fix(core): collapse child agent initial input into a single user message

[Rome-1]

Fixes #113

child_initial_input (strix/core/inputs.py) built a child agent's first turn as two or three consecutive {"role": "user"} messages — an inherited-context block (when a parent history exists), an identity line, and the task. That violates the strict user/assistant alternation that providers like Perplexity (sonar-reasoning) and llama.cpp enforce, aborting scans with PerplexityException - ... user or tool message(s) should alternate with assistant message(s). This collapses those parts into a single {"role": "user"} message joined with blank lines, preserving the exact same content while keeping the request well-formed for strict-alternating backends. The consumer in strix/core/execution.py already treats the return value as an opaque message list, so behavior is unchanged for every other provider.

Testing

• pytest tests/test_inputs.py — 8 passed. New tests assert a single user message (with and without inherited history), that the identity line and task survive, and a parametrized no-consecutive-same-role guard; plus added coverage for the untouched build_root_task builder. Reverting the source change in-tree makes the child-input tests fail (4 failed), confirming they are load-bearing.
• pytest (full suite) — 46 passed, no regressions.
• ruff check, ruff format --check, bandit, and pyupgrade --py310-plus are clean on the changed files. (The repo's pinned pre-commit mypy hook crashes with an internal error on the bundled openai SDK on main as well, so mypy is not asserted here.)

[greptile-apps]

Greptile Summary

This PR fixes child_initial_input in strix/core/inputs.py, which previously emitted 2–3 consecutive {"role": "user"} messages, violating the strict turn-alternation requirement of providers like Perplexity (sonar-reasoning) and llama.cpp. The fix collapses the inherited-context block, the identity line, and the task into a single user message joined with \n\n, preserving all content while satisfying strict alternation.

• strix/core/inputs.py: child_initial_input now builds a parts: list[str], appends each text segment conditionally, and returns a single {"role": "user"} dict. The surrounding call site in execution.py is unaffected.
• tests/test_inputs.py: New test file with 8 tests — single-message contract (with/without inherited history), content-preservation assertions, a parametrized no-consecutive-role guard, and coverage for build_root_task.

Confidence Score: 5/5

Safe to merge — the change is minimal and surgical, touching only the message-building logic in one function with no side effects on callers.

The fix collapses a well-understood multi-message sequence into a single message without dropping or reordering any content. The call site in execution.py treats the return value as an opaque list, so the reduction from 3 messages to 1 is transparent to all consumers. Tests confirm single-message output, content preservation, and the no-consecutive-role invariant in both the empty-history and non-empty-history cases, and the PR author verified the tests fail on a local revert.

No files require special attention.

Important Files Changed

Filename	Overview
strix/core/inputs.py	Collapses 2–3 consecutive user messages in child_initial_input into one; content and ordering are preserved, docstring added — no logic regressions visible.
tests/test_inputs.py	New test file covering child_initial_input (single-message contract, content preservation, no-consecutive-role guard) and build_root_task (empty config, repo/URL targets, diff-scope). Tