Anthropic-native power-user harness: spec-driven, multispec, max-parallel, verified.
View on GitHub →Anthropic-native harnesses, deterministic dark-pattern hooks, and process-integrity agents — Apache-2.0, drop-in, no SaaS. Built for power-user operators who want effective completion and verifiable closeouts, not vibes.
Apache-2.0 · no SaaS · jq / python3-only deps · drop-in
Featured
The flagship projects — the harnesses and hooks most operators reach for first.
Anthropic-native power-user harness: spec-driven, multispec, max-parallel, verified.
View on GitHub →The claudemax harness ported onto Cursor's headless multi-model surface.
View on GitHub →Umbrella suite of Stop hooks suppressing LLM dark-pattern defaults.
View on GitHub →Stop hook blocking false-success closeouts. Single bash file, jq-only.
View on GitHub →Autonomous coding agent: web recon, plan debate, 24/7 loops, $0 local mode.
View on GitHub →MiniMax 2.7 Highspeed Claude Code harness: 100 TPS, 5-tier memory.
View on GitHub →AI coding harnesses
Spec-driven, max-parallel, independently-verified orchestrators that drive coding agents through decompose → fan-out → blind-verify loops on whatever model pool you have.
Anthropic-native power-user harness: spec-driven, multispec, max-parallel, verified.
View on GitHub →The claudemax harness ported onto Cursor's headless multi-model surface.
View on GitHub →MiniMax 2.7 Highspeed Claude Code harness: 100 TPS, 5-tier memory.
View on GitHub →Autonomous coding agent: web recon, plan debate, 24/7 loops, $0 local mode.
View on GitHub →Voice-driven Claude Code orchestration: talk to your agents on the move.
View on GitHub →LLM dark-patterns hook suite
Single-purpose Claude Code Stop hooks. Deterministic out-of-band judges that refuse dark-patterned closeouts — sycophancy, false success, fake citations, permission-loops, training-cutoff confidence — at the textual boundary. jq/python3-only deps, drop-in.
Umbrella suite of Stop hooks suppressing LLM dark-pattern defaults.
View on GitHub →Stop hook blocking false-success closeouts. Single bash file, jq-only.
View on GitHub →Blocks 'want me to continue?' permission-loop endings. Operator already authorized.
View on GitHub →Blocks praise-spam at turn open. Lead with the answer.
View on GitHub →Blocks vibe time estimates and linear-scaling parallelism claims.
View on GitHub →Blocks unsolicited rest/sleep/wellness paternalism. Treats operator as peer.
View on GitHub →Blocks academic citation patterns without a verifiable URL.
View on GitHub →Blocks fabricated-looking percentages and dollar amounts without a source.
View on GitHub →Blocks 'as we discussed earlier' false-memory claims without verbatim quotes.
View on GitHub →Survives context compaction: snapshots working state, rehydrates on start.
View on GitHub →Injects local system clock into every session. Kills 'as of training cutoff'.
View on GitHub →Agent quality & process
Adversarial pre-planning auditors, process-integrity enforcers, design directors, and the deterministic benchmark behind the hook suite — plus a CrewAI-native port of the no-vibes detector.
Deterministic closeout physics engine and benchmark for dark-pattern detection.
View on GitHub →CrewAI Task guardrail blocking verification claims without evidence.
View on GitHub →Mandatory pre-planning auditor for coding agents: 11-step protocol.
View on GitHub →Process-integrity enforcer: did you do the work thoroughly?
View on GitHub →Adversarial design director for coding agents: taste.md + taste.memory.
View on GitHub →Tools & utilities
Date-aware failure troubleshooting, a self-hosted plugin marketplace, and a non-blocking pre-deploy blessing hook.
Date-aware error-troubleshooting hook: researches the current fix, double-validated.
View on GitHub →Portable socx engine + adapters bringing date-aware troubleshooting to any CLI.
View on GitHub →Self-hosted Claude Code marketplace for the dark-pattern hook plugins.
View on GitHub →Non-blocking pre-deploy blessing hook: deploy journal + prod-vs-preview.
View on GitHub →Curated lists & research hubs
The catalog mapping impossible-task classes to dishonest defaults and their candidate hooks — research and methodology, not an installable tool.
Discovery catalog mapping impossible-task classes to dishonest defaults and hooks.
View on GitHub →Open source, all of it
Every tool is Apache-2.0 and drop-in — no SaaS, no telemetry, jq/python3-only deps. Browse the source, file an issue, or send a PR.
23 repositories · Apache-2.0 · Anthropic-native