feat(gsd): implement ADR-005 multi-model provider and tool strategy

Implements all 4 phases of ADR-005 (issue #2790): Phase 1: Provider Capabilities Registry - Declarative ProviderCapabilities interface and PROVIDER_CAPABILITIES registry covering all 12 API providers - Consolidates scattered *-shared.ts knowledge into queryable registry - Unknown providers get permissive defaults (backward compatible) Phase 2: Tool Compatibility Metadata - ToolCompatibility interface (producesImages, schemaFeatures, minCapabilityTier) - compatibility field on ToolDefinition - Tool compatibility registry with pre-populated built-in tools - Auto-registration from registerTool() and MCP tool defaults Phase 3: Tool-Compat Filter + ProviderSwitchReport - ProviderSwitchReport tracks thinking blocks dropped/downgraded, tool call IDs remapped, synthetic results inserted, thought signatures dropped during cross-provider message transformation - isToolCompatibleWithProvider(), filterToolsForProvider(), adjustToolSet() functions in model router - filteredTools field on RoutingDecision - Verbose output for filtered tools in auto-model-selection Phase 4: adjustToolSet Extension Hook - AdjustToolSetEvent and AdjustToolSetResult interfaces - emitAdjustToolSet() on ExtensionAPI and ExtensionRuntime - Default no-op handler in register-hooks.ts Includes 47 new tests (20 provider caps + 10 switch report + 17 tool compat) Closes #2790
2026-04-10 12:33:40 -05:00 · 2026-04-10 12:33:40 -05:00 · b1c0dafc70
commit b1c0dafc70
parent 9016ac362f
17 changed files with 1173 additions and 2 deletions
--- a/docs/dev/ADR-005-multi-model-provider-tool-strategy.md
+++ b/docs/dev/ADR-005-multi-model-provider-tool-strategy.md
@ -0,0 +1,67 @@
+# ADR-005: Multi-Model, Multi-Provider, and Tool Strategy
+
+**Status:** Accepted
+**Date:** 2026-03-27
+**Deciders:** Jeremy McSpadden
+**Related:** ADR-004 (capability-aware model routing), ADR-003 (pipeline simplification), [Issue #2790](https://github.com/gsd-build/gsd-2/issues/2790)
+
+## Context
+
+PR #2755 lands capability-aware model routing (ADR-004), extending the router from a one-dimensional complexity-tier system to a two-dimensional system that scores models across 7 capability dimensions. GSD can now intelligently pick the best model for a task from a heterogeneous pool.
+
+But model selection is only one piece of the multi-model puzzle. The system faces structural gaps as users configure diverse provider pools:
+
+1. **Tool compatibility is assumed, not verified** — Every registered tool is sent to every model regardless of provider capabilities.
+2. **No tool-aware model routing** — ADR-004 scores 7 capability dimensions but none encode whether a model can actually use the tools a task requires.
+3. **Provider failover loses context fidelity** — Cross-provider switches silently degrade conversation quality (thinking blocks dropped, tool IDs remapped).
+4. **Tool availability is static across a session** — The same tools are presented regardless of the selected model's capabilities.
+5. **No provider capability registry** — Provider quirks are scattered across `*-shared.ts` files.
+
+## Decision
+
+Introduce a provider capability registry and tool compatibility layer that integrates with ADR-004's capability-aware model router.
+
+### Design Principles
+
+1. **Layered on ADR-004, not replacing it.** Capability scoring remains primary. This adds tool compatibility as a hard constraint.
+2. **Hard constraints filter; soft scores rank.** Tool support is binary — it filters the eligible set before scoring.
+3. **Provider knowledge is declarative, not scattered.** Provider capabilities move to an explicit registry.
+4. **Tool sets adapt to model capabilities.** Active tool set adjusts when the router selects a different model.
+5. **Graceful degradation preserved.** Unknown providers get full tool access — same as today.
+
+### Implementation Phases
+
+1. **Phase 1:** Provider Capabilities Registry (`packages/pi-ai/src/providers/provider-capabilities.ts`)
+2. **Phase 2:** Tool Compatibility Metadata (extend `ToolDefinition` with `compatibility` field)
+3. **Phase 3:** Tool-compatibility filter in routing pipeline + `ProviderSwitchReport` in `transform-messages.ts`
+4. **Phase 4:** `adjustToolSet` extension hook
+
+## Consequences
+
+### Positive
+- Eliminates silent tool failures when routing to incompatible providers
+- Makes cross-provider routing safe by default
+- Provider knowledge becomes queryable (registry vs scattered code)
+- Cross-provider context loss becomes visible via `ProviderSwitchReport`
+
+### Negative
+- More metadata to maintain (provider capabilities, tool compatibility)
+- Tool filtering adds a pipeline step (sub-millisecond, O(models × tools))
+- Risk of over-filtering (mitigated: opt-in per tool, permissive defaults)
+
+### Neutral
+- Existing behavior unchanged without metadata
+- ADR-004 scoring is unmodified
+- Provider implementations simplify over time as registry replaces scattered workarounds
+
+## Appendix: Architecture Reference
+
+| File | Role |
+|------|------|
+| `packages/pi-ai/src/providers/register-builtins.ts` | Provider registration |
+| `packages/pi-ai/src/providers/*-shared.ts` | Provider-specific handling |
+| `packages/pi-ai/src/providers/transform-messages.ts` | Cross-provider normalization |
+| `packages/pi-ai/src/types.ts` | Core types |
+| `packages/pi-coding-agent/src/core/extensions/types.ts` | ToolDefinition, ExtensionAPI |
+| `src/resources/extensions/gsd/model-router.ts` | Capability scoring (ADR-004) |
+| `src/resources/extensions/gsd/auto-model-selection.ts` | Model selection orchestration |
--- a/packages/pi-ai/src/index.ts
+++ b/packages/pi-ai/src/index.ts
@ -12,7 +12,10 @@ export * from "./providers/google-vertex.js";
 export * from "./providers/mistral.js";
 export * from "./providers/openai-completions.js";
 export * from "./providers/openai-responses.js";
+export * from "./providers/provider-capabilities.js";
 export * from "./providers/register-builtins.js";
+export type { ProviderSwitchReport } from "./providers/transform-messages.js";
+export { createEmptyReport, hasTransformations } from "./providers/transform-messages.js";
 export * from "./stream.js";
 export * from "./types.js";
 export * from "./utils/event-stream.js";
--- a/packages/pi-ai/src/providers/provider-capabilities.test.ts
+++ b/packages/pi-ai/src/providers/provider-capabilities.test.ts
@ -0,0 +1,174 @@
+// GSD-2 — Provider Capabilities Registry Tests (ADR-005 Phase 1)
+import { describe, test } from "node:test";
+import assert from "node:assert/strict";
+
+import {
+  PROVIDER_CAPABILITIES,
+  getProviderCapabilities,
+  getUnsupportedFeatures,
+  mergeCapabilityOverrides,
+  getRegisteredApis,
+} from "./provider-capabilities.js";
+
+// ─── Registry Completeness ──────────────────────────────────────────────────
+
+describe("PROVIDER_CAPABILITIES registry", () => {
+  const EXPECTED_APIS = [
+    "anthropic-messages",
+    "anthropic-vertex",
+    "openai-responses",
+    "azure-openai-responses",
+    "openai-codex-responses",
+    "openai-completions",
+    "google-generative-ai",
+    "google-gemini-cli",
+    "google-vertex",
+    "mistral-conversations",
+    "bedrock-converse-stream",
+    "ollama-chat",
+  ];
+
+  test("covers all expected API providers", () => {
+    for (const api of EXPECTED_APIS) {
+      assert.ok(
+        PROVIDER_CAPABILITIES[api],
+        `Missing capability entry for API: ${api}`,
+      );
+    }
+  });
+
+  test("getRegisteredApis returns all entries", () => {
+    const registered = getRegisteredApis();
+    for (const api of EXPECTED_APIS) {
+      assert.ok(registered.includes(api), `getRegisteredApis missing: ${api}`);
+    }
+  });
+
+  test("all entries have required fields", () => {
+    for (const [api, caps] of Object.entries(PROVIDER_CAPABILITIES)) {
+      assert.equal(typeof caps.toolCalling, "boolean", `${api}.toolCalling`);
+      assert.equal(typeof caps.maxTools, "number", `${api}.maxTools`);
+      assert.equal(typeof caps.imageToolResults, "boolean", `${api}.imageToolResults`);
+      assert.equal(typeof caps.structuredOutput, "boolean", `${api}.structuredOutput`);
+      assert.ok(caps.toolCallIdFormat, `${api}.toolCallIdFormat`);
+      assert.equal(typeof caps.toolCallIdFormat.maxLength, "number", `${api}.toolCallIdFormat.maxLength`);
+      assert.ok(caps.toolCallIdFormat.allowedChars instanceof RegExp, `${api}.toolCallIdFormat.allowedChars`);
+      assert.ok(
+        ["full", "text-only", "none"].includes(caps.thinkingPersistence),
+        `${api}.thinkingPersistence is "${caps.thinkingPersistence}"`,
+      );
+      assert.ok(Array.isArray(caps.unsupportedSchemaFeatures), `${api}.unsupportedSchemaFeatures`);
+    }
+  });
+});
+
+// ─── Provider-specific Values ───────────────────────────────────────────────
+
+describe("provider-specific capabilities", () => {
+  test("Anthropic supports full thinking persistence", () => {
+    assert.equal(PROVIDER_CAPABILITIES["anthropic-messages"].thinkingPersistence, "full");
+  });
+
+  test("Anthropic supports image tool results", () => {
+    assert.equal(PROVIDER_CAPABILITIES["anthropic-messages"].imageToolResults, true);
+  });
+
+  test("Anthropic tool call ID is 64 chars max", () => {
+    assert.equal(PROVIDER_CAPABILITIES["anthropic-messages"].toolCallIdFormat.maxLength, 64);
+  });
+
