mixdog 0.9.1 → 0.9.3

package/src/runtime/agent/orchestrator/providers/anthropic-effort.mjs

@@ -0,0 +1,188 @@
+/**
+ * Anthropic "effort" (extended-thinking budget) request parameter handling.
+ */
+
+export const EFFORT_LEVELS = Object.freeze(['low', 'medium', 'high', 'max']);
+
+export const EFFORT_BETA_HEADER = 'effort-2025-11-24';
+
+export const LEGACY_EFFORT_BUDGET = Object.freeze({
+    low: 1024,
+    medium: 4096,
+    high: 16384,
+    max: 32768,
+});
+const _LOGGED_EFFORT_NORMALIZATION = new Set();
+const _LOGGED_UNKNOWN_EFFORT = new Set();
+
+function normalizeModelId(model) {
+    return String(model || '').toLowerCase().replace(/\./g, '-');
+}
+
+function parseClaudeVersion(model) {
+    const m = normalizeModelId(model);
+    const triple = m.match(/^claude-(haiku|sonnet|opus|fable)-(\d+)-(\d+)/);
+    if (triple) {
+        return { family: triple[1], major: Number(triple[2]), minor: Number(triple[3]) };
+    }
+    const pair = m.match(/^claude-(sonnet|fable)-(\d+)(?:$|[-@])/);
+    if (pair) {
+        return { family: pair[1], major: Number(pair[2]), minor: null };
+    }
+    return null;
+}
+
+function isLegacyAnthropicReasoningModel(model) {
+    const parsed = parseClaudeVersion(model);
+    if (!parsed) return false;
+    if (parsed.family === 'haiku') return true;
+    if (parsed.family === 'fable') return false;
+    if (parsed.family === 'sonnet' || parsed.family === 'opus') {
+        if (parsed.major < 4) return true;
+        if (parsed.major === 4 && parsed.minor !== null && parsed.minor < 6) return true;
+        return false;
+    }
+    return false;
+}
+
+// @[MODEL LAUNCH]: extend allowlist when new Claude models ship with effort support.
+export function modelSupportsEffort(model) {
+    if (isEnvTruthy(process.env.MIXDOG_ANTHROPIC_ALWAYS_ENABLE_EFFORT)) return true;
+    const m = normalizeModelId(model);
+    if (!m.includes('claude')) return false;
+    if (isLegacyAnthropicReasoningModel(model)) return false;
+    if (m.includes('opus-4-6') || m.includes('sonnet-4-6')) return true;
+    if (m.includes('sonnet-5') || m.includes('fable-5')) return true;
+    if (/^claude-opus-4-(6|7|8)(?:$|[-@])/.test(m)) return true;
+    if (/^claude-opus-5-/.test(m)) return true;
+    if (m.startsWith('claude-')) return true;
+    return false;
+}
+
+// @[MODEL LAUNCH]: extend when new Opus models support max effort.
+export function modelSupportsMaxEffort(model) {
+    const m = normalizeModelId(model);
+    if (/^claude-opus-4-(6|7|8)(?:$|[-@])/.test(m)) return true;
+    if (/^claude-opus-5-/.test(m)) return true;
+    return false;
+}
+
+function isEnvTruthy(value) {
+    const s = String(value ?? '').trim().toLowerCase();
+    return s === '1' || s === 'true' || s === 'yes' || s === 'on';
+}
+
+function logOnce(key, write) {
+    if (_LOGGED_EFFORT_NORMALIZATION.has(key)) return;
+    _LOGGED_EFFORT_NORMALIZATION.add(key);
+    try { write(); } catch { /* ignore */ }
+}
+
+/**
+ * Normalize user/config effort to API levels (low/medium/high/max).
+ * Maps legacy xhigh → max (Opus 4.6+) or high; clamps max on non-Opus models.
