bare-agent 0.49.0 → 0.51.0

package/src/thinking-mode.js

@@ -0,0 +1,384 @@
+'use strict';
+
+// Normalized `thinkingMode: 'on' | 'off'` (hamr-approved design: try the known shapes, fail LOUDLY,
+// check every reply). Callers say "on"/"off"; this module owns the per-provider/per-model request
+// field, the retry ladder, the measured per-model facts, and the post-reply signal readers. The
+// provider files stay thin: they ask here for the patch and the reader.
+//
+// Every fact in MODELS below was MEASURED live (poc/thinking-mode-verify.mjs →
+// poc/results/thinking-mode-verify.json, 2026-10-11) EXCEPT Ollama, which is shape-known but not
+// live-verified. A model that is not in the table is 'unverified': the ladder tries the shapes at
+// runtime and fails loudly if every one is rejected.
+//
+// The raw `thinking` passthrough options (BA-7) are a DIFFERENT, untouched field: opaque and verbatim.
+// `thinkingMode` is the normalized switch; setting both at once is a ValidationError (no silent precedence).
+
+const { ValidationError } = require('./errors');
+
+/** @typedef {'anthropic'|'openai'|'deepseek'|'gemini'|'ollama'} LadderKind */
+
+/**
+ * @typedef {object} Shape
+ * @property {string} key - Short stable key (the MODELS table refers to shapes by this).
+ * @property {string} id - Human-readable request-field description (error messages, thinkingSupport).
+ * @property {(body: Record<string, any>) => (Record<string, any>|null)} apply - Returns a NEW body with
+ *   the field set (copy-on-write; never mutates `body`), or `null` when the shape cannot apply to this body.
+ */
+
+/** @type {(key: string, id: string, apply: Shape['apply']) => Shape} */
+const shape = (key, id, apply) => ({ key, id, apply });
+
+// Anthropic `enabled` needs budget_tokens >= 1024 and max_tokens > budget_tokens (measured: haiku-4-5).
+const ANTHROPIC_BUDGET = 2048;
+
+/** @type {Record<LadderKind, { off: Shape[], on: Shape[] }>} */
+const LADDERS = {
+  anthropic: {
+    // sonnet-5 STILL thinks when the field is omitted, so off must be sent explicitly.
+    off: [shape('disabled', "thinking:{type:'disabled'}", (b) => ({ ...b, thinking: { type: 'disabled' } }))],
+    on: [
+      shape('adaptive', "thinking:{type:'adaptive'}", (b) => ({ ...b, thinking: { type: 'adaptive' } })),
+      shape('enabled', "thinking:{type:'enabled',budget_tokens:N}", (b) => {
+        const budget = Math.min(ANTHROPIC_BUDGET, (b.max_tokens || 0) - 1);
+        if (!(budget >= 1024)) return null; // max_tokens too small to leave room for a legal budget
+        return { ...b, thinking: { type: 'enabled', budget_tokens: budget } };
+      }),
+    ],
+  },
+  openai: {
+    // gpt-5.1+ accept 'none' and 400 on 'minimal'; gpt-5 / gpt-5-mini are the reverse.
+    off: [
+      shape('none', "reasoning_effort:'none'", (b) => ({ ...b, reasoning_effort: 'none' })),
+      shape('minimal', "reasoning_effort:'minimal'", (b) => ({ ...b, reasoning_effort: 'minimal' })),
+    ],
+    on: [shape('medium', "reasoning_effort:'medium'", (b) => ({ ...b, reasoning_effort: 'medium' }))],
+  },
+  deepseek: {
+    off: [shape('disabled', "thinking:{type:'disabled'}", (b) => ({ ...b, thinking: { type: 'disabled' } }))],
+    // DeepSeek thinks by default: "on" is the ABSENCE of a field (the body goes out unchanged).
