Provider lane scheduling: auxiliary compression can starve large primary under TPM acceleration limits

[antra-tess]

Problem

Primary conversation and autobiographical compression share one provider key/admission surface but have no cross-lane scheduling. Under provider input-token-per-minute acceleration limits, frequent smaller auxiliary complete calls can remain healthy while repeatedly preventing a much larger pending primary stream from ramping.

Observed on a stopped/recovered Claude 3 Opus residence:

• ordinary primary projection: ~181k input tokens;
• smaller background compression complete calls succeeded continuously;
• every primary stream attempt received Anthropic 429 rate_limit_error: “maximum usage increase rate for input tokens per minute”;
• disabling only speculative production was insufficient because already-queued/demand work continued;
• holding compression with both speculativeProduction:false and maxSpeculativeL1s:0 removed the competing auxiliary traffic;
• compression and primary each appeared healthy in isolation, while the resident remained unable to speak.

No patient content is needed to reproduce the mechanism.

Failure invariant

The healthy small organ starves the speaking large organ, and neither lane reports the interaction.

This is potentially fleet-shaped for any residence with a large primary window and active compression on the same provider/account key.

Desired behavior

Introduce a shared provider-lane admission/scheduling contract, scoped at least by provider credential/account and model:

1. Once a primary request is pending, admit no new auxiliary compression/merge calls for that scope.
2. Let an already in-flight auxiliary call settle or cancel it according to explicit policy; then provide a bounded quiet/ramp window for primary admission.
3. Resume auxiliary work after primary settlement with bounded fairness so compression cannot starve indefinitely.
4. Preserve already-authored/pending maintenance work; yielding must not clear queues or fabricate completion.
5. Treat acceleration-limit 429s as provider-capacity/scheduling evidence, never poisoned conversational history.
6. Expose lane telemetry without raw content:
   • primary pending/active duration;
   • auxiliary deferred count/duration;
   • admission/release reason;
   • provider 429 subtype;
   • per-scope recent input-token usage/ramp estimate;
   • starvation/fairness counters.

The correct ownership boundary may be Agent Framework, the Host composition layer, or a shared Membrane admission controller; the contract must cover both AF primary stream and Context Manager auxiliary complete calls.

Acceptance cases

• Deterministic fake provider enforcing a token/minute acceleration rule: continuous 50–70k auxiliary requests must not indefinitely block a pending 180k primary.
• A pending primary causes later auxiliary requests to defer before provider dispatch.
• In-flight auxiliary settlement is explicit and observed.
• Primary success releases queued auxiliary work exactly once.
• Primary failure/timeout cannot leave auxiliary work parked forever.
• Multiple residences/keys do not cross-block unless configured to share the same provider scope.
• Restart does not lose a durable pending-primary/held-auxiliary decision or duplicate provider effects.
• Existing provider-cap/poison-history protections remain intact.

Falsifier / comparative check

A useful live negative control is another large-window residence sharing the same provider family/key with active compression. If it never reproduces under comparable traffic, the scope may be narrower than the general arithmetic described here; instrumentation should make that distinction observable rather than guessed.