Skip to content

oliverweissl/context-garden / skills/seedbank

seedbank

Verified

Maintain a small, generated AGENTS.md of repository facts (build commands, conventions, invariants, past mistakes) that were rediscovered often or were costly to get wrong. Use at task start to check cached facts, and at task end to log and promote what you had to dig for.

references/guide.md

View on GitHub
# seedbank: operational details

## Automatic observation (Claude Code hook)

The context-garden plugin's `hooks/hooks.json` runs `bin/seedbank hook`
asynchronously after every `Read`, `Grep`, `Glob` and `Bash` (silent; records
into the git repo holding the touched path; a crash shows as a non-blocking
hook error with its traceback, it never blocks the tool call). If you copied just this skill, add to
`.claude/settings.json` (or `~/.claude/settings.json`):
```json
{"hooks": {"PostToolUse": [{"matcher": "Read|Grep|Glob|Bash",
  "hooks": [{"type": "command", "command": "<this-skill-dir>/bin/seedbank hook",
  "async": true, "timeout": 5}]}]}}
```

## End-of-task reminder (Stop hook) and the gardener agent

`bin/seedbank hook stop` runs when Claude is about to stop. Once per
session, and only if this session's observations produced candidates above
`promote_threshold`, it blocks the stop with a reason naming the top
candidates, so Claude either curates them or decides not to and finishes.
It never fires while Claude is already continuing because of a stop hook.
Sessions already reminded are kept in `.seedbank/nudged.json` (local).

The `gardener` agent (`agents/gardener.md`, shipped with the plugin)
preloads this skill and does the curation in its own context: review
candidates, promote with compact text, compile. Copied skill: put the file
in `.claude/agents/` and add
```json
{"hooks": {"Stop": [{"hooks": [{"type": "command",
  "command": "<this-skill-dir>/bin/seedbank hook stop", "timeout": 10}]}]}}
```

## Observing reads and searches by hand

Without the hook:
```
<this-skill-dir>/bin/seedbank observe read <path> --task "<why you read it>" [--scope <topic>]
<this-skill-dir>/bin/seedbank observe search "<pattern>" --task "<why>" [--scope <topic>]
```
Example mistake/fact logs:
```
<this-skill-dir>/bin/seedbank observe mistake "Do not edit src/generated/**; regenerate with tools/codegen.py." \
  --scope invariants --source src/generated/some_file.cpp
<this-skill-dir>/bin/seedbank observe fact "Never weaken convergence tolerances to make tests pass; tolerance is part of the numerical accuracy contract." \
  --scope invariants
```

## Tiers

- **Hot**: inlined in `AGENTS.md`, budget 500 tokens by default.
- **Warm**: `.seedbank/warm/<scope>.md`, only pointed to from `AGENTS.md`;
  open one only when the task is about that topic.
- **Cold**: not stored; the repository itself.

## Import

`import` of a markdown file seeds bullets as lower-confidence candidates;
nothing is promoted until you review them. A JSON bundle from `export`
(another clone's curated facts) is imported as facts, only into the same
project (fingerprint match) and never over an existing fact; see
`references/schema.md`.

## In the target repository

Commit `AGENTS.md` and `.seedbank/warm/*.md`. A new store writes
`.seedbank/.gitignore` ignoring local profiler state (everything except
`warm/`); commit it too, and edit it only if the team decides to share
observation history. An existing `.seedbank/.gitignore` is never
overwritten.

Also commit `CLAUDE.md`: Claude Code reads `CLAUDE.md`, never `AGENTS.md`.
When the repo has none, `compile` creates it as the one line `@AGENTS.md`;
the import avoids a second copy. An existing `CLAUDE.md` is never touched,
`compile` only hints to add that line. A symlink also works but fails on Windows without
admin/Developer Mode.
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 472kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall