aegiscode 6.8.2 → 6.8.4
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/bin/aegiscode.js +35 -5
- package/package.json +1 -1
- package/scripts/predist.mjs +99 -0
- package/src/app.js +40 -8
- package/src/chatflow.js +181 -33
- package/src/commands.js +172 -22
- package/src/deps.js +2 -1
- package/src/events.js +37 -14
- package/src/mcp.js +683 -0
- package/src/panels.js +65 -15
- package/src/screen.js +130 -5
- package/src/screens.js +36 -7
- package/src/signup.js +153 -0
- package/src/title.js +217 -0
- package/src/tokens.js +74 -95
- package/src/update.js +75 -1
- package/vendor/desktop/lib/local/engine.js +24 -0
- package/vendor/desktop/lib/settings.js +67 -0
- package/vendor/desktop/renderer/usage.js +372 -41
- package/vendor/hosts/install.js +638 -0
- package/vendor/hosts/openai-shim.js +419 -0
- package/vendor/hosts/spec.js +164 -0
- package/vendor/hosts/targets.js +346 -0
- package/vendor/mcp/server.js +186 -0
package/src/title.js
ADDED
|
@@ -0,0 +1,217 @@
|
|
|
1
|
+
'use strict';
|
|
2
|
+
|
|
3
|
+
/**
|
|
4
|
+
* The terminal title bar — the session's *topic*.
|
|
5
|
+
*
|
|
6
|
+
* Claude Code keeps the window/tab title pointed at what the session is about
|
|
7
|
+
* ("✳ Fixing the nexus hang") rather than at a static product string, so a row
|
|
8
|
+
* of terminal tabs is readable at a glance. This host shipped the static half
|
|
9
|
+
* only: `chatflow.js`'s `setTitle()` wrote "AEGIS Code" forever, spinning a
|
|
10
|
+
* braille frame while a turn ran and never naming the work. Every tab of every
|
|
11
|
+
* session read the same, which is exactly the problem the topic solves.
|
|
12
|
+
*
|
|
13
|
+
* The topic is derived from the user's own prompt, locally — no model call, so
|
|
14
|
+
* it costs nothing, adds no latency to the first frame, and cannot invent a
|
|
15
|
+
* task the user never asked for. It is re-derived per prompt and adopted only
|
|
16
|
+
* when the ask is about something new (`adoptTopic`), so a follow-up like
|
|
17
|
+
* "now add a test for that" keeps the title it inherited.
|
|
18
|
+
*
|
|
19
|
+
* Everything here is pure and returns strings, so the whole surface is
|
|
20
|
+
* assertable without a TTY; `writeTitle` is the one function that touches a
|
|
21
|
+
* stream, and it swallows a non-writable one.
|
|
22
|
+
*/
|
|
23
|
+
|
|
24
|
+
/** The title the app wears before a topic exists (and after one is cleared). */
|
|
25
|
+
const DEFAULT_TITLE = 'AEGIS Code';
|
|
26
|
+
|
|
27
|
+
/** The braille frames the title spins while a turn runs (the reference's). */
|
|
28
|
+
const TITLE_SPIN = ['⠐', '⠂', '⠄', '⠆', '⠈', '⠠', '⠰', '⠁'];
|
|
29
|
+
|
|
30
|
+
/** Longest topic, in cells, before it is clipped on a word boundary. */
|
|
31
|
+
const TOPIC_MAX = 48;
|
|
32
|
+
|
|
33
|
+
/** Longest string that ever reaches the terminal (topic + " · AEGIS Code"). */
|
|
34
|
+
const TITLE_MAX = 96;
|
|
35
|
+
|
|
36
|
+
/**
|
|
37
|
+
* Politeness and throat-clearing, stripped from the front of a prompt so the
|
|
38
|
+
* title says "Add a retry to the fetcher" instead of "Please can you add a
|
|
39
|
+
* retry to the fetcher". Applied repeatedly — "ok so please can you…" is one
|
|
40
|
+
* habit, not four.
