phionyx-pipeline-mcp

A self-governance MCP server for Claude Code — gates the agent's own "I fixed this / I tested that / this code path changed" declarations against git diff truth and a deterministic gate.

phionyx-pipeline-mcp solves a less-discussed agent-trust problem: when an AI coding agent reports back on its own work, that report is itself a trust object. Without verification, you accept it on faith.

This package gives any MCP-capable host (Claude Desktop, Cursor, Zed, VS Code, JetBrains) a six-tool surface that turns those self-reports into reviewable evidence — coverage of paths claimed vs. paths actually touched, severity-weighted evidence taxonomy, drift tracking across a session, and an explicit pass | regenerate | reject directive before the agent claims "done."

Where this sits in the Phionyx stack

Phionyx ships three distinct things, each with its own version line. This package is the gate:

• Engine — phionyx-core (latest v0.9.0): the deterministic runtime (46-block canonical pipeline, state vector, kill switch, HITL, ethics/safety gates, signed audit chain).

• Gate — phionyx-pipeline-mcp (this package): the inward-facing self-claim gate described here. It verifies what the agent says about its own work against git diff truth and a deterministic gate. v0.3.0 adds opt-in, default-off evidence binding (the require_tool directive + continuity binding) on top of the claim-grounding gate. It is the claim-governance layer of the Phionyx governance stack; phionyx-mcp-server (v0.1.0) is the outward MCP trust boundary.

• Format — the AI Runtime Evidence Protocol (AIREP): an experimental, vendor-neutral open format for an AI decision receipt — one signed, hash-chained, offline-checkable record per AI runtime decision, readable by anyone and tied to no vendor. Phionyx's Reasoned Governance Envelope (RGE) is AIREP's reference producer (the first system that emits AIREP records). AIREP is a proposed format, not a ratified standard. See ai-runtime-evidence-protocol.

These three version lines — engine v0.9.0, gate v0.3.0, and AIREP v0.1 (experimental) — are never cross-attributed.

Related MCP server: hAIve MCP

How it works — three-stage verification

Stage 1: LLM declaration   →   Stage 2: Repo truth          →   Stage 3: Deterministic gate
  "I fixed X, tested Y,         git diff parsed; functions       evidence + recency + revision
   affected paths a,b,c"        extracted; cross-checked         thresholds → directive

The gate is deterministic. Stage 1 (the agent's input) is stochastic. Stage 2 — verifying the agent's path declarations against the actual git diff — is what narrows the gap.

(These three stages are internal to the gate's verification flow. They are unrelated to the broader Phionyx governance-stack layers.)

Where this fits on phionyx.ai

This package surfaces under phionyx.ai/bounded-authority — the safety-first AI providers entry. It is the inward-facing self-claim gate that verifies the agent's own "I fixed / I tested / this code path changed" narration against the repository's actual diff and a deterministic gate, producing a pass | regenerate | reject directive paired with reviewer-runnable evidence.

Companion package: phionyx-mcp-server

This package is the inward-facing layer: it gates what the agent says about its own work.

A companion package, phionyx-mcp-server (v0.1.0), is the outward-facing layer: it sees the host calling a third-party MCP server and signs evidence of that call (descriptor hash, drift detection, audit chain).

When both packages are installed and registered with the same Claude Code host, they agree on a single trace_id per session via PHIONYX_TRACE_ID (with ~/.phionyx/active_trace file fallback). One conversation = one trace = end-to-end view of every third-party tool call AND every agent self-claim gate decision.

phionyx_session_report (this package) surfaces the server-MCP envelope chain head + validity inline, so a reviewer can see both layers in one JSON.

Install

# This package only:
pip install phionyx-pipeline-mcp

# With the outward-facing companion (recommended for full session evidence):
pip install "phionyx-pipeline-mcp[mcp-server-integration]"

Register with Claude Code

Add to .claude/mcp.json in your project:

{
  "mcpServers": {
    "phionyx-pipeline": {
      "command": "phionyx-pipeline-mcp"
    },
    "phionyx-mcp-server": {
      "command": "phionyx-mcp-server"
    }
  }
}

Both servers will share ~/.phionyx/active_trace by default.

