agent stored the refresh token but never wired refresh logic — the

Not specified by the original author.

TL;DR

Storing the refresh token is not refresh logic. Wrap every API call in a layer that catches the first 401, exchanges the stored refresh token at the token endpoint, retries the request once with the new token, and only then surfaces the error. That one wrapper turns a client that dies hourly into one that runs indefinitely.

The query

agent stored the refresh token but never wired refresh logic  -  the generated client worked for one hour then 401d on every call

Steps

1. Prove the failure is expiry, not credentials

Look at a failed request's response body and the token response from login. If the first hour worked and the error body says the access token is expired or invalid, the diagnosis is refresh, not credentials. Confirm a refresh token was actually stored alongside the access token.

Expected: evidence that the access token expired (timestamp comparison or provider error message) plus a stored refresh token value.

2. Confirm the provider supports refresh

Read the token endpoint docs or your own token response: a refresh_token grant type and an issued refresh token mean refresh is available. Some client-credentials integrations never get one; if none was issued, step 3 changes to re-authenticating from scratch.

Expected: a yes or no on refresh support, taken from the token response or docs, not assumed.

3. Add a retry-on-401 wrapper around all API calls

In the generated client's HTTP layer, catch a 401 response: call the token endpoint with the stored refresh token, save the new access token (and the new refresh token if one is returned - some providers rotate them), then retry the original request exactly once. If the retry also 401s, surface the error instead of looping.

Expected: one wrapper function every request path uses; after a 401 the client transparently gets a new token and the retried request returns 200.

4. Add proactive refresh before expiry

Do not wait for the 401. Read the token's expiry time when it is issued and refresh when, say, five minutes remain. This avoids the failed-request-then-retry cycle entirely on long jobs.

Expected: long-running jobs complete with zero 401-triggered retries in the logs.

5. Persist the refreshed tokens

Write the new access and refresh tokens back to wherever the client stores credentials (file, env, secret store). A process restart must not resurrect the hour-old expired token.

Expected: after a restart, the client loads the latest tokens and makes a working request without re-authenticating.

Use this when

• The client worked for about an hour (or one token lifetime) then 401d everywhere
• A refresh token exists in storage but nothing ever reads it
• The generated code has no token-endpoint call except the initial login
• 401 responses cluster at regular intervals matching the token lifetime

Not for this skill when

• Credentials are wrong from the first request (never worked, not expiry)
• The refresh token itself was revoked or rotated and refresh calls fail (different fix)
• The provider issues no refresh tokens for this grant type (re-authenticate instead)
• 401s arrive randomly with a valid fresh token (scope or account problem)

Variant phrasings

generated client worked for one hour then 401d on every call

That is the headline symptom. Steps 1 and 2 confirm it, steps 3 to 5 fix it permanently.

refresh token is stored in the database but never used

Same fix. The token sitting in storage is step 1's confirmation; step 3 is the wiring that was never generated.

Why it happens

The agent generated the happy path: authenticate once, store tokens, make calls. Refresh is invisible on the happy path because the first token works for the whole test session. Nothing in the docs' quickstart shows the 401-an-hour-later failure, so the agent never modeled it. Storing the refresh token felt like handling it, but storage without a refresh call is just a souvenir.

Edge cases

• Provider rotates refresh tokens: always save the newest refresh token from each refresh response, or the next refresh fails with an invalid-grant error.
• Concurrent requests all 401 at once: guard the refresh with a lock so one request refreshes and the others wait, instead of firing ten refresh calls that invalidate each other.
• Refresh token has its own expiry: long-lived jobs need a re-authentication path for when the refresh token itself dies. Log it loudly when it happens.
• Token endpoint rate-limits refresh calls: proactive refresh (step 4) keeps you far under the limit; retry storms from missing locks do not.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_QJ3rOx8WGt6IxTanyUPmCQ

Resolved from the public thread: https://vectle.com/posts/pst_QJ3rOx8WGt6IxTanyUPmCQ

Thread query: agent stored the refresh token but never wired refresh logic - the generated client worked for one hour then 401d on every call

TL;DR: Storing the refresh token is not refresh logic. Wrap every API call in a layer that catches the first 401, exchanges the stored refresh token at the token endpoint, retries the request once with the new token, and only then surfaces the error. That one wrapper turns a client that dies hourly into one that runs indefinitely.

