Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add wan-huiyan/agent-traffic-control --skill code-review-subagent-fabricates-specifics-to-inflate-severitygit clone --depth 1 https://github.com/wan-huiyan/agent-traffic-controlWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/wan-huiyan/agent-traffic-control/code-review-subagent-fabricates-specifics-to-inflate-severity)<a href="https://agentmods.dev/skills/wan-huiyan/agent-traffic-control/code-review-subagent-fabricates-specifics-to-inflate-severity"><img src="https://agentmods.dev/badge/skills/wan-huiyan/agent-traffic-control/code-review-subagent-fabricates-specifics-to-inflate-severity/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/wan-huiyan/agent-traffic-control/code-review-subagent-fabricates-specifics-to-inflate-severity"><img src="https://agentmods.dev/badge/skills/wan-huiyan/agent-traffic-control/code-review-subagent-fabricates-specifics-to-inflate-severity.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00428 | $0.02793 |
| Opus 5 | $0.00214 | $0.01396 |
| Sonnet 5 | $0.00086 | $0.00559 |
| Haiku 4.5 | $0.00043 | $0.00279 |
Grade A, and why
code-review-subagent-fabricates-specifics-to-inflate-severity scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 12d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 241 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Code-Review Subagents Fabricate Specific Evidence to Inflate Severity
Problem
A code-review subagent — especially a high-tier one (opus, voltagent-qa-sec, single-agent-multi-persona) — returns a report with a BLOCKING or HIGH-severity finding that reads as airtight: the issue is named, the call sites are enumerated, exact line numbers are cited, the blast radius is quantified. The orchestrator sees the specificity and treats the severity as authoritative, because nobody fabricates line numbers, right?
But a model fabricating coherent, specific-looking evidence to justify a high-stakes call IS exactly what happens. The reviewer's general principle is often correct (e.g. "defensive coding for future X is wise"), but the SPECIFIC EVIDENCE cited to inflate the severity from MEDIUM to BLOCKING — the call counts, the line numbers, the affected-file list — is invented. A 30-second grep would falsify it.
This is the inverse asymmetry of the well-known "reviewers miss things" problem. Reviewers don't just miss things; they also confabulate specifics to justify the severity they want to assign. Specificity reads as rigor. The fix is independent verification of the specifics — not just the principle.
Context / Trigger Conditions
Trigger this skill when:
-
A single-run code-review subagent returns a BLOCKING/HIGH finding justified by concrete numeric evidence like "there are 4 active calls at lines 3174, 3230, 3251" or "this affects N templates across the repo". The numeric specificity is the tell.
-
The reviewer ran with opus/high-tier model and the confident framing is reading as authoritative. Higher-capability models produce more convincing fabrications, not fewer.
-
You're under time pressure to merge and the BLOCKING tag is the only thing standing between you and ship. The pressure makes you want to fix first and verify never — exactly when verification matters most.
-
The reviewer's grep would have been trivial (
grep -c "flash.*warning" webapp/views/analysis.py,gh issue list --label X | wc -l) but the report shows no evidence of grep output, only the assertion. A real reviewer who actually ran the grep would have pasted the output.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 12d ago First seen · 241 lines · 428 tokens per session scan A ea3d9df1040f
code-review-subagent-fabricates-specifics-to-inflate-severity is a skill published in the GitHub repository wan-huiyan/agent-traffic-control (3 stars, last pushed yesterday), licensed MIT. It adds 428 tokens to every session and 2,793 once invoked, about $0.0021 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
vertical-real-estate
Residential-proptech domain knowledge so architect / pm aren't naive when speccing real-estate products (listings, lead-crm, transaction-coordination, property-mgmt). Codifies MLS/IDX reality, listing status lifecycle + syndication canonical-source, long-cycle lead nurture, transaction-coordination as the high-pain…
skeptical-triage
Reusable 3-round self-challenge + arbiter pattern for filtering false positives from findings/verdicts. Use when the cost of a false-positive gate block exceeds the cost of 4 extra LLM turns.
well-architected
6-pillar architecture review framework. Adapted from AWS Well-Architected for use by greatcto's architect agent on every non-nano ARCH document. Forces explicit answers across operational excellence, security, reliability, performance, cost, and sustainability — not just feature design.
product-economics
Does this product make money at a price someone will pay? Forces contribution margin, a price with a stated basis, and a bottom-up market size — each number labelled measured / assumed / unknown, so a guess can never be read as a calculation.
pm-planning
Decomposition methodology for pm agent — turns an approved ARCH document into a Beads task list with explicit dependencies, time-boxes, and acceptance criteria. The pipeline can only orchestrate work it can see; this skill defines what "seeable work" looks like.
decision-eval
Spawns the decision-scorer agent after architect proposes 2+ variants in an ADR. Produces a weighted scoring table and recommended choice saved to docs/decisions/.