+    on: [shape('omit', '(field omitted — thinks by default)', (b) => ({ ...b }))],
+  },
+  gemini: {
+    off: [shape('budget0', 'generationConfig.thinkingConfig.thinkingBudget:0',
+      (b) => ({ ...b, generationConfig: { ...b.generationConfig, thinkingConfig: { thinkingBudget: 0 } } }))],
+    on: [shape('budget-1', 'generationConfig.thinkingConfig.thinkingBudget:-1',
+      (b) => ({ ...b, generationConfig: { ...b.generationConfig, thinkingConfig: { thinkingBudget: -1 } } }))],
+  },
+  ollama: {
+    // NOT live-verified (no Ollama thinking model was available to measure).
+    off: [shape('think-false', 'think:false', (b) => ({ ...b, think: false }))],
+    on: [shape('think-true', 'think:true', (b) => ({ ...b, think: true }))],
+  },
+};
+
+/**
+ * MEASURED per-model facts. Value `null` ⇒ the mode is known NOT switchable on that model (throws
+ * before any request is sent). A string ⇒ the first shape that worked (tried first).
+ * @type {Record<LadderKind, Record<string, { off: string|null, on: string|null }>>}
+ */
+const MODELS = {
+  anthropic: {
+    'claude-sonnet-5': { off: 'disabled', on: 'adaptive' },
+    'claude-haiku-4-5': { off: 'disabled', on: 'enabled' }, // 400s on adaptive
+    'claude-opus-5-5': { off: null, on: 'adaptive' },       // 400s on disabled → cannot be turned off
+  },
+  openai: {
+    'gpt-5': { off: 'minimal', on: 'medium' },
+    'gpt-5-mini': { off: 'minimal', on: 'medium' },
+    'gpt-5.1': { off: 'none', on: 'medium' },
+    'gpt-5.2': { off: 'none', on: 'medium' },
+    'gpt-5.4': { off: 'none', on: 'medium' },
+    'gpt-5.4-mini': { off: 'none', on: 'medium' },
+    'gpt-5.4-nano': { off: 'none', on: 'medium' },
+    'gpt-5.5': { off: 'none', on: 'medium' },
+  },
+  deepseek: {
+    'deepseek-flash': { off: 'disabled', on: 'omit' },
+    'deepseek-v4-pro': { off: 'disabled', on: 'omit' },
+  },
+  gemini: {
+    'gemini-2.5-flash': { off: 'budget0', on: 'budget-1' },
+    'gemini-3-flash-preview': { off: 'budget0', on: 'budget-1' },
+    'gemini-3.1-flash-lite': { off: 'budget0', on: 'budget-1' },
+    'gemini-3.1-flash-lite-preview': { off: 'budget0', on: 'budget-1' },
+    'gemini-3.5-flash': { off: 'budget0', on: 'budget-1' },
+    'gemini-3.1-pro-preview': { off: null, on: 'budget-1' },            // 400 "Budget 0 is invalid"
+    'gemini-3.1-pro-preview-customtools': { off: null, on: 'budget-1' }, // same
+  },
+  ollama: {},
+};
+
+// Which 400s are "this shape was rejected" (try the next one) vs an ordinary error (surface it). The
+// error text must NAME the field being set; per-provider so an unrelated 400 that merely happens to
+// say "thinking" elsewhere in a different vendor's vocabulary can't trigger a retry. No `.*`, no
+// unbounded quantifiers (ReDoS-safe on provider-supplied text).
+/** @type {Record<LadderKind, RegExp>} */
+const FIELD_NAMED = {
+  anthropic: /\bthinking\b/i,
+  openai: /\b(reasoning_effort|reasoning)\b/i,
+  deepseek: /\bthinking\b/i,
+  gemini: /\b(thinkingBudget|thinkingConfig|thinking_budget|thinking_config|thinking|budget)\b/i,
+  ollama: /\b(thinking|think)\b/i,
+};
+// A 400 that names `temperature` is the temperature fallback's business (e.g. Anthropic's "temperature
+// may only be set to 1 when thinking is enabled" names BOTH) — never a shape rejection.
