agentfootprint 6.25.0 → 6.26.1

package/README.md

@@ -1,17 +1,23 @@
 
+
+<h1 align="center">Agentfootprint</h1>
+
 <p align="center">
-  <picture>
-    <source media="(prefers-color-scheme: dark)" srcset="docs/assets/hero-dark.svg">
-    <source media="(prefers-color-scheme: light)" srcset="docs/assets/hero-light.svg">
-    <img alt="agentfootprint mascot composing context flavors (Skills, Steering, Guardrails, RAG, Tool APIs, Memory) into three structured LLM slots (system, messages, tools) — the central abstraction, visualized." src="docs/assets/hero-light.svg" width="100%"/>
-  </picture>
+  <strong>Your agent picked the wrong tool, gave a wrong answer — and the logs can't tell you why.<br/>Agentfootprint can.</strong>
 </p>
 
-<h1 align="center">Agentfootprint</h1>
+<p align="center">
+  The explainable agent framework: every read, write, decision, and tool call becomes
+  <strong>connected evidence</strong> as your agent runs. When something goes wrong, you don't grep logs — you ask.
+</p>
 
 <p align="center">
-  <strong>We abstract context engineering — and hand back the trace.</strong><br/>
-  <strong>Live</strong> to develop · <strong>offline</strong> to monitor · <strong>detailed</strong> to improve.
+  <a href="https://footprintjs.github.io/agentThinkingUI/">
+    <img src="docs/assets/hero-atui.png" alt="An agent run replayed in AgentThinkingUI — the LLM 'brain' calls the Flight-search tool, the step inspector shows the tool's raw output and the brain's reasoning about it, and the timeline scrubs every step of the run." width="100%">
+  </a>
+</p>
+<p align="center">
+  <sub>A real run, replayed — rendered with <a href="https://github.com/footprintjs/agentThinkingUI"><b>AgentThinkingUI</b></a> (<code>npm i agentthinkingui</code>). Every frame is generated from the run's own trace; <a href="https://footprintjs.github.io/agentThinkingUI/">▶ watch it live</a>.</sub>
 </p>
 
 <p align="center">
@@ -26,7 +32,176 @@
 
 ---
 
-## 1. What we abstract
+## The new error class
+
+For decades, software had two kinds of errors — and developers never needed deep
+domain knowledge to fix either:
+
+| Error class | Where the bug lives | How you find it |
+|---|---|---|
+| **Infrastructure** — crash, timeout, 500 | the system | infra logs, monitoring |
+| **Business logic** — wrong branch, wrong math | the code | stack trace, debugger, `console.log` |
+| **Contextual** — wrong tool chosen, wrong fact believed, stale memory trusted | **what the model was given** | **nothing. Until now.** |
+
+Agents introduced the third class. The code is correct, the infra is healthy — and
+the run is still wrong, because two tool descriptions read alike, or an injected
+fact was misleading, or memory carried last week's truth. Classical logs can't
+explain it: **they record what the code did, never what the context did.**
+
+## The idea
+
+If contextual errors live in what the model was given, then the run itself must be
+structured so context is **evidence** — every injection, read, write, decision, and
+tool call recorded *connected*, the moment it happens. Not logs you grep. Evidence
+you ask.
+
+## How — we abstract context engineering
+
+Every piece of context enters the LLM through one of **3 slots** (`system` ·
+`messages` · `tools`), under one of **4 triggers** — skills, steering, RAG, facts,
+memory, guardrails are all the same move: `Injection = slot × trigger × cache`.
+
+**Because the framework owns that injection point, every piece of context is born
+tracked.** Tracking isn't an add-on you wire up — it's a consequence of the
+abstraction. [The full model ↓](#the-model--what-we-abstract)
+
+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" srcset="docs/assets/hero-dark.svg">
+    <source media="(prefers-color-scheme: light)" srcset="docs/assets/hero-light.svg">
+    <img alt="agentfootprint mascot composing context flavors (Skills, Steering, Guardrails, RAG, Tool APIs, Memory) into three structured LLM slots (system, messages, tools) — the central abstraction, visualized." src="docs/assets/hero-light.svg" width="100%"/>
+  </picture>
+</p>
+
+## What tracking buys you
+
+**See it in 30 seconds** — four questions logs can't answer, each answered by code in this repo from a real run:
+
+```text
+Q: Why did the model pick refund_full instead of refund_partial?
+A: margin 0.02 — ⚠ NARROW: the two tool descriptions read nearly identical
+   (toolChoiceRecorder — and the catalog lint flags the pair before you ever run)
+
+Q: Why was this loan declined?
+A: decision ← [control: "DTI above the 0.43 affordability ceiling"] ← dti 0.52 ← monthlyDebt / income
+   (decide() evidence + the causal slice — every hop is a real recorded edge)
+
+Q: Which piece of context made the answer wrong?
+A: CAUSAL: ablating fact 'vip-override' flipped the outcome in 3/3 seeded reruns
+   (localizeContextBug — ranked proxies, counterfactual proof)
+
+Q: Prove nobody edited this run's record.
+A: verifyAuditBundle → valid: false, brokenAt: #16 — the tampered record, named
+   (hash-chained audit export, offline verification)
+```
+
+And you don't have to read the trace yourself — **we provide the tools for an LLM to track it for you**: the trace toolpack let a debugger model find a planted bug while reading **9.5% of the trace** ([guide](docs/guides/trace-debugging.md)).
+
+---
+
+## Pick your door
+
+| 🔧 Building an agent? | 🐛 Agent misbehaving? | 🏛️ Need audit / compliance? |
+|---|---|---|
+| Typed agents with skills, steering, RAG, memory, guardrails — and the trace for free. | Lint your tool catalog in 5 minutes — works on **any** framework's tool list (plain JSON / MCP / OpenAI / Anthropic shapes). Then causal slices, context bisection, and the debugger-LLM toolpack. | Hash-chained, tamper-evident run records with an offline verifier — record-keeping in the EU-AI-Act shape. |
+| [→ Quick start](#quick-start--runs-offline-no-api-key) | [→ Tool-catalog lint](docs/guides/tool-catalog-lint.md) · [→ Trace debugging](docs/guides/trace-debugging.md) | [→ Tamper-evident audit](docs/guides/security.md) |
+
+---
+
+## Quick start — runs offline, no API key
+
+```bash
+npm install agentfootprint footprintjs
+```
+
+```typescript
+import { Agent, defineTool, mock } from 'agentfootprint';
+
+const weather = defineTool({
+  name: 'weather',
+  description: 'Get current weather for a city.',
+  inputSchema: {
+    type: 'object',
+    properties: { city: { type: 'string' } },
+    required: ['city'],
+  },
+  execute: async ({ city }: { city: string }) => `${city}: 72°F, sunny`,
+});
+
+const agent = Agent.create({
+  provider: mock({ reply: 'I checked: it is 72°F and sunny.' }),
+  model: 'mock',
+})
+  .system('You answer weather questions using the weather tool.')
+  .tool(weather)
+  .build();
+
+const result = await agent.run({ message: 'Weather in Paris?' });
+console.log(result);  // → "I checked: it is 72°F and sunny."
+```
+
+For production, import a real provider from `agentfootprint/llm-providers` and swap it in — `anthropic(...)` / `openai(...)` / `bedrock(...)` / `ollama(...)`. Only the import line changes; the agent code stays the same. (The vendor-SDK providers live on the `agentfootprint/llm-providers` subpath so the main `agentfootprint` barrel stays free of optional peer-dep requires; `mock`, `browserAnthropic`, and `browserOpenai` are on the main barrel.)
+
+### Then add context
+
+A real agent carries more than one prompt and one tool: facts about the user, always-on rules, skills that unlock on demand. Declare each piece — the framework decides **when** it fires and **which slot** it lands in, and every piece is born tracked:
+
+```typescript
+import { defineFact, defineSteering, defineSkill } from 'agentfootprint';
+
+const agent = Agent.create({ provider, model })
+  .system('You are a support agent.')
+  .fact(defineFact({                    // data the model should know — always on
+    id: 'user-profile',
+    data: 'Name: Maya · Plan: Pro · Customer since 2022',
+  }))
+  .steering(defineSteering({            // rules the model must follow — always on
+    id: 'refund-policy',
+    prompt: 'Never promise a refund before checking the policy tool.',
+  }))
+  .skill(defineSkill({                  // guidance + tools — unlocks when the LLM asks
+    id: 'billing',
+    description: 'Use for refunds, charges, billing questions.',
+    body: 'When handling billing: confirm identity first, then…',
+    tools: [refundTool],
+  }))
+  .build();
+```
+
+Same shape for `.instruction()` / `.memory()` / `.rag()` / raw `.injection()` — they're all the one primitive, `Injection = slot × trigger × cache`. [The full model ↓](#the-model--what-we-abstract)
+
+### Then compose control flow
+
+One agent is a `Runner`. So is every composition of agents — four control-flow primitives, and anything that runs composes into anything else:
+
+```typescript
+import { Sequence, Parallel, Conditional } from 'agentfootprint';
+
+const pipeline = Sequence.create()
+  .step('classify', classifyAgent)                  // sequence: step → step
+  .step('review',
+    Parallel.create()                               // parallel: fan out, then merge
+      .branch('legal', legalAgent)
+      .branch('ethics', ethicsAgent)
+      .mergeWithLLM({ provider, model, prompt: 'Synthesize:' })
+      .build())
+  .step('respond',
+    Conditional.create()                            // conditional: one branch runs
+      .when('urgent', (i) => i.message.startsWith('URGENT'), urgentAgent)
+      .otherwise('normal', normalAgent)
+      .build())
+  .build();
+
+await pipeline.run({ message: 'URGENT: refund dispute on order #4411' });
+```
+
+The fourth primitive is `Loop` — `Loop.repeat(agent).until(guard).times(5)`, with a mandatory budget guard. And the named patterns from the research literature ship pre-composed from the same four: `selfConsistency` · `reflection` · `debate` · `mapReduce` · `tot` · `swarm`. Because every composition is a flowchart, the structure you wrote is the structure you see in the UI — and the trace spans the whole pipeline, not one agent at a time. [Designing systems of agents ↓](#how-do-i-design-my-agent-or-system-of-agents)
+
+---
+
+## The model — what we abstract
+
+
 
