nvaldez 846204d9fb fix: cost line ordering (publish after the response) + billing-route provider label
Ordering: post_llm_call fires in the agent worker thread before
the gateway settles the turn and publishes message.complete, so
an immediate notification.show lands ABOVE the response. The
publish now waits (bounded 60s, detached daemon thread so the
timeout-bounded hook returns instantly) until the session's
execution stamp clears - the settlement clears it strictly
after message.complete - making the cost line the LAST line of
every turn. Same hold pattern as the gateway's background-review
delivery.

Label: the line showed the pricing-table match (zai:zai-glm-5-3)
even when the session bills through a custom provider route.
The label now comes from the billing route: the config.yaml
custom_providers name matching the session's billing_base_url
(hermes_cli.config.load_config first, regex fallback for
standalone), else the stored billing_provider when not
generic, else the pricing key. A session served by the custom
provider 'mistral' now shows mistral:zai-glm-5-3.

Also: restore the zai provider in the bundled pricing.json
(dropped by an uncommitted local edit; plugin-local lookup made
the loss visible as default:default).

- tests: completion-stamp wait, billing-route label, execution
  stamp on the notification mock
- version 0.4.3 (package + plugin manifest)

Generated by Mistral Vibe.
Co-Authored-By: Mistral Vibe <vibe@mistral.ai>
2026-10-25 16:57:40 +01:00
2026-10-21 23:18:37 +02:00

agent-cost-estimator

Report token usage and estimated USD cost of an AI agent session, parsed from the harness's local logs. Ships with a Mistral Vibe adapter; other AI harnesses plug in through the adapter interface.

The result is a local estimate priced from pricing.json. The authoritative monthly usage is your provider's console (for Mistral: admin.mistral.ai/subscription).

Install

pip install .        # from this directory
# or run in place:
python -m agent_cost_estimator --help

Usage

agent-cost-estimator                        # mistral:model | session:$X | turn:$Y
agent-cost-estimator -d                     # the full report table
agent-cost-estimator -s                     # session cost only, e.g. 2.1148
agent-cost-estimator -t                     # latest turn cost only
agent-cost-estimator -d --turns             # table plus per-turn breakdown
agent-cost-estimator --session-id 5732da2a  # a specific session id or path
agent-cost-estimator --json                 # machine-readable output
agent-cost-estimator --list-harnesses       # available adapters
agent-cost-estimator --integrate            # install Vibe hooks automatically

With no output flag the tool prints one line: provider:model and costs for the session and latest turn. -d/--details prints the full table; -s/--session and -t/--turn print one bare number each, for scripting.

Use --integrate to automatically set up Vibe CLI hooks:

agent-cost-estimator --integrate

This installs the hook script and configuration, and adds agent-cost-estimator to Vibe's bash allowlist. The turn cost is the most recent turn in the journal window.

Prices come from the bundled agent_cost_estimator/pricing.json, keyed by model (with a default fallback). Bring your own table:

agent-cost-estimator --pricing my-prices.json

Or persist overrides into the pricing table (the built-in one, or the file passed with --pricing):

agent-cost-estimator --model zai-glm-5-3 \
    --input-price 1.2 --cached-input-price 0.12 --output-price 6.0 \
    --update-pricing          # add --dry-run to preview

pricing.json format

{
  "currency": "USD",
  "models": {
    "default":     {"input_price": 1.5, "cached_input_price": 0.15, "output_price": 7.5},
    "zai-glm-5-3": {"input_price": 1.5, "cached_input_price": 0.15, "output_price": 7.5}
  }
}

Prices are USD per million tokens. Lookup order: session's model name, then its alias, then default. On the command line, --input-price, --cached-input-price, and --output-price override the table for that run.

Adding a harness

A harness adapter is a subclass of agent_cost_estimator.harnesses.base.HarnessAdapter implementing three read-only methods:

  • resolve_session_dir(session_arg, home) - find the session directory (or the most recently active one when the argument is None)
  • read_meta(session_dir) - return a SessionMeta (title, model_name, alias)
  • collect_usage(session_dir) - return a SessionUsage (input/cached/output tokens, call count, per-turn list)

Set name, home_env_var, and home_dir_name so --home and the session lookup have a sane default root. Register the adapter either by adding it to BUILTIN_HARNESSES (in-tree) or by exposing it through the agent_cost_estimator.harnesses entry-point group from your own package:

[project.entry-points."agent_cost_estimator.harnesses"]
mytool = "my_package.adapters:MyToolHarness"

Built-ins always win a name clash; a broken plugin is skipped with a warning.

Development

python -m pytest tests/

Standard library only; no runtime dependencies.

S
Description
No description provided
Readme 1.5 MiB
Languages
Python 93.1%
Shell 4.3%
PowerShell 2.6%