+const TEMPERATURE_NAMED = /temperature/i;
+
+/**
+ * Longest-prefix lookup of a model's measured facts. A dated/`-latest` suffix of a listed id counts
+ * (`claude-haiku-4-5-20251001`); any other suffix does NOT (so `gpt-5-chat` never inherits `gpt-5`).
+ * @param {LadderKind} kind
+ * @param {string} model
+ * @returns {{ off: string|null, on: string|null }|null}
+ */
+function lookupModel(kind, model) {
+  const table = MODELS[kind];
+  if (typeof model !== 'string') return null;
+  let best = null;
+  for (const id of Object.keys(table)) {
+    if (model === id || (model.startsWith(id) && /^-(\d{8}|\d{4}-\d{2}-\d{2}|latest)$/.test(model.slice(id.length)))) {
+      if (best === null || id.length > best.length) best = id;
+    }
+  }
+  return best === null ? null : table[best];
+}
+
+/**
+ * @param {any} model
+ * @param {any} baseUrl
+ * @returns {boolean}
+ */
+function isDeepSeek(model, baseUrl) {
+  return (typeof model === 'string' && /^deepseek-/i.test(model)) ||
+    (typeof baseUrl === 'string' && /deepseek\.com/i.test(baseUrl));
+}
+
+/**
+ * @param {string} message
+ * @param {Record<string, any>} [ctx]
+ */
+const invalid = (message, ctx = {}) => new ValidationError(message, { context: { lib: 'bare-agent', ...ctx } });
+
+/**
+ * Validate a `thinkingMode` value. `undefined`/`null` ⇒ `null` (unset); anything but 'on'/'off' throws.
+ * @param {any} value
+ * @param {string} label - provider name for the message
+ * @returns {'on'|'off'|null}
+ */
+function normalizeThinkingMode(value, label) {
+  if (value === undefined || value === null) return null;
+  if (value === 'on' || value === 'off') return value;
+  throw invalid(`[${label}] thinkingMode must be 'on' or 'off' (got ${typeof value === 'string' ? JSON.stringify(value) : `<${typeof value}>`})`, { provider: label });
+}
+
+/**
+ * Resolve the EFFECTIVE mode for one call and reject the raw-`thinking` conflict (no silent precedence).
+ * Per-call overrides the instance default; `null` suppresses it.
+ * @param {object} p
+ * @param {string} p.label
+ * @param {any} p.instanceMode
+ * @param {Record<string, any>} p.options - per-call options
+ * @param {any} [p.instanceThinking] - the instance's raw `thinking` (providers that have one)
+ * @param {boolean} [p.hasRawThinking] - whether this provider has a raw `thinking` option at all
+ * @returns {'on'|'off'|null}
+ */
+function resolveThinkingMode({ label, instanceMode, options, instanceThinking = null, hasRawThinking = false }) {
+  const mode = normalizeThinkingMode(options.thinkingMode !== undefined ? options.thinkingMode : instanceMode, label);
+  if (mode && hasRawThinking) {
+    const raw = options.thinking !== undefined ? options.thinking : instanceThinking;
+    if (raw) {
+      throw invalid(`[${label}] both the raw 'thinking' option and thinkingMode '${mode}' are set — they target the same request field. Use one (thinkingMode is the normalized switch; raw 'thinking' is the verbatim passthrough).`, { provider: label });
+    }
+  }
+  return mode;
+}
+
+/**
+ * @param {LadderKind} kind
+ * @param {any} err
+ * @returns {boolean}
+ */
+function isShapeRejection(kind, err) {
+  const msg = err && typeof err.message === 'string' ? err.message : '';
+  return !!err && err.status === 400 && FIELD_NAMED[kind].test(msg) && !TEMPERATURE_NAMED.test(msg);
+}
+
+/**
+ * Send a request with the `thinkingMode` field applied, walking the shape ladder on a 400 that names
+ * the field. Order: the instance cache (a shape that already worked for this mode+model), else the
+ * table's known-good shape, else the ladder's first; then the remaining shapes in ladder order. A
+ * known-not-switchable model throws BEFORE any request. Every shape rejected ⇒ a loud ValidationError.