+ */
+export function normalizeAnthropicEffortInput(raw, model, logTag = 'anthropic') {
+    if (raw === undefined || raw === null || raw === '') return undefined;
+    let level = String(raw).trim().toLowerCase();
+    if (level === 'extra-high' || level === 'very-high') level = 'xhigh';
+
+    if (level === 'xhigh') {
+        const mapped = modelSupportsMaxEffort(model) ? 'max' : 'high';
+        logOnce(`xhigh:${mapped}`, () => {
+            process.stderr.write(
+                `[${logTag}] effort "xhigh" normalized to "${mapped}" (official levels: ${EFFORT_LEVELS.join(', ')})\n`,
+            );
+        });
+        return mapped;
+    }
+
+    if (level === 'max' && !modelSupportsMaxEffort(model)) {
+        logOnce('max:high', () => {
+            process.stderr.write(`[${logTag}] effort "max" downgraded to "high" for model ${model}\n`);
+        });
+        return 'high';
+    }
+
+    if (EFFORT_LEVELS.includes(level)) return level;
+    return undefined;
+}
+
+export function anthropicEffortUsesOutputConfig(model, opts = {}) {
+    const thinkingBudgetTokens = Number(opts.thinkingBudgetTokens);
+    if (Number.isFinite(thinkingBudgetTokens) && thinkingBudgetTokens > 0) return false;
+    return modelSupportsEffort(model);
+}
+
+export function shouldIncludeEffortBeta(model, opts = {}) {
+    return anthropicEffortUsesOutputConfig(model, opts);
+}
+
+/**
+ * Apply effort / thinking to a Messages API body or SDK params object (mutates).
+ */
+export function applyAnthropicEffortToBody(
+    body,
+    {
+        model,
+        opts = {},
+        maxTokens,
+        clampThinkingBudgetTokens,
+        logTag = 'anthropic',
+    },
+) {
+    if (!body || typeof body !== 'object') return;
+
+    const thinkingBudgetTokens = Number(opts.thinkingBudgetTokens);
+    if (Number.isFinite(thinkingBudgetTokens) && thinkingBudgetTokens > 0) {
+        const budgetTokens = clampThinkingBudgetTokens(thinkingBudgetTokens, maxTokens);
+        if (budgetTokens) {
+            body.thinking = { type: 'enabled', budget_tokens: budgetTokens };
+        }
+        return;
+    }
+
+    const normalized = normalizeAnthropicEffortInput(opts.effort, model, logTag);
+
+    if (modelSupportsEffort(model)) {
+        if (normalized) {
+            const existing = body.output_config && typeof body.output_config === 'object'
+                ? body.output_config
+                : {};
+            body.output_config = { ...existing, effort: normalized };
+        }
+        return;
+    }
+
+    if (normalized && LEGACY_EFFORT_BUDGET[normalized]) {
+        const budgetTokens = clampThinkingBudgetTokens(LEGACY_EFFORT_BUDGET[normalized], maxTokens);
+        if (budgetTokens) {
+            body.thinking = { type: 'enabled', budget_tokens: budgetTokens };
+        }
+        return;
+    }
+
+    if (opts.effort && normalized === undefined && !_LOGGED_UNKNOWN_EFFORT.has(opts.effort)) {
+        _LOGGED_UNKNOWN_EFFORT.add(opts.effort);
+        try {
+            process.stderr.write(
+                `[${logTag}] unknown effort=${opts.effort} ignored (known legacy: ${Object.keys(LEGACY_EFFORT_BUDGET).join(', ')})\n`,
+            );
+        } catch { /* ignore */ }
+    }
+}
+
+export function effortValuesForModel(capabilities, modelId) {
+    const effort = capabilities?.effort;
+    let levels = [...EFFORT_LEVELS];
+    if (!modelSupportsMaxEffort(modelId)) {
+        levels = levels.filter((level) => level !== 'max');
+    }
+    if (!effort) return [];
+    if (effort === true) return levels;
+    const values = levels.filter(
+        (level) => effort?.[level] === true || effort?.[level]?.supported === true,
+    );
+    if (values.length) return values;
+    return effort.supported === true ? levels : [];
+}

package/src/runtime/agent/orchestrator/providers/anthropic-leaked-toolcall.mjs

@@ -0,0 +1,444 @@
+/**
+ * Leaked tool-call recovery for the shared Anthropic SSE parser.
+ *
+ * Opus (via OAuth) and, less often, the API model sometimes emit a tool
+ * call as plain text tags inside `text_delta` (e.g. `<function_calls>`,
+ * `<invoke name="...">`, `<invoke ...>`, or a stray prefix like
+ * `course<invoke ...>`) instead of a native `tool_use` content block. Left
+ * alone those tags stream to the TUI as assistant prose and the tool never
+ * runs.