+  test("Mistral tool call ID is 9 chars max", () => {
+    assert.equal(PROVIDER_CAPABILITIES["mistral-conversations"].toolCallIdFormat.maxLength, 9);
+  });
+
+  test("Mistral has no thinking persistence", () => {
+    assert.equal(PROVIDER_CAPABILITIES["mistral-conversations"].thinkingPersistence, "none");
+  });
+
+  test("Google does not support patternProperties", () => {
+    assert.ok(
+      PROVIDER_CAPABILITIES["google-generative-ai"].unsupportedSchemaFeatures.includes("patternProperties"),
+    );
+  });
+
+  test("Google does not support const", () => {
+    assert.ok(
+      PROVIDER_CAPABILITIES["google-generative-ai"].unsupportedSchemaFeatures.includes("const"),
+    );
+  });
+
+  test("OpenAI Responses does not support image tool results", () => {
+    assert.equal(PROVIDER_CAPABILITIES["openai-responses"].imageToolResults, false);
+  });
+
+  test("OpenAI Responses has text-only thinking persistence", () => {
+    assert.equal(PROVIDER_CAPABILITIES["openai-responses"].thinkingPersistence, "text-only");
+  });
+});
+
+// ─── getProviderCapabilities ────────────────────────────────────────────────
+
+describe("getProviderCapabilities", () => {
+  test("returns known provider capabilities", () => {
+    const caps = getProviderCapabilities("anthropic-messages");
+    assert.equal(caps.toolCalling, true);
+    assert.equal(caps.thinkingPersistence, "full");
+  });
+
+  test("returns permissive defaults for unknown providers", () => {
+    const caps = getProviderCapabilities("unknown-provider-xyz");
+    assert.equal(caps.toolCalling, true);
+    assert.equal(caps.imageToolResults, true);
+    assert.deepEqual(caps.unsupportedSchemaFeatures, []);
+  });
+});
+
+// ─── getUnsupportedFeatures ─────────────────────────────────────────────────
+
+describe("getUnsupportedFeatures", () => {
+  test("returns unsupported features for Google", () => {
+    const unsupported = getUnsupportedFeatures("google-generative-ai", ["patternProperties", "const"]);
+    assert.deepEqual(unsupported, ["patternProperties", "const"]);
+  });
+
+  test("returns empty for Anthropic with any features", () => {
+    const unsupported = getUnsupportedFeatures("anthropic-messages", ["patternProperties", "const"]);
+    assert.deepEqual(unsupported, []);
+  });
+
+  test("returns empty for unknown provider", () => {
+    const unsupported = getUnsupportedFeatures("unknown-xyz", ["patternProperties"]);
+    assert.deepEqual(unsupported, []);
+  });
+});
+
+// ─── mergeCapabilityOverrides ───────────────────────────────────────────────
+
+describe("mergeCapabilityOverrides", () => {
+  test("overrides individual fields", () => {
+    const merged = mergeCapabilityOverrides("openai-responses", {
+      imageToolResults: true,
+    });
+    assert.equal(merged.imageToolResults, true);
+    // Non-overridden fields preserved
+    assert.equal(merged.toolCalling, true);
+    assert.equal(merged.thinkingPersistence, "text-only");
+  });
+
+  test("deep-merges toolCallIdFormat", () => {
+    const merged = mergeCapabilityOverrides("anthropic-messages", {
+      toolCallIdFormat: { maxLength: 128 },
+    });
+    assert.equal(merged.toolCallIdFormat.maxLength, 128);
+    // allowedChars preserved from base
+    assert.ok(merged.toolCallIdFormat.allowedChars instanceof RegExp);
+  });
+
+  test("uses permissive defaults for unknown provider", () => {
+    const merged = mergeCapabilityOverrides("unknown-xyz", {
+      imageToolResults: false,
+    });
+    assert.equal(merged.imageToolResults, false);
+    assert.equal(merged.toolCalling, true); // from default
+  });
+});
--- a/packages/pi-ai/src/providers/provider-capabilities.ts
+++ b/packages/pi-ai/src/providers/provider-capabilities.ts
@ -0,0 +1,215 @@
+// GSD-2 — Provider Capabilities Registry (ADR-005 Phase 1)
+// Declarative registry of what each API provider supports, consolidating
+// scattered knowledge from *-shared.ts files into a queryable data structure.
+
+import type { Api } from "../types.js";
+
+// ─── Types ──────────────────────────────────────────────────────────────────
+
+/**
+ * Declarative capability profile for an API provider.
+ * Used by the model router to filter incompatible models and by the tool
+ * system to adjust tool sets per provider.
+ */
+export interface ProviderCapabilities {
+  /** Whether models from this provider support tool/function calling */
+  toolCalling: boolean;
+  /** Maximum number of tools the provider handles well (0 = unlimited) */
+  maxTools: number;
+  /** Whether tool results can contain images */
+  imageToolResults: boolean;
+  /** Whether the provider supports structured JSON output */
+  structuredOutput: boolean;
+  /** Tool call ID format constraints */
+  toolCallIdFormat: {
+    maxLength: number;
+    allowedChars: RegExp;
+  };
+  /** Whether thinking/reasoning blocks are preserved cross-turn */
+  thinkingPersistence: "full" | "text-only" | "none";
+  /** Schema features NOT supported (tools using these get filtered) */
+  unsupportedSchemaFeatures: string[];
+}
+
+// ─── Registry ───────────────────────────────────────────────────────────────
+
+/**
+ * Built-in provider capability profiles.
+ *
+ * Sources (consolidated from scattered *-shared.ts files):
+ * - anthropic-shared.ts: normalizeToolCallId (64-char, [a-zA-Z0-9_-])
+ * - openai-responses-shared.ts: ID normalization (64-char, fc_ prefix), image-in-tool-result workaround
+ * - google-shared.ts: sanitizeSchemaForGoogle (patternProperties, const), requiresToolCallId
+ * - mistral.ts: MISTRAL_TOOL_CALL_ID_LENGTH = 9
+ * - amazon-bedrock.ts: normalizeToolCallId (64-char, [a-zA-Z0-9_-])
+ */
+export const PROVIDER_CAPABILITIES: Record<string, ProviderCapabilities> = {
+  "anthropic-messages": {
+    toolCalling: true,
+    maxTools: 0,
+    imageToolResults: true,
+    structuredOutput: true,
+    toolCallIdFormat: { maxLength: 64, allowedChars: /^[a-zA-Z0-9_-]+$/ },
+    thinkingPersistence: "full",
+    unsupportedSchemaFeatures: [],
+  },
+  "anthropic-vertex": {
+    toolCalling: true,
+    maxTools: 0,
+    imageToolResults: true,
+    structuredOutput: true,
+    toolCallIdFormat: { maxLength: 64, allowedChars: /^[a-zA-Z0-9_-]+$/ },
+    thinkingPersistence: "full",
+    unsupportedSchemaFeatures: [],
+  },
+  "openai-responses": {
+    toolCalling: true,
+    maxTools: 0,
+    imageToolResults: false,  // images sent as separate user message, not in tool result
+    structuredOutput: true,
+    toolCallIdFormat: { maxLength: 512, allowedChars: /^.+$/ },
+    thinkingPersistence: "text-only",
+    unsupportedSchemaFeatures: [],
+  },
+  "azure-openai-responses": {
+    toolCalling: true,
+    maxTools: 0,
+    imageToolResults: false,
+    structuredOutput: true,
+    toolCallIdFormat: { maxLength: 512, allowedChars: /^.+$/ },
+    thinkingPersistence: "text-only",
+    unsupportedSchemaFeatures: [],
+  },
+  "openai-codex-responses": {
+    toolCalling: true,
+    maxTools: 0,
+    imageToolResults: false,
+    structuredOutput: true,
+    toolCallIdFormat: { maxLength: 64, allowedChars: /^[a-zA-Z0-9_-]+$/ },
+    thinkingPersistence: "text-only",
+    unsupportedSchemaFeatures: [],
+  },
+  "openai-completions": {
+    toolCalling: true,
+    maxTools: 0,
+    imageToolResults: false,
+    structuredOutput: true,
+    toolCallIdFormat: { maxLength: 64, allowedChars: /^[a-zA-Z0-9_-]+$/ },
+    thinkingPersistence: "text-only",
+    unsupportedSchemaFeatures: [],
+  },
+  "google-generative-ai": {
+    toolCalling: true,
+    maxTools: 0,
+    imageToolResults: true,
+    structuredOutput: true,
+    toolCallIdFormat: { maxLength: 64, allowedChars: /^[a-zA-Z0-9_-]+$/ },
+    thinkingPersistence: "text-only",
+    unsupportedSchemaFeatures: ["patternProperties", "const"],
+  },
+  "google-gemini-cli": {
+    toolCalling: true,
+    maxTools: 0,
+    imageToolResults: true,
+    structuredOutput: true,
+    toolCallIdFormat: { maxLength: 64, allowedChars: /^[a-zA-Z0-9_-]+$/ },
+    thinkingPersistence: "text-only",
+    unsupportedSchemaFeatures: ["patternProperties", "const"],
+  },
+  "google-vertex": {
+    toolCalling: true,
+    maxTools: 0,
+    imageToolResults: true,
+    structuredOutput: true,
+    toolCallIdFormat: { maxLength: 64, allowedChars: /^[a-zA-Z0-9_-]+$/ },
+    thinkingPersistence: "text-only",
+    unsupportedSchemaFeatures: ["patternProperties", "const"],
+  },
+  "mistral-conversations": {
+    toolCalling: true,
+    maxTools: 0,
+    imageToolResults: false,
+    structuredOutput: true,
+    toolCallIdFormat: { maxLength: 9, allowedChars: /^[a-zA-Z0-9]+$/ },
+    thinkingPersistence: "none",
+    unsupportedSchemaFeatures: [],
+  },
+  "bedrock-converse-stream": {
+    toolCalling: true,
+    maxTools: 0,
+    imageToolResults: true,  // Bedrock supports image content blocks in tool results
+    structuredOutput: true,
+    toolCallIdFormat: { maxLength: 64, allowedChars: /^[a-zA-Z0-9_-]+$/ },
+    thinkingPersistence: "text-only",
+    unsupportedSchemaFeatures: [],
+  },
+  "ollama-chat": {
+    toolCalling: true,
+    maxTools: 0,
+    imageToolResults: false,
+    structuredOutput: false,
+    toolCallIdFormat: { maxLength: 64, allowedChars: /^[a-zA-Z0-9_-]+$/ },
+    thinkingPersistence: "none",
+    unsupportedSchemaFeatures: [],
+  },
+};
+
+// ─── Default (permissive) profile for unknown providers ─────────────────────
+
+const DEFAULT_CAPABILITIES: ProviderCapabilities = {
