compact-handoff

Stars12

Forks1

UpdatedJuly 11, 2026 at 09:09

This skill should be used when a session is being compacted/summarized (context is full or the user runs a compact command), or when writing any handoff summary that a later session or subagent will continue from. Defines what to preserve versus drop so the continuation doesn't waste tokens re-reading files and re-deriving state. Part of the HarnessTrim token-economy stack; the OpenCode adapter injects this guidance automatically via experimental.session.compacting.

Installation

Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.

Source

giuliastro

giuliastro/HarnessTrim

View GitHub Repository View Creator Repositories

Download

Related occupationsSOC

Based on SOC occupation classification

Software DevelopersComputer and Mathematical Occupations·SOC 15-1252

SKILL.md

readonly

name	compact-handoff
description	This skill should be used when a session is being compacted/summarized (context is full or the user runs a compact command), or when writing any handoff summary that a later session or subagent will continue from. Defines what to preserve versus drop so the continuation doesn't waste tokens re-reading files and re-deriving state. Part of the HarnessTrim token-economy stack; the OpenCode adapter injects this guidance automatically via experimental.session.compacting.
version	0.1.0
license	MIT

compact-handoff

A compaction summary is the seed context for everything that follows. If it drops the wrong things, the next turns re-read files, re-run searches, and re-decide settled questions — spending far more tokens than the summary ever saved. Optimize the summary for resumption, not brevity.

Preserve

The task and its acceptance criteria — what "done" means, verbatim if the user stated it.
Decisions made and alternatives rejected — with the reason, so they aren't re-litigated.
Hard-won specifics — exact file paths, symbol names, commands that worked, config values, version numbers, API signatures discovered. These are the expensive-to-rediscover facts.
Current state — what is done, what is verified (and how), what is still failing and the last error seen.

Drop

Step-by-step narration of what was already done ("first I opened X, then I ran Y").
Raw tool output that has already been acted on (test logs, file dumps, search results).
Anything trivially reconstructable from the current diff or a single cheap command.
Restatements of the codebase's structure that the next session can read directly.

Shape

Write the summary as short labeled sections (Task / Decisions / Key facts / State), not prose. A resuming agent should be able to act from it without re-reading the whole transcript.

More from this repository

same repository

delegate-bulk

giuliastro/HarnessTrim

This skill should be used when deciding whether to hand volumetric or noisy work to an isolated subagent (or another harness) instead of doing it in the main context — e.g. reading many files, digesting long logs/docs, wide searches, or bulk mechanical generation. It gives the rule for when isolation saves tokens versus when it just multiplies them. Part of the HarnessTrim token-economy stack.

2026-07-1112

scaffold-fast

giuliastro/HarnessTrim

This skill should be used for mechanical, low-novelty coding work — generating boilerplate, wiring up a component from an existing pattern, adding a CRUD endpoint like the others, writing repetitive tests, or applying a rote transformation across files. It keeps reasoning effort low and output terse for work whose shape is already decided. Part of the HarnessTrim token-economy stack; the lean-scaffold preset pairs it with low reasoning effort.

2026-07-1112

debug-log-slim

giuliastro/HarnessTrim

This skill should be used before pasting, printing, or otherwise embedding test-runner output, build logs, or CI output into the conversation — whenever a test suite, linter, or build was just run and its raw output is large. Filters logs down to failure/error signal before they enter context. Part of the HarnessTrim token-economy stack (see packages/core/src/reducers/test-output-slim.ts for the reference implementation the OpenCode adapter automates this with).

2026-07-1112

delta-response

giuliastro/HarnessTrim

This skill should be used for every response produced in a coding harness (Claude Code, Codex, OpenCode, Pi) — not only when explicitly requested. Enforces terse, structured, information-dense output instead of narrated verbosity, to cut output-token cost. Part of the HarnessTrim token-economy stack (see delta-response/references/examples.md for before/after samples).

2026-07-1112

review-delta

giuliastro/HarnessTrim

This skill should be used when asked to review code, review a pull request, or review a diff. Focuses the review on problems worth fixing rather than restating or summarizing what changed. Part of the HarnessTrim token-economy stack.

2026-07-1112

name	compact-handoff
description	This skill should be used when a session is being compacted/summarized (context is full or the user runs a compact command), or when writing any handoff summary that a later session or subagent will continue from. Defines what to preserve versus drop so the continuation doesn't waste tokens re-reading files and re-deriving state. Part of the HarnessTrim token-economy stack; the OpenCode adapter injects this guidance automatically via experimental.session.compacting.
version	0.1.0
license	MIT

compact-handoff

Preserve

The task and its acceptance criteria — what "done" means, verbatim if the user stated it.
Decisions made and alternatives rejected — with the reason, so they aren't re-litigated.
Hard-won specifics — exact file paths, symbol names, commands that worked, config values, version numbers, API signatures discovered. These are the expensive-to-rediscover facts.
Current state — what is done, what is verified (and how), what is still failing and the last error seen.

Drop

Step-by-step narration of what was already done ("first I opened X, then I ran Y").
Raw tool output that has already been acted on (test logs, file dumps, search results).
Anything trivially reconstructable from the current diff or a single cheap command.
Restatements of the codebase's structure that the next session can read directly.

Shape

Write the summary as short labeled sections (Task / Decisions / Key facts / State), not prose. A resuming agent should be able to act from it without re-reading the whole transcript.