ask: tell the worker to write answer.md, matching the check

[elRafa]

Fixes #102.

The ask packet preamble never tells the worker about the answer.md contract that the check (test -s answer.md) and expect_files demand — and with max_attempts=1 the retry that would surface the filename never fires. Any engine that answers inline (i.e. all of them, absent an instruction to write a file) fails every run. The existing tests can't catch it because their mock worker writes answer.md unconditionally.

Change: one sentence in the preamble instructing the worker to write the final answer to answer.md and also print it. The print keeps the raw-output/token-regex path unchanged.

Verified: opencode engine, openrouter/z-ai/glm-5.2 — deterministic FAIL before (two runs, both machines), PASS after (both machines; worker tool-calls the file write on attempt 1). Existing test_ask_command.py suite unaffected since the mock worker path doesn't parse the preamble.

🤖 Generated with Claude Code

[elRafa]

Hey Nate — friendly ping on this one after a month, since it's been quietly earning its keep over here.

We've been running this patch in production on two Macs since early August — every ask goes through it, 40+ logged runs across a couple of engines (GLM via OpenCode, and lately Apple's on-device Foundation Models through a little custom engine), no regressions. I rebased the branch onto current main today (474c4db), so it should apply clean.

If you'd rather see the contract handled differently — say, injecting the filename from expect_files instead of a sentence in the preamble — happy to rework it, just point me in a direction.

No rush at all; the fork is easy to maintain. Mainly flagging it because the bug still bites anyone trying ask out of the box on a real engine (#102 has the full writeup).

🤖 Generated with Claude Code