peptide-repo/prautoblogger
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(ci): add Neurometric PR-Summarizer as 4th QA voice (#10)

Browse Source 
Neurometric markets a task-specialized PR-Summarizer SLM (Qwen3-4B +
specialist system prompt) on a free tier. Adding it as a 4th voice in
the existing A/B reviewer panel to see whether a specialized small
model catches things the 3 general-purpose voices miss.

The SLM does not honor OpenAI tool-calling, so it cannot return our
structured submit_review schema. Wired as a "prose voice":

- MODELS tuple gains a 6th provider_type field
  ("openrouter_tools" | "neurometric_prose")
- call_model() dispatches; new call_neurometric_prose() path hits
  api.neurometric.ai/v1 without tools, embeds raw markdown output,
  and infers verdict from keyword heuristic
- Render branch in PR comment marks the verdict as heuristic and
  notes tokens/cost are not tracked (free tier)
- Never the primary voice; heading still driven by Haiku

Also hardens the shared render path against models that omit the
optional 'summary' field (rev.get('summary') fallback).

Requires new repo secret NEUROMETRIC_API_KEY (already set on all four
peptiderepo repos). Voice self-skips with inline error if missing.

Agent-Session: cowork-neurometric-qa-voice-20260415

Co-authored-by: peptiderepo <peptiderepo@users.noreply.github.com>
This commit is contained in:
peptiderepo
2026-04-15 16:12:16 +07:00
committed by GitHub
parent 018a78000c
commit 2d75051967
1 changed files with 111 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