+ * A 400 that does not name the field (or any non-400) is an ordinary error and propagates untouched.
+ *
+ * @param {object} p
+ * @param {LadderKind} p.kind
+ * @param {'on'|'off'} p.mode
+ * @param {string} p.model
+ * @param {string} p.label - provider class name
+ * @param {Record<string, any>} p.body - the base body (NOT mutated; each attempt gets a patched copy)
+ * @param {(body: Record<string, any>) => Promise<any>} p.request
+ * @param {Map<string, string>} p.cache - per provider instance
+ * @returns {Promise<any>} the reply data of the accepted shape
+ */
+async function requestWithThinkingMode({ kind, mode, model, label, body, request, cache }) {
+  const ladder = LADDERS[kind][mode];
+  const facts = lookupModel(kind, model);
+  if (facts && facts[mode] === null) {
+    throw invalid(`[${label}] thinkingMode '${mode}' is not supported by ${kind} model ${model} (measured: the API rejects every ${mode} shape for it) — no request was sent.`,
+      { provider: kind, model, mode, tried: [], sent: false });
+  }
+  const cacheKey = `${mode}|${model}`;
+  const preferredKey = cache.get(cacheKey) || (facts && facts[mode]) || null;
+  const ordered = [...ladder].sort((a, b) => (a.key === preferredKey ? -1 : b.key === preferredKey ? 1 : 0));
+
+  /** @type {string[]} */
+  const tried = [];
+  let lastMessage = '';
+  for (const s of ordered) {
+    const patched = s.apply(body);
+    if (patched === null) { tried.push(`${s.id} (not applicable to this request)`); continue; }
+    try {
+      const data = await request(patched);
+      cache.set(cacheKey, s.key);
+      return data;
+    } catch (err) {
+      if (!isShapeRejection(kind, err)) throw err;
+      tried.push(s.id);
+      lastMessage = err.message;
+    }
+  }
+  throw invalid(`[${label}] thinkingMode '${mode}' is not supported by ${kind} model ${model} (tried: ${tried.join('; ')})` +
+    (lastMessage ? ` — last API error: ${lastMessage}` : ''),
+    { provider: kind, model, mode, tried, sent: tried.length > 0 });
+}
+
+/** @param {any} v @returns {boolean} */
+const positive = (v) => typeof v === 'number' && v > 0;
+
+/**
+ * Read the thinking signal off a raw reply: 'thinking' | 'none' | 'unknown'. "Absent field ⇒ none" is
+ * used ONLY where measured: Gemini thoughtsTokenCount (given usageMetadata is present), DeepSeek
+ * reasoning_content, Ollama message.thinking. Anthropic counts thinking/redacted_thinking blocks.
+ * OpenAI reasoning_tokens absent ⇒ 'unknown'.
