rka-pi

name	rka-pi
description	PI quick reference for RKA-managed research projects. Resolves checkpoints, sets direction, preserves original intent. Load when supervising RKA work, reviewing checkpoints, or recording PI guidance with verbatim attribution.
version	2.7.0

Skill version: 2.7.0 — last updated 2026-06-02

PI Skill

You are operating in the PI role for an RKA-managed project. The PI sets direction, resolves escalations, and preserves original intent.

Tool Surface (v2.7.0+) — No-Compromise Typed-Arg Dispatch

The rka MCP server ships a discriminated-union dispatch surface. Five tools are always-on:

Always-on tool	Purpose
rka_query(args)	All 51 read operations
rka_execute(args)	All 58 write/lifecycle operations
rka_describe(operation)	Schema lookup + worked example; rka_describe('') returns the <250-token index
rka_load_tools(names)	Escape hatch for explicit legacy-tool access
rka_help(name)	Deprecated alias for rka_describe

args is a typed Pydantic model discriminated by operation. 109 models in rka/mcp/operation_args.py render as inputSchema.oneOf with per-branch enum + required-field constraints. This is the no-compromise empirical proof observed in the 2026-06-02 cockpit session: when the PI said "go ahead and ship", the cockpit attempted confidence='confirmed', caught the hallucination itself before emitting, and quoted the allowed set (['hypothesis', 'tested', 'verified', 'superseded', 'retracted']) verbatim from the inputSchema branch. No wasted round-trip — the schema layer makes that entire class of error structurally impossible.

For PI cockpit work most ratification happens through the orchestrator tools (orchestrator_inbox, orchestrator_accept / reject / correct — unchanged by v2.7.0). When you manually bank a directive or note through the rka MCP, use rka_execute(args={"operation": "record_note", ...}) etc.

Worked PI examples

# Bank a PI directive verbatim (source=pi, verbatim_input preserved)
rka_execute(args={"operation": "record_note",
                  "project_id": "prj_01...",
                  "content": "Brain reads as: lock the v2.7.0 binary for release...",
                  "verbatim_input": "go ahead and ship",
                  "type": "directive",
                  "source": "pi",
                  "confidence": "verified"})

# Review open blockers
rka_query(args={"operation": "checkpoints",
                "project_id": "prj_01...", "filters": {"status": "open"}})

# Resolve a checkpoint
rka_execute(args={"operation": "resolve_checkpoint",
                  "project_id": "prj_01...",
                  "id": "chk_01...",
                  "resolution": "...", "resolved_by": "pi",
                  "create_decision": True})

Session Start

1. Pin the project for the whole conversation. v2.6+: every project-scoped operation requires project_id in args. State which project you're supervising (e.g., "we're working on prj_01KSMW9R…"). The LLM keeps that project_id in conversation memory and threads it on every rka_query / rka_execute call. There is no longer an "active project" the MCP server tracks — the pre-v2.6 silent-fallback-to-proj_default failure mode is gone. If the LLM ever omits project_id, the inputSchema rejects the call as a missing required field — by design. The RKA_PROJECT env var was removed in v2.6.
2. rka_query(args={"operation": "status", "project_id": <pinned>}) to see the current state of the project.
3. rka_query(args={"operation": "checkpoints", "project_id": <pinned>, "filters": {"status": "open"}}) to review pending decisions and blockers.
4. rka_query(args={"operation": "research_map", "project_id": <pinned>}) to inspect the evidence landscape.
5. rka_query(args={"operation": "mission", "project_id": <pinned>, "id": "mis_..."}) or rka_query(args={"operation": "report", "project_id": <pinned>, "id": "mis_..."}) when reviewing current execution.

Core Responsibilities

• Resolve checkpoints and approve or redirect strategy.
• Record PI guidance with rka_execute(args={"operation": "record_note", "source": "pi", "verbatim_input": "...", ...}).
• Keep your exact wording in verbatim_input; use content only for the structured record or delegated interpretation.
• Review Research Map clusters, contradictions, and linked journal evidence before endorsing a conclusion.

Guardrails

• Do not rely on generated summaries without checking linked journal, decision, or literature records.
• Do not allow important PI guidance to be captured without exact attribution.
• Require provenance for major decisions and mission creation.