124
.github/workflows/cto-review.yml vendored
View File

@@ -1,15 +1,20 @@
name: CTO Review
# Automated PR review powered by Claude + DeepSeek + Kimi via OpenRouter.
# Automated PR review powered by Claude + DeepSeek + Kimi (via OpenRouter)
# plus Neurometric PR-Summarizer (a Qwen3-4B task-specialist SLM).
#
# During the A/B window (until further notice), every PR gets reviewed by THREE models
# During the A/B window (until further notice), every PR gets reviewed by FOUR models
# in parallel. The PR comment shows each model's verdict side-by-side so we can compare
# quality/cost on real PRs before picking a permanent default.
#
# Neurometric is a "prose voice": it does not support OpenAI tool-calling, so its
# contribution is raw markdown + a heuristic verdict. It is never the primary voice.
#
# Does NOT merge. Does NOT post formal APPROVE/REQUEST_CHANGES reviews — always COMMENT
# (GITHUB_TOKEN can't formally approve PRs without repo-setting toggle).
#
# Requires repo secret: OPENROUTER_API_KEY
# Requires repo secrets: OPENROUTER_API_KEY, NEUROMETRIC_API_KEY (optional; the
# Neurometric voice self-skips with an inline error note if the key is missing).
#
# Cost cap: tracked in Repository Variables (CTO_REVIEW_SPEND_CENTS, CTO_REVIEW_CAP_CENTS).
# Default cap is 1000 cents ($10/month). Auto-resets on the 1st of each month.
@@ -145,11 +150,12 @@ jobs:
echo "EOF_CONV"
} >> "$GITHUB_OUTPUT"
- name: Call 3 models in parallel via OpenRouter
- name: Call 4 models in parallel (OpenRouter + Neurometric)
if: steps.budget.outputs.skip != 'true'
id: review
env:
OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }}
NEUROMETRIC_API_KEY: ${{ secrets.NEUROMETRIC_API_KEY }}
PR_TITLE: ${{ github.event.pull_request.title }}
PR_BODY: ${{ github.event.pull_request.body }}
PR_AUTHOR: ${{ github.event.pull_request.user.login }}
@@ -181,14 +187,19 @@ jobs:
api_key = os.environ["OPENROUTER_API_KEY"]
# A/B harness: three models. Each gets the same prompt.
# A/B harness: four models. Each gets the same prompt.
# Primary model's verdict drives the heading. Others are labeled.
# Edit this list to change the roster (or drop to 1 after A/B window closes).
#
# provider_type semantics:
# - "openrouter_tools": OpenAI tool-calling via OpenRouter; structured verdict.
# - "neurometric_prose": Neurometric marketplace SLM; prose output + heuristic verdict.
MODELS = [
# (openrouter slug, display name, input_per_M_USD, output_per_M_USD, is_primary)
("anthropic/claude-haiku-4.5", "Claude Haiku 4.5", 1.00, 5.00, True),
("deepseek/deepseek-chat-v3.1", "DeepSeek V3.1", 0.15, 0.75, False),
("moonshotai/kimi-k2-thinking", "Kimi K2 Thinking", 0.60, 2.50, False),
# (slug, display name, input_per_M_USD, output_per_M_USD, is_primary, provider_type)
("anthropic/claude-haiku-4.5", "Claude Haiku 4.5", 1.00, 5.00, True, "openrouter_tools"),
("deepseek/deepseek-chat-v3.1", "DeepSeek V3.1", 0.15, 0.75, False, "openrouter_tools"),
("moonshotai/kimi-k2-thinking", "Kimi K2 Thinking", 0.60, 2.50, False, "openrouter_tools"),
("neurometric/pr-summarizer", "Neurometric PR-Summarizer (Qwen3-4B)", 0.00, 0.00, False, "neurometric_prose"),
]
system = """You are the CTO and lead reviewer for the peptiderepo project. You are reviewing a pull request opened by an autonomous agent or human contributor.
@@ -251,7 +262,28 @@ jobs:
PER_MODEL_TIMEOUT = 60 # seconds; one hung model should not block siblings
def call_model(slug, display, in_rate, out_rate, is_primary):
# Heuristic verdict inference for prose-voice models that can't do tool-calling.
# Keyword order matters: block markers beat approve markers.
PROSE_BLOCK_MARKERS = [
"do not merge", "do not approve", "should not be merged", "reject this",
"critical security", "critical issue", "critical vulnerability",
"blocker", "blocking issue", "request_changes", "request changes",
"**critical**", "sql injection", "must be addressed before",
]
PROSE_APPROVE_MARKERS = [
"safe to merge", "ready to merge", "no blocking issues", "lgtm",
"approve this", "looks good to merge", "no concerns",
]
def infer_verdict_from_prose(text):
t = (text or "").lower()
if any(m in t for m in PROSE_BLOCK_MARKERS):
return "REQUEST_CHANGES"
if any(m in t for m in PROSE_APPROVE_MARKERS):
return "APPROVE"
return "COMMENT"
def call_openrouter_tools(slug, display, in_rate, out_rate, is_primary):
body = {
"model": slug,
"max_tokens": 2048,
@@ -279,6 +311,7 @@ jobs:
except Exception as e:
return {
"slug": slug, "display": display, "is_primary": is_primary,
"provider_type": "openrouter_tools",
"error": str(e)[:300],
"cost_cents": 0,
}
@@ -286,9 +319,9 @@ jobs:
choice = data.get("choices", [{}])[0].get("message", {})
tool_calls = choice.get("tool_calls", [])
if not tool_calls:
# Some non-Anthropic providers occasionally put the tool args in content.
return {
"slug": slug, "display": display, "is_primary": is_primary,
"provider_type": "openrouter_tools",
"error": f"No tool_call in response. Raw content: {str(choice)[:300]}",
"cost_cents": 0,
}
@@ -298,6 +331,7 @@ jobs:
except Exception as e:
return {
"slug": slug, "display": display, "is_primary": is_primary,
"provider_type": "openrouter_tools",
"error": f"Malformed tool args: {e}",
"cost_cents": 0,
}
@@ -309,10 +343,62 @@ jobs:
cost_cents = cost_usd * 100
return {
"slug": slug, "display": display, "is_primary": is_primary,
"provider_type": "openrouter_tools",
"review": review, "prompt_tokens": ptok, "completion_tokens": ctok,
"cost_cents": cost_cents,
}
def call_neurometric_prose(slug, display, is_primary):
nm_key = os.environ.get("NEUROMETRIC_API_KEY", "")
if not nm_key:
return {
"slug": slug, "display": display, "is_primary": is_primary,
"provider_type": "neurometric_prose",
"error": "NEUROMETRIC_API_KEY secret not set on this repo",
"cost_cents": 0,
}
body = {
"model": slug,
"max_tokens": 2048,
"messages": [
{"role": "system", "content": system},
{"role": "user", "content": user_message},
],
}
req = urllib.request.Request(
"https://api.neurometric.ai/v1/chat/completions",
data=json.dumps(body).encode(),
headers={
"Authorization": f"Bearer {nm_key}",
"content-type": "application/json",
},
method="POST",
)
try:
with urllib.request.urlopen(req, timeout=PER_MODEL_TIMEOUT) as resp:
data = json.loads(resp.read())
except Exception as e:
return {
"slug": slug, "display": display, "is_primary": is_primary,
"provider_type": "neurometric_prose",
"error": str(e)[:300],
"cost_cents": 0,
}
prose = data.get("choices", [{}])[0].get("message", {}).get("content") or ""
verdict = infer_verdict_from_prose(prose)
return {
"slug": slug, "display": display, "is_primary": is_primary,
"provider_type": "neurometric_prose",
"review": {"verdict": verdict, "prose": prose},
"prompt_tokens": 0, "completion_tokens": 0,
"cost_cents": 0,
}
def call_model(slug, display, in_rate, out_rate, is_primary, provider_type="openrouter_tools"):
if provider_type == "neurometric_prose":
return call_neurometric_prose(slug, display, is_primary)
return call_openrouter_tools(slug, display, in_rate, out_rate, is_primary)
results = []
with concurrent.futures.ThreadPoolExecutor(max_workers=len(MODELS)) as pool:
futures = [pool.submit(call_model, *m) for m in MODELS]
@@ -342,9 +428,21 @@ jobs:
continue
rev = r["review"]
lines.append("")
lines.append(f"**Verdict:** {rev['verdict']}")
if r.get("provider_type") == "neurometric_prose":
lines.append(f"**Verdict:** {rev['verdict']} _(heuristic, inferred from prose)_")
lines.append("")
# Cap at 6KB to keep PR comments readable; 4B model output is rarely over this
prose = rev.get("prose", "")
if len(prose) > 6000:
prose = prose[:6000] + "\n\n_…output truncated for comment length…_"
lines.append(prose)
lines.append("")
lines.append("_free tier; per-call token usage not tracked_")
lines.append("")
continue
lines.append(f"**Verdict:** {rev.get('verdict', 'COMMENT')}")
lines.append("")
lines.append(rev["summary"])
lines.append(rev.get("summary") or "_(model returned no summary)_")
lines.append("")
if rev.get("concerns"):
lines.append("**Concerns:**")