 When you build an Agentic Application, you collect domain-specific data and instructions, then wire them up based on what your system receives.
 
@@ -70,7 +245,7 @@ That's the whole model: `Injection = slot × trigger × cache`.
 
 ---
 
-## 2. Why we chose this abstraction
+## Why we chose this abstraction
 
 The agent space has many credible primary abstractions:
 
@@ -143,7 +318,7 @@ And a fourth, novel: **the agent can read its own trace.** Six months after the
 
 ---
 
-## 3. How do I design my agent or system of agents?
+## How do I design my agent or system of agents?
 
 Two scales — same alphabet. Four control flows are the entire vocabulary.
 
@@ -305,7 +480,14 @@ Same trick as Beat 1: instead of N libraries for N patterns, we found the M buil
 
 ---
 
-## 4. How do I see what my agent did?
+## How do I see what my agent did?
+
+<p align="center">
+  <img src="docs/assets/lens-run.png" alt="A real agent run in the Lens: the conversation (with live PII redaction), the executed path lit on the merge-tree flowchart, the WHAT-HAPPENED timeline of every iteration/context/LLM turn/route, run stats, and the step inspector — all generated from the run's own trace." width="100%">
+</p>
+<p align="center">
+  <sub>One real run, fully explained — the <a href="https://github.com/footprintjs/agentfootprint-lens"><b>Lens</b></a> (<code>npm i agentfootprint-lens</code>): conversation · executed path · per-step timeline · stats, every pixel from the trace.</sub>
+</p>
 
 Because we own the loop (Beat 2), every decision and execution is captured during traversal — not bolted on. The default capture is the **causal trace**: every stage, read, write, and decision evidence, as a JSON-portable, scrubbable, queryable, exportable artifact. Beyond the default, wire custom recorders for cost, latency, or quality scoring — any observation hook fires on the same stream.
 
@@ -427,42 +609,6 @@ off the hot path.
 
