feat(model-router): close the cap residuals: session pin for inheriting spawns, prompt-size estimate, engine cache TTL
A built-in general-purpose agent with no explicit model, spawned by main
while main sits below the session model, is sent the session model
instead of inheriting the demoted one (narrow: built-in source, no fork,
no workflow, main's own child; other types carry their own frontmatter
model the hook cannot see). The cold-read cap now counts the prompt the
engine's usage may not include yet: code points / 3 of the submitted
text, its context entries and main's typed skill body, summed, added at
the turn's first main step unless usage already grew by half of it;
a document attachment makes the size unknown (blocks). Texts and the
step read the same adjusted size; the served record stays raw. The
cache TTL follows the engine's cache_ttl ('5m' | '1h') from main's
PostModelSwitch events, survives /clear, and /route show prints it.
This commit is contained in:
1 parent
ab963e97dc
commit
1be57a382c
2 files changed
+396
-25
No files matched your search
@@ -667,6 +667,8 @@ const switchTo = (
|
|||||||
from: string,
|
from: string,
|
||||||
to: string,
|
to: string,
|
||||||
source: 'command' | 'auto',
|
source: 'command' | 'auto',
|
||||||
|
cache_ttl: '5m' | '1h' = '5m',
|
||||||
|
agent_id?: string,
|
||||||
) => $.classic.PostModelSwitch({
|
) => $.classic.PostModelSwitch({
|
||||||
from_model: from,
|
from_model: from,
|
||||||
to_model: to,
|
to_model: to,
|
||||||
@@ -674,7 +676,8 @@ const switchTo = (
|
|||||||
source,
|
source,
|
||||||
context_tokens: 0,
|
context_tokens: 0,
|
||||||
prompt_cache_warm: false,
|
prompt_cache_warm: false,
|
||||||
cache_ttl: '5m',
|
cache_ttl,
|
||||||
|
...(agent_id === undefined ? {} : { agent_id }),
|
||||||
estimated_cache_write_usd: 0,
|
estimated_cache_write_usd: 0,
|
||||||
pricing: 'catalog',
|
pricing: 'catalog',
|
||||||
})
|
})
|
||||||
@@ -5188,7 +5191,7 @@ test('config: a quoted "false" mainModelSwitch reads as off with a warning',
|
|||||||
test('text: show names the switch and its cap', async ($, on) => {
|
test('text: show names the switch and its cap', async ($, on) => {
|
||||||
await boot($, on)
|
await boot($, on)
|
||||||
expect(await route($, 'show')).toContain(
|
expect(await route($, 'show')).toContain(
|
||||||
'switch: on (downgrade, cap 120000 tokens cold)')
|
'switch: on (downgrade, cap 120000 tokens cold, ttl 60 min)')
|
||||||
})
|
})
|
||||||
|
|
||||||
test('status: the switch shows only when off', async ($, on) => {
|
test('status: the switch shows only when off', async ($, on) => {
|
||||||
@@ -5205,7 +5208,7 @@ test('toast: a downgrade held by the cap says so', async ($, on) => {
|
|||||||
await bootRun($, on, FABLE, w, 300000)
|
await bootRun($, on, FABLE, w, 300000)
|
||||||
await typeFeat($)
|
await typeFeat($)
|
||||||
expect(toasts).toContain('/feat now runs claude-fable-5-1 at medium, ' +
|
expect(toasts).toContain('/feat now runs claude-fable-5-1 at medium, ' +
|
||||||
`held: ${SKIP}: cold read 300000 tokens over 120000`)
|
`held: ${SKIP}: cold read 300007 tokens over 120000`)
|
||||||
})
|
})
|
||||||
|
|
||||||
// ---- downgrade cap: the classifier's model axis ---------------------------
|
// ---- downgrade cap: the classifier's model axis ---------------------------
|
||||||
@@ -5274,3 +5277,215 @@ test('classify: a guessed sonnet rate-limited falls to opus, not haiku', async (
|
|||||||
await runStep($, highStep())
|
await runStep($, highStep())
|
||||||
expect(lastSeen(c)?.model).toBe(OPUS)
|
expect(lastSeen(c)?.model).toBe(OPUS)
|
||||||
})
|
})
|
||||||
|
|
||||||
|
// ---- cap residuals: session pin, prompt estimate, engine TTL ---------------
|
||||||
|
|
||||||
|
const GP = 'general-purpose'
|
||||||
|
|
||||||
|
/** A demoted main (fable session, mechanical route on haiku) + spawn rig. */
|
||||||
|
async function demotedRig($: Engine, on: On, source = 'built-in') {
|
||||||
|
const { rig } = await bootRigVar($, on, FABLE, 1000)
|
||||||
|
await offerOf($, GP, source)
|
||||||
|
expect(await routedStep($, rig, 'mechanical')).toBe(HAIKU)
|
||||||
|
return rig
|
||||||
|
}
|
||||||
|
|
||||||
|
test('spawn: general-purpose during a demoted turn starts on the session model',
|
||||||
|
async ($, on) => {
|
||||||
|
const logs = recordLogs(on)
|
||||||
|
const rig = await demotedRig($, on)
|
||||||
|
await route($, 'verbose on')
