fix(coding-agent): exclude cacheRead from token display, add per-subagent cost
Token counter (token_total status-line segment, subagent progress tree, session-observer stats line) previously included cacheRead in its cumulative sum. With Anthropic prompt caching, cacheRead per turn equals the full cached context, so summing across N turns gives N*context_size -- a session with a 1M context and 5 turns showed ~5M tokens despite no compaction occurring. Fix: display shows input + output + cacheWrite per turn. cacheWrite is kept because each byte is written once; cacheRead re-reads the same context every turn. Dedicated cache_read/cache_write status-line segments still show cache activity; billing cost is unaffected. Also adds per-subagent cost display (dollar amount, statusLineCost color) accumulated incrementally from message_end events. Hidden when cost is zero (subscription/OAuth providers). Brings token and cost display in line with what Claude Code shows per-agent.
This commit is contained in:
Miroslav Drbal
@@ -8,6 +8,8 @@
|
||||
- Added per-line column cap shared across streaming tool outputs (`bash`, `ssh`, `python`, `js eval`) and the `read` tool. Lines wider than `tools.outputMaxColumns` bytes (default **768**) are ellipsis-truncated at write time and remaining bytes up to the next `\n` are dropped — bounded memory even on multi-MB single-line outputs (e.g. `cat /dev/urandom`). The cap lives on `OutputSink` as the new `maxColumns` option, persists state across chunk boundaries so split-mid-line writes still respect the budget, and exposes `columnDroppedBytes` / `columnTruncatedLines` on `OutputSummary`. Middle-elision byte math subtracts column drops so the "elided from middle" count stays honest. `read` reuses the same setting but trims its already-collected lines via `truncateLine`. Skipped when the read selector is `:raw`. The artifact file (`artifact://<id>`) keeps the full uncapped stream. Set `tools.outputMaxColumns = 0` to disable.
|
||||
- Added Bun HTTP/2 fetch opt-in. Dev scripts (`bun run dev`, `bun run stats`) now pass `bun --experimental-http2-fetch` so every `fetch()` advertises `h2` in the TLS ALPN list and falls back to HTTP/1.1 when the server doesn't select it. Multiplexing collapses parallel requests to the same origin onto one TLS connection. For the installed `omp` binary, export `BUN_FEATURE_FLAG_EXPERIMENTAL_HTTP2_CLIENT=1` in your shell to enable the same behavior (the flag has to be set before Bun starts; `process.env` from inside JS is too late). Requires Bun **1.3.14**.
|
||||
|
||||
- Added per-subagent cost display (`$X.XX` in the task progress tree and the session-observer stats line). Cost is accumulated incrementally from `message_end` events and shown only when non-zero, using the `statusLineCost` theme color. Providers that do not report per-turn cost data (e.g. subscription/OAuth usage) continue to show nothing.
|
||||
|
||||
### Changed
|
||||
|
||||
- Raised the image downscaling default JPEG quality from 75 to 80 in `resizeImage` output generation
|
||||
@@ -19,6 +21,10 @@
|
||||
- Changed search truncation metadata/renderer output from match/result-based limits to file-based limits (`fileLimitReached`, `perFileLimitReached`) and updated truncation labels accordingly
|
||||
- Lowered `read.defaultLimit` default from `500` to `300` lines, and split the per-range context padding into asymmetric `RANGE_LEADING_CONTEXT_LINES = 1` / `RANGE_TRAILING_CONTEXT_LINES = 3` (was symmetric `RANGE_CONTEXT_LINES = 3`). Replay analysis over post-summarizer sessions (`scripts/session-stats/optimize_read_config.py`) showed that bare-path reads are over-provisioned at the median (file p50 = 220 lines) and that most follow-up reads are disjoint hops rather than adjacent extensions — so a smaller default plus narrower leading context reclaims tokens without measurably changing first-cover rate. Trailing context stays at 3 lines to keep anchor-stale recovery on narrow reads. Explicit `read.defaultLimit` overrides in settings are honoured unchanged.
|
||||
|
||||
### Fixed
|
||||
|
||||
- Fixed token display for sessions and subagents inflating far beyond the context window. `token_total` status-line segment and the subagent overlay token counter now show `input + output + cacheWrite` instead of `input + output + cacheRead + cacheWrite`. With prompt caching, `cacheRead` per turn equals the full cached context — summing it across all turns produces a cumulative total that is N×context_size (e.g. a 5-turn session with a 1 M-token context reported ~5 M tokens). Cache activity is still visible via the dedicated `cache_read`/`cache_write` status-line segments; billing cost is unaffected.