+  toolCalling: true,
+  maxTools: 0,
+  imageToolResults: true,
+  structuredOutput: true,
+  toolCallIdFormat: { maxLength: 512, allowedChars: /^.+$/ },
+  thinkingPersistence: "text-only",
+  unsupportedSchemaFeatures: [],
+};
+
+// ─── Public API ─────────────────────────────────────────────────────────────
+
+/**
+ * Get capabilities for a provider API. Returns a permissive default for
+ * unknown providers (preserving existing behavior per ADR-005 principle 5).
+ */
+export function getProviderCapabilities(api: string): ProviderCapabilities {
+  return PROVIDER_CAPABILITIES[api] ?? DEFAULT_CAPABILITIES;
+}
+
+/**
+ * Check if a provider supports all required schema features.
+ * Returns the list of unsupported features (empty if all supported).
+ */
+export function getUnsupportedFeatures(api: string, requiredFeatures: string[]): string[] {
+  const caps = getProviderCapabilities(api);
+  return requiredFeatures.filter(f => caps.unsupportedSchemaFeatures.includes(f));
+}
+
+/**
+ * Deep-merge user-provided capability overrides with built-in defaults.
+ * Partial overrides merge with the built-in profile for the given API.
+ */
+export function mergeCapabilityOverrides(
+  api: string,
+  overrides: Partial<Omit<ProviderCapabilities, "toolCallIdFormat">> & {
+    toolCallIdFormat?: Partial<ProviderCapabilities["toolCallIdFormat"]>;
+  },
+): ProviderCapabilities {
+  const base = getProviderCapabilities(api);
+  return {
+    ...base,
+    ...overrides,
+    toolCallIdFormat: overrides.toolCallIdFormat
+      ? { ...base.toolCallIdFormat, ...overrides.toolCallIdFormat }
+      : base.toolCallIdFormat,
+  };
+}
+
+/**
+ * Get all registered API names in the capability registry.
+ * Used by lint rules to verify all providers in register-builtins.ts
+ * have corresponding capability entries.
+ */
+export function getRegisteredApis(): string[] {
+  return Object.keys(PROVIDER_CAPABILITIES);
+}
--- a/packages/pi-ai/src/providers/transform-messages-report.test.ts
+++ b/packages/pi-ai/src/providers/transform-messages-report.test.ts
@ -0,0 +1,189 @@
+// GSD-2 — ProviderSwitchReport Tests (ADR-005 Phase 3)
+import { describe, test } from "node:test";
+import assert from "node:assert/strict";
+
+import { transformMessages, createEmptyReport, hasTransformations } from "./transform-messages.js";
+import type { ProviderSwitchReport } from "./transform-messages.js";
+import type { Message, Model, AssistantMessage, ToolCall } from "../types.js";
+
+// ─── Helpers ────────────────────────────────────────────────────────────────
+
+function makeModel(overrides: Partial<Model<any>> = {}): Model<any> {
+  return {
+    id: "claude-sonnet-4-6",
+    name: "Claude Sonnet 4.6",
+    api: "anthropic-messages",
+    provider: "anthropic",
+    baseUrl: "",
+    reasoning: false,
+    input: ["text", "image"],
+    cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+    contextWindow: 200000,
+    maxTokens: 8192,
+    ...overrides,
+  } as Model<any>;
+}
+
+function makeAssistantMsg(overrides: Partial<AssistantMessage> = {}): AssistantMessage {
+  return {
+    role: "assistant",
+    content: [],
+    api: "anthropic-messages",
+    provider: "anthropic",
+    model: "claude-sonnet-4-6",
+    usage: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, totalTokens: 0, cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 } },
+    stopReason: "stop",
+    timestamp: Date.now(),
+    ...overrides,
+  };
+}
+
+// ─── createEmptyReport / hasTransformations ─────────────────────────────────
+
+describe("createEmptyReport", () => {
+  test("creates report with zero counters", () => {
+    const report = createEmptyReport("anthropic-messages", "openai-responses");
+    assert.equal(report.fromApi, "anthropic-messages");
+    assert.equal(report.toApi, "openai-responses");
+    assert.equal(report.thinkingBlocksDropped, 0);
+    assert.equal(report.thinkingBlocksDowngraded, 0);
+    assert.equal(report.toolCallIdsRemapped, 0);
+    assert.equal(report.syntheticToolResultsInserted, 0);
+    assert.equal(report.thoughtSignaturesDropped, 0);
+  });
+});
+
+describe("hasTransformations", () => {
+  test("returns false for empty report", () => {
+    const report = createEmptyReport("a", "b");
+    assert.equal(hasTransformations(report), false);
+  });
+
+  test("returns true when any counter is non-zero", () => {
+    const report = createEmptyReport("a", "b");
+    report.thinkingBlocksDropped = 1;
+    assert.equal(hasTransformations(report), true);
+  });
+});
+
+// ─── Report Tracking in transformMessages ───────────────────────────────────
+
+describe("transformMessages with report tracking", () => {
+  test("tracks thinking blocks dropped for redacted cross-model", () => {
+    const model = makeModel({ id: "gpt-5", api: "openai-responses", provider: "openai" });
+    const messages: Message[] = [
+      makeAssistantMsg({
+        content: [
+          { type: "thinking", thinking: "", redacted: true },
+          { type: "text", text: "Hello" },
+        ],
+      }),
+    ];
+    const report = createEmptyReport("anthropic-messages", "openai-responses");
+    transformMessages(messages, model, undefined, report);
+    assert.equal(report.thinkingBlocksDropped, 1);
+  });
+
+  test("tracks thinking blocks downgraded to plain text", () => {
+    const model = makeModel({ id: "gpt-5", api: "openai-responses", provider: "openai" });
+    const messages: Message[] = [
+      makeAssistantMsg({
+        content: [
+          { type: "thinking", thinking: "Let me think about this..." },
+          { type: "text", text: "Here is my answer" },
+        ],
+      }),
+    ];
+    const report = createEmptyReport("anthropic-messages", "openai-responses");
+    transformMessages(messages, model, undefined, report);
+    assert.equal(report.thinkingBlocksDowngraded, 1);
+  });
+
+  test("tracks tool call IDs remapped", () => {
+    const model = makeModel({ id: "claude-sonnet-4-6", api: "anthropic-messages", provider: "anthropic" });
+    const toolCall: ToolCall = {
+      type: "toolCall",
+      id: "original-long-id-that-needs-normalization|with-special-chars",
+      name: "bash",
+      arguments: { command: "ls" },
+    };
+    const messages: Message[] = [
+      makeAssistantMsg({
+        provider: "openai",
+        api: "openai-responses",
+        model: "gpt-5",
+        content: [toolCall],
+      }),
+    ];
+    const normalizer = (id: string) => id.replace(/[^a-zA-Z0-9_-]/g, "_").slice(0, 64);
+    const report = createEmptyReport("openai-responses", "anthropic-messages");
+    transformMessages(messages, model, normalizer, report);
+    assert.equal(report.toolCallIdsRemapped, 1);
+  });
+
+  test("tracks thought signatures dropped", () => {
+    const model = makeModel({ id: "claude-sonnet-4-6", api: "anthropic-messages", provider: "anthropic" });
+    const toolCall: ToolCall = {
+      type: "toolCall",
+      id: "tc_001",
+      name: "bash",
+      arguments: { command: "ls" },
+      thoughtSignature: "some-opaque-signature",
+    };
+    const messages: Message[] = [
+      makeAssistantMsg({
+        provider: "google",
+        api: "google-generative-ai",
+        model: "gemini-2.5-pro",
+        content: [toolCall],
+      }),
+    ];
+    const report = createEmptyReport("google-generative-ai", "anthropic-messages");
+    transformMessages(messages, model, undefined, report);
+    assert.equal(report.thoughtSignaturesDropped, 1);
+  });
+
+  test("tracks synthetic tool results inserted", () => {
+    const model = makeModel();
+    const toolCall: ToolCall = {
+      type: "toolCall",
+      id: "tc_orphan",
+      name: "bash",
+      arguments: { command: "ls" },
+    };
+    // Assistant message with tool call followed by another assistant (no tool result)
+    const messages: Message[] = [
+      makeAssistantMsg({ content: [toolCall, { type: "text", text: "Using bash" }] }),
+      makeAssistantMsg({ content: [{ type: "text", text: "Next message" }] }),
+    ];
+    const report = createEmptyReport("anthropic-messages", "anthropic-messages");
+    transformMessages(messages, model, undefined, report);
+    assert.equal(report.syntheticToolResultsInserted, 1);
+  });
+
+  test("does not count transformations for same-model messages", () => {
+    const model = makeModel();
+    const messages: Message[] = [
+      makeAssistantMsg({
+        content: [
+          { type: "thinking", thinking: "Let me think..." },
+          { type: "text", text: "Answer" },
+        ],
+      }),
+    ];
+    const report = createEmptyReport("anthropic-messages", "anthropic-messages");
+    transformMessages(messages, model, undefined, report);
+    assert.equal(report.thinkingBlocksDowngraded, 0);
+    assert.equal(report.thinkingBlocksDropped, 0);
+  });
+
+  test("works without report parameter (backward compatible)", () => {
+    const model = makeModel();
+    const messages: Message[] = [
+      makeAssistantMsg({ content: [{ type: "text", text: "Hello" }] }),
+    ];
+    // Should not throw
+    const result = transformMessages(messages, model);
+    assert.ok(Array.isArray(result));
+  });
+});
--- a/packages/pi-ai/src/providers/transform-messages.ts
+++ b/packages/pi-ai/src/providers/transform-messages.ts
@ -1,5 +1,54 @@
 import type { Api, AssistantMessage, Message, Model, ToolCall, ToolResultMessage } from "../types.js";

