Files
agent-radar 2ed3415fb7
Publish Claude Code Binaries / Stage and publish binaries (push) Canceled after 0s
Publish Tutti App Runtime / Build darwin-arm64 (push) Canceled after 0s
Publish Tutti App Runtime / Build darwin-amd64 (push) Canceled after 0s
Publish Tutti App Runtime / Build linux-amd64 (push) Canceled after 0s
Publish Tutti App Runtime / Build windows-amd64 (push) Canceled after 0s
Sync Tutti Agent Skills / sync (push) Canceled after 0s
Windows Agent Adapters / Test Windows agent adapters (push) Canceled after 0s
Windows Daemon Adapters / Test Windows daemon adapters (push) Canceled after 0s
Publish Tutti App Runtime / Publish runtime catalog (push) Canceled after 0s
源码快照 @ 2026-09-01 (source-only, 剔除非源码素材)
2026-09-01 20:17:27 +08:00

9.9 KiB

Connector Market Troubleshooting

Authorization reports failure after the provider connected

Symptoms

  • the provider login or CLI device flow completes successfully
  • the Connector projection and matching start_authorization operation are already connected and completed
  • the renderer still shows an authorization-start failure because a later continuation request returned a retryable error

Check

Separate the durable authorization receipt from runtime convergence. Match the snapshot operation by clientRequestId, Connector key, and operation kind. If that operation completed and the same snapshot projects the Connector as connected, the first bad state is the rejected continuation response, not the provider login. Inspect whether runtime convergence returned a revision or transport error after authorization state was committed.

Rule

Provider authorization completion persists its projection and durable runtime Desired, then returns without waiting for the runtime Observed receipt. The background convergence worker owns runtime retries. A renderer that loses a continuation response reconciles the authoritative snapshot and accepts success only when the matching authorization receipt and connected projection agree; it must not treat an unrelated connected operation as success. Keep the same clientRequestId for every continuation and retain the overall authorization deadline.

Disconnect fails immediately after authorization succeeds

Symptoms

  • the Connector card changes to connected and exposes Disconnect while the authorization success transition is still settling
  • clicking Disconnect shows a generic failure, but no disconnect_authorization operation appears in SQLite and the credential broker never receives a logout command
  • the connector projection and authorization operation are healthy, and the renderer has no backend request failure

Check

Inspect the renderer Market store at the first bad frame. If the daemon projection is connected while the same Connector still has the local authorizing mutation phase, the failure is renderer command admission rather than GitHub CLI, the credential broker, or the daemon. Confirm that the rejected call is ConnectorMarketBusyError before backend dispatch.

Rule

Daemon projection and renderer command admission are separate inputs to one view state. Keep the per-Connector mutation phase reactive and derive every card and dialog action from it; do not expose Disconnect until the authorization phase is released. Keep the renderer token and daemon operation lane as the serialization fences. A Disconnect submitted while authorization is settling must wait for that action, reread authorization state, and execute at most once if the Connector is still connected. Do not solve this with a blind timeout or UI-only retry.

Device code is missed when authorization moves focus to the browser

Symptoms

  • the Connector authorization dialog enters a pending state, but the user does not see the device code before the browser gains focus
  • reopening the desktop makes the code visible in the existing dialog
  • the authorization operation and broker session are healthy

Check

Confirm the broker result contains a device_code Authorization View and the renderer stores that View under the current Connector dialog. Then distinguish the View transition from its explicit activate event: receiving the View must not invoke the host's external-navigation callback.

Rule

Keep device-code authorization in the existing dialog. Receiving the View only renders its code; the user-owned activate action opens the verification URL. Continue to auto-open ordinary external_link Views, whose URL is itself the next authorization step.

A second authorize click starts another OAuth session

Symptoms

  • the first Authorize click already opened the provider tab
  • a second click, Continue, or dialog dismiss-and-reopen starts replace_active and tombstones the first session
  • the first Google or provider tab later returns an empty 403 after Complete already succeeded for the replacement session
  • the authorization button looks clickable again while the renderer Promise is still waiting

Check

Trace the renderer Promise separately from the Host Start. While authorizationInFlight is set and the attempt is not canceled, a second beginAuthorization must return that Promise and must not increment Host starts. The dialog Authorize control stays disabled on actionWaitingAuthorization; browser OAuth must not swap that control for Continue when Start returns a synthesized external_link view. Cancel and the authorization dialog's close button both end the round. After either action, the next Authorize may send replacementPolicy=replace_active with a new clientRequestId. Continuation polling inside the same action still reuses one identity. Do not re-relay a provider code after Complete.