+ *
+ * `scanLeakedToolCalls` is the boundary-gated scanner the parser feeds its
+ * rolling text window. Unlike hermes-agent's strip-only port (which only
+ * DROPS the call), this recovers a real call: it suppresses the tags from
+ * the visible stream AND returns a synthesizable `{ name, arguments }` when
+ * the leaked payload parses to a KNOWN tool. Unknown / unparseable / prose
+ * `<function>` mentions are flushed as ordinary text so no user-visible
+ * content is ever lost.
+ */
+
+import { traceHash, stableTraceStringify } from './trace-utils.mjs';
+
+// Opener strings (lowercased) used only for partial-tail hold detection —
+// so a sentinel split across text_delta chunk boundaries is still caught.
+const OPENERS = [
+    '<function_calls',
+    '<invoke',
+    '<function',            // Gemma-style; name= is gated at match time
+];
+// OpenAI harmony / gpt-oss channel-syntax openers. Only consulted when a
+// caller opts in via `{ harmony: true }` (OpenAI-family stream parsers) so the
+// Anthropic path — and its regression tests — are byte-for-byte unaffected.
+const HARMONY_OPENERS = [
+    '<|channel|>',
+    '<|start|>',
+];
+const LONGEST_OPENER = Math.max(
+    ...OPENERS.map((o) => o.length),
+    ...HARMONY_OPENERS.map((o) => o.length),
+);
+
+// First *definite* leaked-tool-call sentinel. The bare `<function` opener is
+// boundary-gated to require a `name=` attribute so prose like
+// "Use <function> in JavaScript" is never treated as a call.
+// NOTE (Fix 3): a real leak ALWAYS carries the leading `<` bracket
+// (`<invoke ...>`, `<invoke ...>`, `<function_calls>`); the model never
+// emits a bare `antml:invoke` token without it. The old bare-`antml:` opener
+// therefore only ever matched ordinary prose that mentions the literal string,
+// and — because `matchCompleteBlockAt` has no bare form — it forced the scanner
+// to HOLD from that token to end of stream. Removed: only bracketed forms are
+// sentinels now, so prose containing "antml:invoke" mid-sentence streams
+// promptly and is never recovered.
+const SENTINEL_RE = /<(?:antml:)?function_calls\b|<(?:antml:)?invoke\b|<(?:antml:)?function\b[^>]*\bname\s*=/i;
+
+// Harmony channel-syntax sentinel: the `<|channel|>` (optionally preceded by a
+// `<|start|>` role token) that frames a gpt-oss tool call. Only whether it is a
+// *tool* call is decided at block-match time (recipient `to=functions.NAME`).
+const HARMONY_SENTINEL_RE = /<\|(?:channel|start)\|>/i;
+// Terminator of a harmony message. A tool call ends with `<|call|>`; ordinary
+// channel content ends with `<|end|>` / `<|return|>` — all three close a block
+// so benign channel text flushes promptly instead of stalling to stream end.
+const HARMONY_TERMINATOR_RE = /<\|call\|>|<\|end\|>|<\|return\|>/i;
+const HARMONY_RECIPIENT_RE = /to=functions\.([A-Za-z0-9_.\-]+)/i;
+const HARMONY_MESSAGE_RE = /<\|message\|>/i;
+
+function firstSentinelIndex(s, harmony) {
+    const m = SENTINEL_RE.exec(s);
+    let idx = m ? m.index : -1;
+    if (harmony) {
+        const hm = HARMONY_SENTINEL_RE.exec(s);
+        if (hm && (idx === -1 || hm.index < idx)) idx = hm.index;
+    }
+    return idx;
+}
+
+function isPrefixOfAnyOpener(tail, harmony) {
+    const t = tail.toLowerCase();
+    if (OPENERS.some((o) => o.startsWith(t))) return true;
+    return harmony ? HARMONY_OPENERS.some((o) => o.startsWith(t)) : false;
+}
+
+// --- Markdown code-fence / inline-code tracking (Fix 1) --------------------
+// A leaked tool-call tag written INSIDE a fenced code block (``` … ``` or
+// ~~~ … ~~~) or an inline code span (`…`) is a documentation example, not a
+// real call — it must stream as visible text, never be recovered/executed.