+/**
+ * Report of context transformations during a cross-provider switch (ADR-005 Phase 3).
+ * Tracks what was lost or downgraded when replaying conversation history to a different provider.
+ */
+export interface ProviderSwitchReport {
+	/** API of the messages being transformed from */
+	fromApi: string;
+	/** API of the target model */
+	toApi: string;
+	/** Number of thinking blocks completely dropped (redacted/encrypted, cross-model) */
+	thinkingBlocksDropped: number;
+	/** Number of thinking blocks downgraded from structured to plain text */
+	thinkingBlocksDowngraded: number;
+	/** Number of tool call IDs that were remapped/normalized */
+	toolCallIdsRemapped: number;
+	/** Number of synthetic tool results inserted for orphaned tool calls */
+	syntheticToolResultsInserted: number;
+	/** Number of thought signatures dropped (Google-specific opaque context) */
+	thoughtSignaturesDropped: number;
+}
+
+/**
+ * Create an empty provider switch report.
+ */
+export function createEmptyReport(fromApi: string, toApi: string): ProviderSwitchReport {
+	return {
+		fromApi,
+		toApi,
+		thinkingBlocksDropped: 0,
+		thinkingBlocksDowngraded: 0,
+		toolCallIdsRemapped: 0,
+		syntheticToolResultsInserted: 0,
+		thoughtSignaturesDropped: 0,
+	};
+}
+
+/**
+ * Check if a provider switch report has any non-zero transformations.
+ */
+export function hasTransformations(report: ProviderSwitchReport): boolean {
+	return (
+		report.thinkingBlocksDropped > 0 ||
+		report.thinkingBlocksDowngraded > 0 ||
+		report.toolCallIdsRemapped > 0 ||
+		report.syntheticToolResultsInserted > 0 ||
+		report.thoughtSignaturesDropped > 0
+	);
+}
+
 /**
 * Normalize tool call ID for cross-provider compatibility.
 * OpenAI Responses API generates IDs that are 450+ chars with special characters like `|`.
@ -9,6 +58,7 @@ export function transformMessages<TApi extends Api>(
 	messages: Message[],
 	model: Model<TApi>,
 	normalizeToolCallId?: (id: string, model: Model<TApi>, source: AssistantMessage) => string,
+	report?: ProviderSwitchReport,
 ): Message[] {
 	// Build a map of original tool call IDs to normalized IDs
 	const toolCallIdMap = new Map<string, string>();
@ -42,14 +92,20 @@ export function transformMessages<TApi extends Api>(
 					// Redacted thinking is opaque encrypted content, only valid for the same model.
 					// Drop it for cross-model to avoid API errors.
 					if (block.redacted) {
+						if (!isSameModel && report) report.thinkingBlocksDropped++;
 						return isSameModel ? block : [];
 					}
 					// For same model: keep thinking blocks with signatures (needed for replay)
 					// even if the thinking text is empty (OpenAI encrypted reasoning)
 					if (isSameModel && block.thinkingSignature) return block;
 					// Skip empty thinking blocks, convert others to plain text
-					if (!block.thinking || block.thinking.trim() === "") return [];
+					if (!block.thinking || block.thinking.trim() === "") {
+						if (!isSameModel && report) report.thinkingBlocksDropped++;
+						return [];
+					}
 					if (isSameModel) return block;
+					// Downgrade: structured thinking → plain text
+					if (report) report.thinkingBlocksDowngraded++;
 					return {
 						type: "text" as const,
 						text: block.thinking,
@ -71,6 +127,7 @@ export function transformMessages<TApi extends Api>(
 					if (!isSameModel && toolCall.thoughtSignature) {
 						normalizedToolCall = { ...toolCall };
 						delete (normalizedToolCall as { thoughtSignature?: string }).thoughtSignature;
+						if (report) report.thoughtSignaturesDropped++;
 					}