 ---
 
-## Quick start — runs offline, no API key
-
-```bash
-npm install agentfootprint footprintjs
-```
-
-```typescript
-import { Agent, defineTool, mock } from 'agentfootprint';
-
-const weather = defineTool({
-  name: 'weather',
-  description: 'Get current weather for a city.',
-  inputSchema: {
-    type: 'object',
-    properties: { city: { type: 'string' } },
-    required: ['city'],
-  },
-  execute: async ({ city }: { city: string }) => `${city}: 72°F, sunny`,
-});
-
-const agent = Agent.create({
-  provider: mock({ reply: 'I checked: it is 72°F and sunny.' }),
-  model: 'mock',
-})
-  .system('You answer weather questions using the weather tool.')
-  .tool(weather)
-  .build();
-
-const result = await agent.run({ message: 'Weather in Paris?' });
-console.log(result);  // → "I checked: it is 72°F and sunny."
-```
-
-For production, import a real provider from `agentfootprint/llm-providers` and swap it in — `anthropic(...)` / `openai(...)` / `bedrock(...)` / `ollama(...)`. Only the import line changes; the agent code stays the same. (The vendor-SDK providers live on the `agentfootprint/llm-providers` subpath so the main `agentfootprint` barrel stays free of optional peer-dep requires; `mock`, `browserAnthropic`, and `browserOpenai` are on the main barrel.)
-
----
-
 ## Mocks first, production second
 
 Build the entire app against in-memory mocks with **zero API cost**, then swap real infrastructure one boundary at a time.

package/dist/esm/lib/context-bisect/ablation.js

@@ -0,0 +1,183 @@
+/**
+ * Ablation — the counterfactual seam (RFC-003 Part B, D8 stage 4 + the
+ * D9 stats engine).
+ *
+ * Three pieces:
+ *
+ *   1. **Adapters** — `ablationForSuspect` maps a classified suspect to
+ *      the spec that removes it (tool → drop from catalog; injection /
+ *      fact / skill → exclude the `Injection.id`; memory → filter the
+ *      `MemoryEntry.id`; arg → consumer-override note).
+ *
+ *   2. **The seam** — `applyAblations` filters the inputs an agent is
+ *      BUILT from. Documented here because the seam did not previously
+ *      exist: `AgentOptions` has no `ignoredTools` runtime kill-switch, so
+ *      tool ablation happens at construction (the consumer's
+ *      `AblationRunner` rebuilds the agent from filtered inputs). Same for
+ *      injections and memory entries.
+ *
+ *   3. **The probe engine** — `runAblationProbe` calls the consumer's
+ *      runner N seeded times, measures embedding similarity to the
+ *      original output, counts outcome flips, and returns variance —
+ *      never a single-run verdict (D9 discipline).
+ *
+ * §B2: only `runAblationProbe`-derived verdicts are causal claims; every
+ * score elsewhere is a correlational proxy.
+ */
+import { cosineSimilarity } from '../../memory/embedding/cosine.js';
+import { CONTEXT_BISECT_DEFAULTS } from './types.js';
+// ─── Adapters: suspect → spec ────────────────────────────────────────
+/**
+ * The spec that removes one suspect — or `undefined` for kind `'stage'`
+ * (plain pipeline stages have no removable input; re-rank or refactor).
+ */
+export function ablationForSuspect(suspect) {
+    switch (suspect.kind) {
+        case 'tool':
+            return suspect.detail?.toolName !== undefined
+                ? { kind: 'tool', ignoredTools: [suspect.detail.toolName] }
+                : undefined;
+        case 'injection':
+            return suspect.detail?.injectionId !== undefined
+                ? { kind: 'injection', excludeInjectionIds: [suspect.detail.injectionId] }
+                : undefined;
+        case 'memory':
+            return suspect.detail?.injectionId !== undefined
+                ? { kind: 'memory', excludeMemoryIds: [suspect.detail.injectionId] }
+                : undefined;
+        case 'arg':
+            return {
+                kind: 'arg',
+                source: suspect.source,
+                note: `step ${suspect.source} consumed untracked run input ($getArgs()/env) — ` +
+                    `the runner must override the input itself; the library cannot filter it.`,
+            };
+        case 'stage':
+            return undefined;
+    }
+}
+/**
+ * Apply ablation specs to the inputs an agent is constructed from —
+ * THE documented seam (see module docs). Generic over the concrete tool /
+ * injection / memory-entry types so it filters without importing them.
+ *
+ * `'arg'` specs are deliberately NOT handled here: run input belongs to
+ * the consumer's runner (`spec.note` says so).
+ *
+ * @example inside an AblationRunner
+ * ```ts
+ * const { tools, injections } = applyAblations(specs, {
+ *   tools: ALL_TOOLS, injections: ALL_FACTS,
+ * });
+ * const agent = Agent.create({ provider: freshProvider(), model })
+ *   .tools([...tools]);
+ * for (const inj of injections) agent.fact(inj);
+ * ```
+ */
+export function applyAblations(specs, targets) {
+    const ignoredTools = new Set();
+    const excludedInjections = new Set();
+    const excludedMemory = new Set();
+    for (const spec of specs) {
+        if (spec.kind === 'tool')
+            for (const name of spec.ignoredTools)
+                ignoredTools.add(name);
+        if (spec.kind === 'injection')
+            for (const id of spec.excludeInjectionIds)
+                excludedInjections.add(id);
+        if (spec.kind === 'memory')
+            for (const id of spec.excludeMemoryIds)
+                excludedMemory.add(id);
+    }
+    return {
+        tools: (targets.tools ?? []).filter((tool) => !ignoredTools.has(tool.schema.name)),
+        injections: (targets.injections ?? []).filter((injection) => !excludedInjections.has(injection.id)),
+        memoryEntries: (targets.memoryEntries ?? []).filter((entry) => !excludedMemory.has(entry.id)),
+    };
+}
+// ─── The probe engine (D9 stats) ─────────────────────────────────────
+function similarityStats(values) {
+    if (values.length === 0)
+        return { mean: 0, min: 0, max: 0, stdev: 0 };
+    const mean = values.reduce((sum, v) => sum + v, 0) / values.length;
+    const variance = values.reduce((sum, v) => sum + (v - mean) ** 2, 0) / values.length;
+    return {
+        mean,
+        min: Math.min(...values),
+        max: Math.max(...values),
+        stdev: Math.sqrt(variance),
+    };
+}
+/** The default comparator: embedding similarity below the threshold. */
+export function defaultOutcomeComparator(embedder, flipThreshold) {
+    return async (original, ablated) => {
+        const [a, b] = await Promise.all([
+            embedder.embed({ text: original }),
+            embedder.embed({ text: ablated }),
+        ]);
+        return cosineSimilarity(a, b) < flipThreshold;
+    };
+}
+/**
+ * Run ONE probe: call the consumer's runner with `specs` once per seed
+ * (0..samples-1), measure each output's embedding similarity to the
+ * original, and count outcome flips. Variance is always reported.
+ *
+ * `samples` is clamped to ≥ 2 — D9: never single-run verdicts.
+ */
+export async function runAblationProbe(config, specs) {
+    const samples = Math.max(2, config.rerun.samples ?? CONTEXT_BISECT_DEFAULTS.samples);
+    const flipThreshold = config.rerun.flipThreshold ?? CONTEXT_BISECT_DEFAULTS.flipThreshold;
+    const outcomeChanged = config.rerun.outcomeChanged ?? defaultOutcomeComparator(config.embedder, flipThreshold);
+    const similarities = [];
+    let flips = 0;
+    const originalVec = await config.embedder.embed({ text: config.rerun.originalOutput });
+    for (let seed = 0; seed < samples; seed++) {
+        const output = await config.rerun.runner(specs, { seed });
+        const outputVec = await config.embedder.embed({ text: output });
+        similarities.push(cosineSimilarity(originalVec, outputVec));
+        if (await outcomeChanged(config.rerun.originalOutput, output))
+            flips++;
+    }
+    return { samples, flips, similarity: similarityStats(similarities) };
+}
+/** Majority-flip rule shared by D8 verdicts and D9 probes. */
+export function probeFlipped(stats) {
+    return stats.flips * 2 > stats.samples;
+}
+/**
+ * Translate probe evidence into the verdict — the ONLY causal claim tier
+ * (§B2). `baselineStable=false` (the un-ablated scenario itself flipped)
+ * forces `'inconclusive'`: no ablation verdict is trustworthy on an
+ * unstable baseline.
+ */
+export function verdictFor(label, stats, baselineStable) {
+    if (!baselineStable) {
+        return {
+            verdict: 'inconclusive',
+            claim: `INCONCLUSIVE: the un-ablated baseline itself changed outcome across seeded reruns — ` +
+                `no ablation verdict for ${label} is trustworthy on an unstable scenario.`,
+        };
+    }
+    if (probeFlipped(stats)) {
+        return {
+            verdict: 'confirmed',
+            claim: `CAUSAL: ablating ${label} flipped the outcome in ${stats.flips}/${stats.samples} ` +
+                `seeded reruns (mean similarity to original ${stats.similarity.mean.toFixed(3)} ` +
+                `± ${stats.similarity.stdev.toFixed(3)}).`,
+        };
+    }
+    if (stats.flips > 0) {
+        return {
+            verdict: 'inconclusive',
+            claim: `INCONCLUSIVE: ablating ${label} flipped only ${stats.flips}/${stats.samples} seeded ` +
+                `reruns — below majority; raise samples or check scenario stability.`,
+        };
+    }
+    return {
+        verdict: 'not-confirmed',
+        claim: `NOT CONFIRMED: ablating ${label} did not change the outcome in ${stats.samples} seeded ` +
+            `reruns — its ranking remains a correlational proxy only.`,
+    };
+}
+//# sourceMappingURL=ablation.js.map

