Skillers Compact

compact

When to Use

Invoked by the skillers-compactor agent during /skillers compact. Also usable standalone for manual compaction.

Arguments

Parse from $ARGUMENTS:

Data Sources

Skillers reads conversation transcripts from multiple AI tools. Detect which tools are installed and read from all available sources.

Claude Code

Location: ~/.claude/projects/{project-hash}/{session-id}.jsonl

The project hash is derived from the CWD with path separators replaced by dashes. Each transcript is a JSONL file with entries of type: user, assistant, system, progress, file-history-snapshot.

{
  "type": "user",
  "message": { "role": "user", "content": "the user message" },
  "timestamp": "2026-03-09T14:25:05.135Z",
  "sessionId": "uuid",
  "cwd": "/path/to/project"
}

Codex CLI

Location: ~/.codex/sessions/{YYYY}/{MM}/{DD}/rollout-{timestamp}-{uuid}.jsonl

Sessions are organized by date. Each JSONL file starts with a session_meta entry, followed by response_item and event_msg entries. User messages have type: "event_msg" with payload.type: "user_message".

{"timestamp":"...","type":"session_meta","payload":{"id":"uuid","cwd":"...","cli_version":"0.77.0"}}
{"timestamp":"...","type":"event_msg","payload":{"type":"user_message","message":"the user message"}}
{"timestamp":"...","type":"response_item","payload":{"type":"message","role":"user","content":[{"type":"input_text","text":"..."}]}}

Also available: ~/.codex/history.jsonl - compact log with only user inputs per session:

{"session_id":"uuid","ts":1767723836,"text":"the user message"}

OpenCode

Location: ~/.local/share/opencode/opencode.db (SQLite)

Tables: project, session, message. Query sessions and messages with SQL.

SELECT s.id, s.slug, m.time_created, m.content
FROM session s JOIN message m ON m.session_id = s.id
WHERE m.time_created > {cutoff_timestamp}
ORDER BY m.time_created;

Also available: ~/.local/state/opencode/prompt-history.jsonl - user input history only:

{"input":"the user message","parts":[],"mode":"normal"}

On Windows the DB may also be at %APPDATA%/opencode/opencode.db.

Cursor and Kiro do not store conversation history in an accessible local format. Skip during compaction.

Workflow

Phase 1: Resolve Paths

const os = require('os');
const path = require('path');
const fs = require('fs');

const STATE_DIR = process.env.AI_STATE_DIR || '.claude';
const scope = args.scope || 'global';
const days = args.days || 7;

// Knowledge output directories
const globalDir = path.join(os.homedir(), STATE_DIR, 'skillers', 'knowledge');
const repoDir = path.join(process.cwd(), STATE_DIR, 'skillers', 'knowledge');
const knowledgeDirs = scope === 'both' ? [globalDir, repoDir]
  : scope === 'global' ? [globalDir] : [repoDir];

const sources = [];
const claudeDir = path.join(os.homedir(), '.claude', 'projects');
if (fs.existsSync(claudeDir)) sources.push({ tool: 'claude-code', type: 'jsonl', path: claudeDir });
const codexDir = path.join(os.homedir(), '.codex', 'sessions');
if (fs.existsSync(codexDir)) sources.push({ tool: 'codex', type: 'jsonl', path: codexDir });
const opencodePaths = [
  path.join(os.homedir(), '.local', 'share', 'opencode', 'opencode.db'),
  path.join(process.env.APPDATA || '', 'opencode', 'opencode.db')
];
for (const p of opencodePaths) {
  if (fs.existsSync(p)) { sources.push({ tool: 'opencode', type: 'sqlite', path: p }); break; }
}

Phase 2: Find and Read Transcripts

For each detected source, collect transcripts using the appropriate method:

Claude Code (JSONL):

1. List directories under ~/.claude/projects/
2. For each project directory, find .jsonl transcript files
3. Filter by modification time: only files modified within the last --days days
4. Extract entries with type: "user" and type: "assistant"

Codex CLI (JSONL):

1. Walk ~/.codex/sessions/{YYYY}/{MM}/{DD}/ date directories
2. Find .jsonl files within the date range
3. Extract session_meta for session context (cwd, cli_version)
4. Extract event_msg entries with payload.type: "user_message" for user messages
5. Extract response_item entries with payload.role: "assistant" for tool usage patterns

OpenCode (SQLite):

1. Open opencode.db with a SQLite reader (Bash: sqlite3 or node better-sqlite3)
2. Query sessions with messages in the date range
3. Read message content from the message table joined with session
4. Normalize into the same observation extraction format

Common for all sources:

• Read the skillers config to check lastCompactedAt - skip already-processed transcripts
• Record the session ID, source tool, and timestamp for each entry
• Tag observations with source: "claude-code" | "codex" | "opencode" for traceability
• MUST redact every raw line before parsing using lib/sanitize.js::redact() from the skillers plugin root. Users commonly paste API keys, GitHub tokens, AWS keys, and Bearer tokens into chat; those must not enter agent context or persisted knowledge files.

const path = require('path');
const { redact } = require(path.join(PLUGIN_ROOT, 'lib', 'sanitize'));

function readJsonlSafe(filePath) {
  const raw = fs.readFileSync(filePath, 'utf8');
  return raw.split('\n').filter(Boolean).map(line => {
    const safeLine = redact(line);
    try { return JSON.parse(safeLine); } catch { return null; }
  }).filter(Boolean);
}

For OpenCode SQLite, apply redact() to each message.content string after fetching rows and before extracting observations.