|
||||||
|
await $.agent.spawn(spawnOf(GP, { parentModel: HAIKU }))
|
||||||
|
expect(rig.specs).toEqual([FABLE])
|
||||||
|
expect(logsWith(logs, `agent ${GP}`)).toBe(1)
|
||||||
|
expect(logsWith(logs, '(session, engine)')).toBe(1)
|
||||||
|
})
|
||||||
|
|
||||||
|
test('spawn: no route in force leaves general-purpose untouched', async (
|
||||||
|
$, on) => {
|
||||||
|
const { rig } = await bootRigVar($, on, FABLE, 1000)
|
||||||
|
await offerOf($, GP, 'built-in')
|
||||||
|
await $.agent.spawn(spawnOf(GP, { parentModel: FABLE }))
|
||||||
|
expect(rig.specs).toEqual([undefined])
|
||||||
|
})
|
||||||
|
|
||||||
|
test('spawn: a rowed agent keeps its row during a demoted turn', async (
|
||||||
|
$, on) => {
|
||||||
|
const rig = await demotedRig($, on)
|
||||||
|
await $.agent.spawn(spawnOf('feater', { parentModel: HAIKU }))
|
||||||
|
expect(rig.specs).toEqual([SONNET])
|
||||||
|
})
|
||||||
|
|
||||||
|
test('spawn: a fork during a demoted turn is untouched', async ($, on) => {
|
||||||
|
const rig = await demotedRig($, on)
|
||||||
|
await $.agent.spawn(spawnOf(GP, { parentModel: HAIKU, fork: true }))
|
||||||
|
expect(rig.specs).toEqual([undefined])
|
||||||
|
})
|
||||||
|
|
||||||
|
test('spawn: a nested spawn during a demoted turn is untouched', async (
|
||||||
|
$, on) => {
|
||||||
|
const rig = await demotedRig($, on)
|
||||||
|
await $.agent.spawn(spawnOf(GP, { parentModel: HAIKU, parentAgentId: 'a1' }))
|
||||||
|
expect(rig.specs).toEqual([undefined])
|
||||||
|
})
|
||||||
|
|
||||||
|
test('spawn: a type outside the allowlist is untouched', async ($, on) => {
|
||||||
|
const rig = await demotedRig($, on)
|
||||||
|
await offerOf($, 'code-explorer', 'built-in')
|
||||||
|
await $.agent.spawn(spawnOf('code-explorer', { parentModel: HAIKU }))
|
||||||
|
expect(rig.specs).toEqual([undefined])
|
||||||
|
})
|
||||||
|
|
||||||
|
test('spawn: an upgraded main never pins general-purpose', async ($, on) => {
|
||||||
|
const { rig } = await bootRigVar($, on, HAIKU, 1000)
|
||||||
|
await offerOf($, GP, 'built-in')
|
||||||
|
await planRoute($)
|
||||||
|
expect((await stepOn($, rig, HAIKU)).model).toBe(FABLE)
|
||||||
|
await $.agent.spawn(spawnOf(GP, { parentModel: FABLE }))
|
||||||
|
expect(rig.specs).toEqual([undefined])
|
||||||
|
})
|
||||||
|
|
||||||
|
test('spawn: a shadowed general-purpose is never pinned', async ($, on) => {
|
||||||
|
const rig = await demotedRig($, on, 'projectSettings')
|
||||||
|
await $.agent.spawn(spawnOf(GP, { parentModel: HAIKU }))
|
||||||
|
expect(rig.specs).toEqual([undefined])
|
||||||
|
})
|
||||||
|
|
||||||
|
const longText = (points: number) => 'x'.repeat(points)
|
||||||
|
|
||||||
|
/** Cap 110000, classifier off, a context of `tokens`; estimates need both. */
|
||||||
|
function estimateRig($: Engine, on: On, tokens: number) {
|
||||||
|
const home = { classify: false, downgradeMaxTokens: 110000 }
|
||||||
|
return bootRigVar($, on, FABLE, tokens, newWorld({ home }))
|
||||||
|
}
|
||||||
|
|
||||||
|
test('cap: a pasted prompt counts before the first step', async ($, on) => {
|
||||||
|
const logs = recordLogs(on)
|
||||||
|
const { rig } = await estimateRig($, on, 100000)
|
||||||
|
await typed($, longText(60000))
|
||||||
|
expect(await routedStep($, rig, 'implement')).toBe(FABLE)
|
||||||
|
expect(logsWith(logs, `${SKIP}: cold read 120000 tokens over 110000`))
|
||||||
|
.toBe(1)
|
||||||
|
})
|
||||||
|
|
||||||
|
test('cap: the estimate is skipped once usage counted the prompt', async (
|
||||||
|
$, on) => {
|
||||||
|
const first = await estimateRig($, on, 100000)
|
||||||
|
await typed($, longText(60000))
|
||||||
|
first.usage.set(110000)
|
||||||
|
expect(await routedStep($, first.rig, 'implement')).toBe(SONNET)
|
||||||
|
})
|
||||||
|
|
||||||
|
test('cap: usage growth under half the estimate still adds it', async (
|
||||||
|
$, on) => {
|
||||||
|
const logs = recordLogs(on)
|
||||||
|
const { rig, usage } = await estimateRig($, on, 100000)
|
||||||
|
await typed($, longText(60000))
|
||||||
|
usage.set(109999)
|
||||||
|
expect(await routedStep($, rig, 'implement')).toBe(FABLE)
|
||||||
|
expect(logsWith(logs, 'cold read 129999 tokens over 110000')).toBe(1)
|
||||||
|
})
|
||||||
|
|
||||||
|
test('cap: the first main step consumes the estimate', async ($, on) => {
|
||||||
|
const { rig } = await estimateRig($, on, 100000)
|
