diff --git a/changelog.d/3132-output-cap-clamp.fixed.md b/changelog.d/3132-output-cap-clamp.fixed.md new file mode 100644 index 000000000..860510b5c --- /dev/null +++ b/changelog.d/3132-output-cap-clamp.fixed.md @@ -0,0 +1 @@ +A model call no longer sends an output-token cap above the 12288 a provider endpoint accepts. A model row that declares more than that ceiling — or a caller asking for more — is clamped to it, so the request is answered instead of being refused whole with a 400 that lost the step's answer. A cap at or under the ceiling is sent unchanged, and a model declaring none is never given one. diff --git a/core/src/host/models.js b/core/src/host/models.js index 8a8fdac62..74cfe0db1 100644 --- a/core/src/host/models.js +++ b/core/src/host/models.js @@ -8,6 +8,11 @@ const ROUTE_WAIT_MS = 30_000 const TEST_TIMEOUT_MS = 30_000 const TEST_MAX_TOKENS = 16 const TEST_PROMPT = 'Reply with the single word OK.' +// What a provider's endpoint takes as an output cap: above this DashScope refuses the +// whole request with a 400 (`max_completion_tokens … expected a value <= 12288`), so the +// answer is lost even though the rest of the call was fine (#3132). A model row can +// declare more than the endpoint accepts, so the cap is clamped on the way out. +const OUTBOUND_MAX_TOKENS = 12_288 // The outcome of a connection test as a code only (#409). A provider's own text can // echo part of a key, so none of it leaves the host: the client words each code. @@ -38,6 +43,25 @@ export function install(rt) { const reasoningCache = new Map() ctx.on?.('llm/adapters-updated', () => { catalogCache = { at: 0, value: [] }; reasoningCache.clear() }) + // The clamp sits on `llm/stream`, the seam every model call crosses — a Bot's turn and + // its checkpoint, a workflow step, the host's own request. `next` cannot take new + // options (the constraint secrets.js and images.js document), so a call that needs the + // cap set starts over with a copy and passes this listener the second time, carrying + // 12288. The caller's object is never written to: DSH compares a prepared call's config + // and throws INVALID_PREPARED_CALL when one changes underneath it. + ctx.on?.('llm/stream', (options, next) => { + const asked = options?.maxTokens + if (Number.isFinite(asked)) return asked > OUTBOUND_MAX_TOKENS ? ctx.llm.stream({ ...options, maxTokens: OUTBOUND_MAX_TOKENS }) : next() + return (async function* clampDeclaredCap() { + let declared + try { declared = (await ctx.llm.resolveModelInfo?.(options.provider, options.model))?.defaultMaxTokens } catch { declared = undefined } + // A cap at or under the ceiling, and a model that declares none, pass through: + // a clamp lowers a cap and never raises or invents one. + if (!Number.isFinite(declared) || declared <= OUTBOUND_MAX_TOKENS) { yield* next(); return } + yield* ctx.llm.stream({ ...options, maxTokens: OUTBOUND_MAX_TOKENS }) + })() + }) + // The thinking levels a model serves: `{efforts: [{id, name, description?}], // defaultEffort?}`, or undefined when it has none or the provider cannot say. async function reasoningOf(ref) { diff --git a/core/test/install.test.js b/core/test/install.test.js index 53acddead..dfccd61f6 100644 --- a/core/test/install.test.js +++ b/core/test/install.test.js @@ -42,6 +42,7 @@ test('apply registers listeners, tools, and effects in a fixed order', (t) => { 'effect ds-bot: diagnostics log', 'effect ds-bot: state lock', 'on llm/adapters-updated', + 'on llm/stream', 'on session/event', 'on session/event', 'on agent/status', diff --git a/core/test/llm-output-cap.test.js b/core/test/llm-output-cap.test.js new file mode 100644 index 000000000..18c97ede3 --- /dev/null +++ b/core/test/llm-output-cap.test.js @@ -0,0 +1,95 @@ +import assert from 'node:assert/strict' +import { test } from 'node:test' +import { install } from '../src/host/models.js' +import { fakeContext } from './fake-ctx.js' + +// #3132: DashScope refuses `max_completion_tokens` above 12288 with a 400, and a model +// row may declare more than the endpoint takes. The request is clamped at the seam every +// model call crosses; nothing in this file asserts on the model or frontend packages. + +const CEILING = 12_288 +const MODEL = 'opencsitool/DeepSeek-V4.1-Flash-1017' + +const user = text => ({ role: 'user', source: { kind: 'user' }, content: [{ type: 'text', text }] }) +const request = (overrides = {}) => ({ + provider: 'opencsitool', + model: 'DeepSeek-V4.1-Flash-1017', + messages: [user('Review this pull request')], + sessionId: 's-review', + ...overrides, +}) + +// DSH runs `llm/stream` as a waterfall: a listener calls `next()` for the rest of the +// chain, and a listener that restarts a request calls `ctx.llm.stream(copy)`, which runs +// the chain again from the top. `next` takes no arguments, so this stand-in advances on +// `next()` only — a copy reaches the provider through `stream`, exactly as in the host. +function bench({ info = {}, resolveError } = {}) { + const ctx = fakeContext() + const requests = [] + ctx.llm = { + listProviders: () => [], + resolveModelInfo: async (provider, model) => { + if (resolveError !== undefined) throw resolveError + return info[`${provider}/${model}`] + }, + stream: options => run(options), + } + const run = (options, index = 0) => { + const chain = ctx.listeners.get('llm/stream') ?? [] + if (index < chain.length) return chain[index](options, () => run(options, index + 1)) + requests.push(options) + return (async function* answered() { yield { type: 'finish', reason: { kind: 'stop' } } })() + } + install({ ctx, state: {} }) + const ask = async asked => { + for await (const chunk of ctx.llm.stream(asked)) void chunk + return { sent: requests.at(-1), count: requests.length } + } + return { ctx, requests, ask } +} + +test('a model row declaring more than the endpoint takes is asked for 12288, and the call is not rewritten in place', async () => { + const { ask } = bench({ info: { [MODEL]: { defaultMaxTokens: 16_384 } } }) + const asked = request() + const { sent, count } = await ask(asked) + assert.equal(sent.maxTokens, CEILING) + assert.equal(count, 1, 'one provider request, not two') + assert.equal(asked.maxTokens, undefined, 'the caller\'s object is untouched') +}) + +test('a model row under the ceiling keeps what it declares, and nothing is invented', async () => { + const { ask } = bench({ info: { [MODEL]: { defaultMaxTokens: 4_096 } } }) + const { sent } = await ask(request()) + assert.equal('maxTokens' in sent, false, 'the declared default is filled by DSH, not written here') +}) + +test('a caller that asks for more than the ceiling is clamped to it, and its own object keeps its value', async () => { + const { ask } = bench({ info: { [MODEL]: { defaultMaxTokens: 16_384 } } }) + const asked = request({ maxTokens: 20_000 }) + const { sent, count } = await ask(asked) + assert.equal(sent.maxTokens, CEILING) + assert.equal(count, 1) + assert.equal(asked.maxTokens, 20_000) +}) + +test('a caller at or under the ceiling is passed through unchanged', async () => { + const { ask } = bench({ info: { [MODEL]: { defaultMaxTokens: 16_384 } } }) + for (const maxTokens of [CEILING, 4_096, 1]) { + const { sent } = await ask(request({ maxTokens })) + assert.equal(sent.maxTokens, maxTokens) + } +}) + +test('nothing declared and nothing asked: no cap appears on the request', async () => { + const { ask } = bench({ info: { [MODEL]: {} } }) + const { sent, count } = await ask(request()) + assert.equal('maxTokens' in sent, false) + assert.equal(count, 1) +}) + +test('a model lookup that fails leaves the request alone instead of breaking the call', async () => { + const { ask } = bench({ resolveError: new Error('no adapter') }) + const { sent, count } = await ask(request()) + assert.equal('maxTokens' in sent, false) + assert.equal(count, 1) +}) diff --git a/docs/subsystems/core-host.md b/docs/subsystems/core-host.md index 5aa589897..dca92f4ba 100644 --- a/docs/subsystems/core-host.md +++ b/docs/subsystems/core-host.md @@ -42,6 +42,7 @@ The server side of the patched `ds-bot` core (`core/`): the part of [ds-bot-core | `core/src/host/text.js` | The pure text helpers the host modules share: message text, names, model refs, cutting and estimating text, and the clock format of chat anchors. Since #1895 `codePoints(text)` (a step per code point, no array built since #1898) is how the host counts a limit — one character costs one whatever its UTF-16 width — and the bounds this change puts through it are: a memory entry of 300 and a `MEMORY.md` of 4,000 or a topic of 16,000 (the refusal and the size note print the code-point number), a scheduled task's name of 120, a message of 20,000, and the Project memory entry of 600 in both writers, so an emoji is never charged twice. Both cuts land on a code point boundary: `clip` and `excerpt` move back rather than keep half of a surrogate pair, so a clipped emoji reaches a UTF-8 