 					if (!isSameModel && normalizeToolCallId) {
@ -78,6 +135,7 @@ export function transformMessages<TApi extends Api>(
 						if (normalizedId !== toolCall.id) {
 							toolCallIdMap.set(toolCall.id, normalizedId);
 							normalizedToolCall = { ...normalizedToolCall, id: normalizedId };
+							if (report) report.toolCallIdsRemapped++;
 						}
 					}

@ -117,6 +175,7 @@ export function transformMessages<TApi extends Api>(
 							isError: true,
 							timestamp: Date.now(),
 						} as ToolResultMessage);
+						if (report) report.syntheticToolResultsInserted++;
 					}
 				}
 				pendingToolCalls = [];
@ -157,6 +216,7 @@ export function transformMessages<TApi extends Api>(
 							isError: true,
 							timestamp: Date.now(),
 						} as ToolResultMessage);
+						if (report) report.syntheticToolResultsInserted++;
 					}
 				}
 				pendingToolCalls = [];
--- a/packages/pi-coding-agent/src/core/extensions/index.ts
+++ b/packages/pi-coding-agent/src/core/extensions/index.ts
@ -43,6 +43,9 @@ export type {
 	BeforeProviderRequestEventResult,
 	// Context
 	CompactOptions,
+	// Events - Adjust Tool Set (ADR-005)
+	AdjustToolSetEvent,
+	AdjustToolSetResult,
 	// Events - Agent
 	ContextEvent,
 	// Event Results
@ -135,6 +138,7 @@ export type {
 	ToolCallEvent,
 	ToolCallEventResult,
 	// Tools
+	ToolCompatibility,
 	ToolDefinition,
 	// Events - Tool Execution
 	ToolExecutionEndEvent,
--- a/packages/pi-coding-agent/src/core/extensions/loader.ts
+++ b/packages/pi-coding-agent/src/core/extensions/loader.ts
@ -38,6 +38,7 @@ import type { ExecOptions } from "../exec.js";
 import { execCommand } from "../exec.js";
 import { getUntrustedExtensionPaths } from "./project-trust.js";
 export { isProjectTrusted, trustProject, getUntrustedExtensionPaths } from "./project-trust.js";
+import { registerToolCompatibility } from "../tools/tool-compatibility-registry.js";
 import type {
 	Extension,
 	ExtensionAPI,
@ -428,8 +429,9 @@ export function createExtensionRuntime(): ExtensionRuntime {
 		unregisterProvider: (name) => {
 			runtime.pendingProviderRegistrations = runtime.pendingProviderRegistrations.filter((r) => r.name !== name);
 		},
-		// Stub replaced by ExtensionRunner at construction time via bindEmitMethods().
+		// Stubs replaced by ExtensionRunner at construction time via bindEmitMethods().
 		emitBeforeModelSelect: async () => undefined,
+		emitAdjustToolSet: async () => undefined,
 	};

 	return runtime;
@ -459,6 +461,10 @@ function createExtensionAPI(
 				definition: tool,
 				extensionPath: extension.path,
 			});
+			// ADR-005: auto-register tool compatibility metadata
+			if (tool.compatibility) {
+				registerToolCompatibility(tool.name, tool.compatibility);
+			}
 			runtime.refreshTools();
 		},

@ -585,6 +591,10 @@ function createExtensionAPI(
 			return runtime.emitBeforeModelSelect(event);
 		},

+		async emitAdjustToolSet(event: Omit<import("./types.js").AdjustToolSetEvent, "type">): Promise<import("./types.js").AdjustToolSetResult | undefined> {
+			return runtime.emitAdjustToolSet(event);
+		},
+
 		events: eventBus,
 	} as ExtensionAPI;

--- a/packages/pi-coding-agent/src/core/extensions/runner.ts
+++ b/packages/pi-coding-agent/src/core/extensions/runner.ts
@ -11,6 +11,8 @@ import type { KeyAction, KeybindingsConfig } from "../keybindings.js";
 import type { ModelRegistry } from "../model-registry.js";
 import type { SessionManager } from "../session-manager.js";
 import type {
+	AdjustToolSetEvent,
+	AdjustToolSetResult,
 	BeforeAgentStartEvent,
 	BeforeAgentStartEventResult,
 	BeforeModelSelectEvent,
@ -234,6 +236,7 @@ export class ExtensionRunner {
 		this.modelRegistry = modelRegistry;
 		// Bind emit methods into the shared runtime so createExtensionAPI can delegate to them.
 		this.runtime.emitBeforeModelSelect = (event) => this.emitBeforeModelSelect(event);
+		this.runtime.emitAdjustToolSet = (event) => this.emitAdjustToolSet(event);
 	}

 	bindCore(actions: ExtensionActions, contextActions: ExtensionContextActions): void {
@ -713,6 +716,21 @@ export class ExtensionRunner {
 		return result;
 	}