|
|
41
|
+
*/
|
|
42
|
+
const FILLER =
|
|
43
|
+
/^(?:please|pls|plz|hey|hi|hello|ok(?:ay)?|so|now|just|also|and|then|maybe|well|um|uh|can you|could you|would you|will you|i want you to|i want to|i need you to|i need to|i'?d like you to|i would like you to|i'?d like to|help me|help|let'?s|lets|go ahead and|i'?m trying to|i am trying to|i'?m working on|i am working on)\b[\s,:;-]*/i;
|
|
44
|
+
|
|
45
|
+
/** Leading markdown/list/quote noise: `> `, `- `, `1. `, `## `. */
|
|
46
|
+
const LEADING_NOISE = /^(?:[>#*•·\-–—]+|\d+[.)])\s*/;
|
|
47
|
+
|
|
48
|
+
/** Words that carry no topic signal when comparing two asks. */
|
|
49
|
+
const STOPWORDS = new Set([
|
|
50
|
+
'the', 'and', 'for', 'with', 'that', 'this', 'these', 'those', 'you', 'your', 'yours',
|
|
51
|
+
'can', 'could', 'would', 'should', 'will', 'please', 'pls', 'hey', 'now', 'then',
|
|
52
|
+
'from', 'into', 'onto', 'over', 'under', 'about', 'also', 'just', 'like', 'want',
|
|
53
|
+
'need', 'make', 'made', 'use', 'using', 'get', 'got', 'let', 'lets', 'its', 'it',
|
|
54
|
+
'are', 'was', 'were', 'been', 'has', 'have', 'had', 'but', 'not', 'all', 'any',
|
|
55
|
+
'how', 'what', 'when', 'where', 'which', 'who', 'why', 'there', 'here', 'some',
|
|
56
|
+
'one', 'two', 'add', 'new', 'fix', 'fixing', 'help', 'doing', 'does', 'did', 'me',
|
|
57
|
+
'my', 'our', 'we', 'i', 'a', 'an', 'to', 'of', 'in', 'on', 'at', 'is', 'be', 'do',
|
|
58
|
+
]);
|
|
59
|
+
|
|
60
|
+
/** Strip anything that could break out of the OSC title sequence. */
|
|
61
|
+
function sanitize(text, max = TITLE_MAX) {
|
|
62
|
+
const flat = String(text == null ? '' : text)
|
|
63
|
+
// Control chars (including ESC, BEL, CR, LF, TAB) become a space, so a
|
|
64
|
+
// prompt can never inject an escape sequence into the title bar — the one
|
|
65
|
+
// place in this UI where user text reaches the terminal unsanitized.
|
|
66
|
+
.replace(/[\u0000-\u001f\u007f]/g, ' ')
|
|
67
|
+
.replace(/\s+/g, ' ')
|
|
68
|
+
.trim();
|
|
69
|
+
if (flat.length <= max) return flat;
|
|
70
|
+
return `${flat.slice(0, Math.max(0, max - 1)).trimEnd()}…`;
|
|
71
|
+
}
|
|
72
|
+
|
|
73
|
+
/**
|
|
74
|
+
* A short, single-line topic for a user prompt.
|
|
75
|
+
*
|
|
76
|
+
* Uses the first line that has anything on it (a pasted block's first line is
|
|
77
|
+
* its subject), drops politeness and list noise, collapses whitespace, and
|
|
78
|
+
* clips on a word boundary with an ellipsis. Returns '' when there is nothing
|
|
79
|
+
* to say — the caller then keeps the title it already had.
|
|
80
|
+
*/
|
|
81
|
+
function topicFrom(text, { max = TOPIC_MAX } = {}) {
|
|
82
|
+
if (text == null) return '';
|
|
83
|
+
const lines = String(text).split(/[\r\n]+/);
|
|
84
|
+
let line = '';
|
|
85
|
+
for (const raw of lines) {
|
|
86
|
+
const cleaned = sanitize(raw, 4_000).replace(LEADING_NOISE, '').trim();
|
|
87
|
+
// A pasted block's fence marker is not the topic; the code under it is.
|
|
88
|
+
if (!cleaned || /^```/.test(cleaned)) continue;
|
|
89
|
+
line = cleaned;
|
|
90
|
+
break;
|
|
91
|
+
}
|
|
92
|
+
if (!line) return '';
|
|
93
|
+
let body = line;
|
|
94
|
+
for (let i = 0; i < 6; i++) {
|
|
95
|
+
const next = body.replace(FILLER, '');
|
|
96
|
+
if (next === body) break;
|
|
97
|
+
body = next.replace(LEADING_NOISE, '').trim();
|
|
98
|
+
}
|
|
99
|
+
body = body.replace(/[.!?]+$/, '').trim();
|
|
100
|
+
if (!body) body = line;
|
|
101
|
+
if (/^[a-z]/.test(body)) body = body[0].toUpperCase() + body.slice(1);
|
|
102
|
+
return sanitize(body, max);
|
|
103
|
+
}
|
|
104
|
+
|
|
105
|
+
/** Topic-defining words of a phrase: lowercase, deduped, stopwords dropped. */
|
|
106
|
+
function keywords(phrase) {
|
|
107
|
+
const out = new Set();
|
|
108
|
+
for (const word of String(phrase || '').toLowerCase().match(/[\p{L}\p{N}_]+/gu) || []) {
|
|
109
|
+
if (word.length < 3 || STOPWORDS.has(word)) continue;
|
|
110
|
+
out.add(word);
|
|
111
|
+
}
|
|
112
|
+
return [...out];
|
|
113
|
+
}
|
|
114
|
+
|
|
115
|
+
/**
|
|
116
|
+
* Does `next` read as a *different* task from `prev`?