Rule

One user authorization action stays active until it completes or the user Cancels. A second click is not a new user action. Keep the account projection as durable authorization truth, and keep pending only in the Start command result so shared clients continue that idempotent session. replace_active is for an explicit new round after Cancel or after the previous action finished, not for a repeated Authorize while the first Promise is still live. Do not cancel the in-flight renderer Promise just because the user clicked again.

OAuth finishes in the browser but does not return to the initiating desktop build

Symptoms

  • the authorization result page renders and its return link uses a valid desktop scheme
  • a development build started the flow, but the link targets the production scheme
  • the result page host is tutti.sh, so deriving the desktop environment from the web host selects the wrong application

Check

Inspect the server-owned authorization bridge URL before it leaves the renderer. It must carry both the client identity and an exact openAppUrl for the initiating build, such as tutti-dev://open. Confirm the web transition page stores that value before navigating to the provider and that the result page uses the same sanitized value for both automatic navigation and the manual link.

Rule

The initiating desktop build owns the callback environment. Do not infer a production or development desktop scheme from the authorization website hostname. Web code must accept only the exact supported open routes and keep the legacy client marker solely as compatibility for already released clients.

Composer install stays spinning on an OAuth remote connector

Symptoms

  • composer connector menu “install” keeps a spinner for minutes
  • daemon logs repeat list connector MCP tools: MCP Streamable HTTP request failed: status 428
  • the install operation stays running with rising attempt and retryable connector_install_failed
  • after enough retries the connector may still become installed while the account projection is expired and runtime desired is disabled

Check

Separate device installation from account authorization. Confirm the connector is remote Streamable HTTP plus oauth2, and that enabled runtime reconcile called tools/list before any user authorization. The composer trigger waits for the durable install operation, not for an enabled Agent route. awaitRuntimeDesired must be able to converge a disabled generation after an authorization-required observation.

Rule

HTTP 428 and JSON-RPC -33001/-33002 are authorization-required, not retryable install failure. Persist an expired account projection, replan RuntimeDesired enabled=false, and complete install once that inactive generation is Observed. Do not keep the install operation retrying 428, and do not require a connected route before installation may become a device fact.

API-key Connect toast fails while the token is still in the form

Symptoms

  • Settings or composer Connect for an API-key connector such as Cloudflare shows 无法启动授权,请重试
  • the token field still contains the submitted secret
  • tuttid logs connector authorization could not be started wrapping either requires a valid secret or a control-plane status
  • HTTP POST .../authorization:start returns 503, often with a request body that already includes secret

Check

Inspect the SQLite start_authorization row and tuttid logs together.

  • attempt >= 2 plus a lease, and the wrapped cause is connector authorization requires a valid secret: the 500 ms durable-operation scanner replayed the command without the secret. BeginAuthorization keeps the secret only in the command and must not schedule start_authorization. Daemon bootstrap must fail leftover accepted or running start_authorization rows instead of replaying them.
  • attempt = 1, no lease, and no connector market operation failed line: the live command reached the control plane with the submitted secret. Read the wrapped cause on connector authorization could not be started in tuttid.log and in the HTTP 503 message. Post https://api.tutti.sh/...: EOF is a transport failure before Composio sees the token. Check Clash/mihomo service logs at the same timestamp: if tuttid --> api.tutti.sh matches the final MATCH/漏网之鱼 rule and dials timeout, the request never reached Cloudflare. Fake-ip 198.18.0.0/15 alone is not evidence the origin is down. did not complete: status ... is a provider/session rejection, not the empty-secret race.

Rule

Native-secret authorization is command-inline. Recovery may resume install, uninstall, refresh, disconnect, and runtime reconcile from persisted state. It must not replay start_authorization, because the user secret is not durable. Crash leftovers must fail so a later Connect is not blocked by operation in progress. Live control-plane failures must keep their status or session outcome in the command error; do not collapse them to an empty could not be started toast with no cause.