+ * @param {LadderKind} kind
+ * @param {any} data - raw provider reply
+ * @returns {'thinking'|'none'|'unknown'}
+ */
+function readThinkingSignal(kind, data) {
+  if (!data || typeof data !== 'object') return 'unknown';
+  switch (kind) {
+    case 'anthropic': {
+      if (!Array.isArray(data.content)) return 'unknown';
+      return data.content.some((/** @type {any} */ b) => b && (b.type === 'thinking' || b.type === 'redacted_thinking')) ? 'thinking' : 'none';
+    }
+    case 'openai': {
+      const t = data.usage && data.usage.completion_tokens_details && data.usage.completion_tokens_details.reasoning_tokens;
+      if (typeof t !== 'number') return 'unknown';
+      return t > 0 ? 'thinking' : 'none';
+    }
+    case 'deepseek': {
+      const m = data.choices && data.choices[0] && data.choices[0].message;
+      if (!m || typeof m !== 'object') return 'unknown';
+      return typeof m.reasoning_content === 'string' && m.reasoning_content.length > 0 ? 'thinking' : 'none';
+    }
+    case 'gemini': {
+      const u = data.usageMetadata;
+      if (!u || typeof u !== 'object') return 'unknown';
+      const parts = (data.candidates && data.candidates[0] && data.candidates[0].content && data.candidates[0].content.parts) || [];
+      if (positive(u.thoughtsTokenCount) || (Array.isArray(parts) && parts.some((/** @type {any} */ p) => p && p.thought === true))) return 'thinking';
+      return 'none';
+    }
+    case 'ollama': {
+      const m = data.message;
+      if (!m || typeof m !== 'object') return 'unknown';
+      return typeof m.thinking === 'string' && m.thinking.length > 0 ? 'thinking' : 'none';
+    }
+    default:
+      return 'unknown';
+  }
+}
+
+/**
+ * The `thinkingApplied` verdict for a reply whose request was accepted. 'on' ⇒ `true` (an adaptive
+ * model may legitimately choose not to think, so 'on' is never flagged). 'off' ⇒ `true` when the reply
+ * shows zero thinking, `false` when thinking happened anyway, `null` when the signal is unreadable.
+ * On `false` it warns ONCE per `owner` (the provider instance) and never throws — the round is paid.
+ * @param {object} p
+ * @param {LadderKind} p.kind
+ * @param {'on'|'off'} p.mode
+ * @param {any} p.data
+ * @param {any} p.owner - provider instance (holds the warn-once flag)
+ * @param {string} p.label
+ * @param {string} p.model
+ * @returns {boolean|null}
+ */
+function assessThinkingApplied({ kind, mode, data, owner, label, model }) {
+  if (mode === 'on') return true;
+  const signal = readThinkingSignal(kind, data);
+  if (signal === 'none') return true;
+  if (signal === 'unknown') return null;
+  if (!owner._warnedThinkingNotApplied) {
+    owner._warnedThinkingNotApplied = true;
+    console.warn(`[${label}] thinkingMode 'off' was sent for ${model} but the reply shows thinking happened — the switch did not take effect (the result is kept; thinkingApplied:false). Further occurrences from this provider instance are silent.`);
+  }
+  return false;
+}
+
+/**
+ * What bare-agent knows about switching a model's thinking on/off, with no network. Measured models
+ * answer true/false; a model not in the measured table on a provider with a shape ladder answers
+ * 'unverified' (the ladder will try the shapes at runtime and fail loudly if all are rejected);
+ * clipipe and unknown providers answer false.
+ * @param {'anthropic'|'openai'|'deepseek'|'gemini'|'ollama'|'clipipe'|string} provider
+ * @param {string} [model]
+ * @returns {{ switchable: boolean|'unverified', off: string|null, on: string|null, note: string }}
+ * @when you need to know, before building a request, whether a provider/model can have thinking turned on or off and which request field does it — pure lookup, no network
+ * @fails returns switchable:false for clipipe and unknown providers (never throws); 'unverified' means not measured, not unsupported; Ollama's shapes are not live-verified.
+ * @category providers
+ * @example
+ *   thinkingSupport('anthropic', 'claude-opus-5-5'); // { switchable: false, off: null, on: "thinking:{type:'adaptive'}", note: '...' }
+ */
+function thinkingSupport(provider, model) {
+  if (provider === 'clipipe') {
+    return { switchable: false, off: null, on: null, note: 'CLIPipe has no thinking-mode mechanism; thinkingMode throws a ValidationError.' };
+  }
+  const kind = /** @type {LadderKind} */ (provider);
+  if (typeof provider !== 'string' || !Object.prototype.hasOwnProperty.call(LADDERS, provider)) {
+    return { switchable: false, off: null, on: null, note: `Unknown provider '${String(provider)}'.` };
+  }
+  const idOf = (/** @type {'on'|'off'} */ mode, /** @type {string|null} */ key) =>
+    key === null ? null : (LADDERS[kind][mode].find((s) => s.key === key) || LADDERS[kind][mode][0]).id;
+  const facts = lookupModel(kind, /** @type {string} */ (model));
+  if (!facts) {
+    return {
+      switchable: 'unverified',
+      off: LADDERS[kind].off[0].id,
+      on: LADDERS[kind].on[0].id,
+      note: kind === 'ollama'
+        ? 'Shapes known (think:true/false) but NOT live-verified; the reply signal is message.thinking.'
