fix(openclaw): release the gateway binding when the status callback throws

[chughtapan]

Plan step 2. Based on origin/main, deliberately not on the #943→#951 stack: this bug is pre-existing in production, and per the plan's D3 the legacy path stays reachable in production until the OpenClaw cutover lands. This shortens that exposure window and merges independently of the endpoint chain.

The bug

startGatewayAccountEffect registers the account in activeClients before connecting, then ends with connectGatewayCore(...) and no cleanup.

The host's setStatus is arbitrary caller code. reportConnected invokes it inside Effect.sync, so a throw surfaces as a defect, not a failure — and connectGatewayCore's Effect.catchAll only observes failures. The defect escapes with the binding still registered and the core never disconnected: the account is wedged with no route to recovery.

The fix

Two changes, same defect class:

1. Release the binding on anything escaping the connect path (Effect.onError, which observes defects as well as failures).
2. disconnectAndRemove uses Effect.ensuring instead of Effect.tap, so a failing disconnect still drops the entry. This matches disconnectCoreOnAbort, which already used Effect.ensuring — that inconsistency was the tell.

No public surface change: the diff adds and removes no exported symbol, type, or interface.

Non-vacuity

The new test fails against the pre-fix source with expected +0 to be 1 — closeCalls.count was 0, i.e. the core was never disconnected. Verified by reverting only openclaw-entry.ts to origin/main and re-running with the test in place.

It asserts binding release the same way the existing stopRemovesClient does (sendText → /not connected/i), since activeClients is private.

Gates

• pnpm nx run @moltzap/openclaw-channel:test — 67 passed / 7 files
• pnpm typecheck --skip-nx-cache — green
• pnpm lint — 0 warnings, 0 errors
• git diff origin/main | grep -c ActiveLegacyGateway → 0, and the diff touches exactly 2 source files

The ActiveLegacyGateway<Service> retype from the adapters worktree is deliberately dropped: openclaw-gateway-lifecycle.ts does not exist on origin/main, and every symbol it would retype is deleted by the OpenClaw cutover. Generated MODULE.md / src.mdx churn is line-number shift only.

ADR conformance (R-pass)

Governing outcome	Record	Owner	Binds this diff?	Verdict	Evidence
— none found —	n/a	n/a	n/a	CONFORMS (vacuously)	Discovered from docs/decisions/README.md on main (49 records). No accepted outcome governs internal error handling in the OpenClaw adapter's gateway binding. 20260723-lifecycle-rides-l3 concerns conversation lifecycle at L3, not adapter process lifecycle, and is partially-superseded.

Per R-pass rule 4, "no ADR governs this" is stated as a reviewable claim, not left as silence. The diff changes no wire field, tool catalog, public contract, or trust boundary — the categories ADRs in this repo govern.

🤖 Generated with Claude Code

https://claude.ai/code/session_01P76aaa1STr3WPZ3nDascta

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
moltzap	🟢 Ready	View Preview	Aug 4, 2026, 10:09 PM

💡 Tip: Enable Workflows to automatically generate PRs for you.

[chughtapan]

Closing as obsolete rather than folding into #974 — the bug this fixed is already gone there, by construction.

This patched connectGatewayCore and disconnectAndRemove so a throwing setStatus could not strand an activeClients entry. Both symbols were deleted by the OpenClaw cutover, and this branch conflicts with HEAD.

The guarantee survives without the patch. In #974:

yield* Effect.sync(() => activeHarnessClients.set(ctx.accountId, active));
yield* reportHarnessConnected(client, ctx).pipe(   // the setStatus call
  Effect.zipRight(...),
  Effect.ensuring(finishHarnessClient(activeHarnessClients, ctx.accountId, active)),
);

Effect.ensuring wraps the exact call that invokes setStatus, and the whole thing sits under Effect.scoped, so a throw releases both the map entry and the client. Same outcome, reached by deleting the machinery instead of patching it.

Merging this would resurrect deleted symbols to fix a bug that no longer exists. Reopen if you disagree.