fix(interface): a turn prompt is bounded as metadata at 16 KiB, so a large brief cannot be sent

[drewstone]

Summary

AgentTurnInput.prompt is boundedStringSchema — CONTRACT_MAX_STRING_LENGTH, 16,384 characters. That is the metadata bound. A prompt is not metadata: it is the content of the turn, the same category as the response text and the tool output that were both moved to the content bound in #311 and #312.

The result is that a manager cannot hand a worker a brief longer than 16 KiB. A brief that quotes a paper, embeds a patch, or carries a page of prior evidence is refused at the adapter boundary.

Observed

Measured 2026-09-11 in a live Discovery Lab run (zkp-frontier-20260911d) on agent-provider-tangle 1.1.10 and agent-runtime 0.214.0 — i.e. with both prior bound fixes already in:

retained provider execution requires reconciliation before replacement: caused by ZodError: [
  {
    "origin": "string",
    "code": "too_big",
    "maximum": 16384,
    "inclusive": true,
    "path": [ "prompt" ],
    "message": "Too big: expected string to have <=16384 characters"
  }
]

2 of that run's 10 child deaths. The director was transferring a 27 KB patch to an independent-checker child so it could rebuild and re-verify a measured result; the transfer could not be expressed as a prompt.

Where

• packages/agent-interface/src/environment-runtime.ts:58 — prompt: boundedStringSchema.optional()
• also runtime-control.ts:402, portable-context-continuation.ts:81, environment-interactive-control.ts:549

boundedStringSchema is CONTRACT_MAX_STRING_LENGTH (16,384). boundedEventContentStringSchema already exists for exactly this kind of value and is bounded at CONTRACT_MAX_JSON_BYTES (1 MiB).

Why this is the same class as #311 and #312

Three times now, agent-authored content has been held to the bound written for identifiers and accounting fields:

	field	was	now
#311	toolInvocations[].result	16 KiB metadata	content
#312	same, on throw-vs-truncate	discarded a paid turn	truncated
this	prompt	16 KiB metadata	?

The distinguishing test is simple and worth stating once in the contract: if an agent or a tool produced it, it is content; if it names, routes, or accounts for the turn, it is metadata. A prompt is produced for the agent, and its size is set by the work, not by the protocol.

Note on severity

Unlike #312 this one fails before the turn runs, so it does not discard paid work. It refuses the turn outright, which is the honest failure — but it silently caps what a manager can delegate, and the ceiling is invisible until a brief crosses it.

Suggested change

Move prompt to boundedEventContentStringSchema at the four sites above. Truncation is not appropriate here (a silently shortened instruction is worse than a refusal), so the refusal stays — it just moves to 1 MiB, where the producer's own limits already sit.