@touchskyer/opc

OPC — One Person Company

A full team in a single Claude Code skill. You're the CEO — OPC is everyone else.

16 specialist agents (PM, Designer, Security, Devil's Advocate, and more) that build, review, and evaluate your code through a digraph-based pipeline with code-enforced quality gates.

What's Different in v0.8

Compound eval quality gate (D2). 11-layer substance check on every eval — thin content, missing code refs, low uniqueness, fabricated references, aspirational claims, change scope coverage, etc. ≥3 layers tripped → hard FAIL (enforce by default); --no-strict downgrades to shadow mode. thinEval substance exemption: short evals with complete reasoning/fix/refs are exempt. Evaluator guidance: when D2 triggers, evaluatorGuidance output tells the evaluator exactly which layers failed and how to fix.

Iteration escalation (D3). Persistent eval warnings across ≥2 iterations auto-escalate to FAIL. No more infinite loops of shallow reviews.

Task Scope Registry. Loop mode plans require ## Task Scope with SCOPE-N items. The harness validates at init and blocks completion if any scope item is uncovered — preventing the #1 failure mode where LLM decomposition silently drops requirements.

Pipeline E2E lint. Tasks containing pipeline keywords (cron, webhook, CI/CD) must have an e2e-live-trigger acceptance criterion. Proxy evidence (unit tests) ≠ live evidence.

Evaluator prompt hardening (D6). 5 evidence standards baked into the evaluator protocol: cite evidence, address anomalies, no aspirational claims, distinguish root cause vs symptom, cover change scope.

What's Different in v0.7

Third-party extension authoring. docs/extension-authoring.md (7800+ words) + examples/extensions/_starter/ (30-min walkthrough). Hardened via DX litmus: an independent agent built an extension using only the doc + starter.

What's Different in v0.6

Digraph engine. Tasks flow through typed nodes (build → review → gate → ...) with mechanical verdict routing. No more linear pipelines.

Autonomous loop. opc loop decomposes a feature into units, schedules a durable cron, and runs 8-16 hours unattended — with code-enforced guardrails that survive context compaction.

Code-enforced, not honor-system. 29 test suites verify: tamper detection (write nonce), atomic state writes, review independence, oscillation detection, tick limits, scope coverage, compound defense, and JSON crash recovery.

External validator integration. Pre-commit hooks, test suites, Playwright E2E, and CI pipelines are formally part of the quality architecture — the agent is supervised by tools it doesn't control.

How It Works

One principle: the agent that does the work never evaluates it.

Task → Flow Selection → Node Execution → Gate Verdict → Route Next
                              ↑                              ↓
                              └──────── ITERATE/FAIL ────────┘

• Task inference — reads your request, picks a flow template (review, build-verify, full-stack, pre-release), and enters at the right node.

• Typed nodes — each node has a type (discussion, build, review, execute, gate) with specific protocols. Build nodes produce commits. Review nodes dispatch parallel subagents. Gate nodes compute verdicts from code, not LLM judgment.

• Mechanical gates — verdicts are computed by opc-harness synthesize: any red = FAIL, any yellow = ITERATE, all green = PASS. No LLM gets to decide if a finding is "important enough."

• Cycle limits — max 3 loops per edge, 5 re-entries per node, 20-30 total steps depending on flow. Oscillation detection catches A↔B loops.

Quick Start

Install

npm install -g @touchskyer/opc

Skill files are automatically copied to ~/.claude/skills/opc/.

Manual install (no npm)

git clone https://github.com/iamtouchskyer/opc.git
cp -r opc ~/.claude/skills/opc

Use it

# Review — dispatches 2-5 role agents in parallel
/opc review the auth changes

# Build — implements + independent review + gate
/opc implement user authentication with email/password

# Autonomous loop — decomposes, schedules cron, runs unattended
/opc loop build features F1-F4 from PLAN.md

# Interactive mode — asks clarifying questions first
/opc -i redesign the onboarding flow

# Explicit roles
/opc security devil-advocate

# Flow control
/opc skip          # skip current node
/opc pass          # force-pass gate
/opc stop          # terminate, preserve state
/opc goto build    # jump to node

Extensions

OPC has a capability-routed extension surface. Extensions live in ~/.claude/skills/opc-extension/<name>/ — each with ext.json (capability declarations) + hook.mjs exporting any of promptAppend / verdictAppend / executeRun / artifactEmit hooks. No fork, no rebuild. Hooks are sandboxed via per-extension timeouts + circuit breakers, so a broken third-party extension can't take down the harness.

The companion repo opc-extensions ships 4 extensions: design-intelligence (theme injection + design coverage + VLM visual eval), git-changeset-review, memex-recall, and session-logex.

Full authoring guide: docs/extension-authoring.md — zero-OPC-context quickstart + reference, plus a starter template at examples/extensions/_starter/.