|
||||
|
||||
## [15.0.0] - 2026-05-13
|
||||
### Breaking Changes
|
||||
|
||||
|
||||
@@ -267,7 +267,10 @@ export class SessionObserverOverlayComponent extends Container {
|
||||
if (progress.toolCount > 0) stats.push(`${formatNumber(progress.toolCount)} tools`);
|
||||
if (progress.tokens > 0) stats.push(`${formatNumber(progress.tokens)} tokens`);
|
||||
if (progress.durationMs > 0) stats.push(formatDuration(progress.durationMs));
|
||||
return stats.length > 0 ? theme.fg("dim", stats.join(theme.sep.dot)) : "";
|
||||
const parts: string[] = [];
|
||||
if (stats.length > 0) parts.push(theme.fg("dim", stats.join(theme.sep.dot)));
|
||||
if (progress.cost > 0) parts.push(theme.fg("statusLineCost", `$${progress.cost.toFixed(2)}`));
|
||||
return parts.join(theme.sep.dot);
|
||||
}
|
||||
|
||||
#buildTranscriptLines(messageEntries: SessionMessageEntry[], lines: string[]): void {
|
||||
|
||||
@@ -216,8 +216,11 @@ const tokenOutSegment: StatusLineSegment = {
|
||||
const tokenTotalSegment: StatusLineSegment = {
|
||||
id: "token_total",
|
||||
render(ctx) {
|
||||
const { input, output, cacheRead, cacheWrite } = ctx.usageStats;
|
||||
const total = input + output + cacheRead + cacheWrite;
|
||||
// Excludes cacheRead: that field re-reads the full cached context every
|
||||
// turn, making the cumulative sum N×context_size. The dedicated cache_read
|
||||
// segment handles cache monitoring; the cost segment handles billing.
|
||||
const { input, output, cacheWrite } = ctx.usageStats;
|
||||
const total = input + output + cacheWrite;
|
||||
if (!total) return { content: "", visible: false };
|
||||
|
||||
const content = withIcon(theme.icon.tokens, formatNumber(total));
|
||||
|
||||
@@ -379,21 +379,29 @@ function firstNumberField(record: Record<string, unknown>, keys: string[]): numb
|
||||
}
|
||||
|
||||
/**
|
||||
* Normalize usage objects from different event formats.
|
||||
* Tokens for progress display: input + output + cacheWrite per turn.
|
||||
*
|
||||
* Deliberately excludes cacheRead. With prompt caching, cacheRead in each turn
|
||||
* equals the full cached context (potentially hundreds of KB), so summing it
|
||||
* across all turns produces a cumulative total that is N×context_size — far
|
||||
* larger than the context window and misleading as a "work done" metric.
|
||||
* cacheWrite is kept because each byte is written once, not repeated per turn.
|
||||
* The cost segment handles billing; dedicated cache_read/cache_write segments
|
||||
* handle cache-specific monitoring.
|
||||
*/
|
||||
function getUsageTokens(usage: unknown): number {
|
||||
if (!usage || typeof usage !== "object") return 0;
|
||||
const record = usage as Record<string, unknown>;
|
||||
|
||||
const totalTokens = firstNumberField(record, ["totalTokens", "total_tokens"]);
|
||||
if (totalTokens !== undefined && totalTokens > 0) return totalTokens;
|
||||
|
||||
const input = firstNumberField(record, ["input", "input_tokens", "inputTokens"]) ?? 0;
|
||||
const output = firstNumberField(record, ["output", "output_tokens", "outputTokens"]) ?? 0;
|
||||
const cacheRead = firstNumberField(record, ["cacheRead", "cache_read", "cacheReadTokens"]) ?? 0;
|
||||
const cacheWrite = firstNumberField(record, ["cacheWrite", "cache_write", "cacheWriteTokens"]) ?? 0;
|
||||
|
||||
return input + output + cacheRead + cacheWrite;
|
||||
const computed = input + output + cacheWrite;
|
||||
if (computed > 0) return computed;
|
||||
// Fallback for providers that only surface a pre-summed total without individual
|
||||
// field breakdown. This total includes cacheRead, but returning it is still better
|
||||
// than silently showing 0 for those providers.
