opencode-continual-learning

Automatically and incrementally keeps AGENTS.md up to date by mining the current session's conversation for high-signal learnings.

Features

• Automatic learning: after a configurable number of completed turns and elapsed time, the plugin silently injects a prompt that tells the AI to run the continual-learning skill and update AGENTS.md
• /continual-learning command: manually trigger a learning pass at any time
• AGENTS.md-aware updates: the AI reads existing entries and updates them in place rather than appending blindly, keeping the file clean and deduplicated
• Noise-resistant: only high-signal, reusable information is written—recurring preferences and durable workspace facts only; one-off instructions and transient details are excluded
• Configurable cadence: tune the trigger thresholds via environment variables
• Standalone skill + plugin automation: the canonical SKILL.md lives in skills/continual-learning/ for npx skills installs, and the plugin mirrors it into .opencode/skills/ on first load for OpenCode

Requirements

• OpenCode v1.2.15 or later

Installation

Standalone skill (npx skills)

Install the reusable skill into any supported agent:

npx skills add https://github.com/baradghimire/opencode-continual-learning --skill continual-learning

OpenCode plugin (automation)

Add to your OpenCode config (~/.config/opencode/opencode.json):

{
  "plugin": ["opencode-continual-learning"]
}

Restart OpenCode and you're ready to go.

Optionally, pin to a specific version for stability:

{
  "plugin": ["[email protected]"]
}

Usage

The plugin runs silently in the background. After the cadence threshold is met (default: 10 completed turns and 120 minutes since the last run), it injects a prompt that asks the AI to invoke the continual-learning skill. The AI reads AGENTS.md, mines the session conversation, and writes back only:

• ## Learned User Preferences — recurring corrections and stated preferences
• ## Learned Workspace Facts — durable facts about the project (patterns, conventions, tech choices)

To trigger manually, run /continual-learning (the skill is available as a slash command).

What gets written to AGENTS.md

The AI only stores items that meet all of these criteria:

| Criterion | Example | |---|---| | Actionable in future sessions | "Always use pnpm instead of npm" | | Stable across sessions | "The API layer lives in packages/api/src/" | | Repeated or stated as a broad rule | User has corrected the same thing multiple times | | Non-sensitive | No credentials, personal data |

Items that are never stored: secrets, one-off task instructions, transient details (branch names, commit hashes, temporary error messages).

Configuration

Cadence (environment variables)

| Variable | Default | Description | |---|---|---| | CONTINUAL_LEARNING_MIN_TURNS | 10 | Minimum completed turns before triggering | | CONTINUAL_LEARNING_MIN_MINUTES | 120 | Minimum minutes since last run |

State file

Per-project cadence state is stored at:

<project>/.opencode/state/continual-learning.json

Delete this file to reset the turn counter and timer.

Skill file

The plugin checks for the skill in this order:

1. Global: ~/.agents/skills/continual-learning/SKILL.md — if present, skips local copy
2. Project-local: <project>/.opencode/skills/continual-learning/SKILL.md — created only if global doesn't exist

If you install the skill globally (via npx skills add), the plugin won't create project-local copies. This keeps projects clean when you have the skill available globally.

Contributing

git clone https://github.com/baradghimire/opencode-continual-learning
cd opencode-continual-learning
bun install

Symlink the plugin to your OpenCode config for local development:

mkdir -p ~/.config/opencode/plugins
ln -sf "$(pwd)/src/plugin.ts" ~/.config/opencode/plugins/continual-learning.ts

Run the type checker:

bun run typecheck

Releasing

This repository uses an automated release flow with minimal manual steps.

Workflow Overview

| File | Trigger | Purpose | |------|---------|---------| | ci.yml | PRs + pushes to main | Type checking and validation | | release-please.yml | Push to main | Opens release PR with version bump and changelog | | release.yml | release.published | Publishes to npm with provenance |

Step-by-Step

1. Open a feature branch from main
2. Make changes, commit with Conventional Commits
3. Push and open a PR — ci workflow runs typecheck
4. Merge to main — requires PR review (direct pushes blocked by ruleset)
5. release-please creates/updates a Release PR — includes version bump and changelog
6. Review and merge the Release PR — creates a v* tag and a published GitHub Release
7. release workflow runs automatically — builds, attests, uploads tarball, and publishes to npm with provenance

Direct pushes to main are blocked

The repository has a ruleset enforcing PR-based changes with required review. All changes must go through pull requests.

GitHub Actions Details

• ci: Validates TypeScript on every PR and push to main
• release-please: Uses googleapis/release-please-action to manage releases. Parses Conventional Commits to determine semver bumps.
• release: Publishes to npm with --provenance flag. Uses Bun for dependencies, npm for publish (npm provenance not yet available in Bun).
• oc-zen-free: Reusable workflow that selects free OpenCode models first, falling back to opencode/kimi-k2.

Required Secret for Full Release Automation

Set a repository secret named RELEASE_PLEASE_TOKEN (PAT or GitHub App token with repo/workflow permissions).

Using the default GITHUB_TOKEN can prevent downstream PR-triggered workflows from firing on release-please-created PRs.

Manual Release Trigger

You can manually trigger release-please from the GitHub Actions UI if needed:

1. Go to Actions → release-please → Run workflow
2. Select branch main and run

Attribution

This plugin is inspired by and draws directly from:

• cursor/plugins — the continual-learning concept, cadence logic, SKILL.md format, and AGENTS.md output contract. Key inspiration at commit 9c39b57:

  • continual-learning/skills/continual-learning/SKILL.md — skill definition and workflow
  • continual-learning/hooks/continual-learning-stop.ts — cadence state management logic

• joshuadavidthomas/opencode-handoff — this repository was forked from Josh Thomas' work. It provided the OpenCode plugin architecture patterns, @opencode-ai/plugin and @opencode-ai/sdk usage, and the event hook structure. Key inspiration at release v0.5.0 (commit 2a9b523).

License

opencode-continual-learning is licensed under the MIT license. See the LICENSE file for more information.

opencode-continual-learning is not built by, or affiliated with, the OpenCode team.

OpenCode is ©2025 Anomaly.