pat-eason · August 12, 2026 17:05
diff --git a/README.md b/README.md
diff --git a/adversary.py b/adversary.py
 #!/usr/bin/env python3
 """Adversarial PR review helper — calls one litellm model with the review prompt.

 Usage:
  python3 adversary.py <model_id> "<lens instruction>" <prompt_file> [output_file] [--temperature <n>]

 Reads the litellm API key from ~/.config/opencode/opencode.json.
 Temperature is OMITTED by default (each model uses its own default — some models,
 e.g. gpt-5.6-sol, reject non-default temperature values). Pass --temperature to override.

 Writes JSON to stdout (or output_file if given):
  {"model": ..., "findings": [...]} or {"model": ..., "error": "...", "findings": []}
 """
 import json
 import os
 import sys
 import urllib.request

 LITELLM_URL = "https://litellm.luxurycoders.com/v1/chat/completions"

 OUTPUT_FORMAT = """Return ONLY a JSON array of findings. Each finding:
 {
  "concern": "correctness|maintainability|security|performance|reliability|testability|operability",
  "location": "file:line or concept description",
  "severity": 0-5,
  "what": "one-line description",
  "why": "why it's a problem, with surrounding-code reasoning",
  "fix": "what it should be changed to",
  "blast_radius": "what else would be affected by the fix",
  "confidence": "low|med|high"
 }

 Severity scale: 0 = critical (security hole, data loss, prod-breaking, spec violation),
 1 = high (likely bug, missing critical test), 2 = medium (maintainability/coupling, perf under load),
 3 = low (improvement, minor edge case), 4 = nit (style, naming), 5 = optional (suggestion, non-blocking).

 If you find no issues, return []."""


 def main():
    if len(sys.argv) < 4:
        print(json.dumps({"error": "usage: adversary.py <model> <lens> <prompt_file> [output_file]"}))
        sys.exit(1)

    model = sys.argv[1]
    lens = sys.argv[2]
    prompt_file = sys.argv[3]
    output_file = sys.argv[4] if len(sys.argv) > 4 and not sys.argv[4].startswith("--") else None

    temperature = None
    args = sys.argv[4:]
    for i, a in enumerate(args):
        if a == "--temperature" and i + 1 < len(args):
            try:
                temperature = float(args[i + 1])
            except ValueError:
                pass

    config_path = os.path.expanduser("~/.config/opencode/opencode.json")
    try:
        with open(config_path) as f:
            config = json.load(f)
        api_key = config["provider"]["litellm"]["options"]["apiKey"]
    except Exception as e:
        print(json.dumps({"model": model, "error": f"config read: {e}", "findings": []}))
        sys.exit(0)

    with open(prompt_file) as f:
        prompt = f.read()

    full_prompt = f"""{prompt}

 YOUR REVIEW LENS (in addition to the 7 concerns): {lens}

 {OUTPUT_FORMAT}"""

    payload = {
        "model": model,
        "messages": [{"role": "user", "content": full_prompt}],
        "max_tokens": 8000,
    }
    if temperature is not None:
        payload["temperature"] = temperature
    body = json.dumps(payload).encode()

    req = urllib.request.Request(
        LITELLM_URL,
        data=body,
        headers={
            "Authorization": f"Bearer {api_key}",
            "Content-Type": "application/json",
            "x-litellm-tags": "harness=opencode-skill=pr-code-review",
        },
    )

    try:
        with urllib.request.urlopen(req, timeout=180) as resp:
            data = json.loads(resp.read())
        content = data["choices"][0]["message"]["content"]
        start = content.find("[")
        end = content.rfind("]") + 1
        findings = json.loads(content[start:end]) if start >= 0 and end > start else []
        result = {"model": model, "findings": findings}
    except Exception as e:
        result = {"model": model, "error": str(e), "findings": []}

    out = json.dumps(result)
    if output_file:
        with open(output_file, "w") as f:
            f.write(out)
    print(out)


 if __name__ == "__main__":
    main()
diff --git a/SKILL.md b/SKILL.md
Concern	One-line test
Correctness	Is the spec itself right? Does the solution match what the code actually does?
Maintainability & extensibility	Can the next change be made safely and cheaply?
Security	Can only the right principals do or see this? (IDOR, vacuous auth, unauth endpoints)
Performance	Is the access pattern right for the read/write shape, and does it survive N× load?
Reliability & failure modes	What happens when a dependency is slow, down, or half-done?
Testability & verifiability	Can this code prove itself correct without a human reading every line?
Operability	Can we run this as a team and with agents? Do we know if something is actually failing?
Role	Model	Lens
Adversary 1	DeepSeek-v4-pro	Logic correctness, edge cases, data flow
Adversary 2	Kimi-k3	Cross-file impact, long-range dependencies
Adversary 3	Qwen3.7-plus	Maintainability, extensibility, coupling, boundaries
Adversary 4	GPT-5.6-sol	Security + reliability/failure modes
Adjudicator	Orchestrator (inline) or Claude Opus 5 (optional)	Consensus, dedupe, rank, controlled-entropy check
File	Purpose
`SKILL.md`	The skill definition — opencode/Claude Code reads this to run the review
`adversary.py`	Helper script that calls one litellm model with the review prompt (stdlib only)
`README.md`	This file
	#!/usr/bin/env python3
	"""Adversarial PR review helper — calls one litellm model with the review prompt.

	Usage:
	python3 adversary.py <model_id> "<lens instruction>" <prompt_file> [output_file] [--temperature <n>]

	Reads the litellm API key from ~/.config/opencode/opencode.json.
	Temperature is OMITTED by default (each model uses its own default — some models,
	e.g. gpt-5.6-sol, reject non-default temperature values). Pass --temperature to override.