||||||
|
await typed($, longText(60000))
|
||||||
|
expect(await routedStep($, rig, 'implement')).toBe(FABLE)
|
||||||
|
expect((await stepOn($, rig, FABLE)).model).toBe(SONNET)
|
||||||
|
})
|
||||||
|
|
||||||
|
test('cap: a mid-turn prompt is counted at the next turn first step', async (
|
||||||
|
$, on) => {
|
||||||
|
const { rig } = await estimateRig($, on, 100000)
|
||||||
|
await queued($, 'composer', longText(60000))
|
||||||
|
await endTurn($)
|
||||||
|
expect(await routedStep($, rig, 'implement')).toBe(FABLE)
|
||||||
|
})
|
||||||
|
|
||||||
|
test('cap: an estimate no step consumed is carried over the turn end', async (
|
||||||
|
$, on) => {
|
||||||
|
const { rig } = await estimateRig($, on, 100000)
|
||||||
|
await typed($, longText(60000))
|
||||||
|
await endTurn($)
|
||||||
|
expect(await routedStep($, rig, 'implement')).toBe(FABLE)
|
||||||
|
})
|
||||||
|
|
||||||
|
test('cap: a document attachment blocks the first step switch', async (
|
||||||
|
$, on) => {
|
||||||
|
const logs = recordLogs(on)
|
||||||
|
const { rig } = await estimateRig($, on, 1000)
|
||||||
|
await $.prompt.submit({
|
||||||
|
text: 'see the file',
|
||||||
|
wait: false,
|
||||||
|
origin: { kind: 'composer' },
|
||||||
|
attachments: [{ type: 'document' }],
|
||||||
|
})
|
||||||
|
expect(await routedStep($, rig, 'implement')).toBe(FABLE)
|
||||||
|
expect(logsWith(logs, `${SKIP}: context size unknown`)).toBe(1)
|
||||||
|
expect((await stepOn($, rig, FABLE)).model).toBe(SONNET)
|
||||||
|
})
|
||||||
|
|
||||||
|
test('cap: a typed skill body counts', async ($, on) => {
|
||||||
|
const { rig } = await estimateRig($, on, 100000)
|
||||||
|
await typed($, '/analyze')
|
||||||
|
await $.skill.prompt({ skill: 'analyze', text: longText(60000) })
|
||||||
|
expect(await routedStep($, rig, 'implement')).toBe(FABLE)
|
||||||
|
})
|
||||||
|
|
||||||
|
test('cap: a preload during a live spawn adds nothing', async ($, on) => {
|
||||||
|
const { rig } = await estimateRig($, on, 100000)
|
||||||
|
await typed($, 'hello')
|
||||||
|
await $.agent.spawn(spawnOf('feater'))
|
||||||
|
await $.skill.prompt({ skill: 'analyze', text: longText(60000) })
|
||||||
|
expect(await routedStep($, rig, 'implement')).toBe(SONNET)
|
||||||
|
})
|
||||||
|
|
||||||
|
/** A sonnet record of 100000 tokens, then `age` min later a 210000 context. */
|
||||||
|
async function ttlCase($: Engine, on: On, ttl: '5m' | '1h', age: number) {
|
||||||
|
const { rig, usage } = await bootRigVar($, on, FABLE, 100000)
|
||||||
|
await switchTo($, FABLE, FABLE, 'command', ttl)
|
||||||
|
await routedStep($, rig, 'implement')
|
||||||
|
await endTurn($)
|
||||||
|
await stepOn($, rig, FABLE)
|
||||||
|
await endTurn($)
|
||||||
|
usage.set(210000)
|
||||||
|
await rig.clock.advance(age * 60000)
|
||||||
|
return routedStep($, rig, 'implement')
|
||||||
|
}
|
||||||
|
|
||||||
|
test('ttl: the engine 5m makes a 6 min old record cold', async ($, on) => {
|
||||||
|
expect(await ttlCase($, on, '5m', 6)).toBe(FABLE)
|
||||||
|
})
|
||||||
|
|
||||||
|
test('ttl: the engine 1h keeps a 59 min old record warm', async ($, on) => {
|
||||||
|
expect(await ttlCase($, on, '1h', 59)).toBe(SONNET)
|
||||||
|
})
|
||||||
|
|
||||||
|
test('ttl: a sub-agent switch event moves neither TTL nor session', async (
|
||||||
|
$, on) => {
|
||||||
|
const rig = await bootRig($, on)
|
||||||
|
await route($, 'switch off')
|
||||||
|
await stepOn($, rig, FABLE)
|
||||||
|
await switchTo($, FABLE, OPUS, 'command', '5m', 'a1')
|
||||||
|
const text = await route($, 'show')
|
||||||
|
await route($, 'mechanical')
|
||||||
|
const line = mainLine(await route($, 'show'))
|
||||||
|
expect(line).toContain(FABLE)
|
||||||
|
expect(line).not.toContain(OPUS)
|
||||||
|
expect(text).not.toContain('ttl 5 min')
|
||||||
|
})
|
||||||
|
|
||||||
|
test('ttl: /clear keeps the engine TTL', async ($, on) => {
|
||||||
|
on('session.end', ($, e) => ({ sessionId: e.sessionId }))
|
||||||
|
await bootRig($, on)
|
||||||
|
await switchTo($, FABLE, FABLE, 'command', '5m')
|
||||||
|
await clearSession($)
|
||||||
|
expect(await route($, 'show')).toContain('ttl 5 min')
|
||||||
|
})
|
||||||
|
|
||||||
|
test('ttl: show says 60 min by default and 5 min after a 5m event', async (
|
||||||
|
$, on) => {
|
||||||
|
await boot($, on)
|
||||||
|
expect(await route($, 'show')).toContain('ttl 60 min)')
|
||||||
|