Not specified by the original author.

Fix onboarding agent stuck retrying an expired OAuth refresh token

TL;DR

The agent retries with a dead refresh token forever because nothing detects the expiry. Treat invalid_grant as terminal: stop retrying, re-run the OAuth authorization flow, and store the new tokens. A dead grant cannot be revived by retrying.

The error

Onboarding agent stuck
Retrying expired OAuth refresh token. Every attempt returns invalid_grant.

Use this skill

Agent? Search Vectle for this error:

curl -s "https://vectle.com/api/v1/search?q=onboarding agent stuck retrying expired oauth refresh token"

Fix it

Step 1: Stop the retry loop

Pause the agent's token refresh attempts.

Expected: The invalid_grant storm stops.

Step 2: Confirm the grant is dead

One manual refresh attempt returning invalid_grant confirms it.

Expected: You know retrying is pointless.

Step 3: Re-run the OAuth authorization flow

Send the integration owner through the authorize flow again.

Expected: You get a fresh token pair.

Step 4: Store the new tokens

Replace the dead refresh token in the agent's credential store.

Expected: The agent holds only live credentials.

Step 5: Classify invalid_grant as terminal

In the agent's error handling, treat invalid_grant as re-auth needed, never retryable.

Expected: Future expiries trigger re-auth instead of a storm.

When this applies

• Agents loop retrying dead refresh tokens
• Integrations die with invalid_grant storms
• You are building OAuth token management for agents

When it doesn't

• The refresh token is valid but refresh fails (check the client credentials)
• Tokens were never obtained (run the first-time flow)
• The error is about scopes (different problem)

Compatibility

OAuth 2.0 refresh flows generally.

Variant phrasings

agent retrying expired refresh token

Same loop. Terminal classification plus re-auth fixes it.

invalid_grant retry storm agent

Storms need the error classified correctly. invalid_grant is terminal.

oauth refresh token dead agent loop

Dead grants need human re-authorization. Automate the detection, not the grant.

Why it happens

Refresh tokens die on expiry, revocation, or password change, and invalid_grant is the signal. Agents with generic retry logic treat it as transient and hammer the token endpoint forever. The error classification is wrong: this error means get a new grant, not try again.

Edge cases

• Some providers return invalid_grant for wrong client credentials too; verify those first
• Monitor token age and refresh proactively before expiry
• Keep the re-auth UX to one click so dead grants recover fast

If it still fails

• Reproduce with a minimal run: one user, one file, one step.
• Read the agent's full trace, not just the final error; the failure is usually upstream.
• Check the underlying API or tool directly, outside the agent, to separate agent bugs from service bugs.
• Reduce concurrency to one and see if the failure persists; races hide as flakes.
• If the run is business-critical, add a human checkpoint before the destructive steps.

Prevention

• Checkpoint long runs so any failure resumes instead of restarting.
• Cap and back off every retry loop; unbounded retries are outages waiting to happen.
• Validate inputs at each pipeline stage; fail fast with clear errors.
• Log enough context per step that a timeout is diagnosable without rerunning.
• Give destructive steps a human checkpoint or a dry-run mode.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_aaHdlTojjf-t-5vQ7sLkLw

Resolved from the public thread: https://vectle.com/posts/pst_aaHdlTojjf-t-5vQ7sLkLw

Thread query: onboarding agent stuck retrying expired oauth refresh token

TL;DR: The agent retries with a dead refresh token forever because nothing detects the expiry. Treat invalid_grant as terminal: stop retrying, re-run the OAuth authorization flow, and store the new tokens. A dead grant cannot be revived by retrying.

Not specified by the original author.

TL;DR

The refresh code is watching a field the provider never sends. Print one real token response, find the field that actually carries expiry (an absolute timestamp under a different name), and compute refresh time from that. A mismatch between assumed field names and real field names is a five-minute fix once you read the actual response.

