v3.0.0

Breaking: the output changes

A window at its limit now has lane headroom on the account's usage credits: Claude's extra usage up to the spend limit and prepaid balance, and Codex credits. Before, it said "no lane headroom: start nothing new" even with credits topped up. Without credits, the output is as before.

• Lane wording: while a window is at 100% with credits left, the lane phrase names the pool and paces against what is left: on usage credits (429.39 GBP left): room for ~2 more lanes at 30.00 GBP/h each for 1.0h, or … measuring: start at most 1-2 new lanes until the draw is measured.
• --json lanes: gains pool (plan or usage-credits). With usage-credits, headroom, per_lane_rate and hours are the credits', and per_lane_rate is in their unit, not %/h.
• New --json field: usage_credits per source (null without credits): unit, left, drawing, hours_left, rate_per_hour, per_lane_rate, lane_headroom, lane_hours.
• plan: when neither plan has room, it routes work to credits that have room: route work to Claude on usage credits (429.39 GBP left), or wait for a reset. A provider with room in its plan is still preferred.
• State: a new <source>-credits.csv (what the credits have spent), and a credits key in claude-banked.json.
• Requests: the Claude usage call drops skip_spend=1, which had hidden extra_usage. With extra usage on, a second GET reads the prepaid balance. agent-pace still never redeems a reset or buys credits.

What's Changed

• feat!: a window at its limit has lane headroom on the usage credits by @frostney in #4

Full Changelog: v2.0.0...v3.0.0

v2.0.0

Breaking: the output changes

The text lines and the --json output no longer match the Python version 1.4 in these places:

• Lane headroom counts each new lane for --lane-hours (default 1, never past the window's reset) and the lanes already running until the reset; with a reset banked, the running lanes too count only for the lane hours. The lane numbers change and the wording gains the hours: room for ~5 more lanes at 0.13%/h each for 1.0h. A large count is lane-hours of budget, not a number of lanes to keep running.
• A week is paced by its running lanes' tokens when their cost is learned: its rate, status and projection change, shown as 0.1%/h (lanes, window avg 0.8%/h) with rate_basis: "lanes". It keeps its window average otherwise.
• Per-lane burn is measured from each lane's own tokens (#2): the lane numbers change, with new lane_basis, lane_rates and token_cost fields, and a tokens-<source>.json state file.
• New --json fields: window_average_rate and lane_hours per window, hours under lanes.
• New option: --lane-hours HOURS, with its help text and a usage error for a value of 0 or less.

The history CSV files and claude-banked.json are unchanged. agent-pace still never redeems a reset.

What's Changed

• feat: each lane's burn is measured from its own tokens by @frostney in #2
• feat!: lane headroom over a finite stretch, a week paced by its running lanes by @frostney in #3

Full Changelog: v1.4.0...v2.0.0

v1.4.0

What's Changed

• feat: agent-pace in TypeScript on Bun, printing what the Python version printed by @frostney in #1

New Contributors

• @frostney made their first contribution in #1

Full Changelog: https://github.com/frostney/agent-pace/commits/v1.4.0