+        : `Model not in the measured table; the shape ladder is tried at runtime (off: ${LADDERS[kind].off.map((s) => s.id).join(' → ')}; on: ${LADDERS[kind].on.map((s) => s.id).join(' → ')}) and fails loudly if all are rejected.`,
+    };
+  }
+  const switchable = facts.off !== null && facts.on !== null;
+  return {
+    switchable,
+    off: idOf('off', facts.off),
+    on: idOf('on', facts.on),
+    note: switchable
+      ? 'Measured: both modes work.'
+      : `Measured: ${facts.off === null ? "'off'" : "'on'"} is rejected by the API for this model — requesting it throws before any request is sent.`,
+  };
+}
+
+module.exports = {
+  thinkingSupport,
+  // internal (provider files + tests)
+  normalizeThinkingMode,
+  resolveThinkingMode,
+  requestWithThinkingMode,
+  assessThinkingApplied,
+  readThinkingSignal,
+  isShapeRejection,
+  isDeepSeek,
+  lookupModel,
+};

package/types/index.d.ts

@@ -134,6 +134,15 @@ export interface GenerateResult {
    * temperature, not the one requested — callers reporting an effective temperature must honor this.
    */
   temperatureDropped?: boolean;
+  /**
+   * Present ONLY when `thinkingMode` ('on'|'off') was set on the provider/call. The post-reply check:
+   * `true` — a shape was accepted and, for 'off', the reply shows zero thinking signal (for 'on' always
+   * `true` once accepted: an adaptive model may legitimately choose not to think); `false` — 'off' was
+   * requested but the reply shows thinking happened (the provider warns once; the round is paid and kept,
+   * never thrown); `null` — cannot tell (the reply's thinking signal is unreadable, e.g. OpenAI without
+   * `reasoning_tokens`). Surfaced on `Loop.run()`'s return, sticky-false: any round reporting `false` makes the run `false`, else the latest round's value.
+   */
+  thinkingApplied?: boolean | null;
   /**
    * BA-27 — present ONLY when the model emitted a tool call whose `function.arguments` was
    * syntactically-broken JSON (an extra brace, a truncated object). The billed round already
@@ -143,12 +152,15 @@ export interface GenerateResult {
    * repaired. Absent on a clean round. OpenAI-compatible + Ollama string-arguments only; Anthropic
    * arrives pre-parsed and cannot hit this.
    *
-   * `rawArguments`/`rawTruncated` (BA-27 follow-up) are present only when the provider was
+   * `rawArguments` (BA-27 follow-up) is present only when the provider was
    * constructed with `exposeMalformedArgs: true` (default off, mirrors `exposeErrorBody`) AND the
-   * raw `function.arguments` was a string: the raw value verbatim, capped at 500 chars, with
-   * `rawTruncated: true` added only when it was longer.
+   * raw `function.arguments` was a string: the FULL raw value verbatim, uncapped (model-generated; may be
+   * large or sensitive).
    */
-  malformedToolCall?: { name: string | undefined; error: string; rawArguments?: string; rawTruncated?: true };
+  malformedToolCall?: { name: string | undefined; error: string; rawArguments?: string;
+    /** @deprecated never set since 0.50.0 (rawArguments is now uncapped) */
+    rawTruncated?: true;
+  };
   /**
    * BA-7 — provider-native content blocks the normalized `{text, toolCalls}` shape cannot express
    * (Anthropic `thinking` / `redacted_thinking`), captured opaquely so the Loop can put them on the