rea.run

2026-10-12

What changed in rea-agents 6.4.0 (operator guide)

Upstream published rea-agents 6.4.0 on 2026-10-11 (release tag). This is an operator-length guide: capability matrix, cited community timings, checklists, and compliance — not a changelog paste. Pin the version you intend; run doctor before trusting social screenshots.

Why operators should care

6.4.0 is where Android device work, official compact MCP schemas, and right-sized Evidence views become pin-able. Community Linux benches also show the catalog grew from 139 → 170 tools, so tools/list can get slower even when later jobs improve. Treat list latency and first useful finding as different metrics.

Capability matrix: 6.3.0 → 6.4.0

Area6.3.06.4.0 (2026-10-11)Operator note
Advertised tools (community list)139170 (+31)Source: discussion #1998
Android device lanePrior static; some fixture scans failedADB inspect / split-APK / logs / screenshots / lifecycle + ApktoolYour device/emulator only
JEBNot headlineCaller-started local JEB MCPBYO licensed JEB
MCP schemasFull; community proxies for capsOfficial REA_MCP_INPUT_SCHEMA_PROFILE=compactFirst-party for Kimi/Moonshot
Large Evidence / graphsOften full payloads in contextRetained Evidence discovery; summaries & selected viewsPair with token FAQ
JS analysisPriorOwned process + heap/timeout; retain facts/unknownsStage large ASAR
GhidraPriorImport language/compiler specs; seed before auto-analysisStill BYO

Cited startup numbers (do not invent). @CassieFox23 on 2026-10-12 published Linux medians of 5 timed runs after warm-up (6.4.0 and main tip 13e03e943). Tool-list seconds:

TargetToolsNode 22Node 24Node 26
6.3.01391.90 (1.88–1.91)1.59 (1.56–1.61)1.38 (1.37–1.42)
6.4.01702.16 (2.12–2.17)1.75 (1.72–1.80)1.55 (1.52–1.58)
main 13e03e9431701.78 (1.76–1.80)1.47 (1.46–1.50)1.30 (1.28–1.33)

Author deltas: 6.4.0 is 0.17–0.26 s slower on list than 6.3.0 across Node 22/24/26 while adding 31 tools; that main tip is 16–18% faster than 6.4.0 on list (0.26–0.38 s). Full methodology: /blog/rea-linux-startup-benchmarks/. If unreproduced locally, label community report, pending your replay.

1. Android ADB checklist

Release highlights: device inspection, split-APK, logs, screenshots, file transfers, lifecycle via ADB, Apktool decode (#1552, #1556).

  1. Install platform-tools; adb devices must show a device/emulator you authorize.
  2. Pin npx -y [email protected]; run upstream Android/ADB doctor checks.
  3. Prefer disposable emulators for untrusted APKs.
  4. Targets: apps you own, debug builds you built, authorized fixtures only.
  5. Deepen /docs/android-adb/; matrix /faq/what-can-i-analyze/.
  6. Compliance: ADB ≠ extract any store DRM.

In #1998 Node-24 first-job pass, Android scan failed on 6.3.0 (Coffee path-collision) while 6.4.0 succeeded (~2.03 s median reply; ~0.66 s job alone). Qualitative upgrade — not a license for arbitrary store packages. News hook (not star worship): @Z3nlyte.

2. JEB — BYO license

Caller-started local JEB MCP browses units and decompiles selections (#1536). REA does not ship JEB. License your copy, start JEB MCP per vendor docs, point REA at it, doctor-green before demos. Operator page: /docs/jeb/.

3. Compact MCP schema

Set REA_MCP_INPUT_SCHEMA_PROFILE=compact for size-capped clients such as Kimi/Moonshot (#1514, #1580, #1484).

export REA_MCP_INPUT_SCHEMA_PROFILE=compact
npx -y [email protected] doctor --json
# reconnect MCP client after env change

For curated domain groups keep /faq/mcp-domain-split/ — also /faq/mcp-schema-size/. Compact is compatibility, not permission to corrupt schemas.

4. Large Evidence without drowning tokens

Discover retained Evidence without loading every payload; summaries/selected views for JS graphs, layouts, dossiers (#1682+). Loop: identity → summary views → disk dumps → slices into context → /faq/token-efficient-rea/ · /faq/electron-asar-timeouts/. Never fabricate summaries to hide unknowns.

5. Setup mythbust

Setup registers MCP; may offer Hopper with approval; Ghidra/IDA stay BYO. REA does not ship a decompiler (@cai_smart; @Thomas3cnl) — /faq/setup-vs-providers/, /docs/mcp-clients/.

Compliance

  • ADB = your device/emulator only.
  • JEB/Ghidra/IDA = licensed BYO; no cracks or DRM bypass how-tos.
  • Compact/domain-split = compatibility tools.
  • Targets: own / authorized / open / clean-room — /compliance/, /faq/legal-boundary/.
  • Primary CTAs on rea.run; GH releases/discussions are footnotes.

Sources

Operator FAQ for 6.4.0

QuestionShort answerWhere
Does setup install Ghidra?No — BYO/faq/setup-vs-providers/
Does compact replace domain-split?No — different knobs/faq/mcp-schema-size/
Is ADB for any Play Store app?No — your device/authorized only/docs/android-adb/
Why is tools/list slower?170 vs 139 tools in #1998/blog/rea-linux-startup-benchmarks/

Pinning tip: prefer npx -y [email protected] (or a hash you trust) over floating latest when writing runbooks. Re-run doctor --json after every pin change and keep the JSON in the ticket.

Windows verification narrative: community lab rea-lab pins 6.4.0 + Ghidra 12.1.4 for doctor-green workflows before seeded challenges — still exercises/own/allowed CTF only.

Change log vs operator guide

The GitHub release body for 6.4.0 is long (tens of KB of PR links). Operators usually need five decisions: (1) pin or wait, (2) enable ADB lab or not, (3) set compact schema or not, (4) change Evidence prompting or not, (5) update client doctor checks. This page is organized around those decisions rather than every merged PR. Version index: /changelog/. Field-level breaks from 6.0 through 6.4.0 (absolute MCP paths, normalized_result, capture accounting, websocket_connections): /docs/breaking/.

Suggested pin policy

EnvironmentSuggestion
Teaching lab / CI smokePin 6.4.0 explicitly
Daily driverPin 6.4.0 until you replay #1998-style list timing on your box
Tracking mainRecord commit SHA (example tip in #1998: 13e03e943) and expect drift

When communicating upgrades to a team, attach: release URL, doctor --json excerpt, and whether REA_MCP_INPUT_SCHEMA_PROFILE is set. That trio prevents “it works on my agent” threads.