Agent context crystallizer
Skill stewie-sh/agent-crystallize/skills/agent-context-crystallizer
Local-first checkpoints and crystals for long-running AI coding agent sessions.
npx -y skills add stewie-sh/agent-crystallize --skill agent-context-crystallizerAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
- 5 stars5 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
Preserve high-signal work context with local checkpoints and session crystals using the agent-crystallize CLI. Use before compaction, handoff, session end, cross-agent transfer, after major decisions/findings/failures, or when the user asks to checkpoint, crystallize, preserve context, save open loops, or create a resume handoff.
SKILL.md
3.8 KB, 830 tokens by cl100k_base, as published. Nobody here has run it
Agent Context Crystallizer
Use the agent-crystallize CLI as the source of truth. This skill is a thin
workflow wrapper; do not reimplement the artifact format by hand unless the CLI
is unavailable.
Resolve The CLI Before Writing
Do not assume the agent harness inherited the user's interactive-shell PATH.
Resolve one working invocation in this order:
- Use a repo-provided
agent-crystallizepackage script when present. - Use
agent-crystallizewhencommand -v agent-crystallizesucceeds. - If
AGENT_CRYSTALLIZE_CLInames a readable JavaScript entry point, invoke it with the current Node runtime:node "$AGENT_CRYSTALLIZE_CLI" .... - Use a verified package-manager or harness-provided bundled runtime path. Never guess a machine-specific path or copy one from another user.
- Only when no CLI is available, write a plainly labelled non-validated
emergency handoff outside the canonical
.agent-crystals/checkpoints/and.agent-crystals/sessions/directories. Replace it with a CLI-generated, validated artifact when the runtime is restored.
The emergency handoff preserves continuity; it is not a valid crystal and must not silently enter the manifest or normal validation set.
Workflow
- Inspect the local state that matters:
AGENTS.md/CLAUDE.mdif present,git status --short, relevant changed files, recent test/build/deploy results, and existing.agent-crystals/manifest.jsonwhen present. - Choose the lightest useful artifact:
- checkpoint: during active work, before risky edits, before compaction, or after a high-signal decision/finding/failure;
- session crystal: before handoff/session end, or to roll up recent checkpoints.
- Prefer structured flags over vague prose:
--decision,--finding,--open-loop,--test,--next-action,--evidence,--memory-candidate,--topic, and--relation type:target. - Add safe provenance when available:
--agent-body,--harness,--session-id,--transcript-uri, and--source-ref. Never dump broad environment variables or secrets. - Use
--continuity-tailonly for short recent turns where order/nuance matters after compaction. It is raw continuity evidence, not durable truth. - Validate important artifacts before relying on them.
Commands
Fast checkpoint:
agent-crystallize checkpoint \
--body "<current state, decision, open loop, next action>" \
--decision "<decision and authority>" \
--finding "<runtime finding>" \
--open-loop "<unfinished item>" \
--next-action "<next concrete action>"
Session crystal from recent checkpoints:
agent-crystallize now \
--from-checkpoints latest \
--body "<synthesis since the latest checkpoint and handoff state>"
Validate:
agent-crystallize validate --fail-on-warnings
Hook/continuity check:
agent-crystallize doctor --codex --claude --hooks
Guardrails
- Preserve work context, not hidden chain-of-thought.
- Raw evidence is not truth; decisions/findings should keep provenance.
- Keep generated
.agent-crystals/local/private unless intentionally reviewed and sanitized. - Prefer paths, ids, commits, and source refs over copying large logs or transcripts.
- If hooks are configured but continuity feels broken, ask the user to verify
the host
/hooksview. Codex may skip new or changed hooks until trusted. - Never treat a hand-written emergency handoff as schema-valid merely because it is readable Markdown.
What ships with it: 1 file
287 B alongside SKILL.md
agents/
- openai.yaml287 B
Gives 0 of the 12 instructions most context ai engineering skills give in 830 tokens
Counted across 1,193 of the 1,976 authors here whose files we hold, read 2026-08-07
- Dispatch a fresh implementer subagent per taskin 48 of 1193, across 19 files
- Dispatch a final code reviewer after all tasksin 33 of 1193, across 8 files
- Provide full task text to the subagentin 30 of 1193, across 9 files
- Review spec compliance before code qualityin 27 of 1193, across 10 files
- Make the hook script executablein 26 of 1193, across 8 files
- Re-snapshot after navigation or DOM changesin 25 of 1193, across 19 files
- Read files before editing themin 22 of 1193, across 11 files
- Answer subagent questions before proceedingin 22 of 1193, across 7 files
- Mark task complete in TodoWrite after approvalin 22 of 1193, across 6 files
- Merge hook into existing settingsin 21 of 1193, across 3 files
- Ask if installation is global or projectin 20 of 1193, across 2 files
- Copy the hook script to target locationin 20 of 1193, across 2 files
Said here and by no other author read
- use the CLI as source of truth
- resolve CLI invocation before writing artifacts
- inspect local state before creating artifacts
- choose the lightest useful artifact type
- prefer structured CLI flags over prose
- add safe provenance data when available
Grouped from the skills themselves: near-identical wordings counted once, and counted by distinct author, so one author publishing three of these counts once. Length counted with cl100k_base; the agent that loads this file may tokenize it differently.