await switchTo($, FABLE, FABLE, 'command', '5m')
|
||||||
|
expect(await route($, 'show')).toContain('ttl 5 min)')
|
||||||
|
})
|
||||||
@@ -56,6 +56,12 @@ type Loop = {
|
|||||||
last?: Plan // the plan this loop last ran, once it was told
|
last?: Plan // the plan this loop last ran, once it was told
|
||||||
cause?: string // who set the effort in force, for the log lines
|
cause?: string // who set the effort in force, for the log lines
|
||||||
}
|
}
|
||||||
|
/** A submitted prompt's size where `session.usage` may not count it yet. */
|
||||||
|
type Estimate = {
|
||||||
|
tokens: number
|
||||||
|
usage: number | undefined // context size read when the first entry came
|
||||||
|
unknown: boolean // carries a non-text item: the size cannot be guessed
|
||||||
|
}
|
||||||
type State = {
|
type State = {
|
||||||
cfg: Config
|
cfg: Config
|
||||||
rules: Rule[]
|
rules: Rule[]
|
||||||
@@ -101,6 +107,9 @@ type State = {
|
|||||||
turnEngine: NonNullable<Effort> | null // same, for this turn
|
turnEngine: NonNullable<Effort> | null // same, for this turn
|
||||||
served: Map<string, Served> // canonical id -> main's last step on it
|
served: Map<string, Served> // canonical id -> main's last step on it
|
||||||
capSaid: boolean // the downgrade skip was logged for this cap crossing
|
capSaid: boolean // the downgrade skip was logged for this cap crossing
|
||||||
|
cacheTtlMs: number // the prompt-cache TTL a cold read is measured against
|
||||||
|
promptEstimate: Estimate | null // the submitted prompt, not in usage yet
|
||||||
|
pendingEstimate: Estimate | null // same, typed mid-turn: next turn's
|
||||||
classifyFails: number // classifier failures in a row
|
classifyFails: number // classifier failures in a row
|
||||||
classifyUntil: number // classifier paused until (ms); 0 = running
|
classifyUntil: number // classifier paused until (ms); 0 = running
|
||||||
asking: string | null // key of the one first-use dialog in flight
|
asking: string | null // key of the one first-use dialog in flight
|
||||||
@@ -111,6 +120,8 @@ type Log = (text: string) => void
|
|||||||
type StepIn = Readonly<TurnStepInput>
|
type StepIn = Readonly<TurnStepInput>
|
||||||
type PromptIn = {
|
type PromptIn = {
|
||||||
text: string
|
text: string
|
||||||
|
attachments?: readonly { type: string }[]
|
||||||
|
context?: readonly string[]
|
||||||
turnId?: string
|
turnId?: string
|
||||||
origin: { kind: string }
|
origin: { kind: string }
|
||||||
}
|
}
|
||||||
@@ -155,6 +166,11 @@ const TYPED_ORIGINS: ReadonlySet<string> = new Set([
|
|||||||
const PROJECT_SOURCES: ReadonlySet<string> = new Set([
|
const PROJECT_SOURCES: ReadonlySet<string> = new Set([
|
||||||
'projectSettings', 'localSettings',
|
'projectSettings', 'localSettings',
|
||||||
])
|
])
|
||||||
|
// Built-in agent types with no frontmatter model: an undefined `model` makes
|
||||||
|
// them inherit the spawning loop's model.
|
||||||
|
// shortcut: the frontmatter model of other types is invisible to the hook;
|
||||||
|
// widen the set when the engine exposes it on agent.offer.
|
||||||
|
const INHERITING_AGENTS: ReadonlySet<string> = new Set(['general-purpose'])
|
||||||
// PostModelSwitch sources a person chose (auto and resume are not).
|
// PostModelSwitch sources a person chose (auto and resume are not).
|
||||||
const USER_SWITCH: ReadonlySet<string> = new Set([
|
const USER_SWITCH: ReadonlySet<string> = new Set([
|
||||||
'command', 'picker', 'sdk',
|
'command', 'picker', 'sdk',
|
||||||
@@ -219,9 +235,16 @@ const CLASSIFY_MAX = 2000 // code points sent whole; longer: head + tail
|
|||||||
const CLASSIFY_HEAD = 500
|
const CLASSIFY_HEAD = 500
|
||||||
const CLASSIFY_TAIL = 1500
|
const CLASSIFY_TAIL = 1500
|
||||||
const CLASSIFY_FAILS = 3 // failures in a row that pause the classifier
|
const CLASSIFY_FAILS = 3 // failures in a row that pause the classifier
|
||||||
// The account's prompt-cache TTL (1 h here; an engine-reported `cache_ttl`
|
// The account's prompt-cache TTL: a model that served main within it is
|
||||||
// would supersede it, TODO): a model that served main within it is warm.
|
// warm. 1 h default; replaced by the engine's `cache_ttl` at the first
|
||||||
|
// model-switch event.
|
||||||
const CACHE_TTL_MS = 60 * 60000
|
const CACHE_TTL_MS = 60 * 60000
|
||||||
|
// The engine's two `cache_ttl` values (a closed enum), in ms.