package/dist/esm/lib/context-bisect/ablation.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"ablation.js","sourceRoot":"","sources":["../../../../src/lib/context-bisect/ablation.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;GAyBG;AAEH,OAAO,EAAE,gBAAgB,EAAE,MAAM,kCAAkC,CAAC;AAWpE,OAAO,EAAE,uBAAuB,EAAE,MAAM,YAAY,CAAC;AAErD,wEAAwE;AAExE;;;GAGG;AACH,MAAM,UAAU,kBAAkB,CAAC,OAAgB;IACjD,QAAQ,OAAO,CAAC,IAAI,EAAE,CAAC;QACrB,KAAK,MAAM;YACT,OAAO,OAAO,CAAC,MAAM,EAAE,QAAQ,KAAK,SAAS;gBAC3C,CAAC,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,YAAY,EAAE,CAAC,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,EAAE;gBAC3D,CAAC,CAAC,SAAS,CAAC;QAChB,KAAK,WAAW;YACd,OAAO,OAAO,CAAC,MAAM,EAAE,WAAW,KAAK,SAAS;gBAC9C,CAAC,CAAC,EAAE,IAAI,EAAE,WAAW,EAAE,mBAAmB,EAAE,CAAC,OAAO,CAAC,MAAM,CAAC,WAAW,CAAC,EAAE;gBAC1E,CAAC,CAAC,SAAS,CAAC;QAChB,KAAK,QAAQ;YACX,OAAO,OAAO,CAAC,MAAM,EAAE,WAAW,KAAK,SAAS;gBAC9C,CAAC,CAAC,EAAE,IAAI,EAAE,QAAQ,EAAE,gBAAgB,EAAE,CAAC,OAAO,CAAC,MAAM,CAAC,WAAW,CAAC,EAAE;gBACpE,CAAC,CAAC,SAAS,CAAC;QAChB,KAAK,KAAK;YACR,OAAO;gBACL,IAAI,EAAE,KAAK;gBACX,MAAM,EAAE,OAAO,CAAC,MAAM;gBACtB,IAAI,EACF,QAAQ,OAAO,CAAC,MAAM,mDAAmD;oBACzE,0EAA0E;aAC7E,CAAC;QACJ,KAAK,OAAO;YACV,OAAO,SAAS,CAAC;IACrB,CAAC;AACH,CAAC;AAwBD;;;;;;;;;;;;;;;;;GAiBG;AACH,MAAM,UAAU,cAAc,CAK5B,KAA8B,EAC9B,OAAyD;IAMzD,MAAM,YAAY,GAAG,IAAI,GAAG,EAAU,CAAC;IACvC,MAAM,kBAAkB,GAAG,IAAI,GAAG,EAAU,CAAC;IAC7C,MAAM,cAAc,GAAG,IAAI,GAAG,EAAU,CAAC;IACzC,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE,CAAC;QACzB,IAAI,IAAI,CAAC,IAAI,KAAK,MAAM;YAAE,KAAK,MAAM,IAAI,IAAI,IAAI,CAAC,YAAY;gBAAE,YAAY,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;QACvF,IAAI,IAAI,CAAC,IAAI,KAAK,WAAW;YAC3B,KAAK,MAAM,EAAE,IAAI,IAAI,CAAC,mBAAmB;gBAAE,kBAAkB,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;QACxE,IAAI,IAAI,CAAC,IAAI,KAAK,QAAQ;YAAE,KAAK,MAAM,EAAE,IAAI,IAAI,CAAC,gBAAgB;gBAAE,cAAc,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;IAC7F,CAAC;IACD,OAAO;QACL,KAAK,EAAE,CAAC,OAAO,CAAC,KAAK,IAAI,EAAE,CAAC,CAAC,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,CAAC,YAAY,CAAC,GAAG,CAAC,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC;QAClF,UAAU,EAAE,CAAC,OAAO,CAAC,UAAU,IAAI,EAAE,CAAC,CAAC,MAAM,CAC3C,CAAC,SAAS,EAAE,EAAE,CAAC,CAAC,kBAAkB,CAAC,GAAG,CAAC,SAAS,CAAC,EAAE,CAAC,CACrD;QACD,aAAa,EAAE,CAAC,OAAO,CAAC,aAAa,IAAI,EAAE,CAAC,CAAC,MAAM,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC,cAAc,CAAC,GAAG,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC;KAC9F,CAAC;AACJ,CAAC;AAED,wEAAwE;AAExE,SAAS,eAAe,CAAC,MAAyB;IAChD,IAAI,MAAM,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,EAAE,IAAI,EAAE,CAAC,EAAE,GAAG,EAAE,CAAC,EAAE,GAAG,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,CAAC;IACtE,MAAM,IAAI,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,CAAC,EAAE,EAAE,CAAC,GAAG,GAAG,CAAC,EAAE,CAAC,CAAC,GAAG,MAAM,CAAC,MAAM