Claude Code. Any Model. The most powerful AI coding agent now speaks every language.
Error-shape probe — parasail and writer
Measured 2026-08-14. Bodies fetched live with a deliberately invalid bearer token; every verdict below comes from calling claudish's own classifier functions on the captured body, not from reading them.
parasail
HTTP status: 401
content-type:
application/json;charset=ISO-8859-1body:
Unauthorized. Invalid token.\nJSON.parsesucceeds: false (composed-handler catches; no crash path)extractProviderMessage→Unauthorized. Invalid token.\nisTerminalError→ true (remapped to HTTP 400, surfaced inline)surfaced line the user sees:
Parasail error (HTTP 401): Check API key / OAuth credentials. — Unauthorized. Invalid token.ensureAnthropicErrorFormat→{"type":"error","error":{"type":"api_error","message":"Unauthorized. Invalid token."}}isRetryableError(chain level) → true — the fallback chain advances to the next candidatein-stream retry ladder (3s/15s/30s) reachable: no — it is gated on HTTP 200 + stream format openai-responses-sse|devin; this row is openai-sse and answered HTTP 401
writer
HTTP status: 401
content-type:
application/jsonbody:
{"tpe":"fail.auth","errors":[{"description":"Authentication Error, Invalid proxy server token passed. Received API Key = sk-...0000, Key Hash (Token) =<redacted: sha256 of the invalid probe key>. Unable to find token in cache orLiteLLM_VerificationTokenTable","key":"fail.auth.generic","extras":"key"}],"extras":null}JSON.parsesucceeds: true (composed-handler catches; no crash path)extractProviderMessage→{"tpe":"fail.auth","errors":[{"description":"Authentication Error, Invalid proxy server token passed. Received API Key = sk-...0000, Key Hash (Token) =<redacted: sha256 of the invalid probe key>. Unable to find token in cache orLiteLLM_VerificationTokenTable","key":"fail.authisTerminalError→ true (remapped to HTTP 400, surfaced inline)surfaced line the user sees:
Writer error (HTTP 401): Check API key / OAuth credentials. — {"tpe":"fail.auth","errors":[{"description":"Authentication Error, Invalid proxy server token passed. Received API Key = sk-...0000, Key Hash (Token) =<redacted: sha256 of the invalid probe key>. Unable to find token in cache or `LiteLLM_VerificationTokenTable`","key":"fail.auth.generic","extras":"key"}],"extras":null}ensureAnthropicErrorFormat→{"type":"error","error":{"type":"authentication_error","message":"{\"tpe\":\"fail.auth\",\"errors\":[{\"description\":\"Authentication Error, Invalid proxy server token passed. Received API Key = sk-...0000, Key Hash (Token) =<redacted: sha256 of the invalid probe key>. Unable tisRetryableError(chain level) → true — the fallback chain advances to the next candidatein-stream retry ladder (3s/15s/30s) reachable: no — it is gated on HTTP 200 + stream format openai-responses-sse|devin; this row is openai-sse and answered HTTP 401
groq
HTTP status: 401
content-type:
application/jsonbody:
{"error":{"message":"Invalid API Key","type":"invalid_request_error","code":"invalid_api_key"}}\nJSON.parsesucceeds: true (composed-handler catches; no crash path)extractProviderMessage→Invalid API KeyisTerminalError→ true (remapped to HTTP 400, surfaced inline)surfaced line the user sees:
Groq error (HTTP 401): Check API key / OAuth credentials. — Invalid API KeyensureAnthropicErrorFormat→{"type":"error","error":{"message":"Invalid API Key","type":"invalid_request_error","code":"invalid_api_key"}}isRetryableError(chain level) → true — the fallback chain advances to the next candidatein-stream retry ladder (3s/15s/30s) reachable: no — it is gated on HTTP 200 + stream format openai-responses-sse|devin; this row is openai-sse and answered HTTP 401
Verdict
Both vendors are correctly classified as TERMINAL, and no retry budget is burned.
The claim in research.md — "the only degradation is that a failure message reaches
the user less prettily" — survives the examination, for reasons that are structural
rather than lucky.
Why the classification cannot go wrong here
Every load-bearing decision for these two bodies is made on the HTTP status, not on the body shape:
| Decision | Function | Input used | Result |
|---|---|---|---|
| terminal vs transient | isTerminalError |
status === 401 → true, first line of the function |
terminal for both |
| status the client sees | composed-handler step 7 | terminal → 400 invalid_request_error, message surfaced inline |
no "API error · Retrying" banner |
| chain advance | isRetryableError |
status === 401 → true |
next candidate is tried, as for any vendor |
| in-stream 3s/15s/30s ladder | settleResponsesStreamHead / settleDevinStreamHead |
gated on HTTP 200 and stream format openai-responses-sse / devin |
structurally unreachable: these rows are openai-sse and answered 401 |
So the specific failure the review was worried about — a terminal auth failure treated
as retryable and burning 48s of backoff — cannot occur for either vendor. It could only
occur for a body that arrives inside an HTTP 200, and that path is not wired for the
openai-sse stream format at all.
The body shape is never trusted anywhere
composed-handler.ts reads await response.text() and then attempts JSON.parse inside
a try/catch, so a body that is not JSON produces parsedErrorBody = undefined and the
code falls back to { error: { type: "api_error", message: errorText } }. Parasail's
reply is the interesting case precisely because its content-type lies — it says
application/json;charset=ISO-8859-1 over plain text — and claudish never consults the
content-type, so the lie costs nothing. Measured above: JSON.parse succeeds: false, no
throw escapes, and the rendered line is clean.
What is actually degraded, stated precisely
- Writer's surfaced message is the whole JSON document.
extractProviderMessagelooks forerror.message/message/error/error.detail/detailand Writer supplies none of them (its message lives aterrors[0].description), so the function falls through toJSON.stringify(body). The user sees ~450 characters including a key hash instead of one sentence. It is bounded —buildSurfacedErrorMessagetruncates at 600 chars — and it does contain the real reason, so this is legibility, not information loss. providerErrorTypeisundefinedfor both. composed-handler readsparsed?.error?.type || parsed?.type || parsed?.code; parasail is not JSON at all and Writer nests its code aterrors[0].key. Consequence is confined to telemetry granularity (report_error,recordStats) — no user-visible behaviour depends on it.- Parasail is arguably better than OpenAI-shaped.
Unauthorized. Invalid token.renders as a complete sentence; the control row (groq) rendersInvalid API Key. The "less prettily" concern applies to Writer, not to Parasail.
One thing worth knowing that the shape gave away
Writer's error names LiteLLM_VerificationTokenTable. Its OpenAI-compatibility endpoint
is a LiteLLM proxy, which means the odd {"tpe":…,"errors":[…]} envelope is Writer's
own auth gateway in front of it; body shapes past the auth boundary are likely to be
LiteLLM's — i.e. OpenAI-shaped. This is an observation from one measured response, not a
verified claim about the served path, and nothing in the catalog depends on it.
Recommendation
Keep both rows. Neither misclassifies, neither burns a retry, neither can crash a parse, and the degradation is one long line for one vendor.
Optional follow-up, not done here (it is a change to shared error code, outside the
"one data row" boundary this feature is being judged on): teach extractProviderMessage
about errors[0].description / errors[0].message. That is a two-line addition to the
existing candidates array in handlers/shared/anthropic-error.ts and would make
Writer's line read as one sentence. It should be judged on its own merits — the array
already carries a FastAPI-specific detail[].msg case, so the precedent exists — and
NOT bundled into the catalog fill.