cfdoctor v0.3.0

Highlights

• Add approval-gated, read-only Wrangler snapshots for Worker and Pages deployed state.
• Default snapshots to metadata-only collection; source/config download requires the separate --include-source-config opt-in.
• Require an already installed project-pinned Wrangler executable—no package runners, installs, or Cloudflare mutation commands.
• Capture every traffic-bearing Worker version, bindings/runtime limits, deployment metadata, and secret names/types in private hashed manifests.
• Add targeted account-read guidance for DNS, rulesets/WAF, Access, R2, Queues, usage, and billing without introducing a universal account collector.
• Harden scanner 0.3.5 with bounded imported/computed Durable Object keys, per-Queue consumer matching, alarm-condition precision, linked Stream preload checks, indirect self-fetch detection, observability sampling, webhook idempotency, modern Wrangler product inventory, and generated-corpus exclusion.
• Add current pricing-source bundles, conflict handling, stronger evidence provenance, 29 claim-ledger records, and first-party Cloudflare reliability postmortems.
• Strengthen semantic/adversarial eval oracles for approval, no-install behavior, Static Assets, package-runner rejection, and identifier minimization.

Evaluation

A matched three-way GPT-5.5 evaluation used the same manifest/oracles and three fresh interleaved runs per visible case: 279 answers and 243 primary judgments.

Metric	v0.3.0 skill	Previous GitHub main	No skill
Objective	89.15%	81.79%	72.97%
Combined	89.69%	83.00%	69.24%
Judge mean	0.9507	0.9333	0.7684

The Wrangler slice scored 93.01% versus 63.96% on previous main and 60.53% without the skill. Legacy behavior remained statistically compatible with main. Pricing guidance improved over main but did not establish lift over no skill and remains an explicit efficiency target.

Validation

• 60/60 offline unit tests
• 39/39 trigger cases
• 25/25 detection fixtures
• 14 Wrangler subprocess-boundary tests
• 60-check scanner registry / 66 coverage rows
• 29 evidence records / 23 scenarios / 25 fixtures
• Skill Eval Harness v0.6.0: 37 cases / 6 materialized ablations
• Clean static self-scan and installable-skill boundary
• npm pack --dry-run: 21 release files; no eval, research, docs, tests, or repository scripts in the package boundary
• Both GitHub Actions workflows passed on the release commit

Install

pi install https://github.com/adewale/cfdoctor@v0.3.0

See CHANGELOG.md for the complete change list and disclosed limitations.

cfdoctor v0.2.2

Changes

• Move the runtime Cloudflare Doctor skill from repo root to skills/cloudflare-doctor/.
• Keep only SKILL.md, runtime references, and the static scanner inside the installable skill boundary.
• Keep evals, research, docs, fixtures, and saved results repo-only.
• Update Pi/package metadata, install commands, eval manifest paths, scanner paths, and CI.
• Add install-boundary CI.

Validation

• GitHub CI passed: existing validate job and install-boundary workflow.
• Local: npm test.
• Local: python3 ../skill-eval-harness/skill_benchmark.py validate evals/shared-benchmark.json.
• Local: npm pack --dry-run --json confirmed no eval/research/docs artifacts in the package bundle.

v0.2.1 — Multi-agent Agent Skill support

Multi-agent Agent Skill support.

• Adds Agent Skills compatibility metadata for Codex, OpenCode, Pi, Gemini CLI, and Claude Code.
• Documents per-agent install/use paths in the README.
• Adds or updates package.json skill and pi.skills metadata for Git-installed packages.

v0.2.0

Highlights

• Added optional dead cross-boundary RPC audit path for TypeScript DurableObject, WorkerEntrypoint, WorkflowEntrypoint, RpcTarget, and Agents SDK classes.
• Added scanner lead CFDOC-REL-CROSS-BOUNDARY-RPC-DEAD, a detection fixture, trigger coverage, and gated deadlint guidance.
• Added stable scanner check IDs, JSON scanner output, fixture-backed detection evals, coverage-matrix consistency checks, and CI validation for clean self-scans.
• Added v0.3 shared eval quality oracles and updated shared benchmark guidance.
• Refreshed README usage modes, validation commands, recipes, lessons learned, official source maps, and Wrangler config guidance.

Validation

• CI passed on 2ec391c.
• Local npm test passed.
• Local git diff --check -- . ':(exclude)evals/results' passed.

Install

pi install https://github.com/adewale/cfdoctor@v0.2.0

v0.1.0

Highlights

• Created the public cfdoctor repository.
• Added MIT license metadata and root LICENSE.
• Added CI validation for JSON manifests, Python helpers, trigger evals, self-scan, and diff whitespace.
• Clarified that wrangler.jsonc is Cloudflare's recommended config format for new Wrangler projects while JSON and legacy TOML remain supported audit inputs.
• Expanded README and recipes to explain the Python scanner as a read-only triage layer, not proof or a replacement for sourced audit judgment.
• Added Lessons Learned documentation for eval contract coverage, Wrangler config examples, scanner scope, fixture self-scan noise, and Pi skill bundling.
• Added the Cloudflare Doctor Agent Skill (SKILL.md) with source-backed audit workflow and safety rules.
• Added read-only static scanner for Cloudflare repos.
• Added deterministic trigger evals with current proxy result: 38/38 = 100%.
• Added reference docs for Cloudflare product fit, cost footguns, performance/reliability, account evidence sharing, official docs sourcing, and war-story-derived scenarios.
• Added Coey Durable Objects, Dynamic Workers, Agents SDK, Artifacts, logging, and TCP/database scenario coverage.