,CAAC;IACnE,MAAM,QAAQ,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,CAAC,EAAE,EAAE,CAAC,GAAG,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC,GAAG,MAAM,CAAC,MAAM,CAAC;IACrF,OAAO;QACL,IAAI;QACJ,GAAG,EAAE,IAAI,CAAC,GAAG,CAAC,GAAG,MAAM,CAAC;QACxB,GAAG,EAAE,IAAI,CAAC,GAAG,CAAC,GAAG,MAAM,CAAC;QACxB,KAAK,EAAE,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC;KAC3B,CAAC;AACJ,CAAC;AAED,wEAAwE;AACxE,MAAM,UAAU,wBAAwB,CACtC,QAAkB,EAClB,aAAqB;IAErB,OAAO,KAAK,EAAE,QAAQ,EAAE,OAAO,EAAE,EAAE;QACjC,MAAM,CAAC,CAAC,EAAE,CAAC,CAAC,GAAG,MAAM,OAAO,CAAC,GAAG,CAAC;YAC/B,QAAQ,CAAC,KAAK,CAAC,EAAE,IAAI,EAAE,QAAQ,EAAE,CAAC;YAClC,QAAQ,CAAC,KAAK,CAAC,EAAE,IAAI,EAAE,OAAO,EAAE,CAAC;SAClC,CAAC,CAAC;QACH,OAAO,gBAAgB,CAAC,CAAC,EAAE,CAAC,CAAC,GAAG,aAAa,CAAC;IAChD,CAAC,CAAC;AACJ,CAAC;AAQD;;;;;;GAMG;AACH,MAAM,CAAC,KAAK,UAAU,gBAAgB,CACpC,MAAmB,EACnB,KAA8B;IAE9B,MAAM,OAAO,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,MAAM,CAAC,KAAK,CAAC,OAAO,IAAI,uBAAuB,CAAC,OAAO,CAAC,CAAC;IACrF,MAAM,aAAa,GAAG,MAAM,CAAC,KAAK,CAAC,aAAa,IAAI,uBAAuB,CAAC,aAAa,CAAC;IAC1F,MAAM,cAAc,GAClB,MAAM,CAAC,KAAK,CAAC,cAAc,IAAI,wBAAwB,CAAC,MAAM,CAAC,QAAQ,EAAE,aAAa,CAAC,CAAC;IAE1F,MAAM,YAAY,GAAa,EAAE,CAAC;IAClC,IAAI,KAAK,GAAG,CAAC,CAAC;IACd,MAAM,WAAW,GAAG,MAAM,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,EAAE,IAAI,EAAE,MAAM,CAAC,KAAK,CAAC,cAAc,EAAE,CAAC,CAAC;IACvF,KAAK,IAAI,IAAI,GAAG,CAAC,EAAE,IAAI,GAAG,OAAO,EAAE,IAAI,EAAE,EAAE,CAAC;QAC1C,MAAM,MAAM,GAAG,MAAM,MAAM,CAAC,KAAK,CAAC,MAAM,CAAC,KAAK,EAAE,EAAE,IAAI,EAAE,CAAC,CAAC;QAC1D,MAAM,SAAS,GAAG,MAAM,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,CAAC,CAAC;QAChE,YAAY,CAAC,IAAI,CAAC,gBAAgB,CAAC,WAAW,EAAE,SAAS,CAAC,CAAC,CAAC;QAC5D,IAAI,MAAM,cAAc,CAAC,MAAM,CAAC,KAAK,CAAC,cAAc,EAAE,MAAM,CAAC;YAAE,KAAK,EAAE,CAAC;IACzE,CAAC;IACD,OAAO,EAAE,OAAO,EAAE,KAAK,EAAE,UAAU,EAAE,eAAe,CAAC,YAAY,CAAC,EAAE,CAAC;AACvE,CAAC;AAED,8DAA8D;AAC9D,MAAM,UAAU,YAAY,CAAC,KAAuB;IAClD,OAAO,KAAK,CAAC,KAAK,GAAG,CAAC,GAAG,KAAK,CAAC,OAAO,CAAC;AACzC,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,UAAU,CACxB,KAAa,EACb,KAAuB,EACvB,cAAuB;IAEvB,IAAI,CAAC,cAAc,EAAE,CAAC;QACpB,OAAO;YACL,OAAO,EAAE,cAAc;YACvB,KAAK,EACH,sFAAsF;gBACtF,2BAA2B,KAAK,0CAA0C;SAC7E,CAAC;IACJ,CAAC;IACD,IAAI,YAAY,CAAC,KAAK,CAAC,EAAE,CAAC;QACxB,OAAO;YACL,OAAO,EAAE,WAAW;YACpB,KAAK,EACH,oBAAoB,KAAK,2BAA2B,KAAK,CAAC,KAAK,IAAI,KAAK,CAAC,OAAO,GAAG;gBACnF,8CAA8C,KAAK,CAAC,UAAU,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG;gBACjF,KAAK,KAAK,CAAC,UAAU,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,IAAI;SAC7C,CAAC;IACJ,CAAC;IACD,IAAI,KAAK,CAAC,KAAK,GAAG,CAAC,EAAE,CAAC;QACpB,OAAO;YACL,OAAO,EAAE,cAAc;YACvB,KAAK,EACH,0BAA0B,KAAK,iBAAiB,KAAK,CAAC,KAAK,IAAI,KAAK,CAAC,OAAO,UAAU;gBACtF,qEAAqE;SACxE,CAAC;IACJ,CAAC;IACD,OAAO;QACL,OAAO,EAAE,eAAe;QACxB,KAAK,EACH,2BAA2B,KAAK,kCAAkC,KAAK,CAAC,OAAO,UAAU;YACzF,0DAA0D;KAC7D,CAAC;AACJ,CAAC"}

