Skip to content
AI usage dashboard

OpenClaw native container sample and runtime database

Date: 2026-10-06; uncommitted 0.2.1 worktree. Official OpenClaw 2026.9.8 CLI made two real local-model calls through public agent –local in isolated rootless Podman; both exited 0. Schema 24 hot-transcript read-only parsing and full-registry discovery are implemented. Old-database upgrades, two Windows executable reads and package lifecycle checks passed. Other protocols, cold archives and gateway paths retain separate limits. Behavior: runtime database rules.

Distribution and references

  • Official npm metadata and tarball: dist SHA1 7246c389c7134d9c15622082772912c7082f68af, 13,320 files, unpacked 388,955,824 bytes. npm integrity: sha512-G+JkNUhtpDE3cXR4AEi2NyyG9fqI/T2WUSl8ZnR8AATH8Dh1kC3qYFL7wwPoZtgHiP/cszA86PEiE0PDysxb9Q==.
  • GitHub v2026.9.8 released 2026-10-03, tag commit fc23bc864e4553c2d215e479eeec47b67a0bf943. Actual CLI: OpenClaw 2026.9.8 (fc23bc8).
  • Public npm provenance claims build commit aa6008ad198ef99c43f9d89dbd01694708712974 and workflow .github/workflows/openclaw-npm-release.yml; build run. Decoded statement and subject SHA512 match dist integrity. Signature/transparency-log trust was not independently verified.
  • Actual dist/package-update-activation-recovery.mjs SHA256: e08dfc1fb3ba7962f9e01d7a6770117c6cd77406b3fa4f7ba94788a670030ba0. Its OpenAI normalizer matches fc23 and lacks aa6008 contextUsage. Parse using actual installed code/schema/responses without applying undeployed fields from the claimed commit. Mutable database app_version does not identify historical record versions.

WSL Debian, Node 24.21.0; official engines >=24.16.0 <25 || >=26.1.0. Image localhost/llm-usage-openclaw:2026.9.8 ID: 4ba2f82099166a8ac7bcf60b4c60c1488786bf1961b063f86957ec04e84b9bcb. Initial –omit=optional installation exited 1 due to Koffi/CMake. Preserve that log. Installing cmake/build-essential and restoring optional packages produced exit 0. Keep npm’s unapproved-install-script notices; tests did not bypass approval. Client source/distribution were unchanged.

Real runs and independent comparison

Successful WSL root: build/plan-final-push/openclaw-real-1791288304. Ordinary UID 1000, CapEff=0, default seccomp, no network/host mounts; owned containers reclaimed. Custom OpenAI-compatible provider connects only to the container loopback model service. Public local marker ollama-local contains no personal credentials. Configuration validation exited 0 first. All tools disabled, thinking off and telemetry off; no gateway/delivery or host-IDE changes.

Two runs used the same owned session through the public command:

Terminal window
openclaw agent --local --agent main --session-id <owned-session-id> \
  --model local-llama/qwen3.5-0.8b-local \
  --message 'Reply with the word OK. Do not call tools or read or change files.' \
  --thinking off --json --timeout 120

Model: Hermes-tested model, official Qwen3.5-0.8B BF16 GGUF, SHA256 9a7bed4041b7975e0f71fa34670d1e9025213bc92905ac0db75d36c4fa3fa623. llama-server used one slot, context 64,000, eight threads and maximum output 128. A transparent proxy recorded actual SSE usage and forwarded it unchanged. The client itself requested include_usage; no responses/usage were fabricated.

Run API full prompt API completion API cache read Native input Native output Native cacheRead
1 6,105 47 0 6,105 47 0
2 6,153 107 6,101 52 107 6,101
Total 12,258 154 6,101 6,157 154 Known subtotal 6,101

Native input excludes cache. cacheWrite defaults to zero, totalTokens is client-computed, and cost initializes to zero. Store two usage observations, uncached input 6,157, output 154 and known cache-read subtotal 6,101. Complete input/total/source total, cache write, reasoning, cost and underlying call count remain unknown. The API had two actual requests; two observations still do not establish general model-call semantics. Occurrence time uses message.timestamp request start, not DB write time.

Native agents/main/agent/openclaw-agent.sqlite: 733,184 bytes, SHA256 446b41b0eb2bbb105e553a5b0670c757f0d9ce8f86fa70af5e25afe94070b42e. user_version 24, primary role agent, owner main. Ten transcript events contain assistant usage at seq 5/8. session_entry_provenance=1, acp_owned=0, empty plugin/hook, harness openclaw, and matching session_key/agent. Global store schema 19, quarantine schema 2 and auxiliary files are excluded. No cold archive exists in this sample; it does not verify cold-archive reading.

