JevCalc

Explainable agent task cost simulator

Agent cost simulatorIMPORTED RECORDS
Sanitized repository review · 12 requests · recorded costs · Privacy details
Local logs · JSON / JSONL / CSV

Review this task’s execution

Review suggested routes, adjust them in the canvas, and replay the same recorded tool path.

Query preview · browser memory only · up to 4,000 characters
Suggested routes · manual overrides available · 0 model calls made
Model assignments · 9 Jev / 3 LLM scenario calls

Suggested routes are ready to replay. Override any tool decision below or in the canvas. Jev assumes bounded choices with code-supplied arguments; final responses keep the recorded model.

2 recorded request(s) · Script execution, writes, edits or unrecognized tools may need generated arguments. Keep the LLM; tool names alone do not prove Jev suitability.
9 recorded request(s) · One successful read-only tool call: try Jev for bounded action selection, assuming arguments are supplied by code. This does not establish equivalent quality.
1 recorded request(s) · Final response stays on the LLM. Free-text generation is not a bounded choice.
12 requests · 3 adjacent metadata group(s)
Scenario settings · illustrative Jev prices and token budgets

Defaults reuse the simulator’s illustrative Jev tariff, not live prices. LLM costs come from the log; original model identities are not inferred. Changing these fields updates the estimate immediately.

Jev input = rounded-up recorded input × input share + criteria tokens. Fallback charges both Jev and the original LLM request, with unchanged LLM prompt/cache cost and no escalation overhead. Downstream context and tools stay fixed. Quality and latency are unmeasured.

Recorded path + suggested model assignmentsHeuristic scenarios, not measured Jev confidence or execution latency
Inspecting request 1 · log fully shown
Execution path
Scenario
↻ Viewing request 1 · tool loop
Recorded history · 12 / 12 requests
Request 1 · bashok

Suggested route: LLM
Script execution, writes, edits or unrecognized tools may need generated arguments. Keep the LLM; tool names alone do not prove Jev suitability.

Recorded: Model 2 · $0.000527

Scenario: $0.000527

1,408 recorded input117 recorded output tokens

Cache read / write: 0 / 0

Recorded tools

bash · ok

Turn number not recorded. Request duration not recorded.

Recorded outcome
A model response was recorded.Scenario quality is unverified. Recommendations use safe tool metadata, not query semantics or measured Jev confidence. Request order may include multiple branches.
REVIEW THIS IMPORTED TASK

Log diagnostics

Recorded evidence

Full selected task, independent of playback. Select evidence to inspect it in the canvas; model recommendations and overrides are directly above the canvas.

Recorded evidence

Largest recorded request cost

Request 10 accounts for 26.7% of reported model cost ($0.005324 of $0.019945).

Inspect the request and its context before changing it. Cost concentration alone does not prove unnecessary work.

Recorded evidence

Largest recorded input

Request 12 contains 15,427 input tokens including cache. 12/12 inputs are known.

Review context size and tool payloads in the original log. Usage totals cannot separate system instructions, history and tool results.

Recorded evidence

Cache-read share of input

48.0% of recorded input tokens are cache reads (37,672 / 78,513).

Token share is not cost savings. Provider cache pricing and cache reuse after switching models are not inferred.