|
|
117
|
+
*
|
|
118
|
+
* The overlap is measured against the shorter keyword set, so a long follow-up
|
|
119
|
+
* that repeats the subject ("also add a test for the config loader") is not
|
|
120
|
+
* mistaken for a new task, while a genuinely unrelated ask ("now write the
|
|
121
|
+
* changelog") is. An empty `next` is never a shift — it means "no topic", and
|
|
122
|
+
* the caller keeps what it had.
|
|
123
|
+
*/
|
|
124
|
+
function titleShift(prev, next, { threshold = 0.34 } = {}) {
|
|
125
|
+
if (!next) return false;
|
|
126
|
+
if (!prev) return true;
|
|
127
|
+
const a = keywords(prev);
|
|
128
|
+
const b = keywords(next);
|
|
129
|
+
if (!a.length || !b.length) return true;
|
|
130
|
+
const set = new Set(a);
|
|
131
|
+
let hits = 0;
|
|
132
|
+
for (const word of b) if (set.has(word)) hits++;
|
|
133
|
+
return hits / Math.min(a.length, b.length) < threshold;
|
|
134
|
+
}
|
|
135
|
+
|
|
136
|
+
/** The topic after this prompt: the new one only when it is a new task. */
|
|
137
|
+
function adoptTopic(prev, prompt, opts) {
|
|
138
|
+
const next = topicFrom(prompt);
|
|
139
|
+
if (!next) return prev || '';
|
|
140
|
+
return titleShift(prev, next, opts) ? next : prev || '';
|
|
141
|
+
}
|
|
142
|
+
|
|
143
|
+
/**
|
|
144
|
+
* The composed title.
|
|
145
|
+
*
|
|
146
|
+
* working, with topic `⠂ Fixing the config loader`
|
|
147
|
+
* idle, with topic `Fixing the config loader · AEGIS Code`
|
|
148
|
+
* no topic `AEGIS Code`
|
|
149
|
+
*
|
|
150
|
+
* The idle form keeps the product name so a tab is still identifiable as this
|
|
151
|
+
* CLI; the working form is the topic alone, which is what Claude Code shows
|
|
152
|
+
* while a turn runs.
|
|
153
|
+
*/
|
|
154
|
+
function titleText({ topic = '', frame = null, app = DEFAULT_TITLE } = {}) {
|
|
155
|
+
const clean = sanitize(topic, TOPIC_MAX);
|
|
156
|
+
if (!clean) return sanitize(app, TITLE_MAX) || DEFAULT_TITLE;
|
|
157
|
+
const head = frame ? `${frame} ` : '';
|
|
158
|
+
return sanitize(frame ? `${head}${clean}` : `${clean} · ${app}`, TITLE_MAX);
|
|
159
|
+
}
|
|
160
|
+
|
|
161
|
+
/** Truthiness for the env flags: `1`, `true`, `yes`, `on` (any case). */
|
|
162
|
+
function flagOn(value) {
|
|
163
|
+
if (value === true) return true;
|
|
164
|
+
if (value == null || value === false) return false;
|
|
165
|
+
return /^(?:1|true|yes|on)$/i.test(String(value).trim());
|
|
166
|
+
}
|
|
167
|
+
|
|
168
|
+
/**
|
|
169
|
+
* Is the title bar ours to write?
|
|
170
|
+
*
|
|
171
|
+
* On by default — a title that follows the work is the feature. Off via
|
|
172
|
+
* `AEGIS_DISABLE_TERMINAL_TITLE=1` (and Claude Code's own
|
|
173
|
+
* `CLAUDE_CODE_DISABLE_TERMINAL_TITLE`, so a shell configured for the
|
|
174
|
+
* reference behaves the same here), or `terminalTitle: false` in config.json.
|
|
175
|
+
* `AEGIS_TERMINAL_TITLE=1` wins over the config, for a one-off run.
|
|
176
|
+
*/
|
|
177
|
+
function titleEnabled(env = {}, config = {}) {
|
|
178
|
+
const e = env || {};
|
|
179
|
+
if (flagOn(e.AEGIS_TERMINAL_TITLE)) return true;
|
|
180
|
+
if (flagOn(e.AEGIS_DISABLE_TERMINAL_TITLE)) return false;
|
|
181
|
+
if (flagOn(e.CLAUDE_CODE_DISABLE_TERMINAL_TITLE)) return false;
|
|
182
|
+
const cfg = config || {};
|
|
183
|
+
if (cfg.terminalTitle === false || cfg.terminalTitleBar === false) return false;
|
|
184
|
+
return true;
|
|
185
|
+
}
|
|
186
|
+
|
|
187
|
+
/**
|
|
188
|
+
* Write one title to a stream as an OSC 0 sequence. `''` clears it (what the
|
|
189
|
+
* teardown wants, so the shell that gets the terminal back is not left wearing
|
|
190
|
+
* this app's title). Never throws: a stubbed, closed or non-TTY stream is a
|
|
191
|
+
* missing nicety, not a reason to lose a turn.
