Skip to content

Latest commit

 

History

History

workflows

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 

README.md

workflows/ — agent-workflow scripts

Scripts for the Claude Code Workflow tool (not CLIs — they run with an ambient args binding and orchestrate subagents; no filesystem access inside the script itself).

adversarial-review.mjs

Adversarially reviews a commitwork audit packet into a coherent, dependency-ordered remediation plan. Four phases: Dimension (bifocal edge-walk/fractal + foureyes void analysis) → Adversarial (breakers stress security/perf/correctness/sequencing) → Overloop (critique → revise, ≤4 rounds, independent critic each round) → Synthesize (final ordered plan).

Args: { packet, draft, project, reviewCommands? } — the producer is bin/audit.mjs, which emits audit-packet.json + remediation-draft.md.

Tool-first against internal-d — one source of truth, no prompt-level copies: agents call internal-d's bifocal-plan and overloop MCP tools when that server is connected, and otherwise read the canonical methodology files from internal-d's .claude/commands/ (bifocal-plan.md, eye-foureyes.md, fang-breakers.md, hand-overloop.md); override the location with args.reviewCommands.

The other half of the internal-d seam lives in the monitor: monitor/export-overwatch.mjs pushes each rollup into the overwatch-layer (memory-layer server) at the end of every sweep so internal-d's overlook/meta_audit can reason over audit history — health-gated, rate-limit-aware, idempotent, disabled with SUBSTRATE_EXPORT=0.

Known open defects against this script are in the reconciled audit queue: a machine-specific default for the methodology location, a failed critic counted as approval, a null round-1 revision reaching Synthesize, and an unvalidated args.draft. Queue ids are positions and renumber on regeneration, so find them by file rather than by id.

Cost note (2026-07-30)

Workflow runs burn tokens fast — a 25-agent audit consumed 3.3M tokens and spent ~80% of its 7-hour wall-clock paused on usage-window resets. Prefer default effort + a smaller model for mechanical scan stages (opts.model/opts.effort), partition inputs instead of having every agent re-read the whole surface, and start heavy runs on a fresh usage window.