+// Because the scanner runs on a rolling buffer, the fence state is tracked as
+// a small struct threaded across chunks (a fence opened in one delta and
+// closed in a later one persists). `advanceFenceState` folds a text slice into
+// the state; `isInCode` reports whether a position sits inside code.
+function initialFenceState() {
+    return { fenceChar: null, fenceLen: 0, inlineLen: 0, atLineStart: true };
+}
+function isInCode(s) {
+    return !!s && (s.fenceChar !== null || s.inlineLen > 0);
+}
+// Fold `text` into fence state and return the new state (input not mutated).
+// Fenced blocks (``` / ~~~ at line start, length ≥ 3) persist across lines
+// until a closing fence of the same char (length ≥ opener) at line start.
+// Inline spans (backtick run not opening a fence) are line-local: reset at
+// newline so a stray unmatched backtick in prose cannot poison later lines
+// (which would wrongly suppress a genuine leaked call).
+function advanceFenceState(state, text) {
+    let { fenceChar, fenceLen, inlineLen, atLineStart } = state || initialFenceState();
+    let i = 0;
+    while (i < text.length) {
+        const ch = text[i];
+        if (ch === '`' || ch === '~') {
+            let j = i;
+            while (j < text.length && text[j] === ch) j++;
+            const runLen = j - i;
+            if (fenceChar) {
+                if (ch === fenceChar && atLineStart && runLen >= fenceLen) {
+                    fenceChar = null; fenceLen = 0;
+                }
+            } else if (inlineLen) {
+                if (ch === '`' && runLen === inlineLen) inlineLen = 0;
+            } else if (atLineStart && runLen >= 3) {
+                fenceChar = ch; fenceLen = runLen;
+            } else if (ch === '`') {
+                inlineLen = runLen;
+            }
+            atLineStart = false;
+            i = j;
+            continue;
+        }
+        if (ch === '\n') {
+            atLineStart = true;
+            inlineLen = 0; // inline spans do not cross lines here
+            i++;
+            continue;
+        }
+        if (atLineStart && (ch === ' ' || ch === '\t')) { i++; continue; }
+        atLineStart = false;
+        i++;
+    }
+    return { fenceChar, fenceLen, inlineLen, atLineStart };
+}
+
+// Index from which the tail of `s` could be the *start* of a sentinel that
+// has not fully arrived yet. Everything before it is safe to flush now;
+// from it onward must be held for the next chunk. Returns s.length when
+// nothing needs holding (normal text streams promptly).
+function partialTailIndex(s, harmony) {
+    const start = Math.max(0, s.length - LONGEST_OPENER);
+    for (let i = start; i < s.length; i++) {
+        if (isPrefixOfAnyOpener(s.slice(i), harmony)) return i;
+    }
+    return s.length;
+}
+
+// Coerce a leaked `<parameter>` value. apply_patch-style raw bodies and plain
+// strings pass through verbatim; JSON-looking scalars/objects are parsed so
+// the synthesized arguments match a native tool_use input shape.
+function coerceValue(raw) {
+    const trimmed = raw.trim();
+    if (trimmed === '') return raw;
+    const first = trimmed[0];
+    if ('{[0123456789tfn"-'.includes(first)) {
+        try { return JSON.parse(trimmed); } catch { /* fall through to raw */ }
+    }
+    return raw;
+}
+
+function parseParams(text) {
+    const out = {};
+    const re = /<(?:antml:)?parameter\b[^>]*\bname\s*=\s*["']?([^"'>\s]+)["']?[^>]*>([\s\S]*?)<\/(?:antml:)?parameter\s*>/gi;
+    let m;
+    while ((m = re.exec(text)) !== null) {
+        out[m[1]] = coerceValue(m[2]);
+    }
+    return out;
+}
+
+function parseInvoke(text) {
+    const nameM = /<(?:antml:)?invoke\b[^>]*\bname\s*=\s*["']?([^"'>\s]+)["']?/i.exec(text);
+    if (!nameM) return null;
+    return { name: nameM[1], arguments: parseParams(text) };
+}
+
+// Match a COMPLETE harmony channel block starting at `idx` (which points at a
+// `<|channel|>` or `<|start|>` token). Returns `{ end, call }` where `call` is
+// `{ name, arguments }` when the block is a `to=functions.NAME` tool call, or
+// null when it is ordinary channel content (flushed as visible text). Returns
+// null (undecided → caller holds) until a terminator token has arrived.