|
|
192
|
+
*/
|
|
193
|
+
function writeTitle(text, out) {
|
|
194
|
+
const stream = out || process.stdout;
|
|
195
|
+
if (!stream || typeof stream.write !== 'function') return false;
|
|
196
|
+
try {
|
|
197
|
+
stream.write(`\x1b]0;${sanitize(text, TITLE_MAX)}\x07`);
|
|
198
|
+
return true;
|
|
199
|
+
} catch {
|
|
200
|
+
return false;
|
|
201
|
+
}
|
|
202
|
+
}
|
|
203
|
+
|
|
204
|
+
module.exports = {
|
|
205
|
+
DEFAULT_TITLE,
|
|
206
|
+
TITLE_SPIN,
|
|
207
|
+
TOPIC_MAX,
|
|
208
|
+
TITLE_MAX,
|
|
209
|
+
sanitize,
|
|
210
|
+
topicFrom,
|
|
211
|
+
keywords,
|
|
212
|
+
titleShift,
|
|
213
|
+
adoptTopic,
|
|
214
|
+
titleText,
|
|
215
|
+
titleEnabled,
|
|
216
|
+
writeTitle,
|
|
217
|
+
};
|
package/src/tokens.js
CHANGED
|
@@ -3,92 +3,36 @@
|
|
|
3
3
|
/**
|
|
4
4
|
* Token + cost estimation. Tokens are estimated from text length (~4
|
|
5
5
|
* chars/token, the usual rule of thumb) and priced against the per-model rates
|
|
6
|
-
*
|
|
6
|
+
* in the SHARED table. Everything is labeled approximate in the UI.
|
|
7
7
|
*
|
|
8
|
-
*
|
|
8
|
+
* The rate table itself is not here — see the block below.
|
|
9
9
|
*/
|
|
10
10
|
|
|
11
|
-
//
|
|
11
|
+
// ── The rate table is NOT defined here ─────────────────────────────────────
|
|
12
12
|
//
|
|
13
|
-
//
|
|
14
|
-
//
|
|
15
|
-
//
|
|
16
|
-
//
|
|
17
|
-
//
|
|
13
|
+
// This file used to carry its own copy of `RATES`, and that copy is how the
|
|
14
|
+
// DeepSeek row came to be wrong in one host and right in the other: the table
|
|
15
|
+
// existed twice (here and desktop/renderer/usage.js), so a rate read off the
|
|
16
|
+
// vendor's pricing page reached whichever copy someone happened to edit. The
|
|
17
|
+
// CLI's copy still said $0.14/$0.28 — the retired V4-Flash tier — while the
|
|
18
|
+
// desktop's had been corrected, and `/cost` would have disagreed with the GUI
|
|
19
|
+
// about the same turn on the same engine.
|
|
18
20
|
//
|
|
19
|
-
//
|
|
20
|
-
//
|
|
21
|
-
|
|
22
|
-
|
|
23
|
-
|
|
24
|
-
|
|
25
|
-
|
|
26
|
-
|
|
27
|
-
|
|
28
|
-
|
|
29
|
-
//
|
|
30
|
-
// This row was MISSING, and usageCost fell through to RATES.sonnet for every
|
|
31
|
-
// DeepSeek turn: $3.00/$15.00 per M against a real $0.14/$0.28 — 21x the
|
|
32
|
-
// input rate and 54x the output rate, on every direct DeepSeek call. That is
|
|
33
|
-
// the number that made a local turn look like it cost a fraction of the
|
|
34
|
-
// cloud's, when most of the reported gap was the meter itself.
|
|
35
|
-
//
|
|
36
|
-
// cacheRead is 0.1x input (aegis1 services/pricing.py
|
|
37
|
-
// CACHE_READ_FRACTION_BY_COMPANY["deepseek"] = 0.1 -> $0.014/M).
|
|
38
|
-
// cacheWrite is the input rate: DeepSeek bills a cache write as ordinary
|
|
39
|
-
// input tokens and charges no separate write premium, unlike Anthropic's
|
|
40
|
-
// 1.25x. A zero here would understate a session that writes cache.
|
|
41
|
-
deepseek: { input: 0.14, output: 0.28, cacheRead: 0.014, cacheWrite: 0.14 },
|
|
42
|
-
};
|
|
21
|
+
// So the table lives in exactly one place. `desktop/renderer/usage.js` is
|
|
22
|
+
// already the declared single source of truth for turning a provider's usage
|
|
23
|
+
// object into a display number (cli/src/deps.js resolves it, and
|
|
24
|
+
// test/cli-tools.test.mjs asserts the function identity), so the rates belong
|
|
25
|
+
// there too rather than in a second table that can only drift.
|
|
26
|
+
//
|
|
27
|
+
// `sharedpaths.js` is used rather than `deps.js` because deps.js eagerly
|
|
28
|
+
// requires the engine, the client and the tool registry — pulling all of that
|
|
29
|
+
// in to read a number would be a require cycle for anything in that graph.
