fix: generic interrupt() resumes with the raw user value, not {"decisions"} (gh #99)

[dkedar7]

Problem

#82 and #95 fixed displaying a generic interrupt(...) payload. This closes the matching resume gap.

The CLI built the resume envelope unconditionally as {"decisions": [...]} — the deepagents/langchain tool-review protocol — and forwarded it verbatim as command.resume, which becomes the return value of interrupt(). So the canonical LangGraph "collect input from a human" pattern:

name = interrupt("What is your name?")
return AIMessage(content=f"Hello, {name}!")

received {"decisions": [{"type": "approve"}]} and produced the visibly wrong Hello, {'decisions': [{'type': 'approve'}]}!. Even "Provide custom decision (JSON)" wrapped the typed value inside the envelope, so there was no workaround.

Fix

Branch the resume on the same signal #82/#95 use to render (_is_generic_interrupt, mirroring format_interrupt_request):

• deepagents tool-review interrupt (an action_request dict keyed action, or legacy tool) → keeps the {"decisions": [...]} envelope, unchanged.
• generic / scalar interrupt(value) (a bare string like "What is your name?", or a plain dict without an action key) → resumes with the raw value the user enters, unwrapped. JSON is parsed when it parses (so a number/list/dict survives); otherwise the text is used verbatim as a string.
• --no-interactive on a generic interrupt → resumes with an empty value rather than injecting a fake approval it never asked for, keeping the run-to-completion contract.

Tests

New tests/test_generic_interrupt.py drives real agents through the CLI resume path:

• interrupt("What is your name?") resumed with Alice → Hello, Alice! (and no decisions envelope leaks).
• a structured JSON answer survives as the parsed value.
• a deepagents action_requests interrupt still resumes with {"decisions": [...]} (no HITL approval shows nothing for a generic LangGraph interrupt(...) — bare string → empty banner, non-action dict → 1. unknown; the interrupt payload is dropped (post-#69 gap) #82/HITL: a plain-string interrupt("...") shows "(no action details provided)" — the human approves blind #95 regression).
• the --no-interactive generic path avoids the envelope.
• unit coverage of the _is_generic_interrupt classifier boundary.

test_no_interactive_approve.py's fixture now uses a genuine tool-review interrupt so it keeps testing auto-approve precisely; the generic auto path is covered in the new file.

Full suite: 151 passed. ruff check and ruff format --check clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_011HWCfJii6gXd3XL3Gq3W8B