feat!: upgrade all packages to AI SDK v7 stable

[ThomasK33]

Summary

Upgrades every package in the monorepo from AI SDK v6 (and canary-tagged harness packages) to the stable AI SDK v7 line:

Dependency	Before	After
ai	^6.0.206	^7.0.19 (peer ranges → ^7.0.0)
@ai-sdk/provider	3.0.10	4.0.3
@ai-sdk/react	3.0.208	4.0.20
@ai-sdk/anthropic	^3.x	^4.0.11
@ai-sdk/openai-compatible	^2.0.51	^3.0.7
@ai-sdk/harness / harness-claude-code / tui / provider-utils	canary pins	stable ^1.0.23 / ^1.0.23 / ^1.0.19 / ^5.0.7

Breaking-change migrations

• agent: CoderLanguageModel now natively implements LanguageModelV4 (specificationVersion: "v4"). V4's stream parts / usage / messages are structurally identical to V3; the substantive change is file-part data becoming the tagged SharedV4FileData union — inline data/text variants are handled (text is new), url/reference are rejected with actionable errors. Also: stepCountIs → isStepCount, the new toolsContext setting is surfaced on CoderAgentSettings and forwarded to ToolLoopAgent, the runtime ai-major guard is bumped 6 → 7, and engines.node is raised to >=22 (matching the rest of the repo).
• provider: CoderProvider surface swapped to LanguageModelV4 / EmbeddingModelV4.
• release-please-ai: generateObject's system: option renamed to instructions:.
• sandbox: canary → stable harness packages; the Claude Code thinking option is now a tagged config object ("adaptive" → { type: "adaptive" }, "off" → { type: "disabled" }). Also added an E2E_PORT override to scripts/e2e-claude.ts since port 4000 is often taken inside shared workspaces.
• workspace: the canary-era @ai-sdk/harness/ai pnpm overrides are gone — the whole repo now resolves one coherent stable ai@7 chain. The single-zod pin stays (same duplicate-provider-utils type-identity rationale as before).
• docs: READMEs updated off canary tags and ai v6 wording; install/requirements now say Node ≥ 22 and ai v7.

Validation

Static + unit gates (all green):

• pnpm check — oxfmt, oxlint, tsc --noEmit in all 4 packages
• pnpm -r test — 152 agent + 21 provider + 11 release-please + 118 sandbox tests
• pnpm build, publint ("All good!"), attw ("No problems found") on the published packages
• Runtime smoke on built dist/ artifacts: generateText() through ai@7 against createCoder with a stubbed fetch; CoderAgent construction through the new version guard; sandbox provider satisfies harness-sandbox-v1

Live end-to-end against dev.coder.com (all green):

• agent e2e (vitest run test/e2e): 5/5 passed against real chatd — generate, stream, and tool round-trips through ai@7's ToolLoopAgent
• provider: live generateText/streamText on the Anthropic surface (claude-sonnet-4-6) and OpenAI-compatible surface (gpt-5.1), including v7's new inputTokenDetails/outputTokenDetails usage shape
• sandbox verify-real: 23/23 checks against a started workspace (exec, file I/O, spawn streaming, WebSocket over ssh -L)
• sandbox e2e-claude: real Claude Code turn inside a workspace on stable @ai-sdk/harness@1.0.23 — asked for uname -sr, answered exactly Linux 6.8.0-124-generic (29s)

---

Generated with mux • Model: anthropic:claude-fable-5 • Thinking: xhigh