package/dist/esm/lib/context-bisect/bisect.js

@@ -0,0 +1,129 @@
+/**
+ * bisectCulprits — multi-culprit bisection over the ranked suspect set
+ * (RFC-003 Part B, block D9). The "git bisect" of the localizer.
+ *
+ * When single-suspect ablations don't flip the outcome — redundant causes
+ * (two facts that EACH justify the wrong answer), or interacting ones —
+ * the culprit is a SET. This harness finds a minimal culprit set by
+ * recursive halving over the ranked suspects (delta-debugging style,
+ * Zeller's ddmin specialized to two-way splits), then keeps searching the
+ * remainder for INDEPENDENT culprits until the remainder stops flipping.
+ *
+ * Probe semantics (the D9 discipline):
+ *  - every probe = N seeded reruns of the consumer's `AblationRunner`
+ *    with the probe's combined specs; "flipped" = MAJORITY of runs
+ *    changed outcome; similarity mean ± spread is always reported —
+ *    never single-run verdicts;
+ *  - probe 0 is the BASELINE (no ablation): if it flips, the scenario
+ *    itself is unstable and the result is honestly `'inconclusive'`;
+ *  - probes are cached by spec-set, and budgeted (`maxProbes`) — running
+ *    out of budget yields `'inconclusive'`, never a partial claim
+ *    dressed up as a finding.
+ *
+ * §B2: the returned `verdict`/`culprits` are CAUSAL claims — they rest
+ * exclusively on counterfactual reruns. The input ranking only chooses
+ * the SEARCH ORDER (better ranking = fewer probes), it never decides the
+ * outcome.
+ */
+import { probeFlipped, runAblationProbe } from './ablation.js';
+import { CONTEXT_BISECT_DEFAULTS } from './types.js';
+import { suspectLabel } from './localize.js';
+// ─── The harness ─────────────────────────────────────────────────────
+class ProbeBudgetExceeded extends Error {
+    constructor() {
+        super('probe budget exceeded');
+    }
+}
+/**
+ * Find minimal culprit set(s) by seeded counterfactual bisection. See
+ * module docs for semantics and the §B2 claim tier.
+ */
+export async function bisectCulprits(options) {
+    const candidates = options.suspects.filter((suspect) => suspect.ablation !== undefined && suspect.ablation.kind !== 'arg');
+    const maxProbes = options.maxProbes ?? CONTEXT_BISECT_DEFAULTS.maxProbes;
+    const maxCulprits = options.maxCulprits ?? CONTEXT_BISECT_DEFAULTS.maxCulprits;
+    const config = { rerun: options.rerun, embedder: options.embedder };
+    const probes = [];
+    const cache = new Map();
+    let runsUsed = 0;
+    const keyOf = (set) => set
+        .map((suspect) => suspectLabel(suspect))
+        .sort()
+        .join('|');
+    async function probe(set) {
+        const key = keyOf(set);
+        const cached = cache.get(key);
+        if (cached !== undefined)
+            return cached;
+        if (probes.length >= maxProbes)
+            throw new ProbeBudgetExceeded();
+        const stats = await runAblationProbe(config, set.flatMap((suspect) => (suspect.ablation !== undefined ? [suspect.ablation] : [])));
+        runsUsed += stats.samples;
+        const flipped = probeFlipped(stats);
+        probes.push({ ablated: set.map((suspect) => suspectLabel(suspect)), stats, flipped });
+        cache.set(key, flipped);
+        return flipped;
+    }
+    /**
+     * Minimal subset of `candidates` that — together with `context` — flips
+     * the outcome. Precondition: probe(candidates ∪ context) flipped.
+     * Two-way ddmin: try each half; on interference (neither half alone
+     * suffices) minimize each half with the other as context.
+     */
+    async function minimize(set, context) {
+        if (set.length <= 1)
+            return [...set];
+        const mid = Math.ceil(set.length / 2);
+        const top = set.slice(0, mid); // ranked order: the likelier half first
+        const rest = set.slice(mid);
+        if (await probe([...top, ...context]))
+            return minimize(top, context);
+        if (await probe([...rest, ...context]))
+            return minimize(rest, context);
+        // Interference: parts of BOTH halves are needed jointly.
+        const fromTop = await minimize(top, [...rest, ...context]);
+        const fromRest = await minimize(rest, [...fromTop, ...context]);
+        return [...fromTop, ...fromRest];
+    }
+    try {
+        // Baseline: an unstable scenario invalidates everything downstream.
+        // ZERO-TOLERANCE (review Finding 1): a single un-ablated flip marks the
+        // scenario unstable — the majority-rule probeFlipped() gate would let a
+        // 1-in-3-flaky scenario through to a 'confirmed' CAUSAL verdict, which
+        // violates the §B2 honest-claims discipline. Same gate localize.ts uses.
+        {
+            const baselineStats = await runAblationProbe(config, []);
+            runsUsed += baselineStats.samples;
+            const unstable = baselineStats.flips > 0;
+            probes.push({ ablated: [], stats: baselineStats, flipped: unstable });
+            cache.set(keyOf([]), probeFlipped(baselineStats));
+            if (unstable) {
+                return { verdict: 'inconclusive', culprits: [], probes, runsUsed };
+            }
+        }
+        // Reproduction gate: the full ranked set must flip at all.
+        if (candidates.length === 0 || !(await probe(candidates))) {
+            return { verdict: 'not-reproducible', culprits: [], probes, runsUsed };
+        }
+        // Find minimal sets; then keep searching the remainder for
+        // INDEPENDENT culprits until it stops flipping.
+        const culprits = [];
+        let remaining = candidates;
+        for (let round = 0; round < maxCulprits; round++) {
+            const found = await minimize(remaining, []);
+            culprits.push(found);
+            const foundKeys = new Set(found.map((suspect) => suspectLabel(suspect)));
+            remaining = remaining.filter((suspect) => !foundKeys.has(suspectLabel(suspect)));
+            if (remaining.length === 0 || !(await probe(remaining)))
+                break;
+        }
+        return { verdict: 'confirmed', culprits, probes, runsUsed };
+    }
+    catch (error) {
+        if (error instanceof ProbeBudgetExceeded) {
+            return { verdict: 'inconclusive', culprits: [], probes, runsUsed };
+        }
+        throw error;
+    }
+}
+//# sourceMappingURL=bisect.js.map

package/dist/esm/lib/context-bisect/bisect.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"bisect.js","sourceRoot":"","sources":["../../../../src/lib/context-bisect/bisect.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;GA0BG;AAEH,OAAO,EAAE,YAAY,EAAE,gBAAgB,EAAoB,MAAM,eAAe,CAAC;AAEjF,OAAO,EAAE,uBAAuB,EAAE,MAAM,YAAY,CAAC;AACrD,OAAO,EAAE,YAAY,EAAE,MAAM,eAAe,CAAC;AAiD7C,wEAAwE;AAExE,MAAM,mBAAoB,SAAQ,KAAK;IACrC;QACE,KAAK,CAAC,uBAAuB,CAAC,CAAC;IACjC,CAAC;CACF;AAED;;;GAGG;AACH,MAAM,CAAC,KAAK,UAAU,cAAc,CAAC,OAA8B;IACjE,MAAM,UAAU,GAAG,OAAO,CAAC,QAAQ,CAAC,MAAM,CACxC,CAAC,OAAO,EAAE,EAAE,CAAC,OAAO,CAAC,QAAQ,KAAK,SAAS,IAAI,OAAO,CAAC,QAAQ,CAAC,IAAI,KAAK,KAAK,CAC/E,CAAC;IACF,MAAM,SAAS,GAAG,OAAO,CAAC,SAAS,IAAI,uBAAuB,CAAC,SAAS,CAAC;IACzE,MAAM,WAAW,GAAG,OAAO,CAAC,WAAW,IAAI,uBAAuB,CAAC,WAAW,CAAC;IAC/E,MAAM,MAAM,GAAgB,EAAE,KAAK,EAAE,OAAO,CAAC,KAAK,EAAE,QAAQ,EAAE,OAAO,CAAC,QAAQ,EAAE,CAAC;IAEjF,MAAM,MAAM,GAAqB,EAAE,CAAC;IACpC,MAAM,KAAK,GAAG,IAAI,GAAG,EAAmB,CAAC;IACzC,IAAI,QAAQ,GAAG,CAAC,CAAC;IAEjB,MAAM,KAAK,GAAG,CAAC,GAAuB,EAAU,EAAE,CAChD,GAAG;SACA,GAAG,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,YAAY,CAAC,OAAO,CAAC,CAAC;SACvC,IAAI,EAAE;SACN,IAAI,CAAC,GAAG,CAAC,CAAC;IAEf,KAAK,UAAU,KAAK,CAAC,GAAuB;QAC1C,MAAM,GAAG,GAAG,KAAK,CAAC,GAAG,CAAC,CAAC;QACvB,MAAM,MAAM,GAAG,KAAK,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;QAC9B,IAAI,MAAM,KAAK,SAAS;YAAE,OAAO,MAAM,CAAC;QACxC,IAAI,MAAM,CAAC,MAAM,IAAI,SAAS;YAAE,MAAM,IAAI,mBAAmB,EAAE,CAAC;QAChE,MAAM,KAAK,GAAG,MAAM,gBAAgB,CAClC,MAAM,EACN,GAAG,CAAC,OAAO,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,CAAC,OAAO,CAAC,QAAQ,KAAK,SAAS,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CACrF,CAAC;QACF,QAAQ,IAAI,KAAK,CAAC,OAAO,CAAC;QAC1B,MAAM,OAAO,GAAG,YAAY,CAAC,KAAK,CAAC,CAAC;QACpC,MAAM,CAAC,IAAI,CAAC,EAAE,OAAO,EAAE,GAAG,CAAC,GAAG,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,YAAY,CAAC,OAAO,CAAC,CAAC,EAAE,KAAK,EAAE,OAAO,EAAE,CAAC,CAAC;QACtF,KAAK,CAAC,GAAG,CAAC,GAAG,EAAE,OAAO,CAAC,CAAC;QACxB,OAAO,OAAO,CAAC;IACjB,CAAC;IAED;;;;;OAKG;IACH,KAAK,UAAU,QAAQ,CACrB,GAAuB,EACvB,OAA2B;QAE3B,IAAI,GAAG,CAAC,MAAM,IAAI,CAAC;YAAE,OAAO,CAAC,GAAG,GAAG,CAAC,CAAC;QACrC,MAAM,GAAG,GAAG,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;QACtC,MAAM,GAAG,GAAG,GAAG,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC,CAAC,wCAAwC;QACvE,MAAM,IAAI,GAAG,GAAG,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC;QAC5B,IAAI,MAAM,KAAK,CAAC,CAAC,GAAG,GAAG,EAAE,GAAG,OAAO,CAAC,CAAC;YAAE,OAAO,QAAQ,CAAC,GAAG,EAAE,OAAO,CAAC,CAAC;QACrE,IAAI,MAAM,KAAK,CAAC,CAAC,GAAG,IAAI,EAAE,GAAG,OAAO,CAAC,CAAC;YAAE,OAAO,QAAQ,CAAC,IAAI,EAAE,OAAO,CAAC,CAAC;QACvE,yDAAyD;QACzD,MAAM,OAAO,GAAG,MAAM,QAAQ,CAAC,GAAG,EAAE,CAAC,GAAG,IAAI,EAAE,GAAG,OAAO,CAAC,CAAC,CAAC;QAC3D,MAAM,QAAQ,GAAG,MAAM,QAAQ,CAAC,IAAI,EAAE,CAAC,GAAG,OAAO,EAAE,GAAG,OAAO,CAAC,CAAC,CAAC;QAChE,OAAO,CAAC,GAAG,OAAO,EAAE,GAAG,QAAQ,CAAC,CAAC;IACnC,CAAC;IAED,IAAI,CAAC;QACH,oEAAoE;QACpE,wEAAwE;QACxE,wEAAwE;QACxE,uEAAuE;QACvE,yEAAyE;QACzE,CAAC;YACC,MAAM,aAAa,GAAG,MAAM,gBAAgB,CAAC,MAAM,EAAE,EAAE,CAAC,CAAC;YACzD,QAAQ,IAAI,aAAa,CAAC,OAAO,CAAC;YAClC,MAAM,QAAQ,GAAG,aAAa,CAAC,KAAK,GAAG,CAAC,CAAC;YACzC,MAAM,CAAC,IAAI,CAAC,EAAE,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,aAAa,EAAE,OAAO,EAAE,QAAQ,EAAE,CAAC,CAAC;YACtE,KAAK,CAAC,GAAG,CAAC,KAAK,CAAC,EAAE,CAAC,EAAE,YAAY,CAAC,aAAa,CAAC,CAAC,CAAC;YAClD,IAAI,QAAQ,EAAE,CAAC;gBACb,OAAO,EAAE,OAAO,EAAE,cAAc,EAAE,QAAQ,EAAE,EAAE,EAAE,MAAM,EAAE,QAAQ,EAAE,CAAC;YACrE,CAAC;QACH,CAAC;QACD,2DAA2D;QAC3D,IAAI,UAAU,CAAC,MAAM,KAAK,CAAC,IAAI,CAAC,CAAC,MAAM,KAAK,CAAC,UAAU,CAAC,CAAC,EAAE,CAAC;YAC1D,OAAO,EAAE,OAAO,EAAE,kBAAkB,EAAE,QAAQ,EAAE,EAAE,EAAE,MAAM,EAAE,QAAQ,EAAE,CAAC;QACzE,CAAC;QAED,2DAA2D;QAC3D,gDAAgD;QAChD,MAAM,QAAQ,GAAgB,EAAE,CAAC;QACjC,IAAI,SAAS,GAAG,UAAU,CAAC;QAC3B,KAAK,IAAI,KAAK,GAAG,CAAC,EAAE,KAAK,GAAG,WAAW,EAAE,KAAK,EAAE,EAAE,CAAC;YACjD,MAAM,KAAK,GAAG,MAAM,QAAQ,CAAC,SAAS,EAAE,EAAE,CAAC,CAAC;YAC5C,QAAQ,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;YACrB,MAAM,SAAS,GAAG,IAAI,GAAG,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,YAAY,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC;YACzE,SAAS,GAAG,SAAS,CAAC,MAAM,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,CAAC,SAAS,CAAC,GAAG,CAAC,YAAY,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC;YACjF,IAAI,SAAS,CAAC,MAAM,KAAK,CAAC,IAAI,CAAC,CAAC,MAAM,KAAK,CAAC,SAAS,CAAC,CAAC;gBAAE,MAAM;QACjE,CAAC;QACD,OAAO,EAAE,OAAO,EAAE,WAAW,EAAE,QAAQ,EAAE,MAAM,EAAE,QAAQ,EAAE,CAAC;IAC9D,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QACf,IAAI,KAAK,YAAY,mBAAmB,EAAE,CAAC;YACzC,OAAO,EAAE,OAAO,EAAE,cAAc,EAAE,QAAQ,EAAE,EAAE,EAAE,MAAM,EAAE,QAAQ,EAAE,CAAC;QACrE,CAAC;QACD,MAAM,KAAK,CAAC;IACd,CAAC;AACH,CAAC"}

package/dist/esm/lib/context-bisect/index.js

@@ -0,0 +1,22 @@
+/**
+ * context-bisect — RFC-003 Part B: the contextual-bug LOCALIZER,
+ * "git bisect for context".
+ *
+ * Assembly over shipped pieces: footprintjs 9.8.0's complete causal DAG
+ * (control edges, honesty markers, `EdgeWeigher` hook) × influence-core
+ * scoring (D6) × consumer-run counterfactual ablation.
+ *
+ *   D7 — `llmEdgeWeigher`     influence-weighted LLM-call slice edges
+ *   D8 — `localizeContextBug` trigger → slice → ranked suspects → ablation
+ *   D9 — `bisectCulprits`     seeded multi-culprit bisection + variance
+ *
+ * §B2 claim tiers (spelled out on every type): weights/scores are
+ * embedding-geometry PROXIES; ablation verdicts are the ONLY causal
+ * claims; slice completeness is bounded by tracking — and says so.
+ */
+export { llmEdgeWeigher, stepOutputText, } from './llmEdgeWeigher.js';
+export { defaultSuspectClassifier, formatContextBugReport, llmCallIdsFromEvents, localizeContextBug, suspectLabel, } from './localize.js';
+export { ablationForSuspect, applyAblations, defaultOutcomeComparator, probeFlipped, runAblationProbe, verdictFor, } from './ablation.js';
+export { bisectCulprits, } from './bisect.js';
+export { CONTEXT_BISECT_DEFAULTS, } from './types.js';
+//# sourceMappingURL=index.js.map

package/dist/esm/lib/context-bisect/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../../../src/lib/context-bisect/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAEH,OAAO,EACL,cAAc,EACd,cAAc,GAIf,MAAM,qBAAqB,CAAC;AAE7B,OAAO,EACL,wBAAwB,EACxB,sBAAsB,EACtB,oBAAoB,EACpB,kBAAkB,EAClB,YAAY,GAKb,MAAM,eAAe,CAAC;AAEvB,OAAO,EACL,kBAAkB,EAClB,cAAc,EACd,wBAAwB,EACxB,YAAY,EACZ,gBAAgB,EAChB,UAAU,GAGX,MAAM,eAAe,CAAC;AAEvB,OAAO,EACL,cAAc,GAIf,MAAM,aAAa,CAAC;AAErB,OAAO,EACL,uBAAuB,GAoBxB,MAAM,YAAY,CAAC"}