|
||||||
|
const CACHE_TTLS: Readonly<Record<string, number>> = {
|
||||||
|
'5m': 5 * 60000,
|
||||||
|
'1h': 60 * 60000,
|
||||||
|
}
|
||||||
|
const ESTIMATE_DIVISOR = 3 // code points per token, conservative
|
||||||
const TIMEOUT: unique symbol = Symbol('classify timeout')
|
const TIMEOUT: unique symbol = Symbol('classify timeout')
|
||||||
|
|
||||||
const DEFAULT_CONFIG: Config = {
|
const DEFAULT_CONFIG: Config = {
|
||||||
@@ -959,6 +982,9 @@ function newState(cfg: Config, source: string): State {
|
|||||||
turnEngine: null,
|
turnEngine: null,
|
||||||
served: new Map(),
|
served: new Map(),
|
||||||
capSaid: false,
|
capSaid: false,
|
||||||
|
cacheTtlMs: CACHE_TTL_MS,
|
||||||
|
promptEstimate: null,
|
||||||
|
pendingEstimate: null,
|
||||||
classifyFails: 0,
|
classifyFails: 0,
|
||||||
classifyUntil: 0,
|
classifyUntil: 0,
|
||||||
asking: null,
|
asking: null,
|
||||||
@@ -988,6 +1014,7 @@ function resetSession(st: State): void {
|
|||||||
loaded: st.loaded,
|
loaded: st.loaded,
|
||||||
toggles: st.toggles,
|
toggles: st.toggles,
|
||||||
engineBase: st.engineBase,
|
engineBase: st.engineBase,
|
||||||
|
cacheTtlMs: st.cacheTtlMs,
|
||||||
classifyFails: st.classifyFails,
|
classifyFails: st.classifyFails,
|
||||||
classifyUntil: st.classifyUntil,
|
classifyUntil: st.classifyUntil,
|
||||||
asking: st.asking,
|
asking: st.asking,
|
||||||
@@ -1185,7 +1212,7 @@ function coldRead(
|
|||||||
): number {
|
): number {
|
||||||
const key = canonical(st.cfg, id)
|
const key = canonical(st.cfg, id)
|
||||||
const hit = st.served.get(key)
|
const hit = st.served.get(key)
|
||||||
if (hit === undefined || now - hit.at > CACHE_TTL_MS) return tokens
|
if (hit === undefined || now - hit.at > st.cacheTtlMs) return tokens
|
||||||
if (tokens < hit.tokens) {
|
if (tokens < hit.tokens) {
|
||||||
st.served.delete(key)
|
st.served.delete(key)
|
||||||
return tokens
|
return tokens
|
||||||
@@ -1366,6 +1393,60 @@ async function readTokens($: Api): Promise<number | undefined> {
|
|||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
|
/**
|
||||||
|
* The context size a cap reads: `raw` plus the pending prompt's estimate,
|
||||||
|
* unless usage already grew by half of it (the prompt was counted). Pure:
|
||||||
|
* the step and every text read the same figure. Unknown = blocked.
|
||||||
|
*/
|
||||||
|
function withEstimate(st: State, raw: number | undefined) {
|
||||||
|
const est = st.promptEstimate
|
||||||
|
if (raw === undefined || est === null) return raw
|
||||||
|
if (est.unknown) return undefined
|
||||||
|
const counted = est.usage !== undefined &&
|
||||||
|
raw - est.usage >= est.tokens / 2
|
||||||
|
return counted ? raw : raw + est.tokens
|
||||||
|
}
|
||||||
|
|
||||||
|
/** Sums an entry into a pending estimate; the first baseline is kept. */
|
||||||
|
function sumEstimate(
|
||||||
|
cur: Estimate | null,
|
||||||
|
tokens: number,
|
||||||
|
usage: number | undefined,
|
||||||
|
unknown: boolean,
|
||||||
|
): Estimate {
|
||||||
|
return {
|
||||||
|
tokens: (cur?.tokens ?? 0) + tokens,
|
||||||
|
usage: cur === null ? usage : cur.usage,
|
||||||
|
unknown: (cur?.unknown ?? false) || unknown,
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
/**
|
||||||
|
* Adds a submitted text (and a flag for a non-text item) to the pending
|
||||||
|
* estimate. A mid-turn entry waits for the next turn, where its baseline is
|
||||||
|
* reset; the first entry of a turn reads the usage as its baseline.
|
||||||
|
*/
|
||||||
|
async function addEstimate(
|
||||||
|
$: Api,
|
||||||
|
st: State,
|
||||||
|
size: { points: number; unknown: boolean },
|
||||||
|
midTurn: boolean,
|
||||||
|
): Promise<void> {
|
||||||
|
const tokens = Math.ceil(size.points / ESTIMATE_DIVISOR)
|
||||||
|
if (midTurn) {
|
||||||
|
st.pendingEstimate = sumEstimate(
|
||||||
|
st.pendingEstimate, tokens, undefined, size.unknown)
|
||||||
|
return
|
||||||
|
}
|
||||||
|
const usage = st.promptEstimate === null ? await readTokens($) : undefined
|
||||||
|
st.promptEstimate = sumEstimate(
|
||||||
|
st.promptEstimate, tokens, usage, size.unknown)
|
||||||
|
}
|
||||||
|
|
||||||
|
/** Code points of a text, as the estimate counts them. */
|
||||||
|
const pointsOf = (texts: readonly string[]): number =>
|
||||||
|
texts.reduce((n, t) => n + Array.from(t).length, 0)
|
||||||
|
|
||||||
type Verdict = { call: Call; wanted: string | undefined }
|
type Verdict = { call: Call; wanted: string | undefined }
|
||||||
|
|
||||||
/**
|
/**
|
||||||