	Writes JSON to stdout (or output_file if given):
	{"model": ..., "findings": [...]} or {"model": ..., "error": "...", "findings": []}
	"""
	import json
	import os
	import sys
	import urllib.request

	LITELLM_URL = "https://litellm.luxurycoders.com/v1/chat/completions"

	OUTPUT_FORMAT = """Return ONLY a JSON array of findings. Each finding:
	{
	"concern": "correctness\|maintainability\|security\|performance\|reliability\|testability\|operability",
	"location": "file:line or concept description",
	"severity": 0-5,
	"what": "one-line description",
	"why": "why it's a problem, with surrounding-code reasoning",
	"fix": "what it should be changed to",
	"blast_radius": "what else would be affected by the fix",
	"confidence": "low\|med\|high"
	}

	Severity scale: 0 = critical (security hole, data loss, prod-breaking, spec violation),
	1 = high (likely bug, missing critical test), 2 = medium (maintainability/coupling, perf under load),
	3 = low (improvement, minor edge case), 4 = nit (style, naming), 5 = optional (suggestion, non-blocking).

	If you find no issues, return []."""


	def main():
	if len(sys.argv) < 4:
	print(json.dumps({"error": "usage: adversary.py <model> <lens> <prompt_file> [output_file]"}))
	sys.exit(1)

	model = sys.argv[1]
	lens = sys.argv[2]
	prompt_file = sys.argv[3]
	output_file = sys.argv[4] if len(sys.argv) > 4 and not sys.argv[4].startswith("--") else None

	temperature = None
	args = sys.argv[4:]
	for i, a in enumerate(args):
	if a == "--temperature" and i + 1 < len(args):
	try:
	temperature = float(args[i + 1])
	except ValueError:
	pass

	config_path = os.path.expanduser("~/.config/opencode/opencode.json")
	try:
	with open(config_path) as f:
	config = json.load(f)
	api_key = config["provider"]["litellm"]["options"]["apiKey"]
	except Exception as e:
	print(json.dumps({"model": model, "error": f"config read: {e}", "findings": []}))
	sys.exit(0)

	with open(prompt_file) as f:
	prompt = f.read()

	full_prompt = f"""{prompt}

	YOUR REVIEW LENS (in addition to the 7 concerns): {lens}

	{OUTPUT_FORMAT}"""

	payload = {
	"model": model,
	"messages": [{"role": "user", "content": full_prompt}],
	"max_tokens": 8000,
	}
	if temperature is not None:
	payload["temperature"] = temperature
	body = json.dumps(payload).encode()

	req = urllib.request.Request(
	LITELLM_URL,
	data=body,
	headers={
	"Authorization": f"Bearer {api_key}",
	"Content-Type": "application/json",
	"x-litellm-tags": "harness=opencode-skill=pr-code-review",
	},
	)

	try:
	with urllib.request.urlopen(req, timeout=180) as resp:
	data = json.loads(resp.read())
	content = data["choices"][0]["message"]["content"]
	start = content.find("[")
	end = content.rfind("]") + 1
	findings = json.loads(content[start:end]) if start >= 0 and end > start else []
	result = {"model": model, "findings": findings}
	except Exception as e:
	result = {"model": model, "error": str(e), "findings": []}

	out = json.dumps(result)
	if output_file:
	with open(output_file, "w") as f:
	f.write(out)
	print(out)


	if __name__ == "__main__":
	main()
name	pr-code-review
description	Comprehensive multi-model adversarial + consensus PR review. Invoke with /pr-code-review <PR URL> or when the user asks for "code review on this PR", "review this PR", "adversarial PR review", "multi-model PR review", "consensus review on this PR", or similar. Dispatches 4 adversary models (DeepSeek-v4-pro, Kimi-k3, Qwen3.7-plus, GPT-5.6-sol) for independent findings, then consensus adjudication with full file access for controlled-entropy validation. Claude Opus 5 is used for adjudication when Anthropic is configured; otherwise the orchestrator model adjudicates inline. Reviews against the 7 Engineering Concerns, retrieves Linear + Notion context, ranks findings 0-5 severity (0 = most severe), and produces an action plan.
argument-hint	<GitHub PR URL or PR number>
Role	Model ID (litellm)	Lens
Adversary 1	`fireworks/models/deepseek-v4-pro`	Logic correctness, edge cases, data flow
Adversary 2	`fireworks/kimi-k3`	Cross-file impact, long-range dependencies, large-context
Adversary 3	`fireworks/qwen3.7-plus`	Maintainability, extensibility, coupling, boundaries
Adversary 4	`gpt-5.6-sol`	Security (IDOR, vacuous auth, unauth endpoints) + reliability/failure modes
Adjudicator	Orchestrator inline (default) or `anthropic/claude-opus-5` subagent (if Anthropic configured)	Consensus, dedupe, resolve disagreements, rank, controlled-entropy check
Argument	Action
GitHub PR URL	Parse owner/repo/number
Bare number	Current repo: `gh pr view <num> --json ...`
Signal in paths	Surface	Extra lens
`auggie-platform`, `dashboard`, `home-search`, `websites`, SSR	Full-stack web	Read-path optimization, white-label constraints (nothing hardcoded per client)
`lpmobileV2`, `lpmobile`, OTA, native modules, deep-link, push	Mobile	Contract changes hit shipped binaries (last quarter's app sends last quarter's query); OTA vs native
contact-ingestion, MLS ETL, notifications, audit-logging, lambdas, queues, subgraphs, gateways, jobs, pipelines	Distributed systems	Idempotency (at-least-once), backpressure, partial-failure, data-store access patterns