Tool surface

Reviewer-runnable CLI (phionyx)

pip install phionyx-pipeline-mcp also installs a phionyx command that exposes the same three-stage verification as the MCP server, but reachable from a plain terminal — no Claude Code / MCP host required.

Four subcommands:

# 1. Verify a self-claim against the gate
phionyx verify-claim \
    --claim "scenario continuation bug is fixed" \
    --evidence "12 scenes played, quest_complete=True" \
    --type integration_test \
    --tested "src/foo/regression.py::regression_handler" \
    --affected "src/foo/regression.py"

# 2. Aggregate the audit chain over the last N days
phionyx audit --days 30
phionyx audit --days 7 --json   # machine-readable

# 3. Replay a specific session's timeline by trace_id (or unique prefix)
phionyx replay --trace <trace_id_or_session_id_prefix>

# 4. The killer demo — reviewer reproduces the failure mode in 5 seconds
phionyx demo broken-test-disabled

The demo is the entry point a reviewer should try first. It runs a scripted scenario in which Claude claims a fix passes and the gate detects that no tested-path coverage exists and the evidence weight is the bottom of the taxonomy — and rejects the claim. The exit code is 2 because the gate did its job.

Exit codes:

CLI and MCP-host invocations write to the same telemetry directory (PHIONYX_PROJECT_ROOT/data/mcp_telemetry/), so phionyx audit and phionyx replay see prior sessions started inside Claude Code or any other MCP-capable host.

Evidence taxonomy

Not all evidence is equal. The gate weights confidence by the type of test that backs a claim:

A claim_fixed action with evidence_type=none will be rejected outright; even high confidence values cannot compensate for missing test evidence.

Pre-commit gate helper

A standalone phionyx-pipeline-check-gate CLI reads the recent telemetry and exits non-zero if the most recent gate directive was reject or regenerate. Drop it into your pre-commit hook to enforce the discipline:

phionyx-pipeline-check-gate --mode pre-commit

Exit codes: 0 pass · 1 failed (recent reject/regenerate) · 2 no telemetry (warn-only).

Shared-trace contract

When phionyx-mcp-server is also installed, this package extends phionyx_session_report's output with:

{
  "trace_id": "<active trace>",
  "mcp_envelope_chain": {
    "trace_id": "<same trace>",
    "count": 7,
    "head_hash": "sha256:...",
    "valid": true,
    "broken_at": null
  }
}

The integration is read-only — this package imports FilesystemEnvelopeStore + verify_chain from phionyx-mcp-server and reads its chain. No cross-package write coupling. When the server package isn't installed, the field reports "phionyx-mcp-server not installed" and the rest of the report continues to work.

Tests

pip install -e ".[test]"
pytest tests/ -q

License

AGPL-3.0-or-later. See LICENSE.

See also

• phionyx.ai/bounded-authority — entry pillar this package surfaces under

• phionyx.ai/evidence — Evidence Matrix: every load-bearing claim paired with a reviewer-runnable command

• ai-runtime-evidence-protocol — AIREP, the experimental vendor-neutral open format for per-decision AI decision receipts (signed, hash-chained, offline-checkable). Phionyx's Reasoned Governance Envelope is its reference producer. A proposed format, not a ratified standard.

• Project hub: github.com/halvrenofviryel/phionyx-research

• Engine (PyPI): phionyx-core — the deterministic runtime (v0.9.0)

• Outward MCP layer: phionyx-mcp-server (v0.1.0) — trust boundary over third-party MCP tool calls

• Inspect AI bridge: phionyx-eval-inspect (v0.1.0) — envelope chain → .eval log

• Framework adapters (alpha): phionyx-langchain-langgraph (v0.1.0a1) · phionyx-openai-agents (v0.1.0a1)