@cruxy/cli 1.2.1 → 1.3.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/agent/context.js +178 -0
- package/dist/agent/index.js +1 -0
- package/dist/agent/loop.js +20 -1
- package/dist/agent/mode.js +103 -0
- package/dist/agent/prompts.js +1 -1
- package/dist/agent/session.js +171 -69
- package/dist/approval/classify.js +204 -0
- package/dist/approval/policy.js +41 -3
- package/dist/approval/prompt.js +49 -22
- package/dist/checkpoint/gate.js +12 -0
- package/dist/cli/commands/run.js +374 -227
- package/dist/cli/commands/usage.js +45 -45
- package/dist/cli/onboard.js +2 -1
- package/dist/cli/program.js +60 -18
- package/dist/cli/repl.js +67 -249
- package/dist/cli/session-commands.js +755 -0
- package/dist/cli/session-factory.js +198 -76
- package/dist/cli/suggest.js +77 -0
- package/dist/components/fuzzy.js +3 -3
- package/dist/components/input.js +17 -2
- package/dist/components/keys.js +27 -3
- package/dist/components/select.js +3 -3
- package/dist/config/project.js +53 -1
- package/dist/config/schema.js +49 -16
- package/dist/jobs/log-renderer.js +47 -0
- package/dist/onboarding/steps.js +13 -22
- package/dist/plan/approve.js +36 -24
- package/dist/plan/execute.js +9 -7
- package/dist/plan/render.js +10 -23
- package/dist/plan/service.js +4 -1
- package/dist/render/capabilities.js +30 -1
- package/dist/render/context-view.js +106 -0
- package/dist/render/diff.js +198 -12
- package/dist/render/index.js +31 -5
- package/dist/render/plain-renderer.js +38 -2
- package/dist/render/plan-view.js +108 -0
- package/dist/render/resize.js +7 -2
- package/dist/render/status-view.js +66 -0
- package/dist/render/test-view.js +89 -0
- package/dist/render/tty-renderer.js +40 -0
- package/dist/routing/index.js +1 -0
- package/dist/routing/router.js +13 -4
- package/dist/routing/session-model.js +109 -0
- package/dist/routing/types.js +14 -0
- package/dist/session/export.js +88 -0
- package/dist/session/index.js +20 -0
- package/dist/session/list.js +137 -0
- package/dist/session/log.js +137 -0
- package/dist/session/paths.js +73 -0
- package/dist/session/replay.js +169 -0
- package/dist/session/resume.js +128 -0
- package/dist/session/types.js +223 -0
- package/dist/subagent/orchestrator.js +23 -0
- package/dist/testing/run-tests-tool.js +8 -0
- package/dist/tools/registry.js +3 -3
- package/dist/tui/app.js +385 -0
- package/dist/tui/approval-overlay.js +160 -0
- package/dist/tui/context-gauge.js +48 -0
- package/dist/tui/git-status.js +63 -0
- package/dist/tui/index.js +10 -0
- package/dist/tui/layout.js +269 -0
- package/dist/tui/overlay.js +105 -0
- package/dist/tui/palette.js +73 -0
- package/dist/tui/panels.js +235 -0
- package/dist/tui/renderer.js +776 -0
- package/dist/tui/supports.js +20 -0
- package/dist/tui/tool-versions.js +129 -0
- package/dist/usage/collect.js +6 -6
- package/dist/usage/index.js +10 -2
- package/dist/usage/report.js +76 -0
- package/dist/usage/summary.js +106 -17
- package/dist/usage/types.js +5 -2
- package/dist/usage/weighted.js +77 -0
- package/dist/utils/git.js +50 -4
- package/package.json +1 -1
- package/dist/usage/cost.js +0 -29
|
@@ -0,0 +1,20 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Whether the environment can host the full-viewport TUI.
|
|
3
|
+
*
|
|
4
|
+
* Three requirements, and they are exactly the axes capability detection now
|
|
5
|
+
* resolves — no re-probing of `process.stdin` or `TERM` here:
|
|
6
|
+
*
|
|
7
|
+
* - `interactive` (stdin is a terminal AND the output takes cursor control):
|
|
8
|
+
* the TUI both reads keys and repaints, so it needs both halves. This is the
|
|
9
|
+
* flag that makes `echo hi | cruxy` fall through to the plain path.
|
|
10
|
+
* - not `screenReader`: a live full-screen region is precisely what
|
|
11
|
+
* screen-reader mode exists to avoid (U.11). Announce-friendly linear output
|
|
12
|
+
* wins over the shell, always.
|
|
13
|
+
*
|
|
14
|
+
* Capability only. WHETHER a given invocation wants a TUI is an entry-mode
|
|
15
|
+
* decision (bare `cruxy` does; `cruxy run "task"` does not) and lives with the
|
|
16
|
+
* caller — see `createRenderer`'s `tui` option.
|
|
17
|
+
*/
|
|
18
|
+
export function supportsTui(caps) {
|
|
19
|
+
return caps.interactive && !caps.screenReader;
|
|
20
|
+
}
|
|
@@ -0,0 +1,129 @@
|
|
|
1
|
+
import { execFile } from "node:child_process";
|
|
2
|
+
/**
|
|
3
|
+
* Host tool versions for the rail (P4 track 5).
|
|
4
|
+
*
|
|
5
|
+
* THE COST IS THE DESIGN. Measured warm on a fast Mac: pnpm 584ms, docker
|
|
6
|
+
* 130ms, tsc 108ms, git 34ms — ~911ms serially, and still ~584ms in parallel
|
|
7
|
+
* because pnpm alone dominates. That is half a second of dead terminal to
|
|
8
|
+
* populate a decorative panel, and it is the optimistic case: Windows process
|
|
9
|
+
* creation is far slower, which this repo has already paid for in CI.