|
|
30
|
+
const path = require('node:path');
|
|
43
31
|
|
|
44
|
-
|
|
45
|
-
|
|
46
|
-
* Anthropic spells its family *inside* the id — "claude-opus-4" matches no
|
|
47
|
-
* prefix — so prefix matching alone cannot see it.
|
|
48
|
-
*/
|
|
49
|
-
const RATE_FAMILIES = Object.keys(RATES).filter((k) => k !== 'default');
|
|
32
|
+
const { resolveShared } = require('./sharedpaths.js');
|
|
33
|
+
const shared = require(resolveShared(path.join('desktop', 'renderer', 'usage.js')));
|
|
50
34
|
|
|
51
|
-
|
|
52
|
-
* The rate row for a model id, provider, or alias. Exact id wins, then the
|
|
53
|
-
* longest matching prefix, then a family named inside the id, then the
|
|
54
|
-
* Sonnet-class default — the same resolution order contextWindowFor uses, so
|
|
55
|
-
* the cost meter and the context meter cannot disagree about which family a
|
|
56
|
-
* model belongs to.
|
|
57
|
-
*/
|
|
58
|
-
function ratesFor(model) {
|
|
59
|
-
const raw = String(model || '').toLowerCase();
|
|
60
|
-
if (!raw) return RATES.default;
|
|
61
|
-
// `/class byok` selects a "provider:model" id, and the provider half is a
|
|
62
|
-
// routing label, not a rate: rating it as written fell through every prefix
|
|
63
|
-
// to RATES.default, so an Opus BYOK turn was priced at Sonnet's $3/$15
|
|
64
|
-
// against the vendor's real $5/$25.
|
|
65
|
-
const ids = [raw];
|
|
66
|
-
const tail = raw.slice(raw.lastIndexOf(':') + 1);
|
|
67
|
-
if (tail && tail !== raw) ids.push(tail);
|
|
68
|
-
for (const id of ids) {
|
|
69
|
-
if (RATES[id]) return RATES[id];
|
|
70
|
-
let best = null;
|
|
71
|
-
let bestLen = 0;
|
|
72
|
-
for (const [prefix, rates] of Object.entries(RATES)) {
|
|
73
|
-
if (id.startsWith(prefix) && prefix.length > bestLen) {
|
|
74
|
-
best = rates;
|
|
75
|
-
bestLen = prefix.length;
|
|
76
|
-
}
|
|
77
|
-
}
|
|
78
|
-
if (best) return best;
|
|
79
|
-
}
|
|
80
|
-
// Longest family-key hit wins, so an overlapping pair resolves the same way
|
|
81
|
-
// twice running rather than by object key order.
|
|
82
|
-
let family = null;
|
|
83
|
-
let familyLen = 0;
|
|
84
|
-
for (const key of RATE_FAMILIES) {
|
|
85
|
-
if (raw.includes(key) && key.length > familyLen) {
|
|
86
|
-
family = RATES[key];
|
|
87
|
-
familyLen = key.length;
|
|
88
|
-
}
|
|
89
|
-
}
|
|
90
|
-
return family || RATES.default;
|
|
91
|
-
}
|
|
35
|
+
const { RATES, RATE_BASIS, UNPRICED_BY_DESIGN, isPeak, ratesFor, costBreakdown, cacheHitRate } = shared;
|
|
92
36
|
|
|
93
37
|
// System prompt + tool definitions overhead, roughly, in tokens.
|
|
94
38
|
const SYSTEM_TOKENS = 18000;
|
|
@@ -127,11 +71,11 @@ const CONTEXT_WINDOWS = {
|
|
|
127
71
|
function contextWindowFor(model) {
|
|
128
72
|
const raw = String(model || '').toLowerCase();
|
|
129
73
|
if (!raw) return CONTEXT_WINDOW;
|
|
130
|
-
// Same two-candidate lookup as ratesFor
|
|
131
|
-
// id ("deepseek:deepseek-v4-flash") carries a routing label in
|
|
132
|
-
// model, and matching it as written left the 1M-token DeepSeek
|
|
133
|
-
// as the 200k default — a meter that would have told a user
|
|
134
|
-
// of a context that was 4% full.
|
|
74
|
+
// Same two-candidate lookup as ratesFor in the shared table, for the same
|
|
75
|
+
// reason: a byok id ("deepseek:deepseek-v4-flash") carries a routing label in
|
|
76
|
+
// front of the model, and matching it as written left the 1M-token DeepSeek
|
|
77
|
+
// window reading as the 200k default — a meter that would have told a user
|
|
78
|
+
// they were at 20% of a context that was 4% full.
