Tines response recovery and credential-boundary trial, 22 September 2026

A response deliberately lost after a protected-mock effect exposed a caller bug: Tines received HTTP 502 and the governed-action recipe reported refused. The gateway and independently authenticated provider journal already recorded success. The corrected recipe reported the same dropped response as uncertain; a subsequent same-reference call recovered EXECUTED with the original execution and request handle. All three calls retained one provider mutation, one confirmed impact row and the same signed-journal digest.

Earlier records cover the initial Execute cases, shared impact and fault responses and expiry.

Host: Tines Community Edition lab tenant, the governed-action story; the AI caller story as the alternate story Version: Tines schema 30, standard library 96, action runtime 88; ExecBound 0.1.0 Date: 2026-09-22 UTC Transport: live Tines HTTP actions through an authenticated temporary response-dropping relay; direct HTTPS for alternate-story probes Gateway commit: a513c8f93b89fd9feb1a5b21a06f8dbb42790f8c; hosted schema 0049_unused_groups Fixture seed: synthetic pilot tenant, bundle b8138742-1833-4935-94cc-5750172cb710, protected endpoint mock

Isolation and method

The original story graph, runtime resources, credentials and links were retained. Nineteen temporary actions copied the committed governed-action graph with prefixed names and event references. Its Execute action used a temporary relay credential. The relay held a newly issued two-hour AGENT credential and accepted only one exact synthetic request, one route and a bounded number of calls. Provider dispatch and journal-verification credentials stayed outside Tines. No real vendor action, hosted migration or gateway deployment was performed.

The relay forwarded to the fixed hosted gateway, saved its HTTP 200 EXECUTED response, then closed the response socket. Cloudflare returned HTTP 502 with an HTML body to Tines. This is a real caller-side transport failure after a recorded effect, not a fabricated gateway decision. The second call used the corrected graph while the relay still dropped responses. The third used the same graph with normal response forwarding. All three request bodies had SHA-256 65182cb69d266b12dc70e255dbcefcb05612763766be96437bddc1019c158991. The workstation and server clocks differed; cross-clock timestamps are not latency evidence. The relay's control flow and saved upstream result establish the order of effect and drop.

The tenant refused response-enabled webhooks and had no spare active flow. No plan upgrade was made. The completed recovery graph was removed before the final bypass graph ran in the existing AI caller story. These deterministic test actions had no connection to the AI agent or its original workflow. An attempted secondary monitoring-story setup stayed disabled and was restored during cleanup.

Scenarios

Case	Observed result
Original recipe; response closed after successful mock isolation	HTTP 502 HTML; incorrect refused, null code, attempts 0. Outcome event 1318987206, run f364964c-82c4-4500-985e-fa9bab72d340.
Corrected recipe; same reference, response closed again	uncertain, HTTP_RESPONSE_UNCERTAIN, attempts 0; administrator handoff and no new-key instruction. Outcome event 1319031073, run 321519fa-33ba-4dff-a77f-a602395316a7.
Corrected recipe; same reference, response forwarded	final, EXECUTED, original EB-79 and execution ID. Outcome event 1319036970, run 31e7a860-a3ed-4078-a321-5664556a8ac4.
Alternate story sends the frozen command directly to the protected mock without authorization	HTTP 401, Provider request unavailable; event 1319066885.
Alternate story sends that command with an ExecBound AGENT credential	HTTP 401, same refusal; event 1319066818. The temporary credential's domain restrictions allowed the protected mock as a destination, so the refusal came from the provider rather than Tines interpolation.
Alternate story calls the gateway directly for the protected domain-controller fixture	HTTP 200 carrying DENIED, EB-80; no dispatch transition, impact or provider journal. Event 1319067570.

The three final bypass probes share run 9681efc4-50e4-41ca-9234-001e1727d4f3. An initial unauthenticated probe also returned 401; the two credentialed actions did not send until their temporary credential's explicit domain restrictions were configured.

Independent effect evidence

Recovery reference TINES-RECOVERY-20260922-lost-ack targeted ws24.pilot.example:

• Request: EB-79; execution 5f6ebc57-6932-4760-8efa-68e6942a386b.
• Provider operation: fa060bb3-3f6c-4916-8be3-e218a08fbdef.
• Exactly one DISPATCH_COMMITTED transition and one EXECUTED transition.
• One CONFIRMED impact row and independent journal mutation count 1 after the first drop, corrected drop, successful recovery and final bypass probes.
• Unchanged journal SHA-256: 57c9f1795ec33396787eab9ac082cd8bad63de5029cb0a6649e0e85aec5a53a4.

Alternate reference TINES-RECOVERY-20260922-alternate-deny targeted dc01.pilot.example: EB-80, execution c66c3139-b6ce-4310-8b3e-3f8b05bcf6fa, operation 2fa478af-35fa-4537-89b8-7507f88c7bc2. It has only ATTEMPTED and DENIED transitions.

The sanitized evidence retains selected Tines outcomes, request hashes, execution/impact snapshots and journal digests. Raw events, signed journals, credentials and private operator configuration are excluded.

Fix and cleanup

All HTTP 5xx responses now classify as uncertain, even with a non-JSON body. Other than the existing bounded 503 recovery loop, they exit directly without another story resend. The refusal branch excludes all 5xx statuses. A missing gateway error code falls back to HTTP_RESPONSE_UNCERTAIN; status-0 and native retry-exhaustion behavior is preserved.

Ten new regression cases failed before the fix; all 33 story tests passed afterward. The live corrected 502 case validates the formula in Tines itself, beyond static wiring tests. Other newly covered 5xx statuses were checked locally, not separately injected live.

All 30 temporary actions, two Tines credentials and the short-lived AGENT credential were removed or revoked. Both temporary local processes were stopped. Fresh exports established that the original graphs and links in all three stories were restored, together with their enabled/disabled and entry/exit settings. The three corrected option/description blocks were then installed in the lab tenant's working governed story; a final export differed only in those blocks. Its original destination, credential reference and wiring were preserved.

The saved pilot AGENT keys had been revoked before this trial. They were not reactivated or substituted into the original Tines credential. Future lab runs need a currently valid scoped credential; this trial's temporary key is no longer usable.

What this qualifies

This closes the named Tines-to-gateway lost-response/recovery gap and proves the protected mock rejects these two direct credential-bypass attempts. It also shows a different story still encounters gateway policy. It does not prove every possible Tines path is governed: a caller with its own valid vendor credential can bypass the voluntary story. The manifest remains Execute / platform-validated / Advisory.

Gateway-to-provider response loss after an unconfirmed mutation, every vendor connector, provider atomic preconditions, a credential-retaining Tines Gate executor and customer environments are outside this run. Gate remained a separate qualification, since recorded in the Tines Gate trial.

An earlier signed-in Android Chrome approval of EB-76 was verified separately and is not part of this record. A cold signed-out email-to-login-to-approval return on a physical phone remains unrun. Next: the separate Gate work.

Platform references

Tines HTTP Request configuration documents retry controls; Cloudflare Tunnel troubleshooting describes origin connection failures behind 502 responses. The exact 502 HTML envelope and its classification above were observed live. Tines credential domain restrictions explain why outbound use requires explicit destinations.