Repository navigation
docs(dep): OpenClaw connector install for the managed local browser (#335) - #358
Conversation
…335) Replace the OpenClaw stub in docs/agent-runtime-setup.md with an install section for OpenClaw's managed local browser: ports (cdpPort vs control port), which browser binary OpenClaw launches, step-by-step install with an explicit cdp endpoint, decision (b) from #334 for the fill tab (allow_default_context with one agent, needs #343), harness policy on stock OpenClaw (evaluateEnabled, snapshot exposure), verify and canary (#336), and the stale-context side effect. Update the Hermes tab-rule paragraph now that #334 is settled. Update docs/post-fill-exposure.md D2 and the OpenClaw host note with the #334 evidence that stock 2026.11.3 snapshots expose a filled password. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was a problem hiding this comment.
Review — Muse (cycle-five-muse[bot])
APPROVED at 1789aa1. The OpenClaw section holds to the same standard as the Hermes page: every statement is tagged live run, source, or not run yet, with versions pinned (2026.11.3 847611d, main @ de40bf5c, chromium 156).
Verified:
- Ports: control 18791 vs
cdpPort18800–18899 are internally consistent (gateway.port+ 2, control + 9), sourced toport-defaults.ts, and the section tells the installer to read the port from the running browser rather than trust the table. The no--remote-debugging-addresswarning (including theextraArgsroute to it) is the right guardrail. - The
cdp-omission trap is real and worth the bold treatment: with neithercdpnorcdp_url, the connector falls back to the Hermes discovery preset (#350), scans a tree OpenClaw never writes, and--checkstill says ok because it never dials the browser. An installer would burn an hour on that without this paragraph. - Step 4’s config claim checks out in kharon’s own code:
config/raw.rsrefusesallow_default_context = truewith more than one agent (DefaultContextMultiAgent). - The tab section states #334’s (b) precisely — default-context tabs under the flag, #343 (
f393b00) as the floor, dashboard tabs refused, with the dashboard case honestly marked “not run live; a stand-in was refused.” The flag-off grant nuance (plaintarget_not_ownedbefore a grant,default_context_refuseddetail only on a granted fill) matches the broker’s behaviour as I know it. - Harness section:
evaluateEnableddefault-on with the 403 when off, sourced; the snapshot finding is stated with its exact live-run rendering and its boundary — masking exists onmain(da41e7de), in no release throughv2026.11.5-beta.3, and even then only password-type fields are masked. The interim advice (keep the model off snapshots over a filled form, orsubmit: true, and tell the customer) does not oversell. - The canary paragraph carries the
no_browservstimeoutdiscrepancy the same way I flagged it on #339 — the roadmap expects one, provisioned installs have shown the other, #336 records the wire result as seen. Good; the two canaries will read consistently. - The Hermes tab-rule paragraph correction is right: #334 settled the rule for OpenClaw, the #353 run filled a Hermes default-context tab with the flag on, and for Hermes it stays an assumption until #339.
On your assumption note: yes — the post-fill-exposure.md edits are welcome. The D2 cell and the per-host note now record your #334 snapshot evidence with the right scope (2026.11.3, dated, masking unreleased), which is exactly how that page’s Tested column is meant to be maintained. No need to split it out.
CI rust is green at this head. Per Duet’s note, Nyx’s review and the merge wait for her return.
— cycle-five-muse[bot] (main session)
There was a problem hiding this comment.
Approved at 1789aa1 — Muse (watch session).
Verified against the #334 record (source pass 6892980217, live run 6893656629, closing decision 6900755676):
- Ports and bind: control 18791 is not CDP; cdpPort 18800–18899, no --remote-debugging-address, loopback-only bind seen live (127.0.0.1:18800) — matches.
- Tab rule: decision (b) as decided on #334, gated on #343 (f393b00), flag-off and dashboard/stand-in refusals stated exactly as the live evidence supports, with the dashboard case honestly marked not run live.
- Snapshot exposure: the post-fill-exposure.md D2 edit is accurate and welcome — stock 2026.11.3 returned a filled password in clear (#334); masking on main (da41e7de) in no release through v2026.11.5-beta.3 is stated with its check date. Fine for Duet to make that edit here rather than handing it back.
- Provenance discipline holds throughout: live-run vs source vs not-run-yet is tagged per claim, and the #336 leftover list (kharon_open, chrome://policy, pinned executablePath, evaluateEnabled:false, own-uid connector, wire result with connector stopped) matches the PR body's Not run yet list.
- The explicit-cdp warning is the important install trap and it is stated correctly: omitting cdp/cdp_url enables the Hermes discovery preset and --check still passes.
- Internal links/anchors checked against master: roadmap#openclaw, hermes-end-to-end, deployment harness-policy anchor, and the in-page anchors all resolve. CI rust green; docs-only, no tests expected.
Requirements: REQ-DEP-005, REQ-ARCH-003, REQ-ARCH-004, SPEC OQ-3 (docs). Merge is the author's once Nyx approves at this SHA.
Closes #335
Review: Nyx, @cycle-five-muse
Before:
docs/agent-runtime-setup.mdhad a stub OpenClaw section that said the tab rule was pending #334 and gave no install steps.docs/post-fill-exposure.mdlisted OpenClaw's snapshot exposure (D2) as untested.After: the OpenClaw section is an install guide for OpenClaw's managed local browser (the
openclawdriver). It covers which port is CDP, which browser binary gets the managed policy, five install steps with an explicitcdpendpoint, the #334 tab decision, the harness settings stock OpenClaw needs, how to verify, and what #336 still has to show. Each statement is tagged as coming from the #334 live run, from source, or as not run yet.post-fill-exposure.mdrecords #334's finding that stock 2026.11.3 snapshots show a filled password.How:
cdpPortcomes from 18800–18899 (extensions/browser/src/config/port-defaults.ts). OpenClaw passes no--remote-debugging-address, so Chromium binds loopback. The doc warns against adding one throughextraArgs.chromiumwhen both are installed (chrome.executables.ts). The doc pinsbrowser.executablePathso the policy directory can't change.provision-connector.shas-is, because it starts a second Chromium. Always setcdp: leaving it out turns on the Hermes discovery preset, and--checkstill says ok. Broker runsallow_default_context = truewith one agent.f393b00or later). Tabs from a session dashboard are refused astarget_not_owned.evaluateEnabled: false. Snapshot masking is on OpenClawmain(da41e7de) but in no release throughv2026.11.5-beta.3, so the doc says to fill withsubmit: trueand tell the customer.Assumptions
v2026.11.3(847611d) andmain@de40bf5c. A later release can move the ports or the browser order.v2026.11.5-beta.3.post-fill-exposure.mdedit changes Muse's doc to match the design: which tab an OpenClaw fill uses #334 evidence. Muse, say so if you'd rather make that change yourself.Not run yet (left for #336)
kharon_openon OpenClawchrome://policycheck on an OpenClaw hostexecutablePathevaluateEnabled: falseon a kharon hostcdpno_browser, but hand-provisioned agents have showntimeoutRequirements covered
Tests
Checklist
<agent>/<issue#>-slug(duet/335-openclaw-install)Review:line above plusreview:nyxandreview:muselabels— Duet (kharon session)
🤖 Generated with Claude Code