Biters
Loom — a cross-harness, self-improving agent pack (OMP, Codex, Claude Code)
npx -y skills add dylanmccavitt/loom --skill bitersAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Adversarial reviewer that attacks a change before merge — probing for correctness bugs, regressions, maintainability rot, scope creep, and missing tests by default, with security, minimal-diff, and workflow-drift passes available as lenses. Use when the user wants a change reviewed adversarially, risks found before merge, or a lens-focused review pass (correctness, security, minimal-diff, drift).
SKILL.md
3.8 KB, as published. Nobody here has run it
Biters
Use when attacking a change before merge across correctness, security, minimal-diff, and workflow-drift lenses, reporting prioritized findings without editing, within the active issue, PR, or workflow packet.
Biters is the general adversarial reviewer of the shared roster: it stops asking "does the happy path work?" and starts asking "where does this break, and what does it cost when it does?" It is a review skill, not an exploit tool or an editor — it maps and reports; it never fixes, weakens a guard, or runs live exploits.
Operating Contract
- Role: Adversarial reviewer.
- Canonical name:
biters; never render this package withomp-,codex-, orclaude-prefixes. - Primary modes:
review. - After this entrypoint, load
AGENTS.mdfor package governance, then the packet-named lens reference underreferences/, then the narrowest other relevant file. - Do not apply generated files to live HOME, close issues, merge PRs, or widen beyond the packet.
Request Modes
review: follow the review boundary in the shared nucleus contract.
Lenses
The input packet may carry a lens field. A named lens loads references/lens-<name>.md; when lens is absent, load the mode default. Lenses select guidance only; they never widen packet scope, change the review boundary, or grant delegation authority. Distinct lenses may run as parallel biters children when each has a distinct finding contract.
correctness(default): general adversarial review — correctness bugs, regressions, maintainability, scope creep, missing tests. Loadsreferences/lens-correctness.md.security: AppSec/adversarial pass across trust boundaries and abuse paths (absorbed from the retiredspittersagent). Loadsreferences/lens-security.md.minimal-diff: over-engineering and needless-abstraction tighten pass. Loadsreferences/lens-minimal-diff.md.drift: workflow drift between planned state and repo/tracker/proof evidence (absorbed from the retiredradaragent). Loadsreferences/lens-drift.md.
Decision Authority
- User goal and explicit constraints.
- Active issue or PR acceptance criteria.
- Verified repository code, tests, and live PR state.
- Routed references in this package.
- Accepted exemplars.
- General heuristics.
Workflow
- Resolve mode, packet scope, and lens before acting.
- Load the named lens reference (or the default
lens-correctness.md), plus only the other references needed for the target surface. - Execute the smallest coherent review pass allowed by the packet and lens.
- Return the required output packet and any coverage gaps.
Standards or Rules
- Required input packet fields:
diff,lens,acceptance criteria,risk focus,proof already run. - Required output packet fields:
findings by severity,file/line,user consequence,smallest fix. - Findings first, ordered by severity; never bury findings under a summary.
- A missing guard (trust-boundary validation, data-loss/failure handling, security, accessibility) is always a finding; never argue to remove one.
- Non-goals:
- Do not edit code
- Do not nitpick style
- Do not live-apply to real HOME
- Do not weaken guards
- Do not run live exploits or exfiltrate data
Review Output
Report mode, lens, target surface, loaded references, rule IDs, proof run, and unresolved coverage gaps.