cometweb-io/cometweb-agent-skills
v2.0.1MIT
CometWeb Labs specialist skills for research, product operations, quality workflows, adversarial review, QA, release readiness, and evidence-based decisions.
Changelog
All notable changes to the CometWeb Agent Skills bundle are documented here.
The format follows Keep a Changelog; repository
tags use vMAJOR.MINOR.PATCH.
[Unreleased]
2.0.1 - 2026-09-22
Added
- Integrated the quality workflow package set: artifact acceptance, briefing, content writing/review, repair, benchmark and rubric design, feedback integration, quality-loop operations, skill auditing, and skill evaluation.
- Added dedicated
content-roaster,science-roaster, andrepo-roasterpackages from Roaster Suite 6.0.0, with the quality bundle's duplicate copies intentionally excluded so each skill ID has one canonical version. - Expanded the canonical marketplace bundle from 18 to 32 skills and refreshed registry metadata, host adapters, routing docs, package manifests, and context budget baselines.
Verification
- The imported packages retain their own references, deterministic helpers, eval fixtures, and tests; the source bundles' root CI and release tooling are not copied into this repository.
- Static package integrity, public-tree safety, registry consistency, and the full local validation suite remain release gates; static checks do not claim live-host behavioral lift.
2.0.0 - 2026-09-19
Added
ebook-publisher,longform-publisher,portfolio-operator, andcometweb-context, bringing the canonical bundle to 18 skills.- Cross-runtime installation and generated adapters for Cursor, Claude Code, Codex, Qwen, Qoder, and Lingma.
- CW-AIP v2 schemas and validators for evidence, decisions, releases, reviews, and multi-skill handoffs, while retaining the documented v1 compatibility surface.
- Approval-gated deterministic skill packages, reviewer bundles, public-tree safety scanning, and history scanning with no bypass flag.
- A 12-gate local validator covering repository integrity, generated adapters, routing, behavior, context budgets, eval strength, security, compilation, and the full test suite.
- Context-budget baselines that fail on unreviewed front-door growth without pretending there is one universal token ceiling.
- Mutation-style eval-strength baselines that measure whether each deterministic rule is actually held by a failing case rather than counting green examples.
tooling/new_skill.py, package-layout contracts, registry consistency checks, executable-bit checks, documentation indexes, issue forms, contributor and security policies.- A deterministic
tooling/whykit_draft.pyboundary that validates finalized CW-AIP v2 evidence and decision envelopes and emits explicit, unreviewed WhyKit drafts without mutating a vault or allocating ledger IDs.
Changed
CometWeb-io/agent-skillsis the single public canonical source. The former private-source/public-mirror workflow and its sync tooling are retired.- Tracked context bindings use public placeholders; local paths belong only in
ignored
*.local.jsonand*.local.txtoverlays. registry/skills.jsonremains the metadata source of truth; marketplace manifests, host adapters, compatibility tables, and generated documentation are checked against it.- Large skill front doors were deduplicated against their own references while preserving binding safety rules and execution order.
- CI remains one Python 3.12 validation job. The duplicate PR-only workflow and recurring dependency bots are removed to avoid repeated runner cost.
- GitHub Actions are pinned to reviewed commit SHAs; full history safety remains a local pre-release gate rather than an every-push workflow.
Fixed
- Local validation now passes under the declared pytest import mode; shared test fixtures no longer depend on importing sibling test modules.
- Routing metadata and examples no longer send long-form publishing work to release readiness or point the orchestrator at a script in another package.
- Context scoring and freshness defaults use UTC rather than the runner's local timezone, and Council recency no longer uses naive UTC datetimes.
- All documented contributor commands now map to tools that actually exist.
- Qwen, Qoder, and Lingma installers retain executable file modes in Git.
- The Codex installer no longer deletes an existing real skill directory; all host installers fail safely until the user moves a collision aside.
- Missing package icons, dead readiness locals, incomplete exception chaining, and low-visibility subprocess coverage were corrected.
- Package case-collision checks run on case-insensitive macOS filesystems instead of skipping the risk they exist to detect.
- The retired
ai-antipattern-writingpackage cannot reappear through registry, adapter, or legacy-package synchronization. - CW-AIP v2 validation remains complete without the optional
jsonschemapackage, loads typed validators without module-name collisions, and rejects ambiguous JSON before rendering a WhyKit draft. - WhyKit draft output treats producer prose as inert data, preserves decision approval metadata in provenance, and refuses to overwrite existing files.
Verification
- The full deterministic test suite passes with zero failures, errors, or skips.
- All 12 local validation gates pass.
- Current-tree and full-history public-safety scans pass.
- Deterministic routing, package, adapter, context-budget, and eval-strength checks are part of the release candidate.
1.0.6 - 2026-08-26
Added
- Envelope validator for multiagent handoffs (
validate_envelope.py). - Multiagent smoke walkthrough and demo Evidence Pack.
Changed
- Multiagent documentation now covers cloud subagents and step validation.
1.0.5 - 2026-08-26
Added
scripts/install-codex.shinstalls all skills into Codex.
1.0.4 - 2026-08-26
Added
skill-orchestrator-multiagentfor isolated specialist execution.
Changed
- Routing distinguishes single-thread and multiagent orchestration.
1.0.3 - 2026-08-26
Added
skill-orchestratorfor multi-skill CW-AIP workflows.
1.0.2 - 2026-08-26
Added
scripts/install-claude.shinstalls all skills into Claude Code.
Changed
- Installation and architecture documentation covers Cursor, Claude Code, and ChatGPT-compatible hosts.
1.0.1 - 2026-08-26
Added
- Branded demo, root installation guide, issue/PR templates, and release-bundle support files.
1.0.0 - 2026-08-25
Added
- Twelve public skills with deterministic kernels and CI.
- CW-AIP v1 interchange protocol and JSON Schemas.
- Routing evaluation suite, validators, installers, safety checks, and release packages.