+	async emitAdjustToolSet(event: Omit<AdjustToolSetEvent, "type">): Promise<AdjustToolSetResult | undefined> {
+		let result: AdjustToolSetResult | undefined;
+		await this.invokeHandlers("adjust_tool_set", () => ({
+			type: "adjust_tool_set" as const,
+			...event,
+		} satisfies AdjustToolSetEvent), (handlerResult) => {
+			if (handlerResult) {
+				result = handlerResult as AdjustToolSetResult;
+				return { done: true }; // first override wins
+			}
+			return { done: false };
+		});
+		return result;
+	}
+
 	async emitBeforeAgentStart(
 		prompt: string,
 		images: ImageContent[] | undefined,
--- a/packages/pi-coding-agent/src/core/extensions/types.ts
+++ b/packages/pi-coding-agent/src/core/extensions/types.ts
@ -331,6 +331,19 @@ export interface ToolRenderResultOptions {
 	isPartial: boolean;
 }

+/**
+ * Tool compatibility metadata for provider-aware tool filtering (ADR-005 Phase 2).
+ * Tools without compatibility metadata are assumed universally compatible.
+ */
+export interface ToolCompatibility {
+	/** Tool produces image content in results (filtered for providers without imageToolResults) */
+	producesImages?: boolean;
+	/** Tool requires schema features that some providers don't support (e.g., ["patternProperties"]) */
+	schemaFeatures?: string[];
+	/** Tool is effective only with models above a minimum capability threshold */
+	minCapabilityTier?: "light" | "standard" | "heavy";
+}
+
 /**
 * Tool definition for registerTool().
 */
@ -347,6 +360,8 @@ export interface ToolDefinition<TParams extends TSchema = TSchema, TDetails = un
 	promptGuidelines?: string[];
 	/** Parameter schema (TypeBox) */
 	parameters: TParams;
+	/** Provider compatibility metadata (ADR-005). Omit for universally compatible tools. */
+	compatibility?: ToolCompatibility;

 	/** Execute the tool. */
 	execute(
@ -619,6 +634,30 @@ export interface BeforeModelSelectResult {
 	modelId: string;
 }

+/**
+ * Fired after model selection to allow extensions to adjust the active tool set (ADR-005 Phase 4).
+ * Extensions can add, remove, or reorder tools based on the selected model's provider capabilities.
+ */
+export interface AdjustToolSetEvent {
+	type: "adjust_tool_set";
+	/** The selected model's API type */
+	selectedModelApi: string;
+	/** The selected model's provider */
+	selectedModelProvider: string;
+	/** The selected model ID */
+	selectedModelId: string;
+	/** Current active tool names */
+	activeToolNames: string[];
+	/** Tools already filtered by provider compatibility */
+	filteredTools: string[];
+}
+
+/** Result from adjust_tool_set event handler. Return { toolNames } to override tool set. */
+export interface AdjustToolSetResult {
+	/** Replacement tool names. If omitted, the default filtering is used. */
+	toolNames?: string[];
+}
+
 // ============================================================================
 // User Bash Events
 // ============================================================================
@ -1069,6 +1108,7 @@ export interface ExtensionAPI {
 	on(event: "user_bash", handler: ExtensionHandler<UserBashEvent, UserBashEventResult>): void;
 	on(event: "input", handler: ExtensionHandler<InputEvent, InputEventResult>): void;
 	on(event: "before_model_select", handler: ExtensionHandler<BeforeModelSelectEvent, BeforeModelSelectResult>): void;
+	on(event: "adjust_tool_set", handler: ExtensionHandler<AdjustToolSetEvent, AdjustToolSetResult>): void;

 	// =========================================================================
 	// Event Emission (for host extensions that orchestrate model selection)
@ -1077,6 +1117,9 @@ export interface ExtensionAPI {
 	/** Emit before_model_select event. Returns override model ID or undefined. */
 	emitBeforeModelSelect(event: Omit<BeforeModelSelectEvent, "type">): Promise<BeforeModelSelectResult | undefined>;

+	/** Emit adjust_tool_set event (ADR-005). Returns override tool names or undefined. */
+	emitAdjustToolSet(event: Omit<AdjustToolSetEvent, "type">): Promise<AdjustToolSetResult | undefined>;
+
 	// =========================================================================
 	// Tool Registration
 	// =========================================================================
@ -1395,6 +1438,8 @@ export interface ExtensionRuntimeState {
 	unregisterProvider: (name: string) => void;
 	/** Emit before_model_select event to all registered handlers. Bound by ExtensionRunner. */
 	emitBeforeModelSelect: (event: Omit<BeforeModelSelectEvent, "type">) => Promise<BeforeModelSelectResult | undefined>;
+	/** Emit adjust_tool_set event to all registered handlers. Bound by ExtensionRunner (ADR-005). */
+	emitAdjustToolSet: (event: Omit<AdjustToolSetEvent, "type">) => Promise<AdjustToolSetResult | undefined>;
 }

 /**
--- a/packages/pi-coding-agent/src/core/tools/index.ts
+++ b/packages/pi-coding-agent/src/core/tools/index.ts
@ -112,6 +112,13 @@ export {
 	lspTool,
 } from "../lsp/index.js";
 export type { LspServerStatus } from "../lsp/client.js";
+export {
+	registerToolCompatibility,
+	getToolCompatibility,
+	getAllToolCompatibility,
+	registerMcpToolCompatibility,
+	resetToolCompatibilityRegistry,
+} from "./tool-compatibility-registry.js";

