Ambiguity Handling · Foundry docs

Ambiguity HandlingAmbiguity Handling

Information

Answer a question from your phone or chat, so a long job never sits waiting, with Production-Grade.

Foundry is a deterministic system: it never guesses. When agent output is ambiguous, incomplete, or deviates from the expected format, Foundry responds with a fixed escalation path rather than proceeding silently.Escalation by Condition

Condition

Response

Agent output asks a question

Generate clarification prompt → retry (up to max_retries)

Agent proposes alternatives

Generate clarification prompt → retry

Acceptance criteria partially met

Helper Agent → Interrogation → retry

Output format deviates from schema

HALT immediately

Required artifact missing

Helper Agent → Interrogation → retry

Retry Mechanism

Validation failures trigger automatic retries up to retry_policy.max_retries.

Ambiguous output triggers a clarification prompt on retry: the agent is told exactly what was missing or wrong.

After max retries are exhausted → task is marked BLOCKED.

The supervisor continues to the next task after blocking: it does not halt the entire pipeline on a single task failure.

Critical Halt ConditionsThe supervisor halts the entire pipeline immediately on:

CURSOR_EXEC_FAILURE: Provider CLI execution failed (process-level error).

BLOCKED: Provider explicitly reported a blocked status.

OUTPUT_FORMAT_INVALID: Output schema mismatch after retries.

RESOURCE_EXHAUSTED: Provider rate limit hit: see Resource Exhaustion for backoff details.

Design IntentFoundry never auto-resolves ambiguity by making assumptions. The operator sets intent via tasks and acceptance criteria: if that intent is unclear to the agent, the system surfaces it rather than inventing an answer. This is strict determinism in practice.