fix: send max-steps instruction as user message, not assistant

[rguliyev]

Fixes #32548

Problem

When an agent hits its configured steps cap mid-turn, the loop forces
a text-only response by appending the "wrap up now" instruction
(MAX_STEPS_PROMPT) to the outgoing message array as a role: "assistant"
message:

messages: [
  ...modelMsgs,
  ...(isLastStep ? [{ role: "assistant" as const, content: MAX_STEPS_PROMPT }] : []),
],

This leaves the request ending on an assistant turn. Anthropic (and other
providers) treat a trailing assistant message as a response prefill, and
Claude models with thinking enabled reject it outright:

400 This model does not support assistant message prefill. The conversation
must end with a user message.

Independently confirmed in production logs across every agent/model
combination in proportion to usage, all with zero cost/output tokens on
the failing turn, all traced back to this exact isLastStep branch —
matches #32548's 100%-reproducible repro steps exactly.

Fix

The max-steps instruction is directed at the model, not written by it —
it belongs on a user turn:

...(isLastStep ? [{ role: "user" as const, content: MAX_STEPS_PROMPT }] : []),

Tool results already present in modelMsgs are untouched; only the role
of the appended instruction changes.

Testing

• Added a regression test in prompt.test.ts that configures
  agent: { build: { steps: 1 } }, drives the loop to the cap, and
  asserts the appended max-steps message has role: "user" (fails if
  reverted to role: "assistant").
• bun run lint clean on both changed files (no new warnings).
• bun turbo typecheck passes (ran via the pre-push hook, 30/30 packages).
• bun test test/session/prompt.test.ts — new test passes; the file has
  2 pre-existing flaky/timeout failures unrelated to this change,
  confirmed to occur identically on a clean checkout before this patch.

[github-actions]

Thanks for your contribution!

This PR doesn't have a linked issue. All PRs must reference an existing issue.

Please:

1. Open an issue describing the bug/feature (if one doesn't exist)
2. Add Fixes #<number> or Closes #<number> to this PR description

See CONTRIBUTING.md for details.

[github-actions]

This PR doesn't fully meet our contributing guidelines and PR template.

What needs to be fixed:

• PR description is missing required template sections. Please use the PR template.

Please edit this PR description to address the above within 2 hours, or it will be automatically closed.

If you believe this was flagged incorrectly, please let a maintainer know.

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Potential duplicate found:

• PR fix(core): send max-steps message as user role, not assistant prefill #34276: fix: send max-steps message as user role, not assistant prefill
  • fix(core): send max-steps message as user role, not assistant prefill #34276

This appears to be addressing the exact same issue — converting the max-steps instruction from role: "assistant" to role: "user" to avoid provider rejection on assistant message prefill. The title and scope match PR #36970 closely, suggesting this may be a prior attempt or related fix for the same problem.

[rguliyev]

Closing in favor of #34276, which is more complete — it also fixes the
parallel occurrence in the V2 loop (packages/core/src/session/runner/llm.ts)
that this PR missed, and has direct verification via captured-request replay.

Leaving our independent confirmation over there as supporting evidence
for #34276, since it's been open a while without a maintainer pass.