Phase 3: Extract Observations

For each transcript, analyze the conversation to identify:

For each identified pattern, create an observation:

{"ts": "ISO timestamp", "t": "pain|repeat|task|wish|workflow", "v": "5 word description", "ctx": "file or area", "session": "session-id", "source": "claude-code|codex|opencode"}

Extract observations based on actual conversation content. Focus on:

• Tasks the user performs repeatedly across different sessions
• Pain points expressed through retries, frustration, or workarounds
• Multi-step workflows that follow the same sequence
• Explicit wishes for automation

Do NOT create observations for:

• One-off tasks that won't recur
• Normal productive work without friction
• Sensitive data (API keys, passwords, credentials, PII)

Phase 4: Cluster by Theme

Group observations by semantic similarity:

1. Extract tokens from v and ctx fields (lowercase, split on spaces and path separators)
2. Build token frequency map across all observations
3. Cluster using shared tokens - observations sharing 2+ tokens belong to the same cluster
4. Name each cluster using its top 2-3 tokens joined with hyphens (e.g., "ci-pr-workflow")
5. Merge small clusters (fewer than 3 observations) into the nearest larger cluster or into an "uncategorized" theme

Phase 5: Calculate Weights

For each theme cluster, calculate a composite weight:

function calculateWeight(observations) {
  const now = Date.now();

  // Frequency: more observations = higher weight
  const frequency = Math.min(observations.length / 20, 1.0); // Cap at 20

  // Recency: recent observations weigh more (exponential decay, 30-day half-life)
  const recencyScores = observations.map(obs => {
    const ageMs = now - new Date(obs.ts).getTime();
    const ageDays = ageMs / 86400000;
    return Math.exp(-0.693 * ageDays / 30); // 0.693 = ln(2)
  });
  const recency = recencyScores.reduce((a, b) => a + b, 0) / observations.length;

  // Cross-session: patterns across multiple sessions weigh disproportionately more
  const sessions = new Set(observations.map(obs => obs.session));
  const crossSession = Math.min(sessions.size / 5, 1.0); // Cap at 5 sessions

  // Pain intensity: "pain" and "wish" types weigh more
  const painCount = observations.filter(obs => obs.t === 'pain' || obs.t === 'wish').length;
  const painBoost = painCount > 0 ? 1.0 + (painCount / observations.length) * 0.5 : 1.0;

  // Composite weight
  const raw = (frequency * 0.3 + recency * 0.3 + crossSession * 0.4) * painBoost;
  return Math.round(Math.min(raw, 1.0) * 100) / 100;
}

Phase 6: Merge with Existing Knowledge

For each theme:

1. Check if knowledge/{theme-name}.json already exists
2. If exists:
   • Read existing file
   • Merge observation lists (deduplicate by timestamp)
   • Recalculate weight with merged data
   • Update lastSeen and sessions count
3. If new: create the file

Knowledge file format:

{
  "theme": "ci-pr-workflow",
  "weight": 0.82,
  "observations": [
    {"ts": "...", "t": "repeat", "v": "create PR check CI", "ctx": "github", "session": "abc", "source": "claude-code"},
    {"ts": "...", "t": "workflow", "v": "merge after CI pass", "ctx": "github", "session": "def", "source": "codex"}
  ],
  "sessions": 8,
  "firstSeen": "2026-02-15T...",
  "lastSeen": "2026-03-09T...",
  "totalOccurrences": 23,
  "typeCounts": {"pain": 2, "repeat": 12, "task": 4, "wish": 1, "workflow": 4}
}

Phase 7: Prune

Remove entries that are:

• Older than 90 days AND weight < 0.1
• Single occurrence AND older than 30 days

Phase 8: Update Config

After successful compaction, update the skillers config:

{
  "lastCompactedAt": "2026-03-09T22:30:00.000Z",
  "lastTranscriptsProcessed": ["session-id-1", "session-id-2"]
}

This prevents re-processing the same transcripts on next compact.

Output Format

Return JSON summary to the calling agent:

{
  "sources": {
    "claude-code": {"transcripts": 12, "observations": 35},
    "codex": {"transcripts": 8, "observations": 12},
    "opencode": {"sessions": 3, "observations": 5}
  },
  "totalTranscripts": 23,
  "totalObservations": 52,
  "themesUpdated": 2,
  "themesCreated": 1,
  "themesPruned": 0,
  "themes": [
    {"name": "ci-pr-workflow", "weight": 0.82, "observations": 23},
    {"name": "testing-patterns", "weight": 0.65, "observations": 15}
  ]
}

Constraints

• MUST detect and read from all installed AI tools (Claude Code, Codex, OpenCode)
• MUST handle malformed JSONL lines and missing DB tables gracefully (skip, log warning)
• MUST deduplicate observations by timestamp when merging
• MUST cap theme name length at 40 characters
• MUST update lastCompactedAt after successful compaction
• NEVER include raw sensitive data in knowledge files
• NEVER create observations from one-off tasks
• NEVER read more than 20 transcripts at once (cap for token efficiency)
• NEVER read more than 500 lines from a single transcript - sample the first 200 and last 300 lines for long sessions