composer-forensics

Composer forensics

Extract Composer client data from a live Chrome profile on disk, validate, and analyze offline — or diagnose and repair live via recovery mode.

Live doctor workflow (user has browser): DOCTOR.md — user opens debug port; agent explores; report in /tmp; confirm before any data change.

Full command reference: COMMANDS.md — locate, extract, validate, probe, automerge, SQL, recovery debug port.

Report template: reports/REPORT-TEMPLATE.md

Scope (v1): macOS + Google Chrome default profile (offline extract). Recovery mode works on any origin with /recovery.html.

When to use

• Live doctor session — user can open /recovery.html and debug port; app broken or slow (DOCTOR.md).
• Inspect, extract, dump, or forensically analyze a Composer profile (offline).
• Debug data loss, corruption, or unexpected state on main.composer.space, labs.composer.space, or preview deploys.
• Offline analysis of identity, spaces, feeds, objects, automerge documents.

Safety

1. Read-only by default — copy blobs out; do not modify Chrome profile files unless asked.
2. Live profile changes require user approval — never run compactDocuments, reset, import, or other writes via debug port without explicit confirmation (DOCTOR.md).
3. Consistency — close Composer tabs before extraction when you need clean integrity_check.
4. Privacy — extracts and reports may contain keys and user content; keep under /tmp; never commit.

Pipeline (always in this order)

locate → extract → validate → probe → (automerge …) → record in MEMORY.md

1. Locate

python3 .agents/skills/composer-forensics/scripts/locate-origin.py \
  --origin https://main.composer.space

2. Extract

python3 .agents/skills/composer-forensics/scripts/extract-opfs-sqlite.py \
  --opfs-dir "<opfs_pool_dir from locate>" \
  --out /tmp/composer-forensics/main.composer.space

3. Validate

bash .agents/skills/composer-forensics/scripts/validate-extract.sh \
  /tmp/composer-forensics/main.composer.space/DXOS.sqlite

4. Probe (JS — uses @dxos packages)

export PROTO_HOME="$HOME/.proto" PATH="$PROTO_HOME/shims:$PROTO_HOME/bin:$PATH"
node .agents/skills/composer-forensics/scripts/probe.js \
  /tmp/composer-forensics/main.composer.space/DXOS.sqlite

5. Automerge — find largest doc

cd .agents/skills/composer-forensics/scripts
node automerge-list.js /tmp/composer-forensics/main.composer.space/DXOS.sqlite

6. Automerge — binary vs JSON size (perf debugging)

node automerge-inspect.js /tmp/.../DXOS.sqlite --largest
node automerge-inspect.js /tmp/.../DXOS.sqlite <document-id>

High binary / JSON ratio + high ops / MiB usually means history bloat: storage and load cost far exceed reified document size.

7. Automerge — mutation analysis

node automerge-inspect.js /tmp/.../DXOS.sqlite <document-id> --mutations

Decodes all changes and reports op action breakdown (dominant set ops → whole-array replacement pattern).

8. Automerge — escalate to maintainers

node automerge-escalate.js /tmp/.../DXOS.sqlite --largest --out-dir /tmp/am-escalation

Produces <document-id>.bin (merged binary) + <document-id>-report.md (stats, hypothesis, repro steps) for Automerge issue reports.

9. Automerge — bench load

node automerge-bench-load.js /tmp/composer-forensics/main.composer.space/DXOS.sqlite --largest
node automerge-bench-load.js /tmp/.../DXOS.sqlite <document-id>

Composer recovery mode (in-app)

When Composer cannot boot (e.g. Automerge bloat), open /recovery.html on the same origin.

Doctor workflow: see DOCTOR.md — user opens Open Debug Port; agent uses composer-recovery.js; maintain report under /tmp/composer-forensics/reports/.

Default: static dxos globals only (dxos.Filter, dxos.Obj, dxos.DXN, …) — no client, plugins, sync, or indexing.

After Boot, dxos.client, dxos.spaces, dxos.halo, dxos.exportProfile(), dxos.recovery.compactDocuments(), etc. match devtools hooks.

Debug port workflow (one-shot — default)

User opens debug port first. Agent does not start or control the user's browser.

No persistent server. Browser polls; agent runs one CLI command per eval.

1. Open /recovery.html → "Open Debug Port" (copy session id from log)
2. node composer-recovery.js --session <uuid> '<js snippet>'  (starts, delivers, prints, exits)
3. Repeat step 2 for each command (browser keeps polling)

cd .agents/skills/composer-forensics/scripts
node composer-recovery.js --session <uuid> 'return dxos.recovery.status()'
node composer-recovery.js --session <uuid> 'await dxos.recovery.boot(); return dxos.spaces?.()'

• stdout — JSON result payload (ok, result / error)
• stderr — progress (Queued, Delivered, One-shot mode — waiting…)
• Exit code — 0 on success, 1 on eval error or timeout
• COMPOSER_RECOVERY_CONNECT_TIMEOUT — ms to wait for browser poll (default 6000, ~3× reconnect interval)
• COMPOSER_RECOVERY_TIMEOUT — ms to wait for eval result (default 120000)
• --interactive — persistent REPL when you need many commands without re-running CLI

Mixed content / HTTPS: CSP cannot override mixed-content. On https:// origins the page fetches https://127.0.0.1:9321:

mkcert -install
mkcert -cert-file .recovery-tls/cert.pem -key-file .recovery-tls/key.pem localhost 127.0.0.1
COMPOSER_RECOVERY_HTTPS=1 node composer-recovery.js --session <uuid> 'return dxos.recovery.status()'

Export/Reset/Boot work without the debug port. Offline forensics on exported SQLite always works.

See LINEAR-tagindex-write-amplification.md for the TagIndex bloat recovery path.

Workflow checklist

Doctor (live): DOCTOR.md checklist.

Offline forensics:

Forensics progress:
- [ ] locate-origin.py
- [ ] extract-opfs-sqlite.py
- [ ] validate-extract.sh
- [ ] probe.js (summary)
- [ ] automerge-list.js (or `automerge list`)
- [ ] automerge-inspect.js for binary vs JSON ratio on slow/large docs
- [ ] automerge-inspect.js --mutations when ratio is high (check op breakdown)
- [ ] automerge-escalate.js if escalating to Automerge maintainers
- [ ] automerge-bench-load.js for slow doc candidates
- [ ] `/recovery.html` if app won't boot — export SQLite before reset
- [ ] `composer-recovery.js` + Open Debug Port for live agent commands
- [ ] MEMORY.md updated; promote findings to LINEAR doc if filing an issue

Known issue pattern: TagIndex write amplification

High binary / JSON ratio (e.g. >50×) with dominant set ops on a small reified doc usually means TagIndex whole-array replacement — see LINEAR-tagindex-write-amplification.md for root cause, evidence, and fix plan.

scripts/src/ modules

Use src/, not lib/ — repo .gitignore ignores lib/.

Architecture

Scripts

Probe package: @dxos/composer-forensics in scripts/package.json (workspace; run pnpm install from repo root).

Additional resources

• DOCTOR.md — live recovery / doctor workflow
• reports/REPORT-TEMPLATE.md — session forensics report
• COMMANDS.md — every command documented
• STORAGE.md — Chrome on-disk layout
• VALIDATION.md — SQL templates
• MEMORY.md — session notes
• LINEAR-tagindex-write-amplification.md — Linear issue draft (root cause + fix plan)