Session distillation pipeline for Claude Code session logs.
Ambix stages, analyzes, and distills Claude Code session JSONL logs into structured summary artifacts suitable for downstream review, search, and retrospective analysis.
Status: Experimental. API may change before v1.0.
- Node.js >= 22
npm install ambixAmbix uses parse-cc for session log parsing.
Distill a session into a structured artifact:
ambix distill /path/to/session.jsonlThis runs the full pipeline (stage, analyze, distill, merge) and writes an artifact to ~/.ambix/sessions/<session-id>/. Use --mock to run with a placeholder runner that skips the API call.
Produce a chronological brief for context recovery:
ambix brief /path/to/session.jsonlThis outputs a per-round XML summary where every tool call and assistant response is tagged with a rehydration index, so an agent can pull full details on demand via ambix query.
Compact a session into a new resumable JSONL:
ambix compact /path/to/session.jsonl --full-recent 10Emits a new session file Claude Code's /resume picks up in the source's project dir. The last --full-recent N rounds are preserved verbatim; older turns collapse into a single user-role message containing an <ambix-compaction-marker> preamble plus a <turns> XML list with per-tool structured children. Small tool_use input fields pass through verbatim; fields over --max-field-bytes get a truncated="<bytes>" attribute + short preview body and are rehydratable on demand via ambix query <orig-session-id> <ix>. Task* tool_use + matched tool_result entries pass through as real entries so CC can rebuild its live task list on resume.
Typical session sizes compact to ~3-10% of source. Alternative to CC's built-in /compact when you want a structured navigable history rather than a narrative summary.
--preserve <kind>:<pattern> (repeatable) exempts matching content from condensation. Two kinds: tool:<glob> keeps matching tool_use/tool_result entries verbatim inside the bundled summary; type:<glob> promotes matching entries to real JSONL pass-through. Primary use case: a tool that IS the conversation channel (e.g., an MCP Telegram plugin) — without --preserve 'tool:mcp__telegram__*', the compacted session would summarize away your actual messages.
Known limitation: CC's "restore conversation and code" rewind relies on file-history-snapshot entries, which ambix drops from the condensed range to save bytes. Rewind-with-code from a compacted session can only reach into the preserved tail; code state before that is only recoverable via ambix query + manual file edits.
| Command | Description |
|---|---|
ambix distill <session> |
Full pipeline: stage, analyze, distill, merge, persist |
ambix analyze <session> |
Deterministic analysis only (JSON to stdout) |
ambix info <session> |
Minimal session summary (metadata + token rollup) |
ambix brief <session> |
Chronological per-round summary for context recovery |
ambix compact <session> |
Emit a resumable compacted JSONL (bundled XML summary + preserved tail) |
ambix stage <session> |
Stage a session into a tmp workspace |
ambix file-at <path> <ix> |
Print a tracked file's content at a given turn index |
ambix query <session> <sub> |
Search within a session log (tool-uses, tool-results, text-search, show) |
Run ambix <subcommand> --help for full flag documentation. See docs/cli.md for the complete CLI reference.
Ambix runs a four-stage pipeline over a Claude Code session log:
- Stage -- copies the session into a tmp workspace and produces a condensed JSONL view with rehydrated file histories and tool result snapshots.
- Analyze -- deterministic pass that computes token totals, tool usage, file churn, bash clusters, failures, and permission events.
- Distill -- an agent reads the staged workspace and produces a structured
Narrative(main task, episodes, decisions, corrections, verifications, friction points, wins, unresolved items). - Merge + Persist -- combines metadata, deterministic analysis, and narrative into a final
Artifactpersisted to~/.ambix.
Compact is a standalone capability that produces a chronological per-round summary (XML or markdown) for context recovery, independent of the distillation pipeline.
See docs/concepts.md for the full pipeline walkthrough, compact format, and artifact schema.
The final artifact written to ~/.ambix/sessions/<session-id>/artifact.json:
| I want to... | Doc |
|---|---|
| See every CLI flag and subcommand | CLI reference |
| Understand the pipeline stages and artifact schema | Concepts |
Full pipeline:
import { run, RealAgentRunner } from "ambix";
const result = await run({
session: "/path/to/session.jsonl",
runner: new RealAgentRunner({ model: "claude-sonnet-4-6" }),
});
if (result.success) {
console.log(`artifact: ${result.artifactPath}`);
}See RunOptions and RunResult types for the full set of options and return values.
Lower-level building blocks:
import { stage, analyze, fileAt } from "ambix";
import { Session } from "parse-cc";
const session = new Session("/path/to/session.jsonl");
const layout = await stage(session, "/tmp/ambix-work");
const results = await analyze(session);
const file = await fileAt({
tmp: "/tmp/ambix-work",
path: "src/foo.ts",
ix: 42,
});npm install
npm run build # compile TypeScript
npm test # run tests (vitest)
npm run lint # biome check
npm run typecheck # tsc --noEmitMIT -- see LICENSE.
{ "schema_version": "1", "session_id": "abc123-...", "generated_at": "2026-04-16T12:00:00Z", "metadata": { "session_id": "abc123-...", "cwd": "/home/user/project", "turn_count": 84, "duration_s": 1200, "end_state": "completed" // ... }, "deterministic": { "tokens": { "totals": { "in": 450000, "out": 38000, "cache_read": 320000, "cache_write": 95000 }, "by_model": { "claude-sonnet-4-6": { /* ... */ } } }, "tools": { "invocations": { "Edit": 12, "Read": 24, "Bash": 8, "Grep": 6 } }, "files": { "touched": [{ "path": "src/main.ts", "reads": 3, "edits": 5, "writes": 1 }] }, "bash_clusters": [{ "pattern": "npm", "count": 4 }], "failures": [], "subagents": [] // ... }, "narrative": { "summary": "Implemented user authentication with JWT tokens and added login/logout endpoints.", "main_tasks": [ { "title": "Add JWT auth", "status": "verified", "description": "...", "refs": [0, 84] } ], "episodes": [ { "title": "Research auth patterns", "kind": "research", "ix_range": [0, 12], "summary": "...", "refs": [2, 8] }, { "title": "Implement login endpoint", "kind": "implementation", "ix_range": [13, 45], "summary": "...", "refs": [15, 30] } ], "decisions": [ { "description": "Chose JWT over session cookies", "rationale": "Stateless, works with API clients", "refs": [5] } ], "corrections": [], "verification": { "was_verified": true, "how": "Ran test suite, all 12 tests pass", "refs": [80] }, "friction_points": [], "wins": [{ "description": "Tests passed on first run", "refs": [80] }], "unresolved": [] } }