+function matchCompleteHarmonyBlockAt(s, idx) {
+    const rest = s.slice(idx);
+    if (!/^<\|(?:channel|start)\|>/i.test(rest)) return null;
+    const termM = HARMONY_TERMINATOR_RE.exec(rest);
+    if (!termM) return null;
+    const end = idx + termM.index + termM[0].length;
+    const block = rest.slice(0, termM.index);
+    const recM = HARMONY_RECIPIENT_RE.exec(block);
+    if (!recM) return { end, call: null };
+    const msgM = HARMONY_MESSAGE_RE.exec(block);
+    let args = {};
+    if (msgM) {
+        const raw = block.slice(msgM.index + msgM[0].length).trim();
+        if (raw) {
+            const coerced = coerceValue(raw);
+            if (coerced && typeof coerced === 'object' && !Array.isArray(coerced)) args = coerced;
+        }
+    }
+    return { end, call: { name: recM[1], arguments: args } };
+}
+
+// Try to match a COMPLETE leaked block starting at `idx`. Returns
+// `{ end, call }` where `end` is the index in `s` just past the block and
+// `call` is `{ name, arguments }` (or null when the block is well-formed XML
+// but carries no parseable invoke). Returns null when the block has not
+// finished arriving yet (caller holds and waits for more chunks).
+function matchCompleteBlockAt(s, idx, harmony) {
+    const rest = s.slice(idx);
+
+    // Harmony channel syntax (opt-in). Routed first because its opener token
+    // (`<|...|>`) never collides with the XML `<function...>` families below.
+    if (harmony && /^<\|(?:channel|start)\|>/i.test(rest)) {
+        return matchCompleteHarmonyBlockAt(s, idx);
+    }
+
+    // <function_calls> ... </function_calls> wrapper
+    const wm = /^<(?:antml:)?function_calls\b[^>]*>/i.exec(rest);
+    if (wm) {
+        const cm = /<\/(?:antml:)?function_calls\s*>/i.exec(rest);
+        if (!cm) return null;
+        const end = idx + cm.index + cm[0].length;
+        const inner = rest.slice(wm[0].length, cm.index);
+        return { end, call: parseInvoke(inner) };
+    }
+
+    // <invoke ...> ... </invoke>
+    const im = /^<(?:antml:)?invoke\b[^>]*>/i.exec(rest);
+    if (im) {
+        const cm = /<\/(?:antml:)?invoke\s*>/i.exec(rest);
+        if (!cm) return null;
+        const blockText = rest.slice(0, cm.index + cm[0].length);
+        return { end: idx + blockText.length, call: parseInvoke(blockText) };
+    }
+
+    // <function name="..."> ... </function>  (Gemma-style)
+    const fm = /^<(?:antml:)?function\b[^>]*\bname\s*=\s*["']?([^"'>\s]+)["']?[^>]*>/i.exec(rest);
+    if (fm) {
+        const cm = /<\/(?:antml:)?function\s*>/i.exec(rest);
+        if (!cm) return null;
+        const end = idx + cm.index + cm[0].length;
+        const inner = rest.slice(fm[0].length, cm.index);
+        return { end, call: fm[1] ? { name: fm[1], arguments: parseParams(inner) } : null };
+    }
+
+    // Sentinel matched (e.g. bare `antml:invoke` with no `<`) but no complete
+    // block form — leave undecided so the caller holds (or flushes on final).
+    return null;
+}
+
+/**
+ * Scan a rolling text window for leaked tool-call syntax.
+ *
+ * @param {string} buffer  Held-back text window.
+ * @param {object} opts
+ * @param {(name:string)=>boolean} opts.isKnownTool  Validates the recovered
+ *        tool name against the tools available to this request.
+ * @param {boolean} opts.final  True on stream end: flush everything, never
+ *        hold (legitimate text is never lost).
