agentleFS
OrionArchitektonOrionArchitekton/orion-skills/skills/pre-compact/SKILL.md

pre-compact

Skill2 starsChanged 12 days agoMITView on GitHub

Use when about to run /compact, when the context window is filling up or near its limit, when wrapping up a long working session, or when handing in-progress work to a fresh/future instance so it can resume without re-deriving state. Triggers, "drop a handoff", "recap before compact", "save context", "context is at 90%".

3 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

5 rules. 0 are common, 5 are distinctive, and 0 are very specific to this project.

  • 0

    Common

    Most of the 3 edited versions we read keep these.

  • 5

    Distinctive

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    No engine, hook or private runtime is required.

    line 37
  2. Distinctive:

    Use a directory you control and the current workspace's ownership rules.

    line 90
  3. Distinctive:

    6. Flag, do not perform, promotions: If the session produced a durable decision/learning, list it under *Candidate promotions* for the …

    line 101
  4. Distinctive:

    If task A and task B share a checkout and B writes the newest global pointer, A's reader opens A's explicit pack.

    line 185
  5. Distinctive:

    If A's old pack names accepted draft v1 and a release next action, but the current request accepts v2 and limits work to local review, …

    line 190

“Common” is measured: kept by at least 6 in 10 of the 3 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Pre-Compact Context Pack

Overview

/compact discards most of the conversation. This skill captures the session into a persistent, queryable context pack so a fresh instance resumes with zero re-derivation, what shipped, what's in flight, what's next, and the operating rules to follow.

Core principle: a pack is handoff evidence, not canon. It is never auto-promoted into any repo, memory store, or vault. Durable decisions get promoted separately, by a human, after review.

The compaction contract (what to keep vs. drop). A pack is a decision- and open-thread-shaped compaction, not a transcript. PRESERVE: architectural/locked decisions + their rationale ("do NOT re-ask"), open threads / next actions, the verified-vs-claimed split, and the operating rails. DROP: redundant tool output, verbose command transcripts, intermediate exploration, and anything a fresh instance can re-derive from disk in one command. When in doubt, keep the decision and the pointer (file:line, SHA, command), not the raw output that produced it. A pack heavy with tool dumps but light on decisions has inverted the contract.

Where packs live (engine-agnostic)

A pack is just a Markdown file in a persistent directory: NOT /tmp (wiped on reboot). Pick a durable, agent-neutral location and keep it consistent, e.g.:

  • tasks/<task-id>/packs/: one immutable timestamped file per checkpoint.
  • Return the explicit pack path for the task being handed off. A directory or branch alone is not task identity: multiple tasks can share both.

Choose a stable, distinct task ID (for example project-fix-import) and record the repository identity and worktree in the body. Reuse that ID only for the same work. A replacement names the previous pack for that task; preserve the old bytes. An optional index or task-local pointer is a convenience, never an authority. No engine, hook or private runtime is required.

A legacy global LATEST.md may belong to another task. Use it only after verifying that its task and repository match the current request and there is exactly one eligible continuation. Missing identity, conflicting candidates, or an ambiguous request means no pickup: request the exact task/path if current authority cannot resolve it. Never infer ownership or acceptance from a timestamp or an idle chat.

The Iron Rule: Verify, Don't Transcribe

Your own session narrative is the least trustworthy input. Before writing the pack, confirm claims against reality. A handoff that says "fixed and committed (a1b2c3d)" when no such commit exists is worse than no handoff, it sends the next instance building on phantom work.

Set verified_state honestly: verified (all key claims disk-confirmed) · mixed · claimed-only. In the body, split ✅ VERIFIED from ⚠️ CLAIMED (unverified).

Process

1. Capture real runtime state (don't rely on memory). For each repo/worktree touched this session:

git -C <repo> status -sb && git -C <repo> log --oneline -8
git worktree list

Note: current branch(es), uncommitted changes, any in-flight background jobs/agents, recently edited files.

2. Verify claims. For every "I did X": does the commit/branch/PR actually exist? Did tests actually run (real output)? Was a deploy independently verified? Downgrade verified_state for anything you can't confirm.

3. Compose the body using the template below. Record the task, repository, accepted baseline plus acceptance evidence (or unknown), last verified outcome plus proof, unfinished step, owner, literal next action, and remaining authority. Acceptance and verification are different facts. Never fill in missing acceptance from recency, tests or your narrative.

4. Write the pack. Prepare a body file with literal Task: <task-id> and Repository: <repository-identity> lines. Set TASK_ID, REPO_ID and PACK_BODY to those values and the body's absolute path. Choose a persistent HANDOFF_ROOT outside temporary storage (the default below is one option). Confirm the body and selection match the current task before running this shell recipe:

set -eu
umask 077
: "${TASK_ID:?set a stable task ID}" "${REPO_ID:?set repository identity}" "${PACK_BODY:?set body path}"
case "$TASK_ID" in ''|.|..|*[!a-zA-Z0-9._-]*) echo 'Invalid task ID' >&2; exit 1 ;; esac
grep -qxF -- "Task: $TASK_ID" "$PACK_BODY" || { echo 'Task/body mismatch' >&2; exit 1; }
grep -qxF -- "Repository: $REPO_ID" "$PACK_BODY" || { echo 'Repository/body mismatch' >&2; exit 1; }
HANDOFF_ROOT="${HANDOFF_ROOT:-$HOME/.local/state/agent-handoffs}"
PACKS="$HANDOFF_ROOT/tasks/$TASK_ID/packs"
mkdir -p -- "$PACKS"
PACK=$(mktemp "$PACKS/$(date -u +%Y-%m-%dT%H%M%SZ)-XXXXXX.md")
cat -- "$PACK_BODY" > "$PACK"
printf 'Read %s and resume only the matching task after verification.\n' "$PACK"

This recipe preserves prior checkpoints even when two writes share a timestamp. Use a directory you control and the current workspace's ownership rules; the recipe does not acquire a task lease or arbitrate concurrent owners.