@@ -1392,7 +1473,8 @@ async function readingFor(
|
|||||||
const needs = wanted !== undefined || st.down.size > 0 ||
|
const needs = wanted !== undefined || st.down.size > 0 ||
|
||||||
st.turnModel !== undefined
|
st.turnModel !== undefined
|
||||||
if (!needs) return { tokens: undefined, now: 0 }
|
if (!needs) return { tokens: undefined, now: 0 }
|
||||||
return { tokens: await readTokens($), now: await $.clock.now() }
|
const raw = await readTokens($)
|
||||||
|
return { tokens: withEstimate(st, raw), now: await $.clock.now() }
|
||||||
}
|
}
|
||||||
|
|
||||||
/**
|
/**
|
||||||
@@ -1602,8 +1684,9 @@ async function snapshot($: Api, st: State) {
|
|||||||
}
|
}
|
||||||
|
|
||||||
/** The downgrade switch and, when on, the cold read it is capped at. */
|
/** The downgrade switch and, when on, the cold read it is capped at. */
|
||||||
const switchWord = (c: Config): string => c.mainModelSwitch
|
const switchWord = (st: State): string => st.cfg.mainModelSwitch
|
||||||
? `on (downgrade, cap ${c.downgradeMaxTokens} tokens cold)`
|
? `on (downgrade, cap ${st.cfg.downgradeMaxTokens} tokens cold, ttl ${
|
||||||
|
Math.round(st.cacheTtlMs / 60000)} min)`
|
||||||
: 'off (downgrade)'
|
: 'off (downgrade)'
|
||||||
|
|
||||||
async function show($: Api, st: State): Promise<string> {
|
async function show($: Api, st: State): Promise<string> {
|
||||||
@@ -1612,7 +1695,7 @@ async function show($: Api, st: State): Promise<string> {
|
|||||||
const s = await snapshot($, st)
|
const s = await snapshot($, st)
|
||||||
return [
|
return [
|
||||||
mainText(st, s),
|
mainText(st, s),
|
||||||
`router: ${routerWord(st)} · switch: ${switchWord(c)} · upgrade: ${
|
`router: ${routerWord(st)} · switch: ${switchWord(st)} · upgrade: ${
|
||||||
flag(c.mainUpgrade)} · verbose: ${
|
flag(c.mainUpgrade)} · verbose: ${
|
||||||
flag(c.verbose)} · classify: ${flag(c.classify)} · spinner: ${
|
flag(c.verbose)} · classify: ${flag(c.classify)} · spinner: ${
|
||||||
flag(c.spinner)}`,
|
flag(c.spinner)}`,
|
||||||
@@ -2837,23 +2920,28 @@ function typedPath(st: State, skill: string): TypedPath | undefined {
|
|||||||
return st.promptAllowed && idle ? 'typed-fallback' : undefined
|
return st.promptAllowed && idle ? 'typed-fallback' : undefined
|
||||||
}
|
}
|
||||||
|
|
||||||
/** Routes main by the typed skill's row; the row, for the first-use ask. */
|
type Armed = { row: Picked | undefined; typed: boolean }
|
||||||
function armTypedSkill($: Api, st: State, skill: string) {
|
|
||||||
|
/**
|
||||||
|
* Routes main by the typed skill's row; the row, for the first-use ask, and
|
||||||
|
* whether the user typed the skill (rowed or not: its body is theirs).
|
||||||
|
*/
|
||||||
|
function armTypedSkill($: Api, st: State, skill: string): Armed {
|
||||||
const row = skillRow(st, skill)
|
const row = skillRow(st, skill)
|
||||||
const path = typedPath(st, skill)
|
const typed = typedPath(st, skill) !== undefined
|
||||||
if (row === undefined || path === undefined) return undefined
|
if (row === undefined || !typed) return { row: undefined, typed }
|
||||||
routeMainBySkill(st, skill, row, true)
|
routeMainBySkill(st, skill, row, true)
|
||||||
refresh($, st)
|
refresh($, st)
|
||||||
return row
|
return { row, typed }
|
||||||
}
|
}
|
||||||
|
|
||||||
/** Bookkeeping that may fail without breaking the hook: fail-open. */
|
/** Bookkeeping that may fail without breaking the hook: fail-open. */
|
||||||
function armed($: Api, st: State, skill: string): Picked | undefined {
|
function armed($: Api, st: State, skill: string): Armed {
|
||||||
try {
|
try {
|
||||||
return armTypedSkill($, st, skill)
|
return armTypedSkill($, st, skill)
|
||||||
} catch {
|
} catch {
|
||||||
warnOnce(st, $, 'skill.prompt', 'bookkeeping')
|
warnOnce(st, $, 'skill.prompt', 'bookkeeping')
|
||||||
return undefined
|
return { row: undefined, typed: false }
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
@@ -2913,9 +3001,31 @@ function inTier($: Api, st: State, agent: string, tier: string) {
|
|||||||
return undefined
|
return undefined
|
||||||
}
|
}
|
||||||
|
|
||||||
|
/**
|
||||||
|
* A built-in inheriting type spawned by main while main sits below the
|
||||||
|
* session model (a demotion or a breaker fallback) would inherit that lower
|
||||||
|
* model: it starts on the session model instead. Forks and workflows are
|
||||||
|
* frozen, a nested spawn inherits an agent's model, never main's, and a
|
||||||
|
* shadowing project or user definition may carry its own frontmatter model.
