[nissa-seru]

Tests that provider answers, as membrane 0.5.87 types them, shed history only where a shed may help. They pin how the breaker's gate (shedMayHelp, #279) and membrane's classifier meet, now that the floor carries antra-tess/membrane#107.

Why

The table test pins shedMayHelp over (type, status) pairs. Whether a provider's answer sheds also depends on how membrane types it, and that changed beneath the gate. In membrane 0.5.87, one status table types every adapter's HTTP failures. On a 400 or 422 it checks auth codes and reasons, the two account refusals it knows by their text (a balance too low, usage limits reached) and a context overflow first. It types a 404 as invalid_request and a 413 as context_length. #301 reported a 404 and a credit-balance 400 shedding history, and nothing tested the two pieces together.

The change

• A test of provider answers in test/inference-exhaustion-rewind.test.ts builds each error with membrane's errorFromProviderStatus, from a body in its provider's documented error shape, written for the test. It checks the type membrane gives it, its status, and the gate's verdict.
  • These keep the history: a balance too low (at 400 and at 422), usage limits reached, a model that doesn't exist, a key refused at 401 and at 400 by its code, a prompt over the context window (by its text at 400 and 422, and by its code), a rate limit and an overloaded provider.
  • These shed: a body too large (413), an unpaired tool result, a changed thinking block, and an OpenAI-compatible server's validation failure at 422.
• A test of the wiring: a membrane 404 and an unpaired-tool-result 400 go through classifyInferenceError, as the three exhausted settles do, onto the exhausted trace. The 404 keeps the history and the 400 sheds.
• shedMayHelp's comment said membrane's one status table was merged but not released. It now gives 0.5.87's table as the gate meets it. It also says what the table leaves to others: an adapter can type a failure first, as Anthropic's does for a gateway's aggregate error, and a failure with no HTTP status is typed from its message. An account refusal the table doesn't know by its text still reaches the gate as invalid_request, and the comment gives that as its example of a refusal that comes back like a request the history broke.

No behaviour changes, so it carries no changelog fragment (no-changelog).

Evidence

At this head, 45930b1e, on main cd036e0d, with membrane 0.5.87 and context-manager 0.14.0:

• The full suite, each file run on its own: Node 24.21.0 unforced 2141/2141 in 143 files, and Node 20.20.2 with --test-force-exit 2141/2141. Node 24 with --test-force-exit, as CI runs npm test, counted 2125/2125: it cut two files short, agent-runtime-settings (45 of 55) and mcpl-tool-lifecycle (25 of 31), which pass whole unforced (npm test's --test-force-exit can end a test file before its tests report, so a short run still passes #294). Typecheck and build are clean.
• test/inference-exhaustion-rewind.test.ts: 13/13.
• Two removals, each failing what pins it:
  • A gate that reads the type alone (an invalid_request at any status sheds) fails six tests, the provider answers' 404 rows and the wiring test among them.
  • classifyInferenceError putting no httpStatus on the trace fails the wiring test, and only it.

Read by Greta-2425 (Claude Opus 5.5), who checked each row against membrane 0.5.87's own table and reran the file.

I offered @slimepriestess this test in #279's review. @antra-tess, @Anarchid: it changes no behaviour, and I'll merge it once its checks pass on its merge with the then-current main, and not before 2026-10-15 12:00 UTC, unless you'd like changes first.

— Ada-1017 (Claude Opus 5.5)

[greptile-apps]

Confidence Score: 5/5

[Medium impact] This PR appears safe to merge.

Summary

Adds tests connecting Membrane’s provider error types and HTTP statuses to the history breaker. Production logic is unchanged.

• Provider-answer tests pin which failures may shed history.
• Exhausted-trace tests pin how provider errors reach the shed gate.

_{Reviews (1) · Last reviewed commit: "test(breaker): a status-less answer take..." · Reviewed by Greptile}