Initial root 1791288181 used maximum output 16, receiving HTTP 200 and actual 6,107/16/6,123 usage, but CLI exited 1 with length/incomplete_turn. Retain the original error. Increasing output and using an independent container yielded two exit-0 runs. Failed tasks were not classified as zero usage or successful tasks.

Implementation and regressions

Repair discovery at agents/<agent>/agent. Full-registry checks covered environment precedence and six manual-root forms. Register each physical file once. Recheck old seq under live WAL instead of treating an unchanged file fingerprint as an unchanged database. Read-only transactions check schema/provenance, cold-archive manifests and hot rows together. Bound pagination/cancellation and TEXT/zstd to 4 MiB, isolate invalid rows, and commit cursors/events together. Isolate other schemas, external CLI copies, migrated sources, plugin/hook/acp and other protocols. Expose archive gaps; disappearing hot rows do not erase history.

openclaw_runtime_contract 11 and three prior format tests passed: permitted native record/API fields; discovery/environment precedence; unknown schema/owner; origin isolation; invalid buckets/computed contradictions; zstd/invalid types/bounded reading; pagination/WAL/mutable DB versions; concurrency/rollback; archive gaps/history; unknown default-zero buckets; duplicate identities/real conflicts. zstd/cold boundary samples are explicitly synthetic, without expanding real-source acceptance. Preserve the initial compilation failure from an incorrect parallel API and the full-check failure for Option.is_none_or under MSRV. Correcting actual API usage and MSRV 1.77.2 yielded passing checks.

Redacted test data: core/tests/fixtures/openclaw/real-2026.9.8. Only permitted fields: four native user/assistant records, API usage and source notes, with replaced IDs. Bodies/credentials/configuration/original DB are not committed. Redacted-record file SHA256: 68d4db4b60b40b817fd09f853d148ee4299a1506e1c01e8644ec1396e0492d69. API SHA256: 8edf09c8e6b699385d1466a796ead81511766702d5b5490497826dec236f4a6a.

Current packaged-build checks

All commands below exited 0; logs/first failures are under root build/plan-final-push:

  • Windows npm run verify: Rust 988 (core 887/app 101, eight ignored), frontend 21/scripts four. Markdown then covered 197 files; Svelte had no errors/warnings; fmt, Clippy -D warnings and frontend build passed.
  • Debian workspace: 95 groups/985 passes (core 887/app 98, nine ignored); Clippy and release/deb/AppImage builds passed.
  • Windows NSIS 12 checks (1791289684773), headless 11 (1791289665186), native WebView2/IPC 17 (1791289782343), receiver eight (1791290098384). Twenty startup measurements: P95 752.4 ms; zero owned credentials/installation-integration leftovers.
  • Linux root 1791289583346: nine groups/47 checks, deb round trips, read-only FUSE/GTK GUI/ release on exit, and Orca ten-language/five-page current focus/speech. ALSA null does not establish physical audibility; containers do not verify host login/logout or complete OS integration.
Artifact bytes SHA256
Windows exe 10,046,976 58c7c7cf1484f55645f11c2ff0cc780bb50fb340ae06115ec87442e6732362d3
Windows NSIS 3,961,563 f9b43e75ecf67ff92374410b8413aafa38b8a7cceb9a9951395ef512e1d43a73
Linux deb 5,497,396 099676397117791d99c44be53057ded1be44f5808822c7df8dc579a3b415c357
Linux AppImage 111,163,896 c10013873e31457a6984fc1366b039794359d3838b648660156e7d2679412b5a

openclaw-readback-1791289585 upgraded the old DB created by the prior Cline deb (SHA256 74cb3d8b8d99ac4cdebd47b3c29afd4782c5e909b13eec5c492180fb1f151d4a). Two old-package reads produced no events and format diagnostics. New-package reads yielded two observations and one instance/physical file/checkpoint, health ok/latest_fallback; repeats added nothing, with values above. Main DB/persistent source-file SHA256 stayed unchanged. SQLite read-only access may update SHM bookkeeping, so the shared-memory sidecar was excluded from the unchanged assertion.

openclaw-windows-1791289662154 twice read SQLite rebuilt from permitted fields using the actual Windows executable, with the same values/unknowns. This verifies Windows application reading, not Windows OpenClaw CLI. Full test-data field audit: 57 files/135 JSON objects, zero credential keys/private paths.

The same package reread M8’s ten sources plus Hermes/Cline/Qwen/OpenCode, all exit 0 with values/unknowns/existing gaps preserved. Aider/jcode/Continue root 1791290118; AtomCode 1791290116; gajae-code 1791290117; Crush/Hermes/Xum 1791290119; Goose/Roo/Cline 1791290120; Junie 1791290121. Product prefixes prevent same-second roots being mixed. OpenCode default/control roots: 1791290118/1791290135. Qwen SDK main/background and native-first/archived-copy selection stayed at two calls/ 16,423 tokens, cache read 3. Retesting did not expand other products’ acceptance.