espresso

Espresso

Less filler. Less unnecessary context. Same model, right effort.

Lightweight workflows for Claude Code, Codex, OMP and Pi. Keep explanations, evidence and useful code comments, without repetitive reminders or automatic changes to your global settings.

Where the savings come from

Espresso does not enlarge the context window or erase existing history. Its policy also occupies context. We have verified the integrations, not measured an end-to-end token reduction. There is no universal savings percentage.

RTK is off by default in the OMP/Pi extension. The integration only considers simple git status and git log --oneline commands with a small flag allowlist. Diffs, tests, patch logs and scripts stay untouched. Missing RTK or a failed rewrite leaves the original command intact. Existing RTK hooks are independent. Checked with RTK 0.51.0. See RTK's measurement definition.

Smaller code: Ponytail included

Espresso bundles the real Ponytail 4.13.0 runtime and all six upstream skills, with its MIT license and pinned source commit in ponytail-upstream.json. No separate installation or startup download is needed. It starts in full mode unless an existing Ponytail configuration or PONYTAIL_DEFAULT_MODE overrides that default.

In OMP and Pi, /ponytail opens a selector: full, lite, ultra, off or status. /ponytail off disables it for the session; /ponytail default off changes the default for future sessions. Use /ponytail default full to restore it. Headless commands show status instead of opening a menu. Claude/Codex use the bundled skills and lifecycle hooks; they do not get the OMP/Pi native selector.

Also included: ponytail-review, ponytail-audit, ponytail-gain, ponytail-debt and ponytail-help. Use their slash commands in OMP/Pi, plugin skill commands in Claude, and skill invocation in Codex. Espresso and Ponytail have independent switches. Avoid enabling a second, standalone Ponytail extension alongside the bundled one.

Its authors' agentic benchmark reports less code and fewer tokens on 12 feature tasks with Haiku 4.5, four runs per task and condition. Gains are largest where custom code can be replaced by native features, and small or absent where the implementation is already minimal. These are Ponytail's results, not Espresso measurements. The authors also warn that reasoning overhead can increase cost on GPT-5.5.

Use it to reduce unnecessary implementation, not to chase the fewest lines. Shorter code is not automatically better code; validation, error handling, security and accessibility must remain intact.

Harness support

In Claude Code, Espresso ships one agent, espresso:scout: read-only (no edit tools), on the session model (model: inherit), at medium effort, in the background. Claude picks it for research whose raw output would flood the conversation (multi-file search, flow tracing with GitNexus, docs, long logs), and you keep talking with the main thread while it works. Below the prompt, each running agent shows its model, effort and tokens.

One exception: for web-only research (searching and reading pages, no code to read), the lead passes model: haiku to the scout. With the same research brief, Opus 5.5 took 62 s and 55 s, Haiku 5.5 took 55 s and 76 s, and the answers were comparable. Haiku 5.5 costs $0.10 / $0.50 per million input / output tokens, against $4 / $20 for Opus 5.5 (pricing). Reading and reviewing code stay on the session model: in a code-review comparison, Opus 5.5 alone found 122 bugs, Opus 5.5 with Haiku 5.5 workers found 103 and reported 14 false positives.

Espresso no longer routes code work to a cheaper model. With a strong session model, adding agents brings no gain or a loss on sequential coding work (260-configuration study), and cheaper workers can cost more per solved task once retries are counted. Delegate only substantial read-only research or work whose output would flood the main context; keep edits and consequential review on the main thread. Explicit model choices and specialist agents are preserved.

In OMP/Pi, the text-only worker runs on the session model and follows the session thinking level (low on OMP). Use /espresso effort low|medium|high to set it for this session; /espresso effort and /espresso status report it. This setting applies only to /espresso worker, not native named agents. Before delegating, announce the agent, resolved model, effort and reason.

Install

Node.js is required for hooks and the adapter installer. Restart the harness after installing or updating.

In Claude Code, the same commands work as /plugin … inside a session.

Codex runs plugin hooks only once you trust them. Review them in Codex before starting a new thread, and again after an update: a changed hook gets a new hash.

Pi and Prime Agent load the extension, the six Ponytail skills and the Espresso skill from the package. Pi also accepts a checkout path (pi install /path/to/espresso). Keep a checkout in place when you install from it.

OMP needs a permanent checkout:

git clone https://github.com/mirkobozzetto/espresso.git
cd espresso
node src/hooks/install.js omp --apply

The OMP extension imports the checkout, so git pull updates it. The skills are copies: the installer refuses to overwrite a copy that differs, so replace ~/.omp/agent/skills/espresso and ~/.omp/agent/skills/ponytail* with the checkout's skills/ folders after pulling.

For Codex skills without the marketplace, run node src/hooks/install.js codex --apply. Omit --apply to preview adapter files. Conflicting files are refused; settings, credentials and model overrides are untouched.

See INSTALL.md for upgrade and removal precautions. Existing Caveman/Ponytail installations and old global rules are not deleted automatically.

OMP / Pi commands

/espresso status
/espresso on
/espresso off
/espresso worker Summarize this supplied text: ...
/espresso rtk-on
/espresso rtk-off

The worker launches the same CLI on the session model, with no tools or recursive extensions. It works only on the supplied text. State is per session: concise policy starts on, RTK starts off. off stops new policy injection and RTK rewriting; it cannot erase messages already sent to the model.

OMP automatic research mode

/espresso auto enables standing authorization for substantial independent read-only research. /espresso manual returns to explicit delegation; /espresso off also clears automatic mode. These commands change session state. The default is manual. A personal OMP wrapper can opt in on every startup by calling the extension factory with {auto: true} as its second argument. The installer refuses to overwrite such a customized wrapper.

The lead decides whether delegation is worthwhile after initial inspection. Instructions limit it to two workers on the session model, no nested delegation, no edits, and no trivial tasks. These are model instructions, not an enforced scheduler or read-only sandbox. Explicit choices and stricter skill instructions win; skills that require fresh consent may still ask. No workers run while idle.

Anthropic’s research system supports the usefulness of parallel independent research, not a token-saving guarantee for Espresso. Espresso’s automatic policy still needs task-level evaluation; command and policy-transition checks are not proof of reliable delegation decisions.

Arsenal and Ship compatibility

Arsenal owns skill selection and transitions. The selected skill owns approval gates, solo flags, specialized agents, model assignments, output schemas, artifact states, verification and commits. These take precedence over Espresso recommendations. Espresso must not create a competing team or extra reviewer, or use an external worker CLI to bypass a controlled runtime.

Ship currently asks before each delegation: automatic Espresso mode does not remove that requirement. Likewise, --no-agents and solo mode remain solo. Existing workflows do not require Espresso and are not rewritten by it. This is instruction-level coexistence, not a verified end-to-end compatibility claim for every Arsenal skill or harness.

Verification scope

Local checks for 2.2.0: claude plugin validate, every Claude hook run directly (per-project Ponytail modes, mode commands, visible status line, silence on Codex), the OMP/Pi extension against a simulated host (string, array and section prompts, same-model worker), selective RTK rewriting with RTK 0.51.0, and a Codex 0.160.1 install in an isolated configuration listing the three bundled hooks. For 2.3.0: the scout agent passes claude plugin validate, skips the Ponytail injection, and the agent status rows render from sample input. How often Claude picks the scout on its own, and comparative token/quality benchmarks, were not measured. For 2.4.1: the hooks inject the web-only Haiku rule, the scout still skips Ponytail, and claude plugin validate passes. The Haiku web-research timings come from two runs per model on one brief, not a benchmark.

MIT License.