encoder as the character and not as U+FFFD ([#1907](https://github.com/ThinkFlowLab/FlyBot/issues/1907)). Since #1912 that rule is one exported helper rather than a private detail: `cut(text, max)` answers `{ text, cut }` with the longest prefix that stops before a split pair, and the archive writers that used to `slice(0, max)` go through it — the `oneLine` of `decisions.js` and `meetings.js` (into `decisions.jsonl` and `decisions/-.md`), `claimLine` in `groups.js`, the report cap in `diagnostics.js`, `cleanLabel` in `flybot-feed.js` and `clean` in `browser.js` — so a decision or a meeting record holds the whole character instead of U+FFFD. Since #1916 the boundary rule itself lives in `core/src/shared/text-boundary.js` (`codePointStart`, `codePointEnd`, `cut`), and `host/text.js` builds its `cut` on it so a shared module can bound text too; the eight remaining `slice(0, max)` writers follow it — the claim cue's fallback reference (60), the task lead's name in a member prompt (80), a failed group Session's error text (300), `read_own_chat`'s `query` (80) and `around` (40), the open-page address and title (2000/300), the diagnostics browser-log line and its code (500/40), and `cleanPeerName` in `shared/remote-label.js` (60) — each keeping the longest prefix that stops before a split pair, `length <= max` and a byte prefix of the input. Since #1918 the memory review's own two bounds go through the same rule: its 800-character context `clip` cuts through `cut`, and its 3,000-character two-ended `ends` spends the budget in characters at both ends — `charIndex` steps whole code points to find the head's end and the tail's start, so the head is the rendered line's first 2,000 characters and the tail its last 1,000, and the text is flattened with `toWellFormed()` first as `cut` did — so the prompt and the log hold whole characters within the same character budget: a line of 3,500 emoji keeps 3,000 characters where unit arithmetic kept 1,503 ([#2553](https://github.com/ThinkFlowLab/FlyBot/issues/2553)); the exported `clip`'s reason is optional there too, so a two-argument call reads `… (cut)` instead of `… (cut; undefined)` while a call with a reason is unchanged. Since #1923 the four bounds whose text crosses an owner boundary go through the same `cut` — `cleanPublishedName` in `hooks.js` (60, into `state.flybot.remote`), the group member line's `displayName` 60 and `ownerName` 40 in `tools.js`, `label` in `reach.js` (60, the `othersInRoom` labels), and the local `clip` in `shared/room-share.js` (its per-field caps included) — so a published, listed or relayed name keeps whole characters and never reaches a UTF-8 encoder as U+FFFD ([#1923](https://github.com/ThinkFlowLab/FlyBot/issues/1923)). The same rule now covers the host's own line bounds: the busy ack's 600-unit task and 1,500-unit message text in `busy-ack.js`, the group line's notice (100), a Bot's job line (120), the group answer echo (300) and the rework line's note (200) in `tools.js`, and `cleanName`'s 60-unit peer name in `delivery.js`, which lands in `fenceUntrusted`'s `from="…"` attribute — so none of them can end in half a surrogate pair. The five bounds `turns.js` keeps go through `cut` as well: a failure's 500-unit message and the 200-unit message of its line (`turnFailure`, `failureLine`), the 79-unit step a running call shows (the `…` is appended only when `cut` dropped something, never by a length guess), an approval card's 2,000-unit description and 4,000-unit command (the card stores the command "Allow once" runs rather than a preview of it, and one that still had to be cut ends with `…` inside that bound, so the card says what it abridged and never drops the fact, [#1992](https://github.com/ThinkFlowLab/FlyBot/issues/1992)), and the 160-unit reason a failed turn leaves for the inbox. Since #1963 `cut` also flattens its input before it bounds — `String(text).toWellFormed()` first, `codePointEnd` on the flat string — so a title, a note or a name that arrives already carrying an unpaired surrogate is stored as U+FFFD instead of the lone surrogate, and U+FFFD is one UTF-16 unit for the one it replaces, so a well-formed input keeps its exact prefix and its `cut` flag ([#1963](https://github.com/ThinkFlowLab/FlyBot/issues/1963)). Since #1979 the six remaining bounds on text another home or another person supplied go through it too — a connector item name (80, `connectors.js`), a role label another home set on a face-room