+ * @param {boolean} [opts.harmony]  Opt-in (OpenAI-family callers only): also
+ *        recognize gpt-oss/harmony `<|channel|>...to=functions.NAME...<|call|>`
+ *        tool syntax. Off by default so the Anthropic path is unaffected.
+ * @param {object} [opts.fenceState]  Running markdown fence/inline-code state
+ *        from the previous chunk (Fix 1). A fence opened in an earlier delta and
+ *        closed in a later one is respected because `fence` reflects the state
+ *        as-of the start of `buffer`. Omit for single-shot use.
+ * @returns {{ emit:string, calls:Array<{name:string,arguments:object}>, rest:string, fenceState:object }}
+ *          `emit` = text safe to forward now, `calls` = recovered known-tool
+ *          calls to synthesize/dispatch, `rest` = text to keep buffered,
+ *          `fenceState` = markdown state as-of the start of `rest` (thread it
+ *          back in on the next call).
+ */
+export function scanLeakedToolCalls(buffer, { isKnownTool, final, harmony = false, fenceState = null }) {
+    let emit = '';
+    const calls = [];
+    const buf = buffer;
+    let pos = 0;
+    // Fix 1: fence state advances over EVERY consumed character in raw order,
+    // so `isInCode` at a sentinel reflects the running markdown context —
+    // including a fence opened in an earlier chunk (caller threads fenceState).
+    let fence = fenceState || initialFenceState();
+    // `asText` controls both whether the slice is emitted AND whether it
+    // advances markdown fence state. A recovered/SUPPRESSED tool-call block
+    // (asText=false) is hidden from the visible stream, so it must NOT feed the
+    // fence tracker either (Fix 2 follow-up / reviewer Medium): an unmatched
+    // backtick inside the hidden XML/harmony args would otherwise leave a fence
+    // "open" and wrongly suppress a subsequent REAL leaked call as in-code.
+    // Fence state only ever reflects text the user actually sees.
+    const consume = (to, asText) => {
+        if (asText) {
+            fence = advanceFenceState(fence, buf.slice(pos, to));
+            emit += buf.slice(pos, to);
+        }
+        pos = to;
+    };
+
+    while (pos < buf.length) {
+        const rel = firstSentinelIndex(buf.slice(pos), harmony);
+        if (rel === -1) {
+            if (final) { consume(buf.length, true); }
+            else {
+                const hold = partialTailIndex(buf.slice(pos), harmony);
+                consume(pos + hold, true);
+            }
+            break;
+        }
+        const idx = pos + rel;
+        // Plain text before the sentinel always streams as visible text.
+        consume(idx, true);
+        const inCode = isInCode(fence);
+        const block = matchCompleteBlockAt(buf, idx, harmony);
+        if (!block) {
+            // Opener present but block not fully arrived.
+            if (final) { consume(buf.length, true); break; }
+            if (inCode) {
+                // Inside a code fence / inline span: this can never become a
+                // real recovered call — emit the opener char and keep scanning
+                // instead of holding it (and the rest of the doc) to stream end.
+                consume(idx + 1, true);
+                continue;
+            }
+            // Outside code: hold from the opener for the next chunk.
+            break;
+        }
+        const { end, call } = block;
+        if (!inCode && call && call.name && isKnownTool(call.name)) {
+            // Recovered a real, known tool call OUTSIDE any code fence:
+            // suppress the tags from the visible stream and synthesize it.
+            consume(end, false);
+            calls.push({ name: call.name, arguments: call.arguments || {} });
+            continue;
+        }
+        // In code, unknown tool, or no parseable invoke: keep the block as
+        // ordinary visible text and continue scanning past it.
+        consume(end, true);
+    }
+
+    return { emit, calls, rest: buf.slice(pos), fenceState: fence };
+}
+
+/**
+ * Stateful convenience wrapper around `scanLeakedToolCalls` for the streaming
+ * parsers. Holds the minimal rolling text window across chunk boundaries and
+ * exposes `push(delta)` / `flush()` that return the visible text safe to
+ * forward now plus any recovered known-tool calls. The Anthropic parser keeps
+ * its own inline wiring; this factory is shared by the OpenAI-family stream
+ * parsers so each only wires provider-specific call synthesis + dispatch.