|
||||||
|
*/
|
||||||
|
function sessionPin(st: State, e: SpawnIn): string | undefined {
|
||||||
|
const turn = st.turnModel
|
||||||
|
const inherits = INHERITING_AGENTS.has(e.subagentType) &&
|
||||||
|
st.offers.get(e.subagentType) === 'built-in'
|
||||||
|
const main = !e.fork && e.workflow === undefined &&
|
||||||
|
e.parentAgentId === undefined
|
||||||
|
if (turn === undefined || st.sessionModel === '' || !inherits || !main) {
|
||||||
|
return undefined
|
||||||
|
}
|
||||||
|
const session = canonical(st.cfg, st.sessionModel)
|
||||||
|
const demoted = ranksAbove(st, session, canonical(st.cfg, turn))
|
||||||
|
return demoted && !st.down.has(session) ? st.sessionModel : undefined
|
||||||
|
}
|
||||||
|
|
||||||
/** The model a spawn is rewritten to; an explicit `model` param wins. */
|
/** The model a spawn is rewritten to; an explicit `model` param wins. */
|
||||||
function spawnTarget($: Api, st: State, e: SpawnIn, route: Route | undefined) {
|
function spawnTarget($: Api, st: State, e: SpawnIn, route: Route | undefined) {
|
||||||
if (e.model !== undefined || route === undefined) return undefined
|
if (e.model !== undefined) return undefined
|
||||||
|
if (route === undefined) return sessionPin(st, e)
|
||||||
if (route.tier === undefined) return resolveRoute(st, route)
|
if (route.tier === undefined) return resolveRoute(st, route)
|
||||||
return inTier($, st, e.subagentType, route.tier)
|
return inTier($, st, e.subagentType, route.tier)
|
||||||
}
|
}
|
||||||
@@ -3145,6 +3255,7 @@ type Handoff = { to: Plan; modelCause: string; effortCause: string }
|
|||||||
function spawnModelCause(st: State, sp: Spawned): string {
|
function spawnModelCause(st: State, sp: Spawned): string {
|
||||||
const { e, route, wanted } = sp
|
const { e, route, wanted } = sp
|
||||||
if (e.model !== undefined) return 'explicit'
|
if (e.model !== undefined) return 'explicit'
|
||||||
|
if (route === undefined && wanted !== undefined) return 'session'
|
||||||
if (route === undefined || wanted === undefined) return 'frontmatter'
|
if (route === undefined || wanted === undefined) return 'frontmatter'
|
||||||
const head = headAlias(st.cfg, route)
|
const head = headAlias(st.cfg, route)
|
||||||
const lead = head === undefined ? undefined : st.cfg.models[head]
|
const lead = head === undefined ? undefined : st.cfg.models[head]
|
||||||
@@ -3161,7 +3272,9 @@ function handoffOf(
|
|||||||
given: Level | undefined,
|
given: Level | undefined,
|
||||||
): Handoff | undefined {
|
): Handoff | undefined {
|
||||||
const { e, route, started } = sp
|
const { e, route, started } = sp
|
||||||
if (route === undefined && e.model === undefined && given === undefined) {
|
const pinned = sp.wanted !== undefined
|
||||||
|
if (route === undefined && e.model === undefined && given === undefined &&
|
||||||
|
!pinned) {
|
||||||
return undefined
|
return undefined
|
||||||
}
|
}
|
||||||
const effort = isHaiku(st.cfg, started.model)
|
const effort = isHaiku(st.cfg, started.model)
|
||||||
@@ -3248,11 +3361,15 @@ async function mainPlan($: Api, st: State, e: StepIn): Promise<Step> {
|
|||||||
const { effort, by } = mainEffort(st, e.effort)
|
const { effort, by } = mainEffort(st, e.effort)
|
||||||
await prune($, st)
|
await prune($, st)
|
||||||
const cur = st.turnModel ?? e.model
|
const cur = st.turnModel ?? e.model
|
||||||
const read = { tokens: await readTokens($), now: await $.clock.now() }
|
const raw = await readTokens($)
|
||||||
const { call } = await decideFor($, st, cur, read)
|
const now = await $.clock.now()
|
||||||
|
const { call } = await decideFor(
|
||||||
|
$, st, cur, { tokens: withEstimate(st, raw), now })
|
||||||
logCall($, st, capLogged(st, call))
|
logCall($, st, capLogged(st, call))
|
||||||
if (call.moved) st.turnModel = call.model
|
if (call.moved) st.turnModel = call.model
|
||||||
noteServed(st, { model: call.model, effort }, read)
|
// the record stays raw: an estimate in it would trip the compaction path
|
||||||
|
noteServed(st, { model: call.model, effort }, { tokens: raw, now })
|
||||||
|
st.promptEstimate = null // consumed by the turn's first main step
|
||||||
return { plan: { model: call.model, effort }, why: { by, call, cur } }
|
return { plan: { model: call.model, effort }, why: { by, call, cur } }
|
||||||
}
|
}
|
||||||
|
|
||||||
@@ -3278,7 +3395,21 @@ function noteMain($: Api, st: State, plan: Plan): void {
|
|||||||
safely(st, $, 'turn.step status', () => refresh($, st))
|
safely(st, $, 'turn.step status', () => refresh($, st))
|
||||||
}
|
}
|
||||||
|
|
||||||
|
/**
|
||||||
|
* An estimate no step consumed, and one typed mid-turn, carry to the next
|
||||||
|
* turn with their baseline reset: added at its first step, never dropped.
