that claimed a ticket tries to verify_ticket it, the server refuses with NO-SELF-CLOSE.

Tools

Tool	What it does
next_ticket	Next pending row (priority asc, FIFO; PARKED skipped)
list_tickets	List by status, or all non-closed
cut_ticket	New ticket; parent_id+origin='spawned' for Rule A finds
claim_ticket	Claim as this actor
mark_done	Done ≠ closed; requires a real result writeup
next_done_ticket	Oldest row awaiting verification
verify_ticket	Pass → close + stamp. Fail → linked correction. Refuses self-close.
block_ticket / unblock_ticket	Human-decision states, question/decision recorded
cut_regression	SYSTEM files "passed before, fails now" with escalated priority
submit_evidence	Attach executable proof per check ({check, result, command, detail})
gate_report	Render the PR-ready markdown receipt for any ticket
ticket_lineage	Tree walk: spawn depth, open descendants, settled?
queue_stats	Counts by status and origin

Evidence enforcement: mark_done is refused on a ticket that declares required_checks but has no evidence; verify_ticket(pass=true) is refused while any required check lacks passing evidence — regardless of actor.

Test

No database needed — the smoke test runs the full lifecycle (including the refused self-close and the auto-cut correction) against an in-memory Postgres:

npm test