|
|
10
|
+
*
|
|
11
|
+
* So nothing here runs at startup, nothing is awaited, and nothing blocks a
|
|
12
|
+
* paint. {@link ToolVersions.start} fires the probes and returns immediately;
|
|
13
|
+
* every row renders "…" until its OWN probe lands, independently of the others.
|
|
14
|
+
* A slow pnpm cannot delay a fast git.
|
|
15
|
+
*
|
|
16
|
+
* The shape is `sandbox/detect.ts`'s, which already solved this correctly:
|
|
17
|
+
* memoized for the process, a hard timeout, ENOENT treated as an answer rather
|
|
18
|
+
* than a throw, and never rejecting. One deliberate difference — that module
|
|
19
|
+
* probes `docker version --format {{.Server.Version}}`, which requires a LIVE
|
|
20
|
+
* DAEMON and reports a stopped Docker as absent. This panel is reporting what
|
|
21
|
+
* is INSTALLED, so it asks the client (`docker --version`), which answers with
|
|
22
|
+
* the daemon down.
|
|
23
|
+
*/
|
|
24
|
+
/** Hard ceiling on one probe; a hung binary must never wedge the panel. */
|
|
25
|
+
const PROBE_TIMEOUT_MS = 5000;
|
|
26
|
+
/**
|
|
27
|
+
* Pull a semantic version out of a tool's `--version` line.
|
|
28
|
+
*
|
|
29
|
+
* Every one of these prints something different — "git version 2.50.1 (Apple
|
|
30
|
+
* Git-155)", "Docker version 29.6.2, build dfc4efb", "Version 5.9.3", a bare
|
|
31
|
+
* "10.33.2" — so the first `x.y.z` in the output is both the simplest rule and
|
|
32
|
+
* the most robust one. `null` when there is no version-shaped thing at all,
|
|
33
|
+
* which is honest: something answered, but not with a version.
|
|
34
|
+
*/
|
|
35
|
+
export function parseVersion(output) {
|
|
36
|
+
return /\d+\.\d+\.\d+(?:[-+][\w.]+)?/.exec(output)?.[0] ?? null;
|
|
37
|
+
}
|
|
38
|
+
/** Default probe: spawn, capture, and treat any failure as "not installed". */
|
|
39
|
+
const spawnProbe = (bin, args) => new Promise((resolve) => {
|
|
40
|
+
try {
|
|
41
|
+
execFile(bin, [...args], { encoding: "utf8", timeout: PROBE_TIMEOUT_MS, windowsHide: true }, (err, stdout, stderr) => {
|
|
42
|
+
// Some tools print their version to stderr; a non-zero exit with a
|
|
43
|
+
// parseable version still tells us the tool is there.
|
|
44
|
+
const text = `${stdout ?? ""}${stderr ?? ""}`;
|
|
45
|
+
if (err && text.trim() === "")
|
|
46
|
+
resolve(null);
|
|
47
|
+
else
|
|
48
|
+
resolve(text);
|
|
49
|
+
});
|
|
50
|
+
}
|
|
51
|
+
catch {
|
|
52
|
+
// execFile can throw synchronously on a malformed binary path.
|
|
53
|
+
resolve(null);
|
|
54
|
+
}
|
|
55
|
+
});
|
|
56
|
+
/**
|
|
57
|
+
* The tools reported, in display order.
|
|
58
|
+
*
|
|
59
|
+
* Node is NOT here: `process.versions.node` is the running interpreter's own
|
|
60
|
+
* version, already in memory. Spawning `node --version` to learn something the
|
|
61
|
+
* process knows about itself would be the most obviously wasteful 55ms in the
|
|
62
|
+
* list — and it would answer for whichever node is on PATH, which need not be
|
|
63
|
+
* the one executing this code.
|
|
64
|
+
*/
|
|
65
|
+
const TOOLS = [
|
|
66
|
+
{ name: "tsc", bin: "tsc", args: ["--version"] },
|
|
67
|
+
{ name: "pnpm", bin: "pnpm", args: ["--version"] },
|
|
68
|
+
{ name: "git", bin: "git", args: ["--version"] },
|
|
69
|
+
// The CLIENT, not `docker version --format {{.Server.Version}}`: this panel
|
|
70
|
+
// reports what is installed, and a stopped daemon does not uninstall Docker.
|
|
71
|
+
{ name: "docker", bin: "docker", args: ["--version"] },
|
|
72
|
+
];
|
|
73
|
+
export class ToolVersions {
|
|
74
|
+
probe;
|
|
75
|
+
versions = new Map();
|
|
76
|
+
started = false;
|
|
77
|
+
onSettled;
|
|
78
|
+
constructor(probe = spawnProbe) {
|
|
79
|
+
this.probe = probe;
|
|
80
|
+
}
|
|
81
|
+
/**
|
|
82
|
+
* Every row in display order — a pure map read, safe on the paint path.
|
|
83
|
+
*
|
|
84
|
+
* Node is resolved here rather than probed: free, exact, and it means the
|
|
85
|
+
* panel is never completely empty even on the very first frame.
|
|
86
|
+
*/
|
|
87
|
+
current() {
|
|
88
|
+
return [
|
|
89
|
+
{ name: "node", version: process.versions.node },
|
|
90
|
+
...TOOLS.map((t) => ({
|
|
91
|
+
name: t.name,
|
|
92
|
+
version: this.versions.get(t.name),
|
|
93
|
+
})),
|
|
94
|
+
];
|
|
95
|
+
}
|
|
96
|
+
/**
|
|
97
|
+
* Begin probing. Idempotent and memoized: the first call starts every probe,
|
|
98
|
+
* later calls do nothing, so a panel that repaints 30 times a second cannot
|
|
99
|
+
* re-spawn anything.
|
|
100
|
+
*
|
|
101
|
+
* Returns immediately. Each probe settles on its own and calls `onSettled`,
|
|
102
|
+
* so a fast git appears while a slow pnpm is still running — the rows are
|
|
103
|
+
* independent, and the panel fills in rather than arriving all at once.
