sublimate-distiller

Sublimate Distiller

Mine the current project's Claude Code session corpus, cluster by tool-call similarity, classify, distill, emit candidates.

Stage 1 — Scope

Determine the session corpus path:

• Default: ~/.claude/projects/$(basename "$PWD" | sed 's|/|-|g; s|^|-|; s|$||') — Claude Code encodes the absolute project path as a directory under ~/.claude/projects/ using - as the separator.
• Override: if $ARGUMENTS contains --project-dir <path>, use that path instead.

Verify the directory exists and contains .jsonl files. Abort with a clear message if empty.

Stage 2 — Mine

Run the Python miner as a Bash subprocess:

mkdir -p /tmp/sublimate-run
cd "${CLAUDE_PLUGIN_ROOT}/src" && python3 -m sublimate.cli <projects-dir> \
  --out /tmp/sublimate-run/ \
  --max-sessions 500 \
  --emit-all

Run as a module (python3 -m sublimate.cli) from inside the src/ directory so Python resolves from . import cluster, emit, parser, prefixspan, shingle relative imports. Invoking the script path directly (python3 path/to/cli.py) raises ImportError: attempted relative import with no known parent package.

The miner walks JSONL files, extracts (role, tool_name, tool_input_hash) per turn, computes k=3 shingles, runs union-find on Jaccard >= 0.35, then PrefixSpan on each cluster.

Output: /tmp/sublimate-run/clusters.json (array of cluster records) plus raw cluster-<id>.workflow.js skeletons.

Stage 3 — Curate

Dispatch the cluster-curator agent via the Task tool. The agent prompt MUST include both:

• the path to /tmp/sublimate-run/clusters.json (dict keyed by cluster id)
• the projects directory determined in Stage 1 (so the curator can resolve <sample_session_id>.jsonl → absolute path)

The curator reads ≥ 1 sample session per cluster from <projects_dir>/<sample_id>.jsonl and writes /tmp/sublimate-run/curation.json with one record per cluster:

{
  "id": "<cluster-id>",
  "verdict": "KEEP | SKIP | NEEDS_MORE_SIGNAL",
  "tier": "NOISE | SUBAGENT | SKILL | WORKFLOW",
  "rationale": "<sentence citing the sample session>",
  "sample_cited": "<jsonl path>"
}

Wait for the curator envelope before advancing.

Stage 4 — Distill prompts

For each cluster where verdict == KEEP, dispatch the prompt-distiller agent with:

• cluster id
• tier
• 2-3 sample session jsonl paths from the curation record

Distiller reads samples, identifies the procedure, writes a coherent meta.name plus per-phase agent({prompt}) body. Overwrites the raw skeleton at /tmp/sublimate-run/cluster-<id>.workflow.js for Workflow tier, or writes /tmp/sublimate-run/cluster-<id>/SKILL.md for Skill tier.

Fan out distillers in parallel — one Task call per KEEP cluster in a single assistant message. Do NOT set run_in_background: true.

Stage 5 — Emit

After all distillers return, finalize artifacts under /tmp/sublimate-run/:

• cluster-<id>.workflow.js for Workflow-tier candidates
• cluster-<id>/SKILL.md for Skill-tier candidates
• cluster-<id>/AGENT.md for Subagent-tier candidates (sketch only)

SKIP and NEEDS_MORE_SIGNAL clusters are not emitted.

Stage 6 — Lint

Reject any emitted .workflow.js that contains the banned non-determinism substrings. Build the regex at runtime so this skill body itself does not contain the literal:

BANNED=$(printf '%s|%s|%s' \
  'D''ate.now' \
  'M''ath.random' \
  'new ''Date()')
for f in /tmp/sublimate-run/cluster-*.workflow.js; do
  if grep -E "$BANNED" "$f" >/dev/null; then
    echo "REJECT $f — non-deterministic API detected"
    rm "$f"
  fi
done

Workflows must be reproducible. Non-determinism breaks the deterministic-checkpoint guarantee.

Stage 7 — Report

Print a final table to the user:

Cluster   Size  Tier      Verdict   File
-------   ----  ----      -------   ----
c-001     12    WORKFLOW  KEEP      /tmp/sublimate-run/cluster-c-001.workflow.js
c-002     7     SKILL     KEEP      /tmp/sublimate-run/cluster-c-002/SKILL.md
c-003     3     -         SKIP      -

Plus the count of emitted files per tier, and a one-line next-step: "Promote a candidate with /sublimate:promote <cluster-id>."

3-tier classification rubric

• Subagent — single repeated role, no orchestration, no fan-out. Cheapest runtime. Emit as agents/<name>.md.
• Skill — repeated single-phase procedure that needs scoped tools + context. Medium runtime. Emit as skills/<name>/SKILL.md.
• Workflow — multi-phase, fan-out, deterministic checkpoints, repeated >= 5 times with consistent phase shape. Highest runtime requirements (Claude Code v2.1.154+, Pro/Max/Team/Enterprise). Emit as .workflow.js.

When uncertain, prefer Skill over Workflow — Workflows carry more runtime constraints.