|
|
135
79
|
const ids = [raw];
|
|
136
80
|
const tail = raw.slice(raw.lastIndexOf(':') + 1);
|
|
137
81
|
if (tail && tail !== raw) ids.push(tail);
|
|
@@ -168,14 +112,24 @@ function transcriptUsage(transcript, model = 'sonnet') {
|
|
|
168
112
|
return { input, output, cacheRead, cacheWrite };
|
|
169
113
|
}
|
|
170
114
|
|
|
171
|
-
/**
|
|
172
|
-
|
|
173
|
-
|
|
174
|
-
|
|
175
|
-
|
|
176
|
-
|
|
177
|
-
|
|
178
|
-
|
|
115
|
+
/**
|
|
116
|
+
* Dollar cost of a usage record at the given model's rates.
|
|
117
|
+
*
|
|
118
|
+
* Returns a number for backward compatibility, but the basis is not lost: use
|
|
119
|
+
* `costBreakdown` when the caller needs to know whether the figure is real.
|
|
120
|
+
* (`usageCost` in the shared module, which also folds the record through
|
|
121
|
+
* `usageBuckets` first — so a wire usage object is normalized identically in
|
|
122
|
+
* both hosts, where this file's old copy read `usage.input` directly.)
|
|
123
|
+
*
|
|
124
|
+
* `model` has no `= 'sonnet'` default here either, and that matters for more
|
|
125
|
+
* than tidiness: a default parameter in THIS file would re-introduce exactly
|
|
126
|
+
* the split the shared module just removed, and only on the CLI side — so
|
|
127
|
+
* `usageCost(u)` would report `basis: 'exact'` in the terminal and
|
|
128
|
+
* `basis: 'default'` in the GUI for the same turn on the same engine. The
|
|
129
|
+
* whole point of delegating is that both hosts answer identically.
|
|
130
|
+
*/
|
|
131
|
+
function usageCost(usage, model, opts = {}) {
|
|
132
|
+
return shared.usageCost(usage, model, opts);
|
|
179
133
|
}
|
|
180
134
|
|
|
181
135
|
/** Full session accounting: per-bucket tokens, cost, and context used %. */
|
|
@@ -186,6 +140,14 @@ function sessionAccounting(transcript, model = 'sonnet') {
|
|
|
186
140
|
const history = usage.input + usage.output;
|
|
187
141
|
const used = system + tools + history;
|
|
188
142
|
const contextWindow = contextWindowFor(model);
|
|
143
|
+
// ONE breakdown, not a cost call plus a separate basis call: ratesFor reads
|
|
144
|
+
// the wall clock for peak/off-peak, so two evaluations straddling a peak
|
|
145
|
+
// boundary would report a peak cost with an off-peak basis (or vice versa) —
|
|
146
|
+
// two different claims about the same number, which is the thing this return
|
|
147
|
+
// shape exists to prevent. aegiscodex-dev's sessionAccounting is written the
|
|
148
|
+
// same way; if one host folds this into two calls again the two will disagree
|
|
149
|
+
// about the same turn, which is how the rate table split in the first place.
|
|
150
|
+
const breakdown = costBreakdown(usage, model);
|
|
189
151
|
return {
|
|
190
152
|
usage,
|
|
191
153
|
system,
|
|
@@ -194,7 +156,12 @@ function sessionAccounting(transcript, model = 'sonnet') {
|
|
|
194
156
|
used,
|
|
195
157
|
contextWindow,
|
|
196
158
|
pct: Math.min(100, Math.round((used / contextWindow) * 100)),
|
|
197
|
-
cost:
|
|
159
|
+
cost: breakdown.cost,
|
|
160
|
+
// Which rate row priced this. 'default' means the model is unpriced and
|
|
161
|
+
// `cost` is a Sonnet-class placeholder — surfaced rather than swallowed so
|
|
162
|
+
// /cost can label it instead of printing a fabricated figure.
|
|
163
|
+
costBasis: breakdown.basis,
|
|
164
|
+
costPriced: breakdown.priced,
|
|
198
165
|
};
|
|
199
166
|
}
|
|
200
167
|
|
|
@@ -204,8 +171,9 @@ function sessionAccounting(transcript, model = 'sonnet') {
|
|
|
204
171
|
* already includes system prompt + tools in cacheRead, so `used` counts
|
|
205
172
|
* input + output + cache instead of re-adding the constants.
|
|
206
173
|
*/
|
|
207
|
-
function accountingFromUsage(usage, model = 'sonnet', { exchanges = 0, real = false, costUsd } = {}) {
|
|
208
|
-
const
|
|
174
|
+
function accountingFromUsage(usage, model = 'sonnet', { exchanges = 0, real = false, costUsd, at } = {}) {
|
|
175
|
+
const breakdown = costBreakdown(usage, model, { at });
|
|
176
|
+
const cost = typeof costUsd === 'number' ? costUsd : breakdown.cost;
|
|
209
177
|
const used = real
|
|
210
178
|
? usage.input + usage.output + usage.cacheRead + usage.cacheWrite
|
|
211
179
|
: SYSTEM_TOKENS + TOOL_TOKENS + usage.input + usage.output;
|
|
@@ -219,6 +187,12 @@ function accountingFromUsage(usage, model = 'sonnet', { exchanges = 0, real = fa
|
|
|
219
187
|
contextWindow,
|
|
220
188
|
pct: Math.min(100, Math.round((used / contextWindow) * 100)),
|
|
221
189
|
cost,
|
|
190
|
+
// A settled charge (costUsd) is the ledger's own truth, so it is 'settled'
|
|
191
|
+
// rather than a table match — the pool's bill carries a margin and a
|
|
192
|
+
// prompt-cache discount the table cannot see, so the two must never be
|
|
193
|
+
// conflated.
