Espresso

Less filler. Less unnecessary context. Right-sized delegation.

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

Mechanism	What it changes	What we can honestly claim
No ordinary-prompt reminders	Removes Espresso's repeated reminder payload	Eliminates that input overhead compared with Espresso 1.x; total session savings are unmeasured
Concise policy	Asks for direct answers and the smallest correct change	Aims to reduce unnecessary prose and code; output length and quality remain model-dependent
Optional RTK	Filters supported command output before it enters context	Can reduce tool-output volume; RTK reports bytes removed and estimated tokens, not total session or subscription savings
Model ladder	Sends suitable bounded work to a lower model tier	Changes which model does the work, not the number of tokens; delegation adds overhead and is not a quota-saving guarantee

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. See RTK's measurement definition.

Smaller code: optional Ponytail

Ponytail is a separate, optional companion for avoiding over-engineering: reuse existing code, prefer the standard library and native platform features, and avoid speculative abstractions. Espresso neither bundles nor installs it automatically.

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 or evidence for Astra/Sol. 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

Harness	Policy	Delegation
Claude Code	SessionStart hook and skill	Automatic ladder for unpinned generic Claude workers
Codex	Trusted SessionStart hook and skill	Explicit named GPT agents
OMP	Extension and skill	Named GPT agents and an explicit text-only worker
Pi	Extension and skill	Explicit text-only worker; core Pi has no subagent tool

GPT tiers: Astra → Sol → Terra → Luna, with Luna as the floor. Named agents are espresso-sol, espresso-terra and espresso-luna. OMP workers use the subscription provider openai-codex; Codex agents inherit their provider. The OMP/Pi text worker checks model availability on the same provider before launching. Explicit model choices and specialist agents are preserved. Keep consequential review on the parent model; do not delegate trivial work or spawn workers without authorization.

Install

Node.js is required for hooks and the adapter installer.

Claude Code (run each command separately):

/plugin marketplace add mirkobozzetto/espresso
/plugin install espresso@espresso

Codex:

codex plugin marketplace add mirkobozzetto/espresso
codex plugin add espresso@espresso

Review and trust the hooks in Codex before starting a new thread.

Pi: pi install git:github.com/mirkobozzetto/espresso

OMP, or named GPT agents for Codex, from a permanent checkout:

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

For Codex's named GPT agents, run node src/hooks/install.js codex --apply. Pi can alternatively use node src/hooks/install.js pi --apply for the extension and skill. Omit --apply to preview adapter files. Conflicting files are refused; settings, credentials and model overrides are untouched. Restart the harness. Keep this checkout in place when using a local extension reference.

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

The worker launches the same CLI on a lower available tier, 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 lower-tier workers, 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 or Astra. 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 covered reminder removal, absence of startup writes, explicit-model preservation, installer idempotence and conflict refusal, toggles and selective RTK rewriting. Real OMP and Pi text workers returned replies using the selected Luna CLI configuration. Claude's manifest and direct hooks were checked; Codex's plugin installation was checked in an isolated configuration. Full Claude/Codex delegation sessions and comparative token/quality benchmarks were not run.

MIT License.