|
|
104
|
+
*/
|
|
105
|
+
start(onSettled) {
|
|
106
|
+
if (onSettled)
|
|
107
|
+
this.onSettled = onSettled;
|
|
108
|
+
if (this.started)
|
|
109
|
+
return;
|
|
110
|
+
this.started = true;
|
|
111
|
+
for (const tool of TOOLS) {
|
|
112
|
+
void this.run(tool);
|
|
113
|
+
}
|
|
114
|
+
}
|
|
115
|
+
async run(tool) {
|
|
116
|
+
let version = null;
|
|
117
|
+
try {
|
|
118
|
+
const output = await this.probe(tool.bin, tool.args);
|
|
119
|
+
version = output === null ? null : parseVersion(output);
|
|
120
|
+
}
|
|
121
|
+
catch {
|
|
122
|
+
// A probe that rejects is "not installed" for display purposes; it must
|
|
123
|
+
// never surface as an unhandled rejection from a decorative panel.
|
|
124
|
+
version = null;
|
|
125
|
+
}
|
|
126
|
+
this.versions.set(tool.name, version);
|
|
127
|
+
this.onSettled?.();
|
|
128
|
+
}
|
|
129
|
+
}
|
package/dist/usage/collect.js
CHANGED
|
@@ -60,12 +60,12 @@ export class UsageCollector {
|
|
|
60
60
|
...(u?.billable_input_tokens !== undefined
|
|
61
61
|
? { billableInputTokens: u.billable_input_tokens }
|
|
62
62
|
: {}),
|
|
63
|
-
// The server's own cost, persisted verbatim and
|
|
64
|
-
//
|
|
65
|
-
//
|
|
66
|
-
//
|
|
67
|
-
//
|
|
68
|
-
// must never be read as "unpriced".
|
|
63
|
+
// The server's own cost, persisted verbatim — and now the ONLY source of
|
|
64
|
+
// the cost `cruxy usage` shows; the user-configured price table it used to
|
|
65
|
+
// compete with is retired. Amount and currency travel together: a figure
|
|
66
|
+
// without its unit is not a cost, and the summary refuses to add two
|
|
67
|
+
// currencies rather than invent a rate. A recorded 0 is a real 0; see
|
|
68
|
+
// UsageEntry.costAmount for why that must never be read as "unpriced".
|
|
69
69
|
...(u?.cost !== undefined
|
|
70
70
|
? { costAmount: u.cost.amount, costCurrency: u.cost.currency }
|
|
71
71
|
: {}),
|
package/dist/usage/index.js
CHANGED
|
@@ -7,9 +7,17 @@
|
|
|
7
7
|
* `@cruxy/sdk` import below is a TYPE-only import (`Usage`), erased at build. A
|
|
8
8
|
* future opt-in remote report would be a new, clearly-named seam — this build
|
|
9
9
|
* ships nothing that sends. Asserted by the runtime + static no-phone-home tests.
|
|
10
|
+
*
|
|
11
|
+
* The guarantee survived this module gaining SERVER-SOURCED figures (the
|
|
12
|
+
* gateway's cost, and weighted tokens in the gateway's own metering unit)
|
|
13
|
+
* because none of them is fetched here: cost rides in on the chat response the
|
|
14
|
+
* agent loop was already making, and the weighted multipliers are a compiled-in
|
|
15
|
+
* constant. Reading a number the server sent is not the same as asking it for
|
|
16
|
+
* one — the second would be the thing this promise forbids.
|
|
10
17
|
*/
|
|
11
18
|
export * from "./types.js";
|
|
12
19
|
export { UsageCollector, accumulateCacheTokens, } from "./collect.js";
|
|
13
|
-
export {
|
|
20
|
+
export { TIER_MULTIPLIERS, multiplierForTier, weightedFor, } from "./weighted.js";
|
|
14
21
|
export { loadUsage, appendRun, usageStorePath } from "./store.js";
|
|
15
|
-
export { summarizeRuns, renderSummary, formatCost
|
|
22
|
+
export { summarizeRuns, renderSummary, formatCost } from "./summary.js";
|
|
23
|
+
export { runCountLabel, selectRuns, usageReport, } from "./report.js";
|
|
@@ -0,0 +1,76 @@
|
|
|
1
|
+
import { renderSummary, summarizeRuns } from "./summary.js";
|
|
2
|
+
/** Narrow the retained runs to `scope`, newest-last order preserved. */
|
|
3
|
+
export function selectRuns(runs, scope = {}) {
|
|
4
|
+
let selected = [...runs];
|
|
5
|
+
if (scope.sessionId !== undefined) {
|
|
6
|
+
selected = selected.filter((r) => r.sessionId === scope.sessionId);
|
|
7
|
+
}
|
|
8
|
+
if (scope.last !== undefined) {
|
|
9
|
+
// Guarded, not `slice(-last)` alone: `slice(-0)` is `slice(0)` and returns
|
|
10
|
+
// EVERYTHING, so a zero would quietly widen the scope to the whole store —
|
|
11
|
+
// the exact opposite of what it asks for.
|
|
12
|
+
selected = scope.last <= 0 ? [] : selected.slice(-scope.last);
|
|
13
|
+
}
|
|
14
|
+
return selected;
|
|
15
|
+
}
|
|
16
|
+
/**
|
|
17
|
+
* Say what the weighted figure IS, every time it is shown.
|
|
18
|
+
*
|
|
19
|
+
* The number is meaningless-to-dangerous without its scope: a reader who assumes
|
|
20
|
+
* it covers their whole account, or reads it as a remaining balance, has been
|
|
21
|
+
* misled by a figure that is itself correct. These lines are the only thing
|
|
22
|
+
* standing between the two readings, which is why they are not optional and not
|
|
23
|
+
* abbreviated for the in-session surface.