seat (80, `face-lead.js`), a cross-home group name (200, `project-team.js`), a capability description from a peer server (2,000, `capabilities.js`), the remote home's own error text (157, `remote-chat.js`) and an abbreviated remote reply (80, same file) — so each keeps the longest prefix that stops before a split pair, is well formed and is a code-point prefix of the input, while `remote-chat.js`'s `…` and trailing-period strip still apply to the cut value ([#1979](https://github.com/ThinkFlowLab/FlyBot/issues/1979)). The two bounds on a friend's Bot's own name in that file — the name its chat is created with (40) and the name a refresh writes (40) — go through `cut` as well, so a refresh cannot store a lone high surrogate and a create no longer hands `roster.createBot` half a pair to flatten into U+FFFD; the budget stays 40 and a shorter name is stored word for word. Since #1993 the last five bounds in `core/src/host/` go through it too — the quoted first line of a checkpoint's kept tail (80, `steward.js`), an approval record's `reason` (500, `approvals.js`), a mention receipt's reason line (240, `hooks.js`), the title given to a DSH Agent's Session (80, `AGENT_SESSION_TITLE_MAX` in `agents.js`) and a schedule run's failure `reason` (200, `schedules.js`) — so each keeps the longest prefix that stops before a split pair, is well formed and is a code-point prefix of the input, and none of the five maxima changes ([#1993](https://github.com/ThinkFlowLab/FlyBot/issues/1993)). Since #2002 the connector's own error text goes through it too: the `reason` in `core/src/host/connectors.js` — the text an MCP server's failure leaves on the owner's connector card — keeps the longest prefix that stops before a split pair, so it is well formed and a code-point prefix of what the server sent; since #2016 that prefix is 2,000 units rather than 300, and one that had to be cut ends with `…` inside the bound, so the card says what it abridged ([#2002](https://github.com/ThinkFlowLab/FlyBot/issues/2002), [#2016](https://github.com/ThinkFlowLab/FlyBot/issues/2016)). Since #2005 the five bounds the Project panels read go through it as well — the assistant card's 77-unit title in `project-assistant.js`, the board's 200-unit one-line value and a decision title's 200 in `project-board.js`, and `capped`'s name, role, persona, model and color caps plus the dossier's 4,000-unit memory head in `project-team.js`, whose head is bounded where `memoryDigest` reads it (`memory.js`), because a cut at the dossier alone could only turn half a pair into U+FFFD and never drop the character whole — so a dossier row or a board card keeps whole characters ([#2005](https://github.com/ThinkFlowLab/FlyBot/issues/2005)). Since [#2321](https://github.com/ThinkFlowLab/FlyBot/issues/2321) a bound also judges in code points, not only counts in them: `fits(text, max)` in `core/src/shared/text-boundary.js` is the one predicate — a character outside the BMP spends one of the budget and not two — and `clip` and `excerpt` judge with it while their cuts keep going through `codePointEnd`/`cut`, so a 300-character text is no longer cut by a 400-character limit on its way to `read_own_chat` and the search rows. Since [#2339](https://github.com/ThinkFlowLab/FlyBot/issues/2339) the two bounds in `core/src/host/` that still judged with `text.length` use it as well — the memory review's 3,000-character two-ended `ends` in `memory.js` and `room-context.js`'s `oneLine` (40/80/24/60) — so a user line or a board line of characters outside the BMP stays whole inside its own limit instead of being cut at half of it; both maxima are unchanged and a board line's cut stays on `cut`, so an ASCII text judges exactly as `length <= max` did. The memory review's own third `clip` is the exported one now, so its cut reads `… (cut)`, and a bound added later that judges with `text.length` fails `core/test/text-bounds.test.js` at its own `file:line`. Both render paths of the memory review flatten before they bound as well: the `fits()` answer of `ends` and the labelled Bot line on its way to `clip` pass through `toWellFormed()` first, so an unpaired surrogate that arrives in a message reaches the review prompt as U+FFFD, whole and in place, instead of as half a pair ([#2556](https://github.com/ThinkFlowLab/FlyBot/issues/2556)). Since [#2691](https://github.com/ThinkFlowLab/FlyBot/issues/2691) `normalize` is the one identity key for a name: it NFC-normalizes, drops