 import type { AgentTool } from "@gsd/pi-agent-core";
 import { type BashToolOptions, bashTool, createBashTool } from "./bash.js";
--- a/packages/pi-coding-agent/src/core/tools/tool-compatibility-registry.ts
+++ b/packages/pi-coding-agent/src/core/tools/tool-compatibility-registry.ts
@ -0,0 +1,83 @@
+// GSD-2 — Tool Compatibility Registry (ADR-005 Phase 2)
+// Maps tool names to their provider compatibility metadata.
+// Used by the model router to filter tools incompatible with the selected provider.
+
+import type { ToolCompatibility } from "../extensions/types.js";
+
+// ─── Registry State ─────────────────────────────────────────────────────────
+
+const registry = new Map<string, ToolCompatibility>();
+
+// ─── Built-in Tool Compatibility (universally compatible) ───────────────────
+// Built-in tools (bash, read, write, edit, grep, find, ls) produce text-only
+// results and use standard JSON Schema — compatible with all providers.
+
+const BUILTIN_TOOLS: Record<string, ToolCompatibility> = {
+  bash: {},
+  read: {},
+  write: {},
+  edit: {},
+  grep: {},
+  find: {},
+  ls: {},
+  lsp: {},
+  hashline_edit: {},
+  hashline_read: {},
+};
+
+// Pre-populate registry with built-in tools
+for (const [name, compat] of Object.entries(BUILTIN_TOOLS)) {
+  registry.set(name, compat);
+}
+
+// ─── MCP Tool Defaults ─────────────────────────────────────────────────────
+// MCP tools may use complex schemas. Default to cautious compatibility.
+
+const MCP_TOOL_DEFAULTS: ToolCompatibility = {
+  schemaFeatures: ["patternProperties"],
+};
+
+// ─── Public API ─────────────────────────────────────────────────────────────
+
+/**
+ * Register compatibility metadata for a tool.
+ * Called automatically by registerTool() for extension tools that include
+ * compatibility metadata in their ToolDefinition.
+ */
+export function registerToolCompatibility(toolName: string, compatibility: ToolCompatibility): void {
+  registry.set(toolName, compatibility);
+}
+
+/**
+ * Get compatibility metadata for a tool.
+ * Returns undefined for unknown tools (treated as universally compatible
+ * per ADR-005 principle: "fail open, don't restrict without data").
+ */
+export function getToolCompatibility(toolName: string): ToolCompatibility | undefined {
+  return registry.get(toolName);
+}
+
+/**
+ * Get all registered tool compatibility entries.
+ */
+export function getAllToolCompatibility(): ReadonlyMap<string, ToolCompatibility> {
+  return registry;
+}
+
+/**
+ * Register an MCP tool with default cautious compatibility.
+ * MCP tools may use complex schemas that some providers don't support.
+ */
+export function registerMcpToolCompatibility(toolName: string, overrides?: Partial<ToolCompatibility>): void {
+  registry.set(toolName, { ...MCP_TOOL_DEFAULTS, ...overrides });
+}
+
+/**
+ * Clear all non-builtin entries (for testing).
+ */
+export function resetToolCompatibilityRegistry(): void {
+  registry.clear();
+  for (const [name, compat] of Object.entries(BUILTIN_TOOLS)) {
+    registry.set(name, compat);
+  }
+}
--- a/packages/pi-coding-agent/src/index.ts
+++ b/packages/pi-coding-agent/src/index.ts
@ -49,6 +49,8 @@ export {
 export { createEventBus, type EventBus, type EventBusController } from "./core/event-bus.js";
 // Extension system
 export type {
+	AdjustToolSetEvent,
+	AdjustToolSetResult,
 	AgentEndEvent,
 	AgentStartEvent,
 	AgentToolResult,
@ -118,6 +120,7 @@ export type {
 	SlashCommandSource,
 	TerminalInputHandler,
 	ToolCallEvent,
+	ToolCompatibility,
 	ToolDefinition,
 	ToolInfo,
 	SortResult,
@ -310,6 +313,12 @@ export {
 	type HashlineReadToolDetails,
 	type HashlineReadToolInput,
 	type HashlineReadToolOptions,
+	// Tool compatibility registry (ADR-005)
+	registerToolCompatibility,
+	getToolCompatibility,
+	getAllToolCompatibility,
+	registerMcpToolCompatibility,
+	resetToolCompatibilityRegistry,
 } from "./core/tools/index.js";
 // Main entry point
 export { main } from "./main.js";
--- a/src/resources/extensions/gsd/auto-model-selection.ts
+++ b/src/resources/extensions/gsd/auto-model-selection.ts
@ -212,6 +212,13 @@ export async function selectAndApplyModel(
                "info",
              );
            }
+            // ADR-005: Report tools filtered due to provider incompatibility
+            if (routingResult.filteredTools && routingResult.filteredTools.length > 0) {
+              ctx.ui.notify(
+                `Tool compatibility: ${routingResult.filteredTools.length} tools filtered for provider — ${routingResult.filteredTools.join(", ")}`,
+                "info",
+              );
+            }
          }
        }
        routingTierLabel = ` [${tierLabel(classification.tier)}]`;
--- a/src/resources/extensions/gsd/bootstrap/register-hooks.ts
+++ b/src/resources/extensions/gsd/bootstrap/register-hooks.ts
@ -446,4 +446,12 @@ export function registerHooks(pi: ExtensionAPI): void {
    // Default: no override — let capability scoring handle selection
    return undefined;
  });
+
+  // Tool set adaptation hook (ADR-005 Phase 4)
+  // Extensions can override tool set after model selection by returning { toolNames: [...] }
+  // Return undefined to let the built-in provider compatibility filtering proceed.
+  pi.on("adjust_tool_set", async (_event) => {
+    // Default: no override — let provider capability filtering handle tool set
+    return undefined;
+  });
 }
--- a/src/resources/extensions/gsd/model-router.ts
+++ b/src/resources/extensions/gsd/model-router.ts
@ -5,6 +5,9 @@
 import type { ComplexityTier, ClassificationResult, TaskMetadata } from "./complexity-classifier.js";
 import { tierOrdinal } from "./complexity-classifier.js";
 import type { ResolvedModelConfig } from "./preferences.js";
+import { getProviderCapabilities, type ProviderCapabilities } from "@gsd/pi-ai";
+import { getToolCompatibility, getAllToolCompatibility } from "@gsd/pi-coding-agent";
+import type { ToolCompatibility } from "@gsd/pi-coding-agent";

