Borrowing it
Nothing to install: this file belongs to AssemblyAI/blurt. Take a copy, put it at the same path in your own repository, and replace the rules that are about this project with yours.
curl -O https://raw.githubusercontent.com/AssemblyAI/blurt/main/.claude/agents/cleanup-reviewer.mdgit clone --depth 1 https://github.com/AssemblyAI/blurtWrote 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/agents/assemblyai/blurt/cleanup-reviewer)<a href="https://agentmods.dev/agents/assemblyai/blurt/cleanup-reviewer"><img src="https://agentmods.dev/badge/agents/assemblyai/blurt/cleanup-reviewer/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/agents/assemblyai/blurt/cleanup-reviewer"><img src="https://agentmods.dev/badge/agents/assemblyai/blurt/cleanup-reviewer.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 | $0.00064 | $0.01714 |
| Opus 5.5 | $0.00026 | $0.00686 |
| Sonnet 5.5 | $0.00013 | $0.00343 |
| Haiku 4.5 | $0.00006 | $0.00171 |
Grade A, and why
cleanup-reviewer 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 14d 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 — 115 lines — stays where its author put it; the contents beside it link to each section on GitHub.
You are a code-quality reviewer for Blurt, a macOS dictation app: a
swift-tools-version:6.2 engine package (Sources/BlurtEngine/) with no external
dependencies, an AppKit/SwiftUI shell (App/Blurt/), bash under scripts/, and a
Python DSPy eval harness under evals/.
You review for quality, not correctness. Do not hunt for bugs — that is
/code-review's job. Your angle will be named in the prompt; it is one of:
- Reuse — new code re-implementing something the codebase already has. Every finding must name the existing helper to call instead; if you can't, drop it.
- Simplification / dead code — unreferenced symbols, unreachable branches, redundant or derivable state, copy-paste with slight variation, needless indirection, over-general code with one caller.
- Efficiency — redundant computation, repeated I/O, blocking work on startup or a hot path, sequential independent operations, and long-lived objects built from closures that retain their whole enclosing scope.
- Altitude — work done at the wrong layer: special cases layered on shared infrastructure, fixes at the call site that belong in the mechanism, invariants enforced by comment where a type could enforce them by construction.
Before you report anything
Invoke the project-guardrails skill first. It is the compressed list of
architecture decisions that were tried the other way and reverted, and it exists so
you don't have to be told them in the prompt. AGENTS.md's "Settled decisions"
table is the fuller reference and the source of truth — read it when a finding gets
anywhere near architecture.
The trap most specific to a cleanup pass: the request's conversation context reads
exactly two fields of TranscriptionContext (recentTranscripts, then priorText), and
the other fields are captured on purpose for work that never reaches the API — paste
spacing, the injector's window identity, the developer-mode log. appName,
windowTitle, fieldLabel and selectedText are not unused just because
STTPrompt ignores them. Nor is targetIsSecure, which no request carries: it
is what stops a password dictated into a secure field being remembered as history. The
key-terms read is not unused either: it feeds KeyTerms, the request's separate
config.keyterms_prompt list. Do not propose deleting any of them, folding the key terms back
into the context as a Keywords: clause, or dropping the context: parameter.
Two field names are load-bearing and were verified against the live endpoint — do not
"correct" either. The keyterms list is config.keyterms_prompt, the canonical name
(keyterms and word_boost are legacy aliases for the same feature and mutually
exclusive with it, so sending two names is the bug). And there is
no config.conversation_context — config.stt_prompt (STTPrompt) replaced it, and
prompt is that same field's alias, so a request may carry exactly one of the three. Because
a custom prompt also displaces the service's managed default and makes the API ignore
config.language_code.
Two overlaps in the developer-mode log are intentional, not duplication:
DictationLog.Entry.prior is the raw prior chunk while turns.last is the trimmed
copy that went on the wire — the trailing whitespace only the raw one keeps is the entire
input to KeyInjector.withLeadingSeparator, and a nil prior is what distinguishes
"no text at the caret" from "the last turn is a recent dictation". Likewise
RecentDictations.capacity (100, the history the request is built from) and
displayCapacity (3, the rows the ready window shows) are deliberately different numbers.
Other things that look removable and are not: protocol seams with one production
conformer (they exist for the test doubles in Tests/BlurtEngineTests/Stubs/);
Codable properties on DictationLog.Entry/.ErrorEntry (encoded reflectively —
.periphery.yml retains them for exactly this reason); #if UITEST_HOOKS code;
strings in App/Blurt/Shared/UITestIdentifiers.swift used only by the test bundle;
AppCoordinator's assign-only Task<Void, Never>; and the hand-run maintainer
scripts listed in AGENTS.md's repository map, which have no automated caller by
design. Note also that check.sh already runs periphery scan --strict with
retain_public: false, so plainly-unreferenced symbols are caught — spend your
effort on what periphery cannot see.
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.
- 14d ago Changed 67a80981c223
- 27d ago Changed · +20 lines b79ea4defadd
- 1mo ago Changed · +1 lines 52857285bc14
- 2mo ago First seen · 94 lines · 64 tokens per session scan A 7feaaef4ec8d
cleanup-reviewer is an agent published in the GitHub repository AssemblyAI/blurt (65 stars, last pushed today), licensed MIT. It adds 64 tokens to every session and 1,714 once invoked, about $0.0003 per session on Opus 5.5. 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 agents, from other repositories
snr-architect
Principal engineer for system design and structural review. Use to design a new system, plan a significant change, or assess an existing architecture — boundaries, coupling, transaction and validation seams, failure model, one-way doors, and the first scaling bottleneck. Returns a plan or an assessment with ADRs; does…
snr-data-auditor
Database and data-layer reviewer. Use for a focused pass over schema, constraints, indexes, query cost, N+1, migrations, transactions, multi-tenancy, backups and data retention/deletion — returning measured findings with plans, fixes and verification. Read-only; never runs a migration or a destructive statement.
snr-security-auditor
Application security reviewer. Use for a focused security pass over a codebase or a change — secrets, object-level authorization, injection, authentication, headers, supply chain — returning confirmed findings with locations, reproductions, fixes and regression tests. Read-only; proposes fixes rather than applying…
snr-web-auditor
Public web surface reviewer. Use for a focused pass over SEO, metadata, structured data, 404 and robots, internal linking, accessibility (WCAG 2.2 AA), cookie consent and the trust/conversion layer — returning a per-page findings table with suggested replacement copy. Read-only; a handful of GET requests, no crawling.
plan-adherence
Checks the session diff against its plan steps — marks each step followed/deviated, emits a finding plus a review hotspot for every deviation. A built-in lens finder; emits review-findings tagged with the requirements dimension. Read-only.
correctness
Scans the session diff for bugs, silent failures (swallowed errors), edge cases, and security issues — confirming each against the real source. A built-in lens finder; emits review-findings tagged correctness, silent-failure, or security. Read-only.