|
|
24
|
+
*/
|
|
25
|
+
function weightedNote(summary, t) {
|
|
26
|
+
if (summary.totalWeightedTokens !== undefined) {
|
|
27
|
+
return [
|
|
28
|
+
t.muted("weighted = (billable input + output) × tier weight — weighted tokens used by\n" +
|
|
29
|
+
"this CLI, over the runs shown. Not a balance or a limit: it cannot see your\n" +
|
|
30
|
+
"plan, or what cruxy on the web or desktop used from the same account."),
|
|
31
|
+
];
|
|
32
|
+
}
|
|
33
|
+
// Absent rather than zero, and the reasons are unguessable from a blank line
|
|
34
|
+
// — so name them instead of leaving the figure silently missing.
|
|
35
|
+
return [
|
|
36
|
+
t.muted([
|
|
37
|
+
"weighted tokens unavailable for these runs — a request can only be weighed",
|
|
38
|
+
"when it records a tier and the gateway's billable input count. Runs from an",
|
|
39
|
+
"older cruxy, or against a non-cruxy provider, carry neither.",
|
|
40
|
+
].join("\n")),
|
|
41
|
+
];
|
|
42
|
+
}
|
|
43
|
+
/**
|
|
44
|
+
* The full usage report for `runs`, as lines to print.
|
|
45
|
+
*
|
|
46
|
+
* Empty `runs` is a first-class outcome, not an early-return the caller has to
|
|
47
|
+
* remember: it renders the "nothing yet" answer plus, when tracking is off, the
|
|
48
|
+
* reason — because "you have run nothing" and "nothing is being recorded" are
|
|
49
|
+
* different facts with the same blank output.
|
|
50
|
+
*/
|
|
51
|
+
export function usageReport(runs, t, opts) {
|
|
52
|
+
if (runs.length === 0) {
|
|
53
|
+
const lines = [t.muted("no usage recorded yet")];
|
|
54
|
+
if (!opts.trackingEnabled) {
|
|
55
|
+
lines.push(t.muted("usage tracking is off (usage.enabled = false)"));
|
|
56
|
+
}
|
|
57
|
+
return lines;
|
|
58
|
+
}
|
|
59
|
+
const summary = summarizeRuns(runs);
|
|
60
|
+
const lines = [
|
|
61
|
+
t.heading(`usage — ${opts.scopeLabel}`),
|
|
62
|
+
renderSummary(summary, t),
|
|
63
|
+
...weightedNote(summary, t),
|
|
64
|
+
];
|
|
65
|
+
if (opts.legacyPriceConfig) {
|
|
66
|
+
lines.push(t.muted([
|
|
67
|
+
"note: usage.prices / usage.currency are no longer used. Cost is now the",
|
|
68
|
+
"gateway's own figure. You can remove them from your config.",
|
|
69
|
+
].join("\n")));
|
|
70
|
+
}
|
|
71
|
+
return lines;
|
|
72
|
+
}
|
|
73
|
+
/** `3 runs` / `1 run` — the countable half of a scope label. */
|
|
74
|
+
export function runCountLabel(n) {
|
|
75
|
+
return `${n} run${n === 1 ? "" : "s"}`;
|
|
76
|
+
}
|
package/dist/usage/summary.js
CHANGED
|
@@ -1,14 +1,38 @@
|
|
|
1
1
|
import { formatTokens } from "../render/state.js";
|
|
2
|
-
import {
|
|
2
|
+
import { weightedFor } from "./weighted.js";
|
|
3
|
+
/**
|
|
4
|
+
* Usage aggregation + rendering (C.22). Aggregation sums only KNOWN figures and
|
|
5
|
+
* counts the requests that reported none separately, so a total is never
|
|
6
|
+
* inflated by a fabricated zero — and the renderer ALWAYS surfaces those counts,
|
|
7
|
+
* so a total can never be misread as complete while requests are silently
|
|
8
|
+
* excluded. That rule now applies on three independent axes, because a request
|
|
9
|
+
* can report tokens but no cost, or a cost but nothing weighable:
|
|
10
|
+
* `requestsWithoutUsage`, `requestsWithoutCost`, `requestsWithoutWeight`.
|
|
11
|
+
*
|
|
12
|
+
* Both derived figures come from the server: cost is the gateway's own quote,
|
|
13
|
+
* weighted tokens are the gateway's own metering formula. Neither is derived
|
|
14
|
+
* from a user-configured price — there is no longer such a thing here.
|
|
15
|
+
*
|
|
16
|
+
* Only tier names ever reach the output (U.8 gag). Pure — no I/O, no network.
|
|
17
|
+
*/
|
|
3
18
|
/** Aggregate one or more run records into a {@link UsageSummary}. */
|
|
4
|
-
export function summarizeRuns(runs
|
|
19
|
+
export function summarizeRuns(runs) {
|
|
5
20
|
const byTier = new Map();
|
|
6
21
|
let totalInputTokens;
|
|
7
22
|
let totalOutputTokens;
|
|
8
23
|
let totalCacheReadTokens;
|
|
9
24
|
let totalCacheCreationTokens;
|
|
25
|
+
let totalBillableInputTokens;
|
|
26
|
+
let totalWeightedTokens;
|
|
27
|
+
let totalCost;
|
|
10
28
|
let requests = 0;
|
|
11
29
|
let requestsWithoutUsage = 0;
|
|
30
|
+
let requestsWithoutWeight = 0;
|
|
31
|
+
let requestsWithoutCost = 0;
|
|
32
|
+
// Every distinct currency the gateway stated. Summing across two of them would
|
|
33
|
+
// require an exchange rate we don't have and would never be told, so the set
|
|
34
|
+
// is kept and a >1 outcome withholds the total rather than guessing.
