fix(claude-print): accept the event-array envelope claude CLI 2.x emits

[fedorovvvv]

Fixes #492.

The bug

claude --print --output-format json on claude CLI 2.x returns a JSON array of stream events, with the envelope as the final {"type": "result", ...} entry. parse_envelope decoded stdout straight into ClaudePrintResponse, and serde deserialises a JSON array into a struct positionally — so the first event ({"type": "system", "subtype": "init", ...}) was fed to is_error: bool:

invalid type: map, expected a boolean at line 1 column 1

Both callers report that as an authentication failure. The claude-code provider appends "Is claude logged in? Try claude login.", so a correctly configured, logged-in session looks like a broken login, and forgeplan reason — with every Standard+ FPF ADI gate behind it — is dead.

The change

parse_envelope now parses stdout into serde_json::Value first and unwraps an array to its last type: "result" event before decoding. The bare-object path is untouched, so older CLI versions keep working — the array branch is additive.

An array carrying no result event returns a named error rather than positionally mis-parsing some other event:

`claude --print --output-format json` returned an event array with no `type: "result"` event

One function changed. parse_envelope is pub(crate) with two consumers, so this fixes both the claude-code LLM provider (llm/mod.rs:247, the reported path) and playbook::dispatch's AgentDispatcher / PluginDispatcher (claude_print::invoke).

Tests

Five added, all failing before the change:

Test	Covers
parses_event_array_envelope	the CLI 2.x shape — the regression itself
parses_bare_object_envelope_via_helper	the old shape still decodes
event_array_preserves_in_band_error	is_error / api_error_status read off the result event, not event 0
event_array_without_result_event_errors_clearly	named error, no silent mis-parse
claude_code_accepts_event_array_envelope	end-to-end through the provider's mock-claude harness

The existing fixtures could not have caught this: json_response() and write_mock_claude() both emit the bare-object shape the CLI produced when #382 landed. The end-to-end test reuses write_mock_claude with an array payload, so the harness now covers the shape real users hit.

cargo fmt --check clean, cargo clippy -p forgeplan-core --lib clean, cargo test -p forgeplan-core --lib claude_print 35 passed / llm:: 34 passed.

Verified against the real CLI

forgeplan 0.37.0, claude CLI 2.1.266, macOS arm64 — same workspace, same artifact, same logged-in session:

# released 0.37.0
$ forgeplan reason RFC-001
Error: ADI reasoning failed: claude-code provider: failed to decode `claude --print` JSON envelope:
invalid type: map, expected a boolean at line 1 column 1. Is `claude` logged in? Try `claude login`.

# patched
$ ./target/debug/forgeplan reason RFC-001
  Analyzing RFC-001 with ADI cycle (claude-code/claude-sonnet-5)...
  { "hypotheses": [ { "id": "H1", ... } ] }     # full ADI cycle completes

Reproduced identically on 0.34.0, so this is not a 0.37.0 regression — it is the CLI's output shape moving out from under the parser.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YB7Md4NFwfDXiCcsy49PvX