Post
301
Anthropic liked my agentic council idea and made their own implementation, Dynamic Workflows.
Lovely, dynamic critic role, mine is static, that can be driven by Claude. Done some upgrades on top.
Got my weekly limit reset from 82% to zero but only 50% used.
Been a day and the API backend is still broken, all sessions in retry loops; so sad, so unreliable.
Almost 9h to fix some tests and cut this new release instead of 30m!
🚀 claude-hooks v1.13.0 — the /consultants council learned to argue with itself
The multi-agent council (planner → researcher w/ project tools + citation linting → critic → synthesizer) was missing one thing: a real adversary.
🛡️ Dynamic adversarial review
A new post-synthesis refuter re-reads the council's answer and tries to break it — not a second opinion, a hostile one. It's a singleton, so it stays safe under x-tier multi-model fan-out (no N×M blow-up). The engine
drives an adversary checkpoint: pauses mid-run, streams the pause over SSE, auto-resumes once the skeptic panel reports. Two revived dials — critic_strictness and adversarial_focus — aim the scrutiny (security? perf?
one specific claim?).
🔁 Workflow-driven council
A committed Workflow driver (consult-with-adversarial-review.mjs) runs the whole loop deterministically: ask → review → skeptic panel → accept | follow-up. The council answers, Claude critiques, and the engine either
marks it accepted or auto-issues a focused follow-up and loops — bounded by a flat max_followups cap. A blocking consult --wait makes it scriptable.
🧩 Also in 1.13
• Throttle-aware proxy retry: jittered backoff, Retry-After, a 90s wall-clock deadline, a cross-session circuit breaker — all live at GET /health, so the proxy stops amplifying Anthropic's edge throttle.
• Per-project config: .claude-hooks/consultants.toml promoted via an override_user_global directive.
🐧 Linux / macOS / Windows. PostgreSQL OR SQLite. Local OR cloud LLMs.
🔗 github.com/mann1x/claude-hooks
Lovely, dynamic critic role, mine is static, that can be driven by Claude. Done some upgrades on top.
Got my weekly limit reset from 82% to zero but only 50% used.
Been a day and the API backend is still broken, all sessions in retry loops; so sad, so unreliable.
Almost 9h to fix some tests and cut this new release instead of 30m!
🚀 claude-hooks v1.13.0 — the /consultants council learned to argue with itself
The multi-agent council (planner → researcher w/ project tools + citation linting → critic → synthesizer) was missing one thing: a real adversary.
🛡️ Dynamic adversarial review
A new post-synthesis refuter re-reads the council's answer and tries to break it — not a second opinion, a hostile one. It's a singleton, so it stays safe under x-tier multi-model fan-out (no N×M blow-up). The engine
drives an adversary checkpoint: pauses mid-run, streams the pause over SSE, auto-resumes once the skeptic panel reports. Two revived dials — critic_strictness and adversarial_focus — aim the scrutiny (security? perf?
one specific claim?).
🔁 Workflow-driven council
A committed Workflow driver (consult-with-adversarial-review.mjs) runs the whole loop deterministically: ask → review → skeptic panel → accept | follow-up. The council answers, Claude critiques, and the engine either
marks it accepted or auto-issues a focused follow-up and loops — bounded by a flat max_followups cap. A blocking consult --wait makes it scriptable.
🧩 Also in 1.13
• Throttle-aware proxy retry: jittered backoff, Retry-After, a 90s wall-clock deadline, a cross-session circuit breaker — all live at GET /health, so the proxy stops amplifying Anthropic's edge throttle.
• Per-project config: .claude-hooks/consultants.toml promoted via an override_user_global directive.
🐧 Linux / macOS / Windows. PostgreSQL OR SQLite. Local OR cloud LLMs.
🔗 github.com/mann1x/claude-hooks