|
|
35
|
+
const currencies = new Set();
|
|
12
36
|
const addKnown = (acc, v) => (v === undefined ? acc : (acc ?? 0) + v);
|
|
13
37
|
for (const run of runs) {
|
|
14
38
|
for (const e of run.entries) {
|
|
@@ -16,10 +40,24 @@ export function summarizeRuns(runs, opts) {
|
|
|
16
40
|
const known = e.inputTokens !== undefined || e.outputTokens !== undefined;
|
|
17
41
|
if (!known)
|
|
18
42
|
requestsWithoutUsage++;
|
|
43
|
+
// Weighed per REQUEST, never from the tier's summed totals: the multiplier
|
|
44
|
+
// is per-tier, but a tier's bucket can mix weighable and unweighable
|
|
45
|
+
// requests, and multiplying a sum that includes the latter would silently
|
|
46
|
+
// weigh tokens that were never eligible.
|
|
47
|
+
const weighted = weightedFor(e.tier, e.billableInputTokens, e.outputTokens);
|
|
48
|
+
if (weighted === undefined)
|
|
49
|
+
requestsWithoutWeight++;
|
|
50
|
+
if (e.costAmount === undefined)
|
|
51
|
+
requestsWithoutCost++;
|
|
52
|
+
else if (e.costCurrency !== undefined)
|
|
53
|
+
currencies.add(e.costCurrency);
|
|
19
54
|
totalInputTokens = addKnown(totalInputTokens, e.inputTokens);
|
|
20
55
|
totalOutputTokens = addKnown(totalOutputTokens, e.outputTokens);
|
|
21
56
|
totalCacheReadTokens = addKnown(totalCacheReadTokens, e.cacheReadTokens);
|
|
22
57
|
totalCacheCreationTokens = addKnown(totalCacheCreationTokens, e.cacheCreationTokens);
|
|
58
|
+
totalBillableInputTokens = addKnown(totalBillableInputTokens, e.billableInputTokens);
|
|
59
|
+
totalWeightedTokens = addKnown(totalWeightedTokens, weighted);
|
|
60
|
+
totalCost = addKnown(totalCost, e.costAmount);
|
|
23
61
|
// Per-tier attribution only for entries that carry a tier. Untiered
|
|
24
62
|
// requests (routing inert) still count toward totals — the total stays
|
|
25
63
|
// honest — but there is no tier label to bucket them under.
|
|
@@ -27,42 +65,53 @@ export function summarizeRuns(runs, opts) {
|
|
|
27
65
|
const b = byTier.get(e.tier) ?? {
|
|
28
66
|
requests: 0,
|
|
29
67
|
requestsWithoutUsage: 0,
|
|
68
|
+
requestsWithoutWeight: 0,
|
|
30
69
|
};
|
|
31
70
|
b.requests++;
|
|
32
71
|
if (!known)
|
|
33
72
|
b.requestsWithoutUsage++;
|
|
73
|
+
if (weighted === undefined)
|
|
74
|
+
b.requestsWithoutWeight++;
|
|
34
75
|
b.inputTokens = addKnown(b.inputTokens, e.inputTokens);
|
|
35
76
|
b.outputTokens = addKnown(b.outputTokens, e.outputTokens);
|
|
36
77
|
b.cacheReadTokens = addKnown(b.cacheReadTokens, e.cacheReadTokens);
|
|
37
78
|
b.cacheCreationTokens = addKnown(b.cacheCreationTokens, e.cacheCreationTokens);
|
|
79
|
+
b.billableInputTokens = addKnown(b.billableInputTokens, e.billableInputTokens);
|
|
80
|
+
b.weightedTokens = addKnown(b.weightedTokens, weighted);
|
|
81
|
+
b.cost = addKnown(b.cost, e.costAmount);
|
|
38
82
|
byTier.set(e.tier, b);
|
|
39
83
|
}
|
|
40
84
|
}
|
|
41
85
|
}
|
|
86
|
+
const costMixedCurrency = currencies.size > 1;
|
|
42
87
|
const perTier = [...byTier.entries()].map(([tier, b]) => ({
|
|
43
88
|
tier,
|
|
44
89
|
inputTokens: b.inputTokens,
|
|
45
90
|
outputTokens: b.outputTokens,
|
|
46
91
|
cacheReadTokens: b.cacheReadTokens,
|
|
47
92
|
cacheCreationTokens: b.cacheCreationTokens,
|
|
93
|
+
billableInputTokens: b.billableInputTokens,
|
|
94
|
+
weightedTokens: b.weightedTokens,
|
|
48
95
|
requests: b.requests,
|
|
49
96
|
requestsWithoutUsage: b.requestsWithoutUsage,
|
|
50
|
-
|
|
97
|
+
requestsWithoutWeight: b.requestsWithoutWeight,
|
|
98
|
+
// Withheld under mixed currencies for the same reason the total is: a tier's
|
|
99
|
+
// requests are not guaranteed to share a currency either.
|
|
100
|
+
cost: costMixedCurrency ? undefined : b.cost,
|
|
51
101
|
}));
|
|
52
|
-
const costs = perTier
|
|
53
|
-
.map((p) => p.cost)
|
|
54
|
-
.filter((c) => c !== undefined);
|
|
55
|
-
const priced = costs.length > 0;
|
|
56
|
-
const totalCost = priced ? costs.reduce((a, c) => a + c, 0) : undefined;
|
|
57
102
|
return {
|
|
58
103
|
perTier,
|
|
59
104
|
totalInputTokens,
|
|
60
105
|
totalOutputTokens,
|
|
61
106
|
totalCacheReadTokens,
|
|
62
107
|
totalCacheCreationTokens,
|
|
63
|
-
|
|
64
|
-
|
|
65
|
-
|
|
108
|
+
totalBillableInputTokens,
|
|
109
|
+
totalWeightedTokens,
|
|
110
|
+
requestsWithoutWeight,
|
|
111
|
+
totalCost: costMixedCurrency ? undefined : totalCost,
|
|
112
|
+
costCurrency: costMixedCurrency ? undefined : [...currencies][0],
|
|
113
|
+
costMixedCurrency,
|
|
114
|
+
requestsWithoutCost,
|
|
66
115
|
requestsWithoutUsage,
|
|
67
116
|
requests,
|
|
68
117
|
runCount: runs.length,
|
|
@@ -70,15 +119,27 @@ export function summarizeRuns(runs, opts) {
|
|
|
70
119
|
}
|
|
71
120
|
/**
|
|
72
121
|
* Format a cost figure honestly: enough precision for small per-run costs,
|
|
73
|
-
* trailing zeros trimmed.
