agentsclimarketplace

Manual compaction reset

Skill zoidbergclawd/coherence/skills/manual-compaction-reset

Five portable Agent Skills for Claude Code that keep humans and coding agents aligned at high velocity.

Install
npx -y skills add zoidbergclawd/coherence --skill manual-compaction-reset

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Use when context is long, automatic compaction is likely, a session resumes after time away, direction changes, or the agent appears to be carrying stale assumptions, to perform a hard context reset by writing and reloading curated project context instead of relying on accumulated chat history.

SKILL.md

4.5 KB, as published. Nobody here has run it

Manual Compaction Reset

When to use this skill

  • Conversation has run long enough that compaction is imminent or has already occurred
  • A session is resuming after hours, days, or a context switch
  • The user has changed direction, scope, or priority mid-session
  • The agent is repeating earlier mistakes or referencing facts that may be stale
  • Multiple unrelated subtasks have accumulated in the same buffer
  • The agent cannot confidently answer "what is the current state of the work?"

If any of these are true, run this skill before continuing the next task.

Core principle

Long conversations accumulate ambient assumptions that cannot be audited. Trusting compacted history is trusting a summary the agent did not write. A manual reset replaces fuzzy chat memory with a small, curated, written context that both the agent and the user can inspect, edit, and reload deterministically.

The reset succeeds when the next session can continue work using only the brief, the source code, and active specs — never by re-reading prior chat.

Procedure

  1. Stop current work. Do not start a new task on stale context.
  2. Write a Reset Brief (template below) to a file the agent can re-read, e.g. .context/reset-brief.md or docs/context/<task>.md.
  3. Compose the brief from current ground truth, not memory:
    • git status, git log -n 5, the active branch
    • The most recent CI run
    • The relevant boundary specs and plans on disk
    • The user's most recent stated goal
  4. Ask the user to confirm or correct the brief. Treat it as ground truth only after confirmation.
  5. End the current session, or explicitly clear the working buffer.
  6. Re-enter context by reading only: the Reset Brief, the relevant source files, the active boundary specs, and the active multi-pass plan. Do not re-read prior chat history.
  7. Restate the next step in the agent's own words and wait for the user to confirm before resuming.

Reset Brief template

# Reset Brief: <task or session name>

## Goal
<one sentence — what we are trying to achieve>

## Scope
- <in-scope item>
- <in-scope item>

## Out of scope
- <explicitly excluded item>

## Current state
- Branch: <name>
- Last commit: <sha> <subject>
- Working: <what is verified to work>
- Broken: <what is verified to fail, with evidence>
- CI: <last status>

## Active specs and plans
- <path to boundary spec>
- <path to multi-pass plan>

## Verified facts
- <claim> — verified by <command or user statement>

## Known unknowns
- <question that must be resolved before resuming>

## Next step
<one concrete action — file, function, or command>

Confirmed by user: yes/no

Required output format

A single markdown file matching the template above, with every section filled in or explicitly marked none. The file must end with Confirmed by user: yes/no. The agent does not act on the brief until the footer reads yes.

Stop conditions / escalation triggers

  • The user cannot confirm a section because the agent's recollection of it is wrong — rewrite that section from source, not memory
  • A Verified facts claim cannot be re-derived from the code or repo state — remove it
  • The brief contradicts the most recent commit, PR, or CI run — the brief is wrong; the repo is right
  • After reload, the agent finds itself referring to messages no longer in context — write another brief; do not improvise
  • Known unknowns is non-empty when the user asks the agent to proceed — surface them first

Short examples

Good Current state section:

Branch: feat/auth-rewrite. Last commit: 7f3a1c2 "replace cookie middleware". Working: login flow on dev (manually verified). Broken: refresh-token endpoint returns 500 in CI run #481. CI: failing on auth.spec.ts.

Bad Current state section (rejected):

We've been working on auth for a while and most of it is done.

Rewrite from git status, git log, and the most recent CI output. The phrase "most of it" is exactly the kind of compacted assumption that erodes coherence.

Good Next step:

Add a failing reproducer test for the 500 in src/auth/refresh.spec.ts, asserting status 401 for an expired token.

Bad Next step:

Continue working on auth.

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.