the Unicode format characters (`\p{Cf}` — zero-width space / non-joiner / joiner, the byte-order mark, soft hyphen, the bidi controls) that `trim()` keeps, trims, lowercases and NFC-normalizes again, so either canonical spelling of one visible name is one key; `findBot` and `findRooms` compare with it, the duplicate guards in `roster.js` ask those same finds, and `createBot`, `updateBot` and `cleanRoomName` store the visible form (the format characters stripped before the 40 / `ROOM_NAME_MAX` cut) while `shared/mentions.js` `nameProblem` refuses a name that is only invisible characters. A name already in `state.json` is never rewritten: the key is computed from what is stored, so an old NFD name still resolves. Since [#2692](https://github.com/ThinkFlowLab/FlyBot/issues/2692) that one key is `nameKey` in `core/src/shared/identity.js` (`normalize` in `host/text.js` is it, so the host, a shared module and the browser bundle cannot drift apart), and three identity sites that used their own `trim().toLowerCase()` compare with it: `shared/task-board.js` `sameRole` — a member's role label and a queued card's `queueRole`, so a card reaches the carrier whose label spells the same role the other way instead of staying in the pool — and `memory.js`'s topic, where the key is also the file stem: `cleanTopic` answers the key, a topic file already on disk keeps the spelling it was written with (the key is computed from the name, so a file `가.md` is still topic `[[가]]` and remember appends to that same file), the file already in key form wins when both spellings exist, and the `## Index` link check and the link a delete removes both compare by key rather than by the text that was written. A decomposed Latin topic, refused while U+0301 was not a `\p{L}`, is accepted as its NFC form. `plugins/tree` keeps its own copy of the same function (`src/identity.js`) because it ships as its own package, and `plugins/tree/test/identity.test.js` holds the two to the same answers. Since [#2935](https://github.com/ThinkFlowLab/FlyBot/issues/2935) a name also has a drawn form: `drawnKey` in the same `shared/identity.js` is `nameKey` with the characters `shared/mentions.js` exports as `NOT_DRAWN` removed — an explicit list of the code points with no glyph of their own (the combining grapheme joiner, the Hangul fillers, the braille blank, the Mongolian free variation selectors and the variation selectors), never the whole `\p{Mn}` class, because a combining mark is drawn by the letter under it and `กิ`/`กี` are two names — so `Dev` written with a trailing U+034F has the drawn form `dev` while its key stays `dev\u034f`. `roster.js` compares it in `createBot` and `updateBot` beside the key, so a second row that reads `Dev` and that no `@Dev` can reach is refused, from either side, with the sentences already in use (`A Bot named … already exists` / `Name … is not available`). Since [#2986](https://github.com/ThinkFlowLab/FlyBot/issues/2986) the stored name also folds a run of whitespace to one plain space — `foldSpaces` in `host/text.js`, used by `createBot` and `updateBot` before the 40-character cut and by the store when a saved team is loaded, the way `cleanRoomName` folds a group's name — because every name row draws with `white-space: nowrap`, so `Dev Ops` and `Dev Ops` are one row on screen while `nameKey` keeps them apart and `@Dev Ops` reaches only the single-space Bot. `Dev Ops` is therefore stored as `Dev Ops`, and the existing sentences refuse it while that name exists (`A Bot named … already exists` / `Name … is not available`); a line break is not folded, because `nameProblem` refuses it by name. A team saved before that fold kept its run, so `load()` folds `bots[].name` once, before the room names derived from the roster are rebuilt: when the folded name is one another Bot already holds, it is left as it was saved and one warning line names that Bot, because renaming a Bot the user can see is worse than the twin; the drawn-form guard `drawnClash` folds both sides too, so no new twin can be created beside a saved one. Nothing is stripped from the stored name and a drawn form nobody holds is still a name, so `#2879`'s guard and a blank beside drawn characters are unchanged. Tests: `core/test/name-identity.test.js`, `core/test/memory.test.js`, `core/test/bot-tasks.test.js`. | | `core/src/host/prompt.js` | The Bot system prompt; a Main Bot's section is headed `You are the Main Bot` and uses that