|
|
74
|
-
*
|
|
75
|
-
*
|
|
122
|
+
* trailing zeros trimmed. The currency is the gateway's own code (e.g. "usd"),
|
|
123
|
+
* so it is appended as a UNIT (`0.0021 USD`) rather than prefixed as a symbol —
|
|
124
|
+
* a code reads as a suffix, and cruxy no longer has a user-configured symbol to
|
|
125
|
+
* prefix with. An absent currency renders the bare number rather than assuming
|
|
126
|
+
* one. Callers pass a cost only when the gateway actually quoted it.
|
|
76
127
|
*/
|
|
77
128
|
export function formatCost(n, currency) {
|
|
78
129
|
const abs = Math.abs(n);
|
|
79
130
|
const decimals = abs >= 1 ? 2 : abs >= 0.01 ? 4 : 6;
|
|
80
131
|
const trimmed = n.toFixed(decimals).replace(/\.?0+$/, "");
|
|
81
|
-
|
|
132
|
+
const amount = trimmed === "" || trimmed === "-0" ? "0" : trimmed;
|
|
133
|
+
return currency ? `${amount} ${currency.toUpperCase()}` : amount;
|
|
134
|
+
}
|
|
135
|
+
/**
|
|
136
|
+
* Weighted tokens at status-line width. Rounded to a whole token because the
|
|
137
|
+
* multipliers are fractional (3.25×, 3.8×) and a `812.5` would imply the meter
|
|
138
|
+
* counts half-tokens; the unrounded value stays on the summary for anything that
|
|
139
|
+
* needs the exact arithmetic.
|
|
140
|
+
*/
|
|
141
|
+
function formatWeighted(n) {
|
|
142
|
+
return formatTokens(Math.round(n));
|
|
82
143
|
}
|
|
83
144
|
/** `↑1.2k ↓340` (worded `up 1.2k down 340` under a screen reader). */
|
|
84
145
|
function tokenText(input, output, t) {
|
|
@@ -98,7 +159,9 @@ export function renderSummary(summary, t) {
|
|
|
98
159
|
for (const p of summary.perTier) {
|
|
99
160
|
const known = p.inputTokens !== undefined || p.outputTokens !== undefined;
|
|
100
161
|
if (known) {
|
|
101
|
-
const cost = p.cost !== undefined
|
|
162
|
+
const cost = p.cost !== undefined
|
|
163
|
+
? ` ${formatCost(p.cost, summary.costCurrency)}`
|
|
164
|
+
: "";
|
|
102
165
|
parts.push(`${p.tier} ${tokenText(p.inputTokens, p.outputTokens, t)}${t.muted(cost)}`);
|
|
103
166
|
}
|
|
104
167
|
else {
|
|
@@ -111,11 +174,19 @@ export function renderSummary(summary, t) {
|
|
|
111
174
|
const totalKnown = summary.totalInputTokens !== undefined ||
|
|
112
175
|
summary.totalOutputTokens !== undefined;
|
|
113
176
|
const totalCost = summary.totalCost !== undefined
|
|
114
|
-
? ` ${formatCost(summary.totalCost, summary.
|
|
177
|
+
? ` ${formatCost(summary.totalCost, summary.costCurrency)}`
|
|
115
178
|
: "";
|
|
116
179
|
parts.push(totalKnown
|
|
117
180
|
? `${t.strong("total")} ${tokenText(summary.totalInputTokens, summary.totalOutputTokens, t)}${t.muted(totalCost)}`
|
|
118
181
|
: `${t.strong("total")} —`);
|
|
182
|
+
// Weighted tokens — the gateway's own metering unit, computed locally from
|
|
183
|
+
// billable input + output. Shown only when at least one request could actually
|
|
184
|
+
// be weighed; the label stays "weighted", never anything implying a balance
|
|
185
|
+
// (`cruxy usage` carries the fuller wording). A reported 0 IS shown, same rule
|
|
186
|
+
// as `cached ↻0`: it is a measurement, not an absence.
|
|
187
|
+
if (summary.totalWeightedTokens !== undefined) {
|
|
188
|
+
parts.push(t.muted(`weighted ${formatWeighted(summary.totalWeightedTokens)}`));
|
|
189
|
+
}
|
|
119
190
|
// Cache reads (phase-A caching, Anthropic dev path): shown ONLY when the
|
|
120
191
|
// provider actually reported cache usage — `undefined` on the cruxy gateway
|
|
121
192
|
// and every OpenAI-compat path, so normal users never see this segment. A
|
|
@@ -130,6 +201,24 @@ export function renderSummary(summary, t) {
|
|
|
130
201
|
if (summary.requestsWithoutUsage > 0) {
|
|
131
202
|
parts.push(t.warning(unreported(summary.requestsWithoutUsage)));
|
|
132
203
|
}
|
|
204
|
+
// Coverage for the two derived figures. Each is emitted only when it says
|
|
205
|
+
// something the line doesn't already: `requestsWithoutUsage` above already
|
|
206
|
+
// explains requests that reported nothing at all, so these fire only for the
|
|
207
|
+
// EXTRA shortfall — requests that did report tokens but still couldn't be
|
|
208
|
+
// weighed (unknown tier / no billable input) or weren't quoted a cost.
