refactor(tool): migrate tool framework + all 18 built-in tools to Effect Schema

[kitlangton]

Stacked on #23244. Fully removes zod from the tool layer.

Framework (tool/tool.ts)

• Tool.Def.parameters typed as Schema.Decoder<unknown> (replaces z.ZodType).
• Runtime validation uses Schema.decodeUnknownEffect + Effect.mapError instead of Effect.try wrapping toolInfo.parameters.parse(args) (matches idiomatic patterns already used in mcp-exa.ts and installation/index.ts; satisfies AGENTS.md's avoid-try/catch rule).
• formatValidationError?(error: unknown) — the hook is no longer zod-specific. No tool currently implements it.

JSON Schema pipeline

Extracted EffectZod.toJsonSchema(schema) into util/effect-zod.ts. Three emit sites now share one implementation:

• session/prompt.ts:407 — every LLM tool registration
• server/routes/instance/experimental.ts:216 — was previously calling z.toJSONSchema(t.parameters) directly; would have failed at runtime once t.parameters became a Schema. Caught by the review pass, fixed by this PR.
• test/tool/parameters.test.ts — snapshot test mirrors the production path, so any future drift is impossible.

Tools (18 files)

Mechanical conversion of each tool's Parameters:

Before (zod)	After (Schema)
z.object({...})	Schema.Struct({...})
z.string().describe("x")	Schema.String.annotate({ description: "x" })
z.string().optional()	Schema.optional(Schema.String)
z.number().int().min(1)	Schema.Number.check(Schema.isInt()).check(Schema.isGreaterThanOrEqualTo(1))
z.enum([...])	Schema.Literals([...])
z.array(X)	Schema.mutable(Schema.Array(X))
z.number().min(1000).max(50000).default(5000)	chained checks + Schema.withDecodingDefault(Effect.succeed(5000))
z.infer<typeof Parameters>	Schema.Schema.Type<typeof Parameters>

Notable simplifications:

• read.ts: dropped z.coerce.number() on offset/limit. The LLM tool-call path receives typed JSON; coercion from strings is only relevant to CLI-facing callers, of which there are none. JSON Schema output is identical (type: "number").
• todo.ts: inlined the three todo fields rather than pulling Todo.Info.shape from the still-zod session/todo.ts. Removes the last zod import from the tool file; LLM-visible JSON Schema is identical.
• question.ts: references Question.Prompt (already a Schema.Class) directly instead of Question.Prompt.zod.

Plugin tools (tool/registry.ts)

Plugin tool args are defined via zod (@opencode-ai/plugin API — external). The registry wraps the derived z.object(def.args) in a Schema.declare(typeGuard) so it slots into the Schema-typed framework, and annotates it with ZodOverride so the walker emits the original zod object when generating JSON Schema for the LLM. Runtime behaviour and JSON Schema for plugin tools are unchanged.

Verification

• bun typecheck clean
• bun run test test/ — 2054 pass / 0 fail (full suite)
• 18 JSON Schema snapshots byte-identical — LLM view of every tool is unchanged. The refactor(tool): migrate tool framework + all 18 built-in tools to Effect Schema #23244 snapshot test caught zero drift.
• SDK byte-identical (types.gen.ts and openapi.json)

Scope notes

• Tool authors write pure Schema. No zod in any tool file.
• The walker (@/util/effect-zod) and the central EffectZod.toJsonSchema helper remain — they convert Schema → zod at the edge where the ai SDK still wants zod-style JSON Schema. These go away entirely when the runtime LLM layer eventually accepts Effect Schema natively (orthogonal future work).
• session/todo.ts still exports a zod Info; migration deferred to a follow-up (not touching it here keeps this PR scoped to the tool layer).