one term, the one the UI shows (#276). The team rules state the idle default (#699): a Bot does nothing until the user addresses it, it is @-mentioned in a group, its parent gives it a task, or unfinished tasks it was already given are still on its board. It does not invent work. A coordinator acts on a user task rather than on its own initiative — an addressed Bot answers as before. | | `core/src/host/health.js` | `GET /api/health` (#2128, P-392): the one answer this instance gives without a session, and exactly four fields — `ok` (always `true`), `version` (`core/package.json`), `commit` (`git rev-parse HEAD` of the clone `core/` sits in, `null` for a package or path install with no `.git`) and `startedAt` (this process's start, from `process.uptime()`, stable until a restart). It is an **exact** WebServer route, which the web server matches before the `/api` prefix whose admission is the sign-in wall, so no cookie is needed and no other `/api` route changes; outside the Web composition there is no `webServer` and the route does not exist. A method other than `GET` answers 405. Tests: `core/test/health.test.js`. | +| `core/src/host/models.js` | The model surface: the provider catalog with its short caches (#409), a Bot's desired model and thinking level, the Session rebinding a group needs, and the connection test (`rt.testModel`). Every model call crosses the `llm/stream` listener installed here, which clamps an outbound output cap to 12288 (#3132): a model row may declare more than a provider endpoint accepts, and a request above that ceiling is refused whole by DashScope with a 400 (`max_completion_tokens … expected a value <= 12288`), so the step loses an answer it would otherwise give. A cap at or under the ceiling, a model that declares none, and a caller with no cap are passed through untouched — the clamp lowers a cap and never raises or invents one — and the request is re-entered with a copy rather than written to, because DSH compares a prepared call's config (`INVALID_PREPARED_CALL`). Tests: `core/test/llm-output-cap.test.js`, `core/test/install.test.js`. | | `core/src/host/room-membership.js` | Typed room participants (human / local bot / granted remote), authorship, membership-window history (#50); rooms with friends' Bots or other owners take granted local Bots only and evict a Bot that stops qualifying (#90). Face-to-face rooms (#83) reuse this model (`origin: 'face-to-face'`, room-only remotes without a person-grant, `upsertFaceRoom` / `syncFaceParticipants` / `applyFaceRoomLine`); a face-room local Bot is `viaCode` (brought by number), `evictUnshared` skips face rooms (membership is the server's), and a Bot that returns to the member list drops the empty local `left` line (#236). Refusing an ungranted local Bot next to a friend's Bot throws `ungrantedRefusal`: a message that names the Bot, the friend and the two ways out, plus `error.details` `{ reason: 'ungranted-local', bots, remotes }`; `channel.js` `fail` forwards a plain-object `details` (empty otherwise) and `client/api.js` puts it on the thrown error, where `client/new-chat.js` draws a card (**Copy invite link for {Bot}** via `__flybot__.social.createInvite`, **Leave {Bot} out**) read through `ungrantedLocalFix` in `shared/remote-label.js`; the same file's `copiedToClipboard` reports whether the invite link really reached the clipboard (none, or a refused write, is `false`), and the card then shows the link to copy by hand instead of saying it was copied. | | `core/src/host/face-mentions.js` | Who "@" calls in a face-to-face room (#287): every alias of a participant (`faceAliases`, kept in `shared/remote-label.js` so the bubble chips the same names, #305; bare name, the picker label `'s Chief` / ` · 's Bot`, ` · `, handle), local and friend's Bots told apart. `faceMentions().handles` is the handle of every Bot a line names (the explicit mention list a sender attaches, #293) and `mentionedByHandle` resolves a received list to my Bots (`known` counts handles the room has). | | `core/src/shared/look-image.js` | The image a Bot's look may carry (#2646): `safeLookImage` keeps a plain `https?://` URL or a `data:image/(png|jpeg|webp|gif);base64,` string and answers `null` for a value carrying a quote, a paren, a backslash, whitespace or a control character; `cssImageUrl` writes the kept value as `url("…")` for the page-wide mention stylesheet. It is read at the entries — hooks.js `cleanRemoteLook` and remote-chat.js `applyPublishedLook` — and again at that sink (`client/mention-style.js`), so a friend's Bot cannot close the declaration or add a rule; a dropped image draws the colour/mask look. `plugins/social/src/look.js` carries the same test (the same formats, gif included since #2649) and `plugins/social/test/look.test.js` keeps the two in step. Tests: `core/test/mention-style.test.js`. | diff --git a/patches/ds-bot/0001-core-fixes.patch b/patches/ds-bot/0001-core-fixes.patch index 52e5c247c..cce7fee57 100644 --- a/patches/ds-bot/0001-core-fixes.patch +++ b/patches/ds-bot/0001-core-fixes.patch @@ -7369,13 +7369,18 @@ diff --git a/src/host/memory.js b/src/host/memory.js diff --git a/src/host/models.js b/src/host/models.js --- a/src/host/models.js +++ b/src/host/models.js -@@ -5,6 +5,28 @@ import { normalize, splitRef } from './text.js' +@@ -5,6 +5,33 @@ import { normalize, splitRef } from './text.js' const CATALOG_TTL_MS = 10_000 const ROUTE_WAIT_MS = 30_000 +const TEST_TIMEOUT_MS = 30_000 +const TEST_MAX_TOKENS = 16 +const TEST_PROMPT = 'Reply with the single word OK.' ++// What a provider's endpoint takes as an output cap: above this DashScope refuses the ++// whole request with a 400 (`max_completion_tokens … expected a value <= 12288`), so the ++// answer is lost even though the rest of the call was fine (#3132). A model row can ++// declare more than the endpoint accepts, so the cap is clamped on the way out. ++const OUTBOUND_MAX_TOKENS = 12_288 + +// The outcome of a connection test as a code only (#409). A provider's own text can +// echo part of a key, so none of it leaves the host: the client words each code. @@ -7398,6 +7403,32 @@ diff --git a/src/host/models.js b/src/host/models.js export function install(rt) { const { ctx, state } = rt +@@ -16,6 +16,25 @@ export function install(rt) { + const reasoningCache = new Map() + ctx.on?.('llm/adapters-updated', () => { catalogCache = { at: 0, value: [] }; reasoningCache.clear() }) + ++ // The clamp sits on `llm/stream`, the seam every model call crosses — a Bot's turn and ++ // its checkpoint, a workflow step, the host's own request. `next` cannot take new ++ // options (the constraint secrets.js and images.js document), so a call that needs the ++ // cap set starts over with a copy and passes this listener the second time, carrying ++ // 12288. The caller's object is never written to: DSH compares a prepared call's config ++ // and throws INVALID_PREPARED_CALL when one changes underneath it. ++ ctx.on?.('llm/stream', (options, next) => { ++ const asked = options?.maxTokens ++ if (Number.isFinite(asked)) return asked > OUTBOUND_MAX_TOKENS ? ctx.llm.stream({ ...options, maxTokens: OUTBOUND_MAX_TOKENS }) : next() ++ return (async function* clampDeclaredCap() { ++ let declared ++ try { declared = (await ctx.llm.resolveModelInfo?.(options.provider, options.model))?.defaultMaxTokens } catch { declared = undefined } ++ // A cap at or under the ceiling, and a model that declares none, pass through: ++ // a clamp lowers a cap and never raises or invents one. ++ if (!Number.isFinite(declared) || declared <= OUTBOUND_MAX_TOKENS) { yield* next(); return } ++ yield* ctx.llm.stream({ ...options, maxTokens: OUTBOUND_MAX_TOKENS }) ++ })() ++ }) ++ + // The thinking levels a model serves: `{efforts: [{id, name, description?}], + // defaultEffort?}`, or undefined when it has none or the provider cannot say. + async function reasoningOf(ref) { @@ -68,7 +68,9 @@ export function install(rt) { // The profile default model of DSH, when it is served; else the first served model. function pickDefault(entries) { @@ -13279,7 +13310,7 @@ diff --git a/test/images.test.js b/test/images.test.js diff --git a/test/install.test.js b/test/install.test.js --- a/test/install.test.js +++ b/test/install.test.js -@@ -29,14 +29,23 @@ function start(state, { config = {}, services } = {}) { +@@ -29,13 +29,23 @@ function start(state, { config = {}, services } = {}) { } // Listeners run, and tools are listed, in the order they are registered. @@ -13296,13 +13327,13 @@ diff --git a/test/install.test.js b/test/install.test.js + 'effect ds-bot: diagnostics log', 'effect ds-bot: state lock', 'on llm/adapters-updated', - 'on session/event', ++ 'on llm/stream', + 'on session/event', -+ 'on agent/status', + 'on session/event', ++ 'on agent/status', + 'on session/event', 'effect ds-bot: agent tool restrict retry', 'on agent/created', - 'effect ds-bot: identity section', @@ -47,6 +47,7 @@ test('apply registers listeners, tools, and effects in a fixed order', (t) => { 'tool stop_bot', 'tool create_bot', diff --git a/patches/ds-bot/files/test/llm-output-cap.test.js b/patches/ds-bot/files/test/llm-output-cap.test.js new file mode 100644 index 000000000..18c97ede3 --- /dev/null +++ b/patches/ds-bot/files/test/llm-output-cap.test.js @@ -0,0 +1,95 @@ +import assert from 'node:assert/strict' +import { test } from 'node:test' +import { install } from '../src/host/models.js' +import { fakeContext } from './fake-ctx.js' + +// #3132: DashScope refuses `max_completion_tokens` above 12288 with a 400, and a model +// row may declare more than the endpoint takes. The request is clamped at the seam every +// model call crosses; nothing in this file asserts on the model or frontend packages. + +const CEILING = 12_288 +const MODEL = 'opencsitool/DeepSeek-V4.1-Flash-1017' + +const user = text => ({ role: 'user', source: { kind: 'user' }, content: [{ type: 'text', text }] }) +const request = (overrides = {}) => ({ + provider: 'opencsitool', + model: 'DeepSeek-V4.1-Flash-1017', + messages: [user('Review this pull request')], + sessionId: 's-review', + ...overrides, +}) + +// DSH runs `llm/stream` as a waterfall: a listener calls `next()` for the rest of the +// chain, and a listener that restarts a request calls `ctx.llm.stream(copy)`, which runs +// the chain again from the top. `next` takes no arguments, so this stand-in advances on +// `next()` only — a copy reaches the provider through `stream`, exactly as in the host. +function bench({ info = {}, resolveError } = {}) { + const ctx = fakeContext() + const requests = [] + ctx.llm = { + listProviders: () => [], + resolveModelInfo: async (provider, model) => { + if (resolveError !== undefined) throw resolveError + return info[`${provider}/${model}`] + }, + stream: options => run(options), + } + const run = (options, index = 0) => { + const chain = ctx.listeners.get('llm/stream') ?? [] + if (index < chain.length) return chain[index](options, () => run(options, index + 1)) + requests.push(options) + return (async function* answered() { yield { type: 'finish', reason: { kind: 'stop' } } })() + } + install({ ctx, state: {} }) + const ask = async asked => { + for await (const chunk of ctx.llm.stream(asked)) void chunk + return { sent: requests.at(-1), count: requests.length } + } + return { ctx, requests, ask } +} + +test('a model row declaring more than the endpoint takes is asked for 12288, and the call is not rewritten in place', async () => { + const { ask } = bench({ info: { [MODEL]: { defaultMaxTokens: 16_384 } } }) + const asked = request() + const { sent, count } = await ask(asked) + assert.equal(sent.maxTokens, CEILING) + assert.equal(count, 1, 'one provider request, not two') + assert.equal(asked.maxTokens, undefined, 'the caller\'s object is untouched') +}) + +test('a model row under the ceiling keeps what it declares, and nothing is invented', async () => { + const { ask } = bench({ info: { [MODEL]: { defaultMaxTokens: 4_096 } } }) + const { sent } = await ask(request()) + assert.equal('maxTokens' in sent, false, 'the declared default is filled by DSH, not written here') +}) + +test('a caller that asks for more than the ceiling is clamped to it, and its own object keeps its value', async () => { + const { ask } = bench({ info: { [MODEL]: { defaultMaxTokens: 16_384 } } }) + const asked = request({ maxTokens: 20_000 }) + const { sent, count } = await ask(asked) + assert.equal(sent.maxTokens, CEILING) + assert.equal(count, 1) + assert.equal(asked.maxTokens, 20_000) +}) + +test('a caller at or under the ceiling is passed through unchanged', async () => { + const { ask } = bench({ info: { [MODEL]: { defaultMaxTokens: 16_384 } } }) + for (const maxTokens of [CEILING, 4_096, 1]) { + const { sent } = await ask(request({ maxTokens })) + assert.equal(sent.maxTokens, maxTokens) + } +}) + +test('nothing declared and nothing asked: no cap appears on the request', async () => { + const { ask } = bench({ info: { [MODEL]: {} } }) + const { sent, count } = await ask(request()) + assert.equal('maxTokens' in sent, false) + assert.equal(count, 1) +}) + +test('a model lookup that fails leaves the request alone instead of breaking the call', async () => { + const { ask } = bench({ resolveError: new Error('no adapter') }) + const { sent, count } = await ask(request()) + assert.equal('maxTokens' in sent, false) + assert.equal(count, 1) +})