Documentation | ConversationArc

ConversationArc troubleshooting

Start with the workflow paused when there is any risk of duplicate or incorrect delivery. Record the workspace slug, workflow ID, approximate time with time zone, and opaque correlation ID. Do not include passwords, agent keys, OAuth tokens, raw email lists, message bodies, or screenshots containing customer data.

Account and invitation problems

Symptom	What it means	Safe action
Invitation expired or already used	One-time link is no longer valid	Ask the inviting owner/operator to rotate and resend it
Login loops back to sign-in	Session is absent, expired, or revoked	Sign in again; if repeated, ask an owner to confirm membership is active
Second-factor code fails	The current authenticator or clock may not match	Check device time, use the current authenticator entry, then ask the platform operator for the audited reset process
`401 unauthorized`	Bearer key is absent, wrong, rotated, expired, or revoked	Create a new personal key from Team and replace the local environment value
`403 agent_key_scope_forbidden`	A read-only key attempted a mutation	Use a read/write key only if the task needs it
`403 forbidden`	The authenticated role cannot perform the operation	Ask a workspace owner to perform it or change the role intentionally
`404 not_found` for another tenant	Tenant isolation intentionally hides cross-workspace records	Switch to the correct workspace/key; do not retry with guessed tenant IDs

Workflow does not activate

Check the validation message, then verify:

the workflow is not archived;
the exact provider account is connected and healthy;
the trigger keyword, scope, destination, private copy, and public variations are valid;
the resource is ready and its destination passed verification; and
the selected provider capability actually supports the requested action.

New, duplicated, and rolled-back workflows require deliberate review. Archive is non-destructive and read-only; permanent deletion is a different exact-confirmation operation.

Instagram comment produced no message

Use Activity and Connections to distinguish these outcomes:

Outcome	Meaning	Safe action
`keyword_miss` or ignored	The event did not match the active workflow policy	Check whole-word keyword, media scope, account, and self-comment guard
`unknown_tenant` or account mismatch	The event cannot map to an active tenant/account	Recheck the connected external account; contact support if the mapping is correct
`reply_window_expired`	Meta's private-reply eligibility window passed	Do not retry a private reply; use an allowed public or off-platform path
connection unhealthy / reauthorization required	Credential or permission health check failed	Reconnect through Connections; never paste a token into support or a workflow
permission missing	Meta did not grant the action-specific scope	Complete or wait for provider approval; local retries cannot bypass it
duplicate / claim not acquired	The event already has a durable claim	Do not send again; inspect the original correlation
failed	Provider rejected a known request	Read the redacted reason and correct the account/policy issue before any new attempt
unknown / ambiguous	Transport ended without a trustworthy provider outcome	Pause the workflow, check Instagram directly, then use bounded recovery only if the provider state is known

An external tenant may complete OAuth but still lack production delivery while Meta Business Verification or App Review is pending.

Resource or waitlist problems

Destination verification fails: confirm HTTPS, redirects, DNS, and that the final page is the intended destination. Scheduled release remains blocked on a failed fresh check.
Feed creates no item: confirm the source is enabled, the URL returns valid RSS/Atom, and inspect the latest sync status/backoff. Existing imported items are matched by stable identity.
Release email stays queued: Delivery may be temporarily unavailable or retrying. The request is preserved rather than discarded.
A delivery failed: this is what the Help check "Resource delivery needs attention" counts. Open Resources, open the resource named in the check, find the failed item in its delivery history, then use Retry on that exact item to queue it again or Cancel to stop it. Retry returns only that one item to the queue and never re-sends anything already delivered; the check clears once no item is left failed.
Fulfillment is ambiguous: do not retry until the provider outcome is known. Use explicit retry or cancel only on the exact queue item.
Subscriber is suppressed: unsubscribe, bounce, or complaint suppression intentionally blocks future claims.
Duplicate signup: hosted waitlists deduplicate the same consented subscriber for the same resource; this is expected.

REST and MCP problems

Send Authorization: Bearer $CORRESPOND_MCP_TOKEN to both REST and MCP.
Use Content-Type: application/json for JSON mutations and MCP requests.
Personal keys are automatically tenant-scoped. Omit tenant_id in MCP unless a tool supports it and use only the key's own tenant.
Exact-confirmation fields must match byte-for-byte: for example confirm_workflow_id, confirm_source_id, confirm_email, or confirm_version_number.
A 409 generally means the current durable state conflicts with the requested transition; fetch the record again instead of forcing it.
A 400 invalid_request means ordinary request/schema validation failed; correct the named fields rather than retrying the same body. A 422 is used only by the narrower resource-template and style-policy validation contracts.
A 429 includes a retry boundary. Back off; do not parallelize retries.

See the complete customer surface in REST API and MCP.

Status and support

Start inside the product. Help in the workspace navigation opens /help, which checks this workspace against five conditions covered by this guide, links each one straight to the screen that fixes it, and lists the recent references you can safely quote in a report.

If it is still wrong after that, /help also holds the report contract and, for an owner, administrator, or support teammate, a one-click download of the redacted diagnostics bundle to attach.

Check https://conversationarc.com/status for customer-facing service availability and https://conversationarc.com/support for the current support intake state.

When configured, solving the Turnstile challenge reveals ThriveDesk's own Assistant contact-form widget, and the browser submits directly to ThriveDesk from there. Correspond never sees or stores the message. If the widget is not configured, the page says so plainly instead of pretending to accept a request it cannot deliver. The page never exposes a private operator address.

For support, include:

workspace slug;
workflow or resource ID;
approximate event time and time zone;
opaque correlation ID; and
the visible, redacted outcome.

Mark suspected cross-tenant exposure or credential compromise as SECURITY or PRIVACY and stop sharing further artifacts.