v0.1.0 — ProDoc Guard

ProDoc Guard v0.1.0 — Release Notes

Version 0.1.0 of ProDoc Guard, dated 2026-10-15.

---

Project memory that lives in Git, not in an AI's head.

ProDoc Guard keeps three canonical documents honest across AI-assisted sessions:

Document	Answers
PRODUCT_REQUIREMENTS.md	What is this, for whom, and what must it do?
DECISION_LOG.md	What was decided, why, and what was rejected?
PROJECT_STATUS.md	What is done, tested, blocked, and next?

What's in 0.1.0

An Agent Skill (SKILL.md) written to the open Agent Skills specification. It
instructs an assistant to read all three documents before substantial work, create
anything missing, record decisions, update status, and report the files changed,
the commit SHA, the current state, and what remains outstanding.

A standalone CLI (scripts/prodoc), Python standard library only, no model and
no API key. Five commands, JSON output, stable exit codes:

python3 scripts/prodoc init                        # create missing documents from templates
python3 scripts/prodoc validate                    # structure, sections, decision records, evidence
python3 scripts/prodoc status                      # repository and document state
python3 scripts/prodoc check                       # history integrity, evidence, secrets
python3 scripts/prodoc update --phase "In Progress" # refresh observed Git metadata

It refuses to guess. A freshly initialized project fails validate on purpose,
naming every piece of missing information. A work item marked Tested without
cited evidence is an error. A release-state claim with no recorded verification is
an error. A release with no remote still works and says so.

One portable artifact, no vendor plugins. Specification fields only in the
portable SKILL.md; client differences are install paths and, at most, a
documented local edit. No vendor SDKs, no forks, no marketplace packages.

A compatibility matrix that separates Specification-conformant from
Runtime-tested, and states which claims were actually executed.

Verified in this release

Check	Result
Test suite	159 total, 158 passed, 0 failed, 0 errored, 1 skipped (the opt-in official-validator test), on CPython 3.8.18 and 3.12.15 in CI
Document gates	3/3 (skill-spec, documents-validate, documents-check)
Official Agent Skills validator (agentskills validate)	Valid skill, exit 0
Functional scenarios in disposable repositories (python3 tests/rc1_functional.py)	39/39
Credential scan (three canonical documents, plus all tracked files)	Clean
Bandit static analysis of scripts/	0 high, 0 medium, 2 low (the subprocess-without-shell=True pair)
Dependency manifests	None — standard library only
OpenCode v2.0.21	Runtime-tested: discovery, invocation, CLI execution
Gemini CLI 0.56.0	Runtime-tested for discovery; invocation not confirmed
GitHub Actions (5 jobs)	Green on Python 3.8.18 and 3.12.15 — https://github.com/gitdamnit/prodoc-guard/actions/runs/38839726677

Not verified in this release

Stated plainly, because the alternative is the failure mode this project exists to
prevent.

• Codex, Claude Code, Cursor and GitHub Copilot were never started. They are
  not installed on the machine used for this release. Their rows are
  documentation-and-conformance claims only.
• Windows is implemented but untested. Not run.
• Python 3.9, 3.10 and 3.11 were not executed. 3.8 and 3.12 both run green in
  CI, so the declared 3.8+ floor is verified and the range in between is not.
• macOS was not executed. Nothing in the tool is platform-specific, but it is
  untested.
• No client can be shown to activate a skill automatically. Discovery and
  invocation were exercised where possible; a client may still decline to use a
  skill on any given prompt.

Security posture

• The CLI opens no sockets, reads no credentials, and requires no API key. The one
  exception is the opt-in check_skill.py --official, which fetches the official
  validator with uvx.
• It runs no mutating Git command: never add, commit, push, checkout,
  reset, merge, branch, tag, or stash.
• It never commits, pushes, tags, or publishes. Those stay with the user.
• It refuses to write through a symbolic link or outside the project you point it
  at, and refuses to clobber a file that changed while it was reading it.
• Documents are treated as data, never as instructions.
• prodoc check scans the three canonical documents for credential patterns and
  redacts matches in its output.

Known limits: the secret scan covers only the three canonical documents; the
decision-history check is an integrity check against HEAD, not tamper-proofing
against rewritten history; the concurrent-write window is two adjacent syscalls.

Requirements

Git 2.x and Python 3.8+. Nothing else. Install by copying the folder:

git clone https://github.com/gitdamnit/prodoc-guard.git /tmp/prodoc-guard
mkdir -p .agents/skills
cp -R /tmp/prodoc-guard .agents/skills/prodoc-guard

.agents/skills/ is the cross-vendor path read by OpenCode, Codex, Claude Code,
Cursor, GitHub Copilot and Gemini CLI. The directory must be named prodoc-guard.

Upgrade notes

There is no earlier release, so nothing is broken by upgrading. Users of the
development branch before this release should note three changes: the managed
block is now prodoc:observed with Observed * fields (old markers are accepted
and migrated); ## Verification Evidence is a newly required section; and
prodoc update --json renamed facts to observed.

Licence

MIT — Copyright (c) 2026 Evan Zellner.