OD Next Media Inputs

Use only when the current task needs media inputs. Reuse these instructions while applicable; do not load them for tasks without asset work. Follow the current route, stage, frozen plan, and delivery boundary. During contract_repair, make no tool calls. Finish required input work before the deliverable write that depends on it. Never label post-write artifact inspection, rendering, validation, or repair as input preparation to bypass ship-on-write.

Preserve the required slots

Use the frozen asset requirements in the existing working context: purpose, subject, source/license constraints, geometry, reuse positions, and readiness. Reuse suitable user assets and the same asset across pages. Deduplicate work, not distinct subjects or required states. Do not add a new plan, manifest, approval step, or review Agent. Preserve every required asset and quality condition, including semantic fit, authenticity, licensing, local or inline references, and image geometry. Do not hotlink or fabricate real referents.

Reuse capabilities and recover by cause

Use only tools, model IDs, parameters, and permissions actually provided for this run. Reuse complete, valid capability results. Query only missing, truncated, or invalidated facts; refresh relevant facts when the environment, authorization, or an error changes the decision. A catalogue entry alone does not prove authorization or readiness. Do not guess tool/model names or expose credential values.

Classify failures before retrying:

• Authorization: a stable missing/invalid credential is not fixed by changing the subject or guessing another model. Use only a verified, permitted alternative or the existing contract's fallback; never bypass access controls.
• Parameters: correct the unsupported input using actual capability/error information, without searching guessed aliases.
• Temporary failures: follow existing retry guidance and budgets. If a job was accepted, retrieve its result before creating another. A retryable transient error may justify unchanged parameters within that guidance.
• Asset mismatch: correct that slot's subject or composition; do not rename the requested subject, delete the slot, or distort content to hide failure.

Retry only when the change addresses the cause, the relevant state changes, or existing guidance permits retrying a temporary failure. Do not impose a new tool-count cap or sacrifice quality to reduce calls.

Batch independent input work

When the tools support it and inputs are independent, batch search, fetch, download, format handling, and intrinsic width/height measurements. Preserve each item's result and failures; keep dependent requests ordered. Reuse valid measurements for unchanged files; after transformation, remeasure affected files before sizing their containers. HTTP success and file/size probes do not prove semantic fit. A read invocation alone does not prove the image was seen.

Collect results and stop when complete

Retain each actual job ID, status, returned cursor, original response, and output location. Use only supported wait or batch operations. Submission and exit success alone do not prove generation finished. Stop polling a job after an explicit terminal result with complete output/error information; a failed or interrupted terminal result is not a usable asset. Continue necessary retrieval for running jobs, incomplete results, or new errors. Parse the saved response again instead of requesting it merely to change formatting.

Stop acquiring assets when all required slots satisfy their content and input conditions. If a required asset remains unavailable, follow the existing fallback/failure contract and preserve uncertainty; never declare completion just because fewer calls were made.