+ *
+ * @param {object} opts
+ * @param {Set<string>|string[]} opts.knownToolNames  Tools offered this request.
+ * @param {boolean} [opts.harmony]  Opt-in gpt-oss/harmony channel detection.
+ * @returns {{ enabled:boolean, push:(delta:string, final?:boolean)=>{text:string,calls:Array<{name:string,arguments:object}>}, flush:()=>{text:string,calls:Array<{name:string,arguments:object}>} }}
+ */
+export function createLeakGuard({ knownToolNames, harmony = false } = {}) {
+    const known = knownToolNames instanceof Set
+        ? knownToolNames
+        : new Set(Array.isArray(knownToolNames) ? knownToolNames : []);
+    const enabled = known.size > 0;
+    const isKnownTool = (name) => known.has(name);
+    let buffer = '';
+    // Running markdown fence/inline-code state threaded across chunks (Fix 1).
+    let fence = initialFenceState();
+    const run = (delta, final) => {
+        if (!enabled) return { text: delta || '', calls: [] };
+        buffer += delta || '';
+        if (!buffer && !final) return { text: '', calls: [] };
+        const { emit, calls, rest, fenceState } = scanLeakedToolCalls(buffer, { isKnownTool, final, harmony, fenceState: fence });
+        buffer = rest;
+        fence = fenceState;
+        return { text: emit, calls };
+    };
+    return {
+        enabled,
+        push: (delta, final = false) => run(delta, final),
+        flush: () => run('', true),
+    };
+}
+
+/**
+ * Shared name+args fingerprint dedupe for tool calls (Fix 2). The leak guard
+ * dispatches a synthesized call the instant it is recovered from text; the
+ * native tool_use / tool_calls path can later surface the SAME call. Without a
+ * cross-path guard both fire and a side-effecting tool double-executes. This
+ * factory holds a per-stream Set of dispatched fingerprints (name + stably
+ * stringified args) — call `shouldDispatch(name, args)` before EITHER a
+ * synthetic or a native dispatch: it returns false (skip) when an identical
+ * fingerprint already fired. Mirrors gemini.mjs's dedupe so all providers
+ * behave identically.
+ */
+export function toolCallFingerprint(name, args) {
+    let a = args;
+    if (a === null || typeof a !== 'object' || Array.isArray(a)) a = {};
+    return traceHash(stableTraceStringify({ name: name || '', args: a }));
+}
+export function createToolCallDedupe() {
+    const seen = new Set();
+    return {
+        // True the first time this (name,args) fingerprint is seen; false on
+        // any later identical call (synthetic-then-native or vice-versa).
+        shouldDispatch(name, args) {
+            const fp = toolCallFingerprint(name, args);
+            if (seen.has(fp)) return false;
+            seen.add(fp);
+            return true;
+        },
+        has(name, args) { return seen.has(toolCallFingerprint(name, args)); },
+    };
+}
+
+/**
+ * Drop duplicate tool calls from a RETURNED `toolCalls` array by name+args
+ * fingerprint (Fix 2, array side). Dispatch-time dedupe (`createToolCallDedupe`)
+ * only suppresses the second `onToolCall`; providers that also RETURN a
+ * `toolCalls` array (which the agent loop executes) can still carry a
+ * synthetic-leaked + identical-native duplicate. Run the final array through
+ * this so the loop never executes a side-effecting tool twice. Order is
+ * preserved; the FIRST occurrence of each fingerprint wins.
+ *
+ * @param {Array<{name?:string,arguments?:object}>} calls
+ * @returns {Array} the array with later duplicate-fingerprint calls removed.
+ */
+export function dedupeToolCallList(calls) {
+    if (!Array.isArray(calls) || calls.length < 2) return calls;
+    const seen = new Set();
+    const out = [];
+    for (const call of calls) {
+        if (!call || typeof call !== 'object') { out.push(call); continue; }
+        const fp = toolCallFingerprint(call.name, call.arguments);
+        if (seen.has(fp)) continue;
+        seen.add(fp);
+        out.push(call);
+    }
+    return out.length === calls.length ? calls : out;
+}
+
+// Exposed for focused unit tests.
+export const _internals = {
+    firstSentinelIndex,
+    partialTailIndex,
+    matchCompleteBlockAt,
+    parseInvoke,
+};