blineadam/agent-plan-and-track

Portable, user-scoped agent rules, skills, and hooks for Claude Code, GitHub
Copilot, and Codex. They stay active through a session and carry over into
the next one.
Uses a .tasks/todo.md to plan non-trivial work, logs your corrections in
.tasks/lessons.md, and verifies changes before calling anything done.
These rules stay in place even as a session's context grows.
Best for iterative work in an existing repo: features, bug fixes,
refactors, anywhere a persistent plan, a completion gate, and a growing
lessons file pay off over time.
Why this exists
Global instruction files (CLAUDE.md, AGENTS.md, copilot-instructions.md)
sit at the top of the context window, and in a long chat the model's
attention to them drifts: it starts forgetting your rules. The fix is
matching each kind of rule to a mechanism that keeps it alive:
| Rule type | Mechanism | Why it sticks |
| --- | --- | --- |
| Short, constant constraints | Instructions file | Always loaded and kept short |
| Procedures | **Skills** | Loaded near the end of context when triggered |
| Core rules | **Hooks** re-injecting a digest | Repeated by the harness throughout the session |
This repo checks its own enforcement. skill-comply measures whether a rule
is followed, while skill-activation measures whether the right skill fires.
Both have caught real gaps in past runs.
Supporting three harnesses accounts for most of the installer's bulk. At
runtime, each session uses only its own harness integration.
What you get
The skills used most often are:
- plan-and-track plans and tracks non-trivial work.
- gateguard asks for concrete facts before the first edit to a file.
- capture-lesson turns corrections into durable rules.
- humanizer cleans up longer user-facing writing before it ships.
Other skills cover autonomous work, docs-first research, maintenance, and
delegation to subagents.
See docs/skills.md for the full catalog, including the
maintenance and design skills. See docs/models.md for model
defaults, tiered subagents, and per-tool notes. See
docs/credits.md for the upstream projects this repo adapts
from.
Install
The hooks run on Node.js; the installer checks for it before touching
anything. The macOS/Linux installer also needs jq for the Claude and Codex
targets (not required for Copilot-only). If jq is missing, the installer
offers to install it through your platform's package manager: an interactive run
prompts for confirmation, while a non-interactive run requires PT_INSTALL_JQ=1
to authorize it. See docs/installers.md#dependencies for
the full probe order and behavior.
On Windows, run the PowerShell installer instead (same targets):
Re-running is safe: each run updates the files this repo manages and leaves
your additions alone. A skill or subagent this repo has renamed or removed is
moved into .plan-and-track-pruned/ rather than deleted. The exact ownership
rules are in the install.sh header.
To update, run git pull and reinstall. Tagged releases (vX.Y.Z on the