|
|
209
|
+
const weightShortfall = summary.requestsWithoutWeight - summary.requestsWithoutUsage;
|
|
210
|
+
if (summary.totalWeightedTokens !== undefined && weightShortfall > 0) {
|
|
211
|
+
parts.push(t.muted(`${weightShortfall} not weighted`));
|
|
212
|
+
}
|
|
213
|
+
const costShortfall = summary.requestsWithoutCost - summary.requestsWithoutUsage;
|
|
214
|
+
if (summary.totalCost !== undefined && costShortfall > 0) {
|
|
215
|
+
parts.push(t.muted(`${costShortfall} without cost`));
|
|
216
|
+
}
|
|
217
|
+
// Two gateways quoting two currencies: refuse to add them, and say why rather
|
|
218
|
+
// than just dropping the cost silently.
|
|
219
|
+
if (summary.costMixedCurrency) {
|
|
220
|
+
parts.push(t.warning("cost omitted: mixed currencies"));
|
|
221
|
+
}
|
|
133
222
|
return parts.join(sep);
|
|
134
223
|
}
|
|
135
224
|
/** `2 requests: usage not reported`. */
|
package/dist/usage/types.js
CHANGED
|
@@ -65,8 +65,11 @@ export const UsageEntrySchema = z
|
|
|
65
65
|
cacheCreationTokens: z.number().int().nonnegative().optional(),
|
|
66
66
|
/**
|
|
67
67
|
* Tokens the gateway actually METERED as fresh input — cache-read tokens
|
|
68
|
-
* excluded, which is what makes the caching win legible
|
|
69
|
-
*
|
|
68
|
+
* excluded, which is what makes the caching win legible. This is the input
|
|
69
|
+
* term of the weighted-pool formula `(billable_input + output) × multiplier`
|
|
70
|
+
* that weighted.ts computes, and `inputTokens` is NEVER an acceptable
|
|
71
|
+
* stand-in for it: substituting a count that includes cache reads overstates
|
|
72
|
+
* precisely the requests caching made cheaper. Reported by the cruxy v2
|
|
70
73
|
* wire on every request INCLUDING as a literal `0`, so `undefined` here means
|
|
71
74
|
* a gateway too old to send it or a provider that never does — unknown, not
|
|
72
75
|
* zero. Never zero-filled.
|
|
@@ -0,0 +1,77 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Weighted tokens (C.22) — the unit the cruxy gateway actually meters a
|
|
3
|
+
* subscriber in.
|
|
4
|
+
*
|
|
5
|
+
* A subscriber is not billed per token and not billed in dollars: the gateway
|
|
6
|
+
* counts `(billable_input + output) × multiplier` per request and checks that
|
|
7
|
+
* against two windows. This module computes THAT number, locally, from figures
|
|
8
|
+
* already on disk — so `cruxy usage` can state consumption in the same unit the
|
|
9
|
+
* meter uses instead of a currency nobody is charged in.
|
|
10
|
+
*
|
|
11
|
+
* WHAT THIS IS NOT: it is not headroom, and it must never be rendered as such.
|
|
12
|
+
* A cap needs the user's plan, and how much of that plan every OTHER surface
|
|
13
|
+
* (web chat, desktop, phone) has already spent — neither of which is on this
|
|
14
|
+
* machine. All this module can honestly say is "this is what the runs in this
|
|
15
|
+
* CLI's own store weigh". Anything phrased as remaining/left/available would be
|
|
16
|
+
* a number this process cannot possibly know.
|
|
17
|
+
*
|
|
18
|
+
* No network: the multipliers are a compiled-in constant, not a fetched one.
|
|
19
|
+
*/
|
|
20
|
+
/**
|
|
21
|
+
* Weighted-token multipliers, keyed by tier — mirroring `multipliers` in the
|
|
22
|
+
* gateway's `internal/budget/config.go`, which is the source of truth.
|
|
23
|
+
*
|
|
24
|
+
* MIRRORED, THEREFORE DRIFTABLE. Nothing checks these against the server at
|
|
25
|
+
* runtime (that would be a network call this module is forbidden to make), so a
|
|
26
|
+
* gateway-side reweighting silently makes an old CLI's arithmetic stale. That
|
|
27
|
+
* is the accepted cost of computing offline, and it is why every rendering of
|
|
28
|
+
* this figure is framed as what THIS CLI reckons it used — never as an
|
|
29
|
+
* authoritative balance. If they drift, the server is right and this is wrong.
|
|
30
|
+
*
|
|
31
|
+
* Only the chat tiers are listed. The gateway also weighs its embed/guard/rerank
|
|
32
|
+
* models at 0.1×, but the CLI's chat path cannot produce a request on one, so
|
|
33
|
+
* including them would model traffic that never reaches this store.
|
|
34
|
+
*/
|
|
35
|
+
export const TIER_MULTIPLIERS = {
|
|
36
|
+
mira: 1.0,
|
|
37
|
+
vaani: 3.25,
|
|
38
|
+
kavi: 3.8,
|
|
39
|
+
};
|
|
40
|
+
/** The multiplier for a tier, or `undefined` when the tier isn't a known one. */
|
|
41
|
+
export function multiplierForTier(tier) {
|
|
42
|
+
return TIER_MULTIPLIERS[tier];
|
|
43
|
+
}
|
|
44
|
+
/**
|
|
45
|
+
* Weighted tokens for one request's figures, or `undefined` when the request
|
|
46
|
+
* cannot be weighed honestly.
|
|
47
|
+
*
|
|
48
|
+
* BOTH terms are required, and this is the whole discipline of the function.
|
|
49
|
+
* The meter's formula has two addends; substituting 0 for an unknown one does
|
|
50
|
+
* not produce an approximate answer, it produces a confidently understated one —
|
|
51
|
+
* and understating consumption is the direction that flatters, so it is exactly
|
|
52
|
+
* the error that would go unnoticed. `costFor` could price one known side
|
|
53
|
+
* because the raw token counts stayed visible beside it; here the weighted
|
|
54
|
+
* figure IS the deliverable, so a partial one is worse than none. Callers count
|
|
55
|
+
* the `undefined`s and surface them instead.