 // ─── Types ───────────────────────────────────────────────────────────────────

@ -37,6 +40,8 @@ export interface RoutingDecision {
  selectionMethod: "tier-only" | "capability-scored";
  /** Capability scores per eligible model (capability-scored path only) */
  capabilityScores?: Record<string, number>;
+  /** Tools filtered out due to provider incompatibility (ADR-005) */
+  filteredTools?: string[];
  /** Task requirement vector used for scoring */
  taskRequirements?: Partial<Record<string, number>>;
 }
@ -536,3 +541,71 @@ function getModelCost(modelId: string): number {
  // Unknown cost — assume expensive to avoid routing to unknown cheap models
  return 999;
 }
+
+// ─── Tool Compatibility Filter (ADR-005 Phase 3) ───────────────────────────
+
+/**
+ * Check if a tool is compatible with a provider's capabilities.
+ * Returns true if the tool can be used with the provider.
+ */
+export function isToolCompatibleWithProvider(
+  toolName: string,
+  providerCaps: ProviderCapabilities,
+): boolean {
+  const compat = getToolCompatibility(toolName);
+  if (!compat) return true;  // no metadata = always compatible
+
+  // Hard filter: provider doesn't support image tool results
+  if (compat.producesImages && !providerCaps.imageToolResults) return false;
+
+  // Hard filter: tool uses schema features provider doesn't support
+  if (compat.schemaFeatures?.some(f => providerCaps.unsupportedSchemaFeatures.includes(f))) {
+    return false;
+  }
+
+  return true;
+}
+
+/**
+ * Filter a list of tool names to only those compatible with a provider.
+ * Used by the routing pipeline to adjust tool sets when switching providers.
+ */
+export function filterToolsForProvider(
+  toolNames: string[],
+  providerApi: string,
+): { compatible: string[]; filtered: string[] } {
+  const providerCaps = getProviderCapabilities(providerApi);
+
+  // Provider doesn't support tool calling at all
+  if (!providerCaps.toolCalling) {
+    return { compatible: [], filtered: toolNames };
+  }
+
+  const compatible: string[] = [];
+  const filtered: string[] = [];
+
+  for (const name of toolNames) {
+    if (isToolCompatibleWithProvider(name, providerCaps)) {
+      compatible.push(name);
+    } else {
+      filtered.push(name);
+    }
+  }
+
+  return { compatible, filtered };
+}
+
+/**
+ * Adjust the active tool set for a selected model's provider capabilities.
+ * Returns tool names that should be active — removes incompatible tools.
+ *
+ * This is a hard filter only — it removes tools that would fail at the
+ * provider level. It does NOT remove tools based on soft heuristics.
+ */
+export function adjustToolSet(
+  activeToolNames: string[],
+  selectedModelApi: string,
+): { toolNames: string[]; removedTools: string[] } {
+  const { compatible, filtered } = filterToolsForProvider(activeToolNames, selectedModelApi);
+  return { toolNames: compatible, removedTools: filtered };
+}
--- a/src/resources/extensions/gsd/tests/tool-compatibility.test.ts
+++ b/src/resources/extensions/gsd/tests/tool-compatibility.test.ts
@ -0,0 +1,199 @@
+// GSD-2 — Tool Compatibility + Model Router Tool Filtering Tests (ADR-005 Phases 2-3)
+import { describe, test, beforeEach } from "node:test";
+import assert from "node:assert/strict";
+
+import {
+  registerToolCompatibility,
+  getToolCompatibility,
+  getAllToolCompatibility,
+  registerMcpToolCompatibility,
+  resetToolCompatibilityRegistry,
+} from "@gsd/pi-coding-agent";
+
+import {
+  isToolCompatibleWithProvider,
+  filterToolsForProvider,
+  adjustToolSet,
+} from "../model-router.js";
+
+import {
+  getProviderCapabilities,
+} from "@gsd/pi-ai";
+
+// ─── Tool Compatibility Registry ────────────────────────────────────────────
+
+describe("tool compatibility registry", () => {
+  beforeEach(() => {
+    resetToolCompatibilityRegistry();
+  });
+
+  test("built-in tools are pre-registered", () => {
+    const builtins = ["bash", "read", "write", "edit", "grep", "find", "ls", "lsp"];
+    for (const name of builtins) {
+      const compat = getToolCompatibility(name);
+      assert.ok(compat !== undefined, `${name} should be pre-registered`);
+    }
+  });
+
+  test("unknown tool returns undefined", () => {
+    assert.equal(getToolCompatibility("nonexistent_tool_xyz"), undefined);
+  });
+
+  test("registerToolCompatibility stores and retrieves metadata", () => {
+    registerToolCompatibility("screenshot_tool", {
+      producesImages: true,
+      minCapabilityTier: "standard",
+    });
+    const compat = getToolCompatibility("screenshot_tool");
+    assert.ok(compat);
+    assert.equal(compat.producesImages, true);
+    assert.equal(compat.minCapabilityTier, "standard");
+  });
+
+  test("registerMcpToolCompatibility sets default schema features", () => {
+    registerMcpToolCompatibility("mcp__test__tool");
+    const compat = getToolCompatibility("mcp__test__tool");
+    assert.ok(compat);
+    assert.ok(compat.schemaFeatures?.includes("patternProperties"));
+  });
+
+  test("registerMcpToolCompatibility allows overrides", () => {
+    registerMcpToolCompatibility("mcp__test__override", { producesImages: true });
+    const compat = getToolCompatibility("mcp__test__override");
+    assert.ok(compat);
+    assert.equal(compat.producesImages, true);
+    assert.ok(compat.schemaFeatures?.includes("patternProperties"));
+  });
+
+  test("getAllToolCompatibility returns all entries", () => {
+    const all = getAllToolCompatibility();
+    assert.ok(all.size >= 10); // at least built-in tools
+    assert.ok(all.has("bash"));
+    assert.ok(all.has("read"));
+  });
+
+  test("resetToolCompatibilityRegistry clears custom entries but keeps builtins", () => {
+    registerToolCompatibility("custom_tool", { producesImages: true });
+    assert.ok(getToolCompatibility("custom_tool"));
+    resetToolCompatibilityRegistry();
+    assert.equal(getToolCompatibility("custom_tool"), undefined);
+    assert.ok(getToolCompatibility("bash")); // built-in preserved
+  });
+});
+
+// ─── isToolCompatibleWithProvider ───────────────────────────────────────────
+
+describe("isToolCompatibleWithProvider", () => {
+  beforeEach(() => {
+    resetToolCompatibilityRegistry();
+  });
+
+  test("tool without compatibility metadata is always compatible", () => {
+    const caps = getProviderCapabilities("anthropic-messages");
+    assert.equal(isToolCompatibleWithProvider("unknown_tool", caps), true);
+  });
+
+  test("built-in tools are compatible with all providers", () => {
+    const providers = ["anthropic-messages", "openai-responses", "google-generative-ai", "mistral-conversations"];
+    const tools = ["bash", "read", "write", "edit"];
+    for (const api of providers) {
+      const caps = getProviderCapabilities(api);
+      for (const tool of tools) {
+        assert.equal(
+          isToolCompatibleWithProvider(tool, caps),
+          true,
+          `${tool} should be compatible with ${api}`,
+        );
+      }
+    }
+  });
+
+  test("image-producing tool filtered for providers without image support", () => {
+    registerToolCompatibility("screenshot", { producesImages: true });
+    const openaiCaps = getProviderCapabilities("openai-responses");
+    assert.equal(isToolCompatibleWithProvider("screenshot", openaiCaps), false);
+
+    const anthropicCaps = getProviderCapabilities("anthropic-messages");
+    assert.equal(isToolCompatibleWithProvider("screenshot", anthropicCaps), true);
+  });
+
+  test("tool with unsupported schema features filtered for Google", () => {
+    registerToolCompatibility("complex_schema_tool", {
+      schemaFeatures: ["patternProperties"],
+    });
+    const googleCaps = getProviderCapabilities("google-generative-ai");
+    assert.equal(isToolCompatibleWithProvider("complex_schema_tool", googleCaps), false);
+
+    const anthropicCaps = getProviderCapabilities("anthropic-messages");
+    assert.equal(isToolCompatibleWithProvider("complex_schema_tool", anthropicCaps), true);
+  });
+});
+
+// ─── filterToolsForProvider ─────────────────────────────────────────────────
+
+describe("filterToolsForProvider", () => {
+  beforeEach(() => {
+    resetToolCompatibilityRegistry();
+  });
+
+  test("all built-in tools pass for any provider", () => {
+    const toolNames = ["bash", "read", "write", "edit", "grep", "find", "ls"];
+    const { compatible, filtered } = filterToolsForProvider(toolNames, "mistral-conversations");
+    assert.deepEqual(compatible, toolNames);
+    assert.deepEqual(filtered, []);
+  });
+
+  test("image tool filtered for OpenAI Responses", () => {
+    registerToolCompatibility("browser_screenshot", { producesImages: true });
+    const toolNames = ["bash", "read", "browser_screenshot"];
+    const { compatible, filtered } = filterToolsForProvider(toolNames, "openai-responses");
+    assert.deepEqual(compatible, ["bash", "read"]);
+    assert.deepEqual(filtered, ["browser_screenshot"]);
+  });
+
+  test("MCP tool with patternProperties filtered for Google", () => {
+    registerMcpToolCompatibility("mcp__repowise__search");
+    const toolNames = ["bash", "read", "mcp__repowise__search"];
+    const { compatible, filtered } = filterToolsForProvider(toolNames, "google-generative-ai");
+    assert.deepEqual(compatible, ["bash", "read"]);
+    assert.deepEqual(filtered, ["mcp__repowise__search"]);
+  });
+
+  test("unknown provider passes all tools (permissive default)", () => {
+    registerToolCompatibility("image_tool", { producesImages: true });
+    registerMcpToolCompatibility("mcp_tool");
+    const toolNames = ["bash", "image_tool", "mcp_tool"];
+    const { compatible, filtered } = filterToolsForProvider(toolNames, "unknown-provider-xyz");
+    assert.deepEqual(compatible, toolNames);
+    assert.deepEqual(filtered, []);
+  });
+});
+
+// ─── adjustToolSet ──────────────────────────────────────────────────────────
+
+describe("adjustToolSet", () => {
+  beforeEach(() => {
+    resetToolCompatibilityRegistry();
+  });
+
+  test("returns all tools for Anthropic (most permissive)", () => {
+    registerToolCompatibility("screenshot", { producesImages: true });
+    const toolNames = ["bash", "read", "screenshot"];
+    const { toolNames: result, removedTools } = adjustToolSet(toolNames, "anthropic-messages");
+    assert.deepEqual(result, toolNames);
+    assert.deepEqual(removedTools, []);
+  });
+
+  test("removes incompatible tools and reports them", () => {
+    registerToolCompatibility("screenshot", { producesImages: true });
+    registerMcpToolCompatibility("mcp_complex");
+    const toolNames = ["bash", "read", "screenshot", "mcp_complex"];
+    const { toolNames: result, removedTools } = adjustToolSet(toolNames, "google-generative-ai");
+    // Google supports images but not patternProperties
+    assert.ok(result.includes("bash"));
+    assert.ok(result.includes("read"));
+    assert.ok(result.includes("screenshot")); // Google supports images
+    assert.ok(!result.includes("mcp_complex")); // patternProperties not supported
+    assert.deepEqual(removedTools, ["mcp_complex"]);
+  });
+});