|
|
194
|
+
costBasis: typeof costUsd === 'number' ? 'settled' : breakdown.basis,
|
|
195
|
+
costPriced: typeof costUsd === 'number' ? true : breakdown.priced,
|
|
222
196
|
exchanges,
|
|
223
197
|
real,
|
|
224
198
|
};
|
|
@@ -235,12 +209,17 @@ function fmtCost(d) {
|
|
|
235
209
|
|
|
236
210
|
module.exports = {
|
|
237
211
|
RATES,
|
|
212
|
+
RATE_BASIS,
|
|
213
|
+
UNPRICED_BY_DESIGN,
|
|
214
|
+
isPeak,
|
|
238
215
|
SYSTEM_TOKENS,
|
|
239
216
|
TOOL_TOKENS,
|
|
240
217
|
CONTEXT_WINDOW,
|
|
241
218
|
CONTEXT_WINDOWS,
|
|
242
219
|
contextWindowFor,
|
|
243
220
|
ratesFor,
|
|
221
|
+
costBreakdown,
|
|
222
|
+
cacheHitRate,
|
|
244
223
|
estimateTokens,
|
|
245
224
|
transcriptUsage,
|
|
246
225
|
usageCost,
|
package/src/update.js
CHANGED
|
@@ -7,7 +7,81 @@
|
|
|
7
7
|
* the same thing for `aegis-desktop`: two hosts, two package names, one
|
|
8
8
|
* checker. Resolved through deps.js so the in-repo and published-vendor
|
|
9
9
|
* layouts both work.
|
|
10
|
+
*
|
|
11
|
+
* On top of the shared checker this module carries the three pieces /upgrade
|
|
12
|
+
* needs — `compareVersions`, `upgradeAdvice` and `latestPublishedVersion` —
|
|
13
|
+
* ported from the reference (`aegiscodex-dev/src/session.js`). They belong
|
|
14
|
+
* here, next to `isNewer`/`updateLine`, so the CLI has ONE idea of "is there a
|
|
15
|
+
* newer published version": the welcome-box notice and /upgrade now answer
|
|
16
|
+
* from the same place and compare versions the same way. The shared module's
|
|
17
|
+
* own functions are spread through unchanged, keeping the identity the
|
|
18
|
+
* update-channel test pins (the shim and the shared file must be the SAME
|
|
19
|
+
* code, not a copy).
|
|
10
20
|
*/
|
|
11
21
|
|
|
22
|
+
const { execFileSync } = require('node:child_process');
|
|
23
|
+
|
|
12
24
|
const { resolveShared } = require('./deps.js');
|
|
13
|
-
|
|
25
|
+
|
|
26
|
+
const shared = require(resolveShared('client/update.js'));
|
|
27
|
+
|
|
28
|
+
const VERSION = require('../package.json').version;
|
|
29
|
+
/** Both hosts publish from npm; the CLI is `aegiscode`. */
|
|
30
|
+
const PKG = shared.PKG || 'aegiscode';
|
|
31
|
+
|
|
32
|
+
// ── /upgrade ─────────────────────────────────────────────────────────────────
|
|
33
|
+
|
|
34
|
+
/** Compare two dotted version strings: -1 | 0 | 1 (prerelease tags ignored). */
|
|
35
|
+
function compareVersions(a, b) {
|
|
36
|
+
const pa = String(a || '0').split('-')[0].split('.').map((n) => Number.parseInt(n, 10) || 0);
|
|
37
|
+
const pb = String(b || '0').split('-')[0].split('.').map((n) => Number.parseInt(n, 10) || 0);
|
|
38
|
+
for (let i = 0; i < Math.max(pa.length, pb.length); i++) {
|
|
39
|
+
const d = (pa[i] || 0) - (pb[i] || 0);
|
|
40
|
+
if (d) return d > 0 ? 1 : -1;
|
|
41
|
+
}
|
|
42
|
+
return 0;
|
|
43
|
+
}
|
|
44
|
+
|
|
45
|
+
/** What /upgrade should say, given the local and published versions. */
|
|
46
|
+
function upgradeAdvice({ current = VERSION, latest = null, packageName = PKG } = {}) {
|
|
47
|
+
if (!latest) {
|
|
48
|
+
return {
|
|
49
|
+
upToDate: null,
|
|
50
|
+
lines: [` Could not read the npm registry — local version is ${current}.`],
|
|
51
|
+
};
|
|
52
|
+
}
|
|
53
|
+
const cmp = compareVersions(latest, current);
|
|
54
|
+
if (cmp > 0) {
|
|
55
|
+
return {
|
|
56
|
+
upToDate: false,
|
|
57
|
+
lines: [
|
|
58
|
+
` Installed: ${current}`,
|
|
59
|
+
` Published: ${latest}`,
|
|
60
|
+
` Upgrade: npm i -g ${packageName}@latest`,
