Problem
Subscription quota (5-hour window, weekly limit, reset times, booster balance) is only visible in the TUI (/usage). ACP clients have no way to show it, so a user running an agent from an IDE/client cannot tell they are about to hit the 5-hour cap until turns start failing with 429.
Related: #2157 (machine-readable quota channel), #1855 (token usage over ACP), #872 (quota in status line).
Proposal
The data already flows through packages/oauth (fetchManagedUsage → /usages) and the SDK auth facade (getManagedUsage). Emit it over ACP attached to the standard usage_update frame as _meta.kimiCode.rateLimits:
{
"sessionUpdate": "usage_update",
"used": 182344,
"size": 1048576,
"_meta": {
"kimiCode": {
"rateLimits": {
"summary": { "window": "1w", "used": 35, "limit": 100, "resetAt": "2026-08-06T13:38:24Z" },
"limits": [{ "window": "5h", "used": 35, "limit": 100, "resetAt": "2026-07-31T14:38:24Z" }],
"booster": { "balanceCents": 500, "totalCents": 1000, "currency": "USD" }
}
}
}
}Same extension pattern claude-agent-acp uses (_meta["_claude/rateLimit"]), so the convention has precedent in the ecosystem. Fetch is rate-limited to once a minute and best-effort (never blocks a turn). Emitted after every turn and on session open.
Implemented in my fork together with the usage_update emission itself (#1855 covers the base frame); can open a PR.
Problem
Subscription quota (5-hour window, weekly limit, reset times, booster balance) is only visible in the TUI (
/usage). ACP clients have no way to show it, so a user running an agent from an IDE/client cannot tell they are about to hit the 5-hour cap until turns start failing with 429.Related: #2157 (machine-readable quota channel), #1855 (token usage over ACP), #872 (quota in status line).
Proposal
The data already flows through
packages/oauth(fetchManagedUsage→/usages) and the SDK auth facade (getManagedUsage). Emit it over ACP attached to the standardusage_updateframe as_meta.kimiCode.rateLimits:{ "sessionUpdate": "usage_update", "used": 182344, "size": 1048576, "_meta": { "kimiCode": { "rateLimits": { "summary": { "window": "1w", "used": 35, "limit": 100, "resetAt": "2026-08-06T13:38:24Z" }, "limits": [{ "window": "5h", "used": 35, "limit": 100, "resetAt": "2026-07-31T14:38:24Z" }], "booster": { "balanceCents": 500, "totalCents": 1000, "currency": "USD" } } } } }Same extension pattern claude-agent-acp uses (
_meta["_claude/rateLimit"]), so the convention has precedent in the ecosystem. Fetch is rate-limited to once a minute and best-effort (never blocks a turn). Emitted after every turn and on session open.Implemented in my fork together with the
usage_updateemission itself (#1855 covers the base frame); can open a PR.