5. Print the explicit pickup line. Give the exact path returned above. On pickup, check the pack's task/repository against the current request, then compare its accepted baseline, proof and next action with current project authority. Re-probe only the referenced live state needed for continuation. A stale pack is historical evidence: preserve it, use the newer accepted baseline, and do not execute a gated action just because its old next-action line names one. An unknown remaining authority stays unknown until resolved.

6. Flag, do not perform, promotions. If the session produced a durable decision/learning, list it under Candidate promotions for the human to later move into memory/canon. Never auto-write canon.

Pack Body Template

# HANDOFF: <title>
Task: <task-id>
Repository: <repository-identity>
**Date:** <YYYY-MM-DD> · **Host:** <box> · **Identity:** <git identity>
**Worktree / branch / HEAD:** <absolute directory / branch / SHA>
**Owner:** <actual owner or unknown; verify lease before resuming>
**Accepted baseline:** <artifact/plan + acceptance evidence, or unknown>
**Last verified outcome:** <observable result + proof and verification time>
**Unfinished step:** <one concrete step>
**Next action:** <literal authorized first move>
**Remaining authority:** <scope granted; additional decisions still needed, or unknown>
**Supersedes:** <prior pack for this same task, or none>
**Pointers:** memory file(s) · authoritative plan · proof packet · audit packs

## TL;DR: where things stand
<2-4 sentences: the mission and current position.>

## Verified state
- ✅ VERIFIED: <claim + how confirmed (SHA, test output, health check)>
- ⚠️ CLAIMED (unverified): <claim + why not confirmed + how to confirm>

## Shipped / merged
- <thing>, <repo>, <SHA / PR#>, <verified?>

## In-flight (current working state)
- Branch/worktree: <name @ path> · Uncommitted: <summary> · Background: <jobs/pipelines>

## Open work (prioritized, actionable)
### 1. <next thing>: <why it matters>
**Branch/worktree:** <where to work>
**Decisions already made (do NOT re-ask):** <locked choices + rationale>
**Scope:** <specific files / lines / steps>
**Blocker to surface to the user:** <question that needs a human, if any>

## Gated / deferred (do NOT start without explicit ask)
- <thing>, <why gated>

## Operating rules (follow when resuming)
- <commit discipline, base branch, force-push ban, worktree discipline, etc.>

## Resume: first action
<The single literal first move for the fresh instance.>

## Candidate promotions (evidence → canon; human review only)
- <durable decision/learning worth promoting to memory or canon later>

The "do NOT re-ask" Discipline

The highest-value lines in a pack are locked decisions. A fresh instance has no memory of why you chose the shim over the rewrite, or that a given repo's base branch is not main. Without "decisions already made (do NOT re-ask)", it re-litigates settled questions or guesses wrong. Capture the decision and its rationale.

Pack management operations (if you build/use an engine)

A pack manager, optional, is just convenience around the file format above. Useful operations: create (timestamp + write + record explicit task path + index), latest (path + next-action, filtered by task identity), list (by status), supersede (link a replacement), set-status (active | done | superseded | stale | archived), sweep (archive superseded packs older than N days, move, never delete). None are required; you can do all of this by hand.

Common Mistakes

  • Transcribing instead of verifying → claimed-vs-verified split skipped; next instance builds on phantom work. Run the git/disk checks first.
  • Writing to /tmp → evaporates on reboot. Always write to a persistent packs directory.
  • Vague next steps ("continue the work") → give the literal first command/file.
  • Omitting locked decisions → re-litigation. Include "do NOT re-ask".
  • Auto-promoting to canon → packs are evidence; flag promotions, let a human move them.
  • Leaving many "active" packs for one arc → a per-dir "newest active pack" selector then surfaces an older pack instead of the head. When a session continues an arc that already has a pack, record that the prior pack is superseded by this task's replacement, preserving the old pack. Never supersede another task's pack.

Other agent runtimes

The format is plain Markdown and tool-agnostic, it works anywhere you can write a file. If your runtime auto-surfaces a pack on resume (via a session-start hook), great; if not, the convention still works, read the explicit task pack manually when starting a fresh session. Any runtime adapter must separately demonstrate matching task selection; these instructions do not install or qualify one.

Worked reader checks

If task A and task B share a checkout and B writes the newest global pointer, A's reader opens A's explicit pack. If given B's pack instead, it identifies the mismatch and does not execute B's next action. A legacy latest pickup is permitted only after an affirmative unique match of task and repository.

If A's old pack names accepted draft v1 and a release next action, but the current request accepts v2 and limits work to local review, resume v2's review. Preserve v1's evidence and the release gate. A recorded next action never grants authority.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.