|
|
56
|
+
*
|
|
57
|
+
* Three ways a request is unweighable, all of them real and all treated alike:
|
|
58
|
+
* - no tier, or a tier this build doesn't know (a non-cruxy provider — those
|
|
59
|
+
* tokens never touched the weighted pool, so weighing them would be fiction);
|
|
60
|
+
* - no `billable_input_tokens` (a gateway too old to send it). Note that
|
|
61
|
+
* `inputTokens` is NOT a fallback: it includes cache-read tokens, which the
|
|
62
|
+
* meter excludes, so it would overstate every cached request — the precise
|
|
63
|
+
* inverse of the caching win the field exists to show;
|
|
64
|
+
* - no output count.
|
|
65
|
+
*
|
|
66
|
+
* A reported `0` on either side is a real measurement and weighs in as 0.
|
|
67
|
+
*/
|
|
68
|
+
export function weightedFor(tier, billableInputTokens, outputTokens) {
|
|
69
|
+
if (tier === undefined)
|
|
70
|
+
return undefined;
|
|
71
|
+
const multiplier = multiplierForTier(tier);
|
|
72
|
+
if (multiplier === undefined)
|
|
73
|
+
return undefined;
|
|
74
|
+
if (billableInputTokens === undefined || outputTokens === undefined)
|
|
75
|
+
return undefined;
|
|
76
|
+
return (billableInputTokens + outputTokens) * multiplier;
|
|
77
|
+
}
|
package/dist/utils/git.js
CHANGED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { spawnSync } from "node:child_process";
|
|
1
|
+
import { execFile, spawnSync } from "node:child_process";
|
|
2
2
|
/** Hard ceiling on a git invocation; a hung git must never stall startup. */
|
|
3
3
|
const GIT_TIMEOUT_MS = 5000;
|
|
4
4
|
/**
|
|
@@ -17,6 +17,35 @@ function runGit(args, cwd) {
|
|
|
17
17
|
}
|
|
18
18
|
return res.stdout;
|
|
19
19
|
}
|
|
20
|
+
/**
|
|
21
|
+
* {@link runGit} without blocking the event loop. Same contract in every other
|
|
22
|
+
* respect: `null` on any failure, never throws, same timeout.
|
|
23
|
+
*
|
|
24
|
+
* This exists because `spawnSync` is measurably expensive — around 45ms warm for
|
|
25
|
+
* the branch+status pair — and the TUI needs this data while a model response is
|
|
26
|
+
* streaming. Blocking the loop for 45ms mid-stream stutters the paint; blocking
|
|
27
|
+
* it from the paint path itself would do so on every frame.
|
|
28
|
+
*/
|
|
29
|
+
function runGitAsync(args, cwd) {
|
|
30
|
+
return new Promise((resolve) => {
|
|
31
|
+
execFile("git", args, { cwd, encoding: "utf8", timeout: GIT_TIMEOUT_MS, windowsHide: true }, (err, stdout) => resolve(err ? null : stdout));
|
|
32
|
+
});
|
|
33
|
+
}
|
|
34
|
+
/**
|
|
35
|
+
* Number of changed paths in `git status --porcelain` output.
|
|
36
|
+
*
|
|
37
|
+
* One line is one path, INCLUDING a rename (`R old -> new`), which is one
|
|
38
|
+
* change and not two. Blank lines are ignored so a trailing newline — always
|
|
39
|
+
* present on non-empty output — cannot inflate the count by one.
|
|
40
|
+
*/
|
|
41
|
+
export function countChanges(status) {
|
|
42
|
+
return status.split("\n").filter((line) => line.trim() !== "").length;
|
|
43
|
+
}
|
|
44
|
+
/** Shared shaping so the sync and async paths cannot disagree. */
|
|
45
|
+
function toGitInfo(branch, status) {
|
|
46
|
+
const changed = countChanges(status);
|
|
47
|
+
return { branch: branch.trim(), dirty: changed > 0, changed };
|
|
48
|
+
}
|
|
20
49
|
/**
|
|
21
50
|
* Branch name plus the raw `git status --porcelain` text for `cwd`, or `null`
|
|
22
51
|
* when it isn't a git repository (or git is unavailable). Backs the `git_status`
|
|
@@ -32,12 +61,29 @@ export function getGitStatus(cwd) {
|
|
|
32
61
|
return { branch: branch.trim(), status };
|
|
33
62
|
}
|
|
34
63
|
/**
|
|
35
|
-
* Compact git context for the system prompt: current branch
|
|
36
|
-
*
|
|
64
|
+
* Compact git context for the system prompt: current branch, whether the working
|
|
65
|
+
* tree has uncommitted changes, and how many paths changed. `null` when not a
|
|
66
|
+
* repo / git missing.
|
|
67
|
+
*
|
|
68
|
+
* The change count costs no extra subprocess — it is derived from the porcelain
|
|
69
|
+
* output {@link getGitStatus} already fetched.
|
|
37
70
|
*/
|
|
38
71
|
export function getGitInfo(cwd) {
|
|
39
72
|
const info = getGitStatus(cwd);
|
|
40
73
|
if (info === null)
|
|
41
74
|
return null;
|
|
42
|
-
return
|
|
75
|
+
return toGitInfo(info.branch, info.status);
|
|
76
|
+
}
|
|
77
|
+
/**
|
|
78
|
+
* {@link getGitInfo} without blocking the event loop — what the TUI rail uses.
|
|
79
|
+
* Identical result, identical `null` semantics.
|
|
80
|
+
*/
|
|
81
|
+
export async function getGitInfoAsync(cwd) {
|
|
82
|
+
const branch = await runGitAsync(["rev-parse", "--abbrev-ref", "HEAD"], cwd);
|
|
83
|
+
if (branch === null)
|
|
84
|
+
return null;
|
|
85
|
+
const status = await runGitAsync(["status", "--porcelain"], cwd);
|
|
86
|
+
if (status === null)
|
|
87
|
+
return null;
|
|
88
|
+
return toGitInfo(branch, status);
|
|
43
89
|
}
|