Flow Templates

Template	Nodes	When
review	code-review → gate	PR review, audit, "find problems"
build-verify	build → code-review → test-design → test-execute → gate	"implement X", "fix bug Y"
full-stack	discuss → build → review → test → acceptance → audit → e2e → gates	Complex/vague requests
pre-release	acceptance → audit → e2e → gates	"verify before release"

Autonomous Loop (v0.6)

/opc loop build the math tutoring app features F1-F4

What happens:

• Runbook lookup — opc-harness runbook match "<task>" checks --dir flag → OPC_RUNBOOKS_DIR → ~/.opc/runbooks/ for a matching recipe. If one hits, its units / flow / tier become the plan; otherwise fall through to step 2. Disable per-invocation with OPC_DISABLE_RUNBOOKS=1. See docs/runbooks.md and examples/runbooks/add-feature.md.
• Decompose (runbook miss only) — breaks task into atomic units (spec, implement, review, fix, e2e)
• Definition of done — establishes verify/eval criteria per unit before any work starts
• Schedule — durable cron (survives process restart) fires every 10 min
• Execute — each tick runs one unit through the appropriate OPC flow
• Guard — opc-harness enforces: git commit required, ≥2 independent reviewers, no plan tampering, no state forgery, artifact freshness, tick limits
• Terminate — auto-stops when plan complete, tick limit hit, or wall-clock deadline reached

Guardrails (code-enforced, not prompt-level)

Guard	Enforcement
Write nonce	Random SHA256 at init; state written by harness only
Atomic writes	write → rename (POSIX atomic); no truncated JSON on crash
Plan integrity	SHA256 hash at init; verified every tick
Review independence	≥2 eval files, identical content rejected, line overlap warned
Git commit required	HEAD must change for implement/fix units
Screenshot required	UI units must produce .png/.jpg artifact
Tick limits	maxTotalTicks (units×3) + 24h wall-clock deadline
Oscillation detection	A↔B pattern over 4-6 ticks = warning/hard stop
Concurrent tick mutex	in_progress status blocks overlapping cron fires
JSON crash recovery	try/catch on all JSON.parse; structured errors, not crashes
External validators	Pre-commit hooks, test suites detected at init and leveraged

Built-in Roles

Product:     pm, designer
User Lens:   new-user, active-user, churned-user
Engineering: frontend, backend, devops, architect, engineer
Quality:     security, tester, compliance, a11y
Specialist:  planner, user-simulator, devil-advocate

Devil's Advocate (the 10th person) is auto-included when consensus is near-unanimous or decisions are irreversible. Comes with an automated verification script that checks its own findings.

Custom Roles

Add a .md file to roles/:

---
tags: [review, build]
---
# Role Name

## Identity
One sentence: who you are and what you care about.

## Expertise
- **Area** — what you know about it

## When to Include
- Condition that triggers this role

Available immediately, no configuration needed.

Testing

bash test/run-all.sh

84 test files covering init-loop, complete-tick, next-tick, review independence, JSON crash recovery, compound defense, scope registry, criteria lint, pipeline E2E lint, D2 calibration, and orchestrator-level E2E flow tests.

Reproducing benchmarks

OPC ships with an extension system (v0.5, Run 1) so you can plug in additional hooks — visual checks, design-system audits, a11y scans — without forking the skill. The extension loader honors three bypasses so a single harness invocation can ignore locally-configured extensions:

# Disable every extension for one harness run
OPC_DISABLE_EXTENSIONS=1 node bin/opc-harness.mjs init --flow review --entry review --dir .harness

# Same effect, CLI flag form
node bin/opc-harness.mjs init --flow review --entry review --dir .harness --no-extensions

# Whitelist specific extensions only
node bin/opc-harness.mjs init --flow review --entry review --dir .harness --extensions visual-check,a11y

Priority order: OPC_DISABLE_EXTENSIONS=1 env var > --no-extensions CLI flag > --extensions foo,bar whitelist > config in ~/.claude/skills/opc-extension/config.json. See docs/specs/2026-04-16-opc-extension-system-design.md for the full contract.

Note: bash test/run-all.sh runs OPC's own internal test suite, which includes tests that intentionally load extensions to exercise the system. Don't set OPC_DISABLE_EXTENSIONS=1 when running the suite — use the bypasses only on real benchmarking / workflow invocations.

Requirements

• Claude Code (CLI, desktop app, or IDE extension)
• Node.js >= 18
• No runtime dependencies, no MCP server, no build step

Works better with memex (optional)

OPC works standalone — pair it with memex for cross-session memory. Memex remembers which roles were useful, which findings were false positives, and your project-specific context.

npm install -g @touchskyer/memex

Community

Using OPC? Share your setup in Discussions → Show and tell. Questions go in Q&A. Feature ideas in Ideas.

License

MIT