|
|
61
|
+
],
|
|
62
|
+
};
|
|
63
|
+
}
|
|
64
|
+
return {
|
|
65
|
+
upToDate: true,
|
|
66
|
+
lines: [` Installed ${current} is the published ${latest} — nothing to upgrade.`],
|
|
67
|
+
};
|
|
68
|
+
}
|
|
69
|
+
|
|
70
|
+
/** Best-effort registry lookup (5s cap). Returns a version string or null. */
|
|
71
|
+
function latestPublishedVersion({ packageName = PKG, timeoutMs = 5000 } = {}) {
|
|
72
|
+
try {
|
|
73
|
+
const out = execFileSync('npm', ['view', packageName, 'version'], {
|
|
74
|
+
encoding: 'utf8', timeout: timeoutMs, stdio: ['ignore', 'pipe', 'ignore'],
|
|
75
|
+
});
|
|
76
|
+
const v = String(out).trim();
|
|
77
|
+
return /^\d+\.\d+\.\d+/.test(v) ? v : null;
|
|
78
|
+
} catch { return null; }
|
|
79
|
+
}
|
|
80
|
+
|
|
81
|
+
module.exports = {
|
|
82
|
+
...shared,
|
|
83
|
+
VERSION,
|
|
84
|
+
compareVersions,
|
|
85
|
+
upgradeAdvice,
|
|
86
|
+
latestPublishedVersion,
|
|
87
|
+
};
|
|
@@ -437,6 +437,18 @@ function createLocalEngine({
|
|
|
437
437
|
settings,
|
|
438
438
|
tools,
|
|
439
439
|
promptBuilder,
|
|
440
|
+
// Phase 20 — the level's recall breadth, injected by the host that owns the
|
|
441
|
+
// avatar ledger (desktop/main.js builds one from lib/avatar/turn.js). It is a
|
|
442
|
+
// SEAM rather than a call into lib/avatar/* because this engine is shared with
|
|
443
|
+
// the CLI (cli/src/deps.js loads this file): the engine must not grow a
|
|
444
|
+
// dependency on the desktop's avatar directory, and a host with no avatar
|
|
445
|
+
// simply passes nothing and gets the exact pre-Phase-20 turn.
|
|
446
|
+
//
|
|
447
|
+
// `recallProvider({ payload, env, system })` may be sync or async and returns
|
|
448
|
+
// either the final system prompt or `{ system }`. It runs once per turn,
|
|
449
|
+
// before the first dispatch, and its result is the system prompt every round
|
|
450
|
+
// of that turn travels with.
|
|
451
|
+
recallProvider,
|
|
440
452
|
env,
|
|
441
453
|
getConfirmMode,
|
|
442
454
|
// Injectable for tests; defaults to the real transport so every existing
|
|
@@ -1532,6 +1544,12 @@ function createLocalEngine({
|
|
|
1532
1544
|
statedMaxTokens: opts.statedMaxTokens ? doubledBudget(opts.statedMaxTokens) : undefined,
|
|
1533
1545
|
});
|
|
1534
1546
|
addUsage(res);
|
|
1547
|
+
// Same guard as line ~1499, for the same reason: an abort landing
|
|
1548
|
+
// during THIS dispatch also resolves as empty text, and without
|
|
1549
|
+
// this check that "no answer" would fall straight into the
|
|
1550
|
+
// empty-turn nudge below instead of ending the turn — a cancelled
|
|
1551
|
+
// request quietly buying a THIRD billed call.
|
|
1552
|
+
if (signal.aborted) return withTurnUsage(res);
|
|
1535
1553
|
}
|
|
1536
1554
|
|
|
1537
1555
|
const calls = toolSchemas.length ? extractToolCalls(res) : [];
|
|
@@ -1562,6 +1580,12 @@ function createLocalEngine({
|
|
|
1562
1580
|
tools: [],
|
|
1563
1581
|
});
|
|
1564
1582
|
addUsage(res);
|
|
1583
|
+
// Same guard again: a cancel landing during the synthesis pass
|
|
1584
|
+
// resolves empty too, and without this check it fell through to
|
|
1585
|
+
// emptyTurnError — a real exception, so the renderer painted a red
|
|
1586
|
+
// "no answer" failure over what was actually just the user hitting
|
|
1587
|
+
// Cancel.
|
|
1588
|
+
if (signal.aborted) return withTurnUsage(res);
|
|
1565
1589
|
if (!assistantText(res)) {
|
|
1566
1590
|
throw emptyTurnError({
|
|
1567
1591
|
cls,
|