feat(claude): expose agent.output_format for structured output

.changeset/a2a-claude-beta-6.md

@@ -0,0 +1,5 @@
+---
+"throng-agent-claude": minor
+---
+
+Move `@col/a2a-claude` to `0.2.1-beta.6`, which adds `claude.outputFormat` — structured JSON output constrained to a caller-supplied JSON Schema. No behaviour changes for agents that do not set it; the field is absent from Throng's generated config unless a manifest asks for it.

.changeset/claude-output-format.md

@@ -0,0 +1,16 @@
+---
+"throng-agent-claude": minor
+---
+
+Add `agent.output_format`, a new optional initialise key that constrains a Claude turn's output to a JSON Schema.
+
+```jsonc
+"agent": {
+  "platform": "claude",
+  "output_format": { "type": "json_schema", "schema": { /* … */ } }
+}
+```
+
+Throng validates the wrapper shape only — `type` must be `json_schema` and `schema` must be an object — and returns a field-level 400 from `POST /api/initialise` when either is wrong. The schema body itself is forwarded verbatim: it is never inspected or key-transformed, so JSON Schema's own vocabulary survives intact, and an invalid schema surfaces from the SDK at turn time rather than at initialise. Omitting the key leaves the wrapper's freeform-text default in place, so no existing manifest changes behaviour.
+
+Two operational notes for anyone enabling it. The structured result arrives as an additional `application/json` data part on the `response` artifact, with the text part still published alongside it — text-only clients are unaffected. That data part is only added when the result is a JSON object, so a schema whose top level is an array or a scalar reaches clients through the text part alone. And an unsatisfiable schema fails the turn with "Structured output retries exhausted." rather than degrading to freeform output; that message usually means a schema bug, though the SDK reports the same exhaustion when structured output is retracted by a model fallback.

README.md

@@ -47,6 +47,10 @@ Throng agent works best when run on a platform such as [E2B.dev](https://e2b.dev
     "permission_mode": "plan", // engine-specific (claude)
     "thinking": { "type": "adaptive" }, // engine-specific (claude)
     "effort": "high",          // engine-specific (claude)
+    "output_format": {         // engine-specific (claude) — structured output
+      "type": "json_schema",
+      "schema": { "type": "object" }
+    },
     "plugins": []              // engine-specific (claude)
   }
 }
@@ -139,6 +143,41 @@ it.
 - **`effort`** is optional and sets the reasoning effort level: one of
   `low`, `medium`, `high`, `xhigh`, `max`.
 
+### `output_format` (claude)
+
+Optional. Constrains the model's turn output to a JSON Schema. Omit it and the
+agent returns freeform text, which is the default.
+
+```jsonc
+"output_format": {
+  "type": "json_schema",   // the only type the SDK supports
+  "schema": { /* any JSON Schema object */ }
+}
+```
+
+Throng validates the wrapper shape only — that `type` is `json_schema` and that
+`schema` is an object. The schema body is forwarded to the SDK **verbatim**; it
+is never inspected or key-transformed, so JSON Schema's own vocabulary
+(`additionalProperties`, `patternProperties`, …) survives intact. A schema that
+is not valid JSON Schema is reported by the SDK at turn time, not at initialise.
+
+Two consequences worth knowing before you enable it:
+
+- **The structured result is an additional part, not a replacement.** On success
+  the parsed object is published as an `application/json` data part on the
+  `response` artifact, and the text part is still published alongside it, first —
+  text-only clients are unaffected. The data part is only added when the result
+  is a JSON **object**: a schema whose top level is an array or a scalar produces
+  no data part, and such a result reaches clients through the text part alone.
+- **An unsatisfiable schema fails the turn.** The SDK retries when output does
+  not match, and on exhaustion the turn fails with *"Structured output retries
+  exhausted."* It does not fall back to freeform text. A schema that no output
+  can satisfy — for example one whose `required` names a property that
+  `properties` never declares while `additionalProperties` is `false` — fails
+  every turn. That message usually means a schema bug, but not always: the SDK
+  reports the same exhaustion when structured output is retracted by a model
+  fallback, so rule out the schema before concluding the model is at fault.
+
 ### `credentials`, `github_token` and `user_identity`
 
 - **`credentials`** is how the agent gets GitHub tokens in production. `git` uses