|
||||
return firstNumberField(record, ["totalTokens", "total_tokens"]) ?? 0;
|
||||
}
|
||||
|
||||
/**
|
||||
@@ -497,6 +505,7 @@ export async function runSubprocess(options: ExecutorOptions): Promise<SingleRes
|
||||
recentOutput: [],
|
||||
toolCount: 0,
|
||||
tokens: 0,
|
||||
cost: 0,
|
||||
durationMs: 0,
|
||||
modelOverride,
|
||||
};
|
||||
@@ -892,6 +901,7 @@ export async function runSubprocess(options: ExecutorOptions): Promise<SingleRes
|
||||
accumulatedUsage.cost.cacheRead += getNumberField(costRecord, "cacheRead") ?? 0;
|
||||
accumulatedUsage.cost.cacheWrite += getNumberField(costRecord, "cacheWrite") ?? 0;
|
||||
accumulatedUsage.cost.total += getNumberField(costRecord, "total") ?? 0;
|
||||
progress.cost = accumulatedUsage.cost.total;
|
||||
}
|
||||
}
|
||||
// Accumulate tokens for progress display
|
||||
|
||||
@@ -306,6 +306,7 @@ export class TaskTool implements AgentTool<TSchema, TaskToolDetails, Theme> {
|
||||
recentOutput: [],
|
||||
toolCount: 0,
|
||||
tokens: 0,
|
||||
cost: 0,
|
||||
durationMs: 0,
|
||||
});
|
||||
}
|
||||
@@ -831,6 +832,7 @@ export class TaskTool implements AgentTool<TSchema, TaskToolDetails, Theme> {
|
||||
recentOutput: [],
|
||||
toolCount: 0,
|
||||
tokens: 0,
|
||||
cost: 0,
|
||||
durationMs: 0,
|
||||
modelOverride,
|
||||
description: taskItem.description,
|
||||
|
||||
@@ -532,6 +532,9 @@ function renderAgentProgress(
|
||||
if (progress.tokens > 0) {
|
||||
statusLine += `${theme.sep.dot}${theme.fg("dim", `${formatNumber(progress.tokens)} tokens`)}`;
|
||||
}
|
||||
if (progress.cost > 0) {
|
||||
statusLine += `${theme.sep.dot}${theme.fg("statusLineCost", `$${progress.cost.toFixed(2)}`)}`;
|
||||
}
|
||||
} else if (progress.status === "completed") {
|
||||
if (progress.toolCount > 0) {
|
||||
statusLine += `${theme.sep.dot}${theme.fg("dim", `${progress.toolCount} tools`)}`;
|
||||
@@ -539,6 +542,9 @@ function renderAgentProgress(
|
||||
if (progress.tokens > 0) {
|
||||
statusLine += `${theme.sep.dot}${theme.fg("dim", `${formatNumber(progress.tokens)} tokens`)}`;
|
||||
}
|
||||
if (progress.cost > 0) {
|
||||
statusLine += `${theme.sep.dot}${theme.fg("statusLineCost", `$${progress.cost.toFixed(2)}`)}`;
|
||||
}
|
||||
}
|
||||
|
||||
lines.push(statusLine);
|
||||
|
||||
@@ -217,7 +217,10 @@ export interface AgentProgress {
|
||||
recentTools: Array<{ tool: string; args: string; endMs: number }>;
|
||||
recentOutput: string[];
|
||||
toolCount: number;
|
||||
/** Cumulative input + output + cacheWrite tokens across all turns. Excludes cacheRead (re-reads cached context every turn, making cumulative sum misleading). */
|
||||
tokens: number;
|
||||
/** Cumulative billing cost in USD, accumulated incrementally from message_end events. */
|
||||
cost: number;
|
||||
durationMs: number;
|
||||
modelOverride?: string | string[];
|
||||
/** Data extracted by registered subprocess tool handlers (keyed by tool name) */
|
||||
@@ -239,6 +242,7 @@ export interface SingleResult {
|
||||
stderr: string;
|
||||
truncated: boolean;
|
||||
durationMs: number;
|
||||
/** Cumulative input + output + cacheWrite tokens across all turns. Excludes cacheRead (re-reads cached context every turn, making cumulative sum misleading). */
|
||||
tokens: number;
|
||||
modelOverride?: string | string[];
|
||||
error?: string;
|
||||
|
||||
Reference in New Issue
Block a user