🍯 honey-stack

An always-on workflow layer for Claude Code and Codex that learns from every session.

honey-stack combines the best of Superpowers, Ponytail, Matt Pocock's skills, pstack and gstack in one small layer, then adds RSI (recursive self-improvement): it learns from your corrections and can apply validated improvements on its own.

• Small: ~2 KB of always-on rules, not ~25 KB of stacked prompts.
• Disciplined: sizes each task, makes the smallest complete change and proves it before "done".
• Self-improving: turns your corrections into rules and evaluated skill patches.
• Local and reversible: Honey state stays in ~/.honey/; honey uninstall restores everything. External history integrations are optional.

Quick start

Needs Claude Code and/or Codex, python3 3.9+ and git on macOS or Linux.

1. Install. This clones the repo and shows a dry run; then run the install command it prints.

curl -fsSL https://raw.githubusercontent.com/umararshad4/honey-stack/main/install.sh | bash

2. Activate. Claude Code: start a new session or run /reload-plugins. Codex: open /hooks and trust the honey entries.

3. Check. Run honey doctor. Replies now start with 🍯 HONEY · direct.

Manual and plugin-only installs, hook troubleshooting and the hooks unverified label are covered in the setup guide.

Everyday use

Work as usual: honey sizes each task and picks the right skill. To steer it, send one of these as the whole message:

Modes change how much is explained, not how carefully work is done. Sizing, verification and safety apply in every mode.

Skills

Flow: grill → plan → execute → verify → review → ship. Bugs start at debug.

Call a skill with /honey:<skill> in Claude Code or $<skill> in Codex. See docs/skills.md for focused workflows and browser QA.

Delta: pick up where you left off

Honey integrates with Delta to find decisions in earlier conversations and carry their context into a new session. Use it to recover a decision, check what was already tried, or attach conversation history to a handoff. Honey keeps working normally without Delta.

Install Delta

Install the optional Delta app and CLI on each computer where you want to use history. On macOS or Linux:

curl -fsSL https://delta.dev/install.sh -o delta-install.sh
sh delta-install.sh
delta auth login
delta --version
honey delta status

Use Delta 0.22.0 or newer. If it is already installed, check for updates with delta update --check and install an available update with delta update. The official installation guide covers other platforms.

Use your history

Ask for /honey:delta in Claude Code or $delta in Codex, or run these commands in your terminal:

Replace THREAD_ID with an ID returned by search, or use a Delta join link. Reads start with one message so short conversations work. Choose a larger range when needed, up to 20 messages per call. History results are JSON and default to at most 12,000 characters; shortened results say truncated: true.

For handoffs, Honey can attach a verified thread reference, the current Git state and test-report fingerprints to your local summary. Restoring it checks what changed and keeps the summary usable if Delta is unavailable. See the Delta guide for attach, restore, readiness fields and troubleshooting.

Record new conversations, optionally

History lookup works in both Claude Code and Codex. To record new Claude Code sessions, install Delta's separate official plugin:

claude plugin marketplace add zed-industries/delta-claude
claude plugin install delta@zed-industries

Start a new Claude session or run /reload-plugins. Then use /delta setup in your repository to choose manual, private or organization capture. Honey can preview the manual policy with honey delta setup --mode manual; only an explicit --apply writes it. Native Codex session recording is not included.

Privacy: installing Honey or reading history does not enable capture. Private sync uploads conversation and checkout data to Delta's servers. Organization capture shares it with that organization. Honey keeps history separate from RSI feedback and fresh verification evidence. See Delta's capture guide before enabling uploads.

RSI: how it learns

1. Observe: hooks log redacted corrections, "still broken" moments and which skills fired.
2. Reflect: after a session with signal, a small model (Haiku, ~$0.002) proposes evidence-backed WHEN → THEN rules.
3. Curate: a rule goes live once two sessions back it, goes stale after 30 days and is archived after 90.
4. Apply: learned rules load automatically. Instruction edits, skill patches and guards wait for /rsi review, or apply after validation when autonomy is on.

In Codex, use $rsi …. Tiers, budgets, autonomy and GitHub publishing are covered in docs/rsi.md.

Configure

All keys in ~/.honey/config.json are optional:

{
  "mode": "direct",
  "autoupdate": true,
  "budgets":  {"enabled": true, "reflection_usd": 0.15, "weekly_usd": 1.5, "monthly_usd": 5},
  "autonomy": {"min_sessions": 2, "holdout_rate": 0}
}

Privacy: Honey has no telemetry and stores redacted learning snippets of up to 300 characters locally. Explicit Delta lookups read your separate Delta account. Capture is opt-in; private Delta synchronization uploads conversations and checkout data to its servers. Honey's history reads and handoff metadata do not upload sessions or import history into RSI.

Update and uninstall

honey checks for a new release at session start, at most once an hour.

honey update                                   # update now; then /reload-plugins in Claude
honey uninstall --dry-run && honey uninstall   # keeps ~/.honey; add --purge to delete it

Docs

Documentation site: honey-stack.vercel.app — every feature, explained with interactive diagrams. The guides below are the in-repo source.

MIT © King Umar · Changelog · Credits and licenses