Prompt Inspector and outbound calls

Prompt Inspector lets you see and control model traffic without pretending that a complex turn is one prompt.

The three views

View	What it shows	Is it authoritative?
Next Prompt	A live forecast of the next Narrator prompt based on the current draft and settings.	No. Later retrieval, planning, trimming, tools, or state changes can alter the real call.
Review Model Call	The complete model-visible request RoleCall has assembled for one logical call, before provider-specific wire serialization.	Yes for the reviewed logical call. No provider receives that call until it is approved.
Outbound Calls	The ordered record of logical calls and provider attempts that actually happened.	Yes. It is recorded at the model egress boundary rather than reconstructed by the UI.

What happens during a turn

A turn may have several phases:

1. Before narration can include Compendium retrieval, Story Director planning, and pre-turn mechanics.
2. Narration contains the initial Narrator call.
3. Interleaved work can contain tool results, Narrator continuations, mechanics questions, or other loops.
4. After narration can include Final Cut, Compendium capture, summaries, and post-turn mechanics.
5. Recovery or manual work can contain an explicitly retried, resumed, or user-started call.

Prompt Inspector never invents possible loop requests before they exist. The current durable runner pauses a resumable job on its first immutable review wave. Calls that become ready later in the same execution are still captured in Outbound Calls, but do not open another blocking wave yet. This prevents the monolithic turn worker from replaying an earlier successful call while phase-output checkpoints are being completed.

Calls that become ready together appear in one review wave. Send all ready calls applies only to the fingerprinted calls visible in that wave. A future runner revision will allow later loop and post-turn waves after every preceding phase can resume from a durable result checkpoint; the current UI does not claim that unsupported behavior.

Review actions

• Send atomically locks the exact payload shown and makes that encrypted request eligible for dispatch. The worker cannot wake between approval and the prompt lock.
• Edit and send atomically locks the edited payload and sends only those reviewed bytes.
• Skip is available only when that subsystem declares a safe fallback. Skipping a required Narrator call is not allowed.
• Stop turn cancels the owning operation. It does not approve later calls.

Closing the review, navigating away, refreshing, or opening another device does not count as approval. The job remains paused until an authorized account resolves it.

After Send or Edit and send, the prompt lock is authoritative for that logical call. A worker retry, hot reload, server reassignment, or randomized preset macro may rebuild a different candidate while reconstructing the surrounding turn, but it must dispatch the locked reviewed request rather than replacing it or asking for approval again.

Logical calls and provider attempts

A logical call is one request made by a RoleCall system, such as Story Director or Narrator. The managed router may make several provider attempts for that one call.

Outbound Calls keeps these levels separate:

• The call row identifies the RoleCall system, phase, assigned model, and final result.
• The provider chain shows attempts in order.
• Failed attempts are marked separately from the successful attempt.
• Provider codenames protect infrastructure details for ordinary accounts. Authorized operator views may show additional routing detail.

Provider fallback does not create a second Narrator call unless RoleCall actually starts a new logical invocation.

Token totals

Before dispatch, input size is an estimate because the provider has not reported usage yet.

After a response:

• A green outline means every displayed token total is exact provider-reported usage.
• A blue outline means at least one displayed token total is estimated.
• Unknown usage remains unknown. It is never displayed or added as zero.
• Input, generated, and total tokens are displayed separately. Generated includes the visible reply, hidden reasoning, and model-created tool-call content. Local estimates count that semantic content, not provider transport framing or metadata.
• Show what was counted expands a numeric-only audit for each call: canonical request, visible reply, hidden reasoning, and tool calls each show their character count and local tokenizer estimate. The generated local subtotal is the sum of those three output components.
• When a provider supplies an exact aggregate total, that total remains authoritative. The component audit stays labeled as a local estimate because most providers do not report a visible/reasoning/tool-call split, and their private or protocol-specific token accounting may differ.
• Systems that ran more than once expose both the individual call split and each call's count audit.

The turn total includes only calls linked to that turn. Tool execution that did not call a model is listed as a non-model operation and does not add model tokens.

Refresh and cross-device recovery

The review decision is attached to the durable job and user account, not to one browser tab. When a call is waiting:

• the scene shows that the turn is paused for Prompt Inspector;
• an authorized device can reopen the pending review;
• refreshing does not send, skip, or cancel it;
• a completed call remains in Outbound Calls after the streaming UI disappears.

If RoleCall cannot prove whether a provider received a request, the call is marked ambiguous and is not automatically retried. This avoids charging twice or generating two conflicting continuations.

A browser refresh is recoverable. A server-process loss after an earlier model call has succeeded is handled conservatively: RoleCall stops instead of regenerating that call. Full automatic continuation after process loss requires the remaining phase-output checkpoints and is not currently promised.

Privacy

Exact prompt bodies and user edits are encrypted for the owning account. History reads are bound to the authenticated owner and the selected chat, group, or job. Public activity logs contain safe summaries, not prompt text, credentials, provider keys, or private routing configuration.

The browser decrypts exact payloads only when you expand them. Token counts, phases, status, and safe provider codenames can be shown without decrypting the prompt.

Account setting

Prompt Inspector is an account preference shared by solo scenes, group scenes, and authorized devices. Browser storage is only an immediate display cache. The server preference remains authoritative.

Turning Prompt Inspector off affects future calls. It does not erase previous Outbound Calls history or approve a call that is already paused.