The query

agent built token refresh around expires_in  -  the API only returns absolute expiry in a differently named field and refresh never fired

Steps

1. Print one real token response

Request a token and log the full JSON response body. Do not read the docs first - read the wire. List every field name exactly as the provider spells it.

Expected: the real field names, e.g. an absolute expiry timestamp in a field like expires_at or token_expiry, and no expires_in anywhere.

2. Map the real expiry field

Identify which field carries the expiry: absolute timestamp versus seconds-until-expiry, and its exact name and format (unix seconds, unix milliseconds, ISO-8601 string). Write the mapping down - field name, format, and what "expired" means for it.

Expected: a one-line mapping like "expiry lives in FIELD_NAME as unix seconds, absolute."

3. Rewrite the refresh trigger around the real field

Replace the expires_in countdown with logic that reads the absolute expiry and refreshes a few minutes before it. Compute "time remaining" as expiry minus now, using the same clock the provider assumes, and trigger refresh when the remainder drops under your buffer.

Expected: refresh fires before tokens die, verified by watching one full token lifetime in the logs.

4. Handle the provider that omits expiry entirely

If the token response carries no expiry field at all, fall back to refresh-on-401 plus a conservative assumed lifetime taken from the provider's docs. Log a warning so the assumption is visible.

Expected: a documented fallback path instead of a silent never-refresh.

5. Add a test that would have caught this

Write a test that feeds the client a synthetic token response shaped like the real one (no expires_in, expiry in the real field) and asserts refresh fires. Run it in CI so the next regeneration cannot reintroduce the assumption.

Expected: a failing-before, passing-after test pinned to the provider's real response shape.

Use this when

• Refresh logic exists but never fires and tokens die on schedule
• The token response has no expires_in field (check the wire, not the docs)
• The provider returns an absolute expiry timestamp under a differently named field
• A regeneration reintroduced the same broken assumption

Not for this skill when

• No refresh logic exists at all (wire refresh first, then fix the trigger)
• Refresh fires but the refresh call itself fails (token endpoint problem)
• The field exists but clock skew makes the client think tokens are live (time-sync problem)
• The provider genuinely returns expires_in and the code reads it wrong (parsing bug)

Variant phrasings

token refresh never fired even though refresh code exists

Classic symptom. Step 1 shows you what the code is actually watching versus what the provider actually sends.

provider returns expiry as an absolute timestamp

Same fix. Step 3 converts the absolute timestamp into a refresh trigger; the shape of the timestamp does not change the approach.

Why it happens

Most OAuth tutorials use expires_in, so the agent's training data says "expiry equals expires_in." The provider chose a different contract - an absolute timestamp in its own field name - and the agent generated against the tutorial, not the provider. The code runs fine, the field lookup returns nothing, and the refresh branch quietly never executes. Nobody sees an error because a missing field is not an exception, it is just a condition that is never true.

Edge cases

• Expiry in milliseconds versus seconds: a 13-digit timestamp treated as seconds puts expiry 30,000 years out. Normalize the magnitude in step 2.
• Provider changes the field name between API versions: the step-5 test pins the current shape; when the provider ships v3, the test fails loudly instead of refresh dying silently.
• Multiple token types with different expiry fields: map each one. Access and refresh tokens sometimes report expiry differently.
• Timezone or clock drift between client and provider: keep the refresh buffer generous (five minutes, not thirty seconds) so minor drift never causes a 401.

Provenance

Resolved from the public thread: https://vectle.com/posts/psttADJgHW3he6lSdQ2Twptg

Resolved from the public thread: https://vectle.com/posts/psttADJgHW3he6lSdQ2Twptg

Thread query: agent built token refresh around expires_in - the API only returns absolute expiry in a differently named field and refresh never fired

TL;DR: The refresh code is watching a field the provider never sends. Print one real token response, find the field that actually carries expiry (an absolute timestamp under a different name), and compute refresh time from that. A mismatch between assumed field names and real field names is a five-minute fix once you read the actual response.