Maintiary - Offfold

================================================================================
                              M A I N T I A R Y
                        Your Quiet AI Co-Maintainer
                          Offfold catalogue / No. 01
================================================================================

PRODUCT STATUS: Version: 1.1.1 | Runtime: PYTHON 3.11+ | Model API: NONE | Storage: LOCAL SQLITE

1. What Maintiary is

Maintiary is one portable Agent Skill for contextual project completeness and durable maintenance memory. Your existing coding model judges what matters; a Python standard-library core collects evidence and validates history.

It writes only the managed project's .maintiary/ directory and its designated global location index. No additional model account, API key, Python package or background service is required.

2. How an assessment runs

  check        collect deterministic evidence: Git state, file hashes, sources
    |
  response_draft   the host model reads the sources and drafts the assessment
    |
  assess       structure the answer, preserve earlier decisions
    |
  validate     verify handbook integrity and stored history

A successful assessment stores a structured profile, supported suggestions, uncertainties, real Git evidence and source hashes. Zero concerns is a valid result. A scan alone is not a completed AI assessment.

3. Capabilities

Capability	Behaviour
Project memory	A multi-file handbook keeps the profile, latest report, historical assessments, paged notes, persistent decisions and archive separate. Exact quotes and complete event histories survive re-entry.
Completeness review	Contextual gaps are reported with the evidence behind them, not as a generic checklist. Unchanged evidence and handbook-only writes are suppressed.
Quiet operation	Activation is explicit, using generated configuration inside `.maintiary/`. File events do not each invoke a model. No daemon or scheduler is installed.
Decision durability	Resolution requires changed related evidence and explicit verification. A deferred concern needs a material-evidence explanation before it is reconsidered.
Portability	Complete bundles support `export`, `restore` and `merge`. Divergent histories keep both branches under `.maintiary/conflicts/` instead of overwriting.
Recovery	SQLite is local authority. `doctor --repair` rebuilds readable projections from a healthy database; `doctor` reports each transport and analysis layer distinctly.

4. Host support

Verified on Windows. The Skill installs into each host's documented Skill directory; activation uses generated configuration inside the managed project.

Host	Skill directory	Verified version	Normal boundary
Codex CLI	`~/.agents/skills/`	0.147.0	Cannot wake a closed or idle Codex process
Claude Code	`~/.claude/skills/`	2.1.284	Cannot wake a closed host
OpenCode	`~/.config/opencode/skills/`	1.18.30	Host must already be running

5. Requirements & installation

Python 3.11+, Git, and a host able to read Skills and execute local tools.

# install from a source checkout
python tools/distribution/install.py --destination "$USERPROFILE/.agents/skills/maintiary"

# operate the installation directly
python "$Skill/scripts/maintiary.py" init     --root PATH
python "$Skill/scripts/maintiary.py" check    --root PATH
python "$Skill/scripts/maintiary.py" assess   --root PATH --input PATH/.maintiary/inputs/assessment.json
python "$Skill/scripts/maintiary.py" validate --root PATH

The host reads relevant project sources and saves the filled response_draft under .maintiary/inputs/. Actual time, Git status and identifiers come from the scripts; a source hash proves which bytes were used.

6. Boundaries

No cloud: no synchronization service and no automatic divergence reconciliation.
No model client: the runtime never calls a model API; the host supplies inference and tools.
No background process: nothing is installed as a daemon or scheduled job.
Unverified surfaces: hosts other than the three listed, other operating systems and desktop/IDE surfaces remain unverified.
Local data: notes and bundles are unencrypted local files. Avoid recording credentials.

7. Contact

- Email: [email protected]
- Catalogue: Offfold