Generation telemetry & cost

Every render leaves a machine-readable receipt: what route ran, how many shots it sent, what the provider charged, how long it took, and how many clips came back. Over time those receipts teach vclaw what your videos actually cost — so the price quote stops being a guess and becomes your own history.

What it does

~$0.27

per Seedance scene, clip ≤10s

static default

~$0.40

per Seedance scene, clip ≤15s

static default

~$0.53

per Seedance scene, clip longer

static default

20

samples feed the historical average

completed seedance-direct runs override the static line

$0

to plan / dry-run

dry-runs are never used as cost samples

• Records a generation.telemetry.recorded event after every submit and every poll — routeId, operation, status, dry-run flag, task count, and the scene indices touched.
• Captures the run's config: aspect ratio, resolution, audio on/off, output count, average clip duration, prompt word count, and reference image/video/audio counts.
• Extracts provider-reported numbers when the provider exposes them: cost.usd, cost.creditsCharged, and generationTimeSec — pulled out of the raw provider response automatically.
• Records outputsIngested after a poll so you can see how many clips each job actually returned.
• Logs issues on blocked, failed, or warning runs — your failure ledger, not just your success ledger.
• Feeds cost-estimate: once real completed seedance-direct runs exist, the estimate switches from static defaults to your own historical average per scene.
• Dry-runs are recorded too, but are never used as cost samples — only live completed runs count.

How to use it

All examples use node dist/cli/vclaw.js ...; the installed binary is the same command as vclaw ....

node dist/cli/vclaw.js video cost-estimate --project my-spot --scenes 6 --clip-duration 10 --narration on

Prints a JSON cost + wall-time estimate for the project (scene count and clip duration are inferred from the storyboard if you omit the flags).

node dist/cli/vclaw.js video cost-estimate --project my-spot

Same estimate with everything inferred — scenes and durations from storyboard, narration from the brief's execution profile, new-character count from the cast.

node dist/cli/vclaw.js video metrics --root . --mode storyboard

Prints portfolio-wide metrics (completion rate, average score, review states, ops status, platform/priority/due-risk breakdowns) across every project under the workspace.

node dist/cli/vclaw.js video execute --project my-spot

Live submit — appends a telemetry event with the route, config, and any provider cost the response carries. Use Execution & runtime to drive the full run.

node dist/cli/vclaw.js video execute-status --project my-spot

Polls the job — appends a second telemetry event with outputsIngested, final status, and any provider-reported credits or USD.

Telemetry is read back through the library helpers readProjectGenerationTelemetry, readPortfolioGenerationTelemetry, and findHistoricalSeedanceCostTelemetry for integration code.

How it flows

Diagram source (live Mermaid)

Artifacts & outputs

• Telemetry ledger — appended to projects/<slug>/events/events.jsonl as generation.telemetry.recorded records (schemaVersion: 1). This is the longitudinal cost/timing/failure history; it does not replace the execution report.
• Cost estimate — printed to stdout as JSON. estimateSource is either static-default or historical-telemetry; when historical, a telemetry block reports sampleCount, matchedRouteId, averageSeedancePerSceneUsd, and lastRecordedAt.
• Portfolio metrics — vclaw video metrics prints the JSON VideoPortfolioMetrics object to stdout (not persisted).

TIP

The static estimate scales by clip length (about $0.27 for ≤10s, $0.40 for ≤15s, $0.53 longer per Seedance scene) and adds small line items for Gemini, Go Bananas new characters, and ElevenLabs narration. Once you have real runs, your own average overrides the Seedance line — the more you ship, the more accurate the quote.

WARNING

Credits (creditsCharged) are recorded but never auto-converted to USD — there is no implicit credit-to-dollar rate. Only completed seedance-direct runs with a positive cost.usd and a positive taskCount feed the historical estimate (last 20 samples); other routes' costs are logged but not used for estimation.

Driving it from an agent

cost-estimate and metrics emit pure JSON on stdout and exit 0 on success — parse them directly. cost-estimate throws (non-zero exit) if it cannot determine a scene count, so either pass --scenes or create a storyboard first. Telemetry is written as a side effect of execute/execute-status, so no separate step is needed to record it.

Related: Execution & runtime · Storyboard · Projects lifecycle · Seedance assets