|
||||||
|
*/
|
||||||
|
function carryEstimate(st: State): void {
|
||||||
|
const left = [st.promptEstimate, st.pendingEstimate]
|
||||||
|
.filter((x): x is Estimate => x !== null)
|
||||||
|
const sum = left.reduce<Estimate | null>(
|
||||||
|
(acc, x) => sumEstimate(acc, x.tokens, undefined, x.unknown), null)
|
||||||
|
st.promptEstimate = sum
|
||||||
|
st.pendingEstimate = null
|
||||||
|
}
|
||||||
|
|
||||||
function endMainTurn($: Api, st: State): void {
|
function endMainTurn($: Api, st: State): void {
|
||||||
|
carryEstimate(st)
|
||||||
st.turnFloor = st.pendingPrompt
|
st.turnFloor = st.pendingPrompt
|
||||||
st.pendingPrompt = null
|
st.pendingPrompt = null
|
||||||
st.turnMain = null
|
st.turnMain = null
|
||||||
@@ -3327,8 +3458,14 @@ type SwitchIn = {
|
|||||||
to_model: string
|
to_model: string
|
||||||
requested_model: string | null
|
requested_model: string | null
|
||||||
source: string
|
source: string
|
||||||
|
cache_ttl?: unknown
|
||||||
|
agent_id?: unknown // set when a sub-agent's model switched
|
||||||
}
|
}
|
||||||
|
|
||||||
|
/** The engine's `cache_ttl` in ms, or undefined for any other value. */
|
||||||
|
const ttlOf = (v: unknown): number | undefined =>
|
||||||
|
typeof v === 'string' && hasKey(CACHE_TTLS, v) ? CACHE_TTLS[v] : undefined
|
||||||
|
|
||||||
/**
|
/**
|
||||||
* The engine switched the model by itself. The model it left is marked, one
|
* The engine switched the model by itself. The model it left is marked, one
|
||||||
* strike, unless it landed where the router already was; the mark targets
|
* strike, unless it landed where the router already was; the mark targets
|
||||||
@@ -3347,9 +3484,15 @@ async function onAutoSwitch($: Api, st: State, e: SwitchIn): Promise<void> {
|
|||||||
markDown($, st, target, 'engine fallback', await $.clock.now())
|
markDown($, st, target, 'engine fallback', await $.clock.now())
|
||||||
}
|
}
|
||||||
|
|
||||||
/** Any model change: keeps `sessionModel`; a user choice clears its mark. */
|
/**
|
||||||
|
* Any model change: keeps `sessionModel` and the cache TTL (main's own
|
||||||
|
* switches only); a user choice clears its mark.
|
||||||
|
*/
|
||||||
async function onModelSwitch($: Api, st: State, e: SwitchIn): Promise<void> {
|
async function onModelSwitch($: Api, st: State, e: SwitchIn): Promise<void> {
|
||||||
st.sessionModel = e.to_model
|
if (e.agent_id === undefined) {
|
||||||
|
st.sessionModel = e.to_model
|
||||||
|
st.cacheTtlMs = ttlOf(e.cache_ttl) ?? st.cacheTtlMs
|
||||||
|
}
|
||||||
if (USER_SWITCH.has(e.source)) st.runMain = null
|
if (USER_SWITCH.has(e.source)) st.runMain = null
|
||||||
if (st.off || e.source === 'resume') return
|
if (st.off || e.source === 'resume') return
|
||||||
if (e.source === 'auto') await onAutoSwitch($, st, e)
|
if (e.source === 'auto') await onAutoSwitch($, st, e)
|
||||||
@@ -3462,7 +3605,11 @@ function registerSkills(on: On, st: State): void {
|
|||||||
on('skill.prompt', async ($, e, next) => {
|
on('skill.prompt', async ($, e, next) => {
|
||||||
await ensureConfig($, st)
|
await ensureConfig($, st)
|
||||||
if (st.off || st.skillCalls > 0) return next(e)
|
if (st.off || st.skillCalls > 0) return next(e)
|
||||||
const row = armed($, st, e.skill)
|
const { row, typed } = armed($, st, e.skill)
|
||||||
|
if (typed) {
|
||||||
|
const size = { points: pointsOf([e.text]), unknown: false }
|
||||||
|
await addEstimate($, st, size, false)
|
||||||
|
}
|
||||||
if (row) await confirmSkill($, st, e.skill, row)
|
if (row) await confirmSkill($, st, e.skill, row)
|
||||||
return next(e)
|
return next(e)
|
||||||
}).catch(($, e, next) => {
|
}).catch(($, e, next) => {
|
||||||
@@ -3755,12 +3902,21 @@ async function classifyPrompt(
|
|||||||
if (routeFromLabel(st, label)) refresh($, st)
|
if (routeFromLabel(st, label)) refresh($, st)
|
||||||
}
|
}
|
||||||
|
|
||||||
|
/** What a prompt adds to the context: its text, context blocks, items. */
|
||||||
|
function promptSize(e: PromptIn): { points: number; unknown: boolean } {
|
||||||
|
return {
|
||||||
|
points: pointsOf([e.text, ...(e.context ?? [])]),
|
||||||
|
unknown: (e.attachments ?? []).some(a => a.type === 'document'),
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
function registerPrompt(on: On, st: State): void {
|
function registerPrompt(on: On, st: State): void {
|
||||||
on('prompt.submit', async ($, e, next) => {
|
on('prompt.submit', async ($, e, next) => {
|
||||||
await ensureConfig($, st, true)
|
await ensureConfig($, st, true)
|
||||||
if (st.off) return next(e)
|
if (st.off) return next(e)
|
||||||
const midTurn = e.turnId !== undefined
|
const midTurn = e.turnId !== undefined
|
||||||
noteTyped(st, e.origin.kind, e.text, midTurn)
|
noteTyped(st, e.origin.kind, e.text, midTurn)
|
||||||
|
await addEstimate($, st, promptSize(e), midTurn)
|
||||||
if (e.origin.kind === 'composer') routeFromPrompt(st, e.text, midTurn)
|
if (e.origin.kind === 'composer') routeFromPrompt(st, e.text, midTurn)
|
||||||
refresh($, st)
|
refresh($, st)
|
||||||
await classifyPrompt($, st, e, next.signal)
|
await classifyPrompt($, st, e, next.signal)
|
||||||
|
|||||||
Reference in new issue
Block a user