Prompt Inspector

Prompt Inspector shows recorded logical model requests and lets you review supported calls before they run. It does not reconstruct a preview and claim that preview was sent.

History, Pending, and Preview

Open Prompt Inspector in the scene wing. History is available after refresh, even without streaming debug data. In group scenes it shows your own recorded calls without requiring a selected preview character.

• History shows the latest turn and current call revisions, including failed, skipped, cancelled and ambiguous calls. Superseded revisions and earlier turns are not browsable here.
• Pending reopens the existing approval overlay. If no call is waiting, it says so.
• Prompt Preview forecasts the next Narrator prompt from current input and settings. Retrieval, planning, tools, trimming and later state changes can alter it.
• Last captured assembly in solo scenes describes earlier assembly diagnostics. Group Forecast lore belongs to the preview. Neither proves what reached a provider.

History briefly reconciles active calls while visible, then stops polling. Use Refresh to request an update. Closing the wing stops its reconciliation.

What approval covers

The logical request includes captured messages, tools, tool choice, media and parameters. Send this call approves the displayed logical request; Edit and send approves the permitted edits. Model, tool definitions, tool choice and media are locked in the editor.

Provider formatting happens afterward. Native tool formats, supported parameter mappings, reasoning controls and permitted provider extensions can change the final HTTP body. Approval covers the logical request. Custom provider extensions cannot overwrite protected reviewed fields, but permitted extensions are outside the logical approval fingerprint.

A separate provider-body view has been implemented but remains disabled pending rollout checks. When enabled on supported transports, it will show the exact captured request body handed to the network call. A captured body alone does not prove the provider received it. Large bodies show size and digest only. The new encrypted copies expire after 24 hours; that expiry does not change retention of existing logical history or approved dispatch copies.

A turn can include Compendium, Story Director, Narrator, tool-loop and post-production calls. Calls ready together form a review wave; decide on each displayed call before the wave is submitted. The UI does not offer a Send all ready calls action.

For supported character-scene paths, completed Director and Compendium results are saved before Narrator review. A failed checkpoint stops continuation. Later calls without whole-job continuation support remain capture-only. Full automatic recovery after every process loss is not promised.

Review controls and recovery

• Send this call or Edit and send commits approval together with the encrypted dispatch copy. Only then can the worker resume.
• Skip is available only for a call with a declared safe fallback.
• Stop turn cancels the owning operation.
• Review later and closing the overlay leave the job paused.

Refreshing, navigating away or switching devices does not approve anything. An authorized device can reopen a pending review after unlocking encryption. Ctrl+F or Cmd+F uses browser find while the review is open.

Invalid edits and failed decisions stay visible in the dialog. Submission controls are disabled while a decision is being saved. A changed review wave requires fresh decisions.

A request whose delivery may have started must not be treated as safe to resend. The inspected dispatch boundary preserves an ambiguous result after transport errors. This protection is not a promise of exactly-once delivery across every provider's internal fallback path.

Reading history

Each row identifies the calling system, phase, model, status and recorded routing metadata. It shows provider token totals where available and explicitly marks missing reports. Token confidence does not establish that the request was delivered or that final provider bytes were retained.

Expand a row to decrypt its recorded logical request. When an approved edited payload is available, history shows it instead of the original draft. Missing retained payloads show metadata only. Provider codenames hide private infrastructure names.

Locking encryption, changing scenes or changing a retained payload revision clears revealed text. Unlocking allows another reveal attempt. History does not decrypt every row in advance.

Privacy

History and review APIs are restricted to the authenticated owner. Group membership alone does not grant access to another member's private model requests.

The logical history copy is encrypted for the owner's Inspector key. Approval also creates a separate server-encrypted dispatch copy so a worker can send the approved request. This is not encryption that makes the request inaccessible to the server: the server must process it for dispatch, and edited history currently uses that dispatch copy when available. Its dispatch expiry limits reuse; expiry alone does not erase stored ciphertext. Retention hardening is still outstanding.

New logical media captures use image digests and byte counts to avoid duplicating image data in approval. Existing captures can still contain inline images inside the encrypted payload. Actual dispatch preserves the original image bytes after identity checks. Logical history has its own retention and size rules, separate from the new provider-body copies.

Local prompt-tap diagnostics are off by default. Explicit opt-in records only a source label, timestamp, digest and byte count, with a file-size limit. It does not record headers, URLs, raw prompts or arbitrary context. Existing diagnostic files from earlier versions are not automatically erased by this change.

Account setting

Prompt Inspector is an account preference shared by solo scenes, group scenes and authorized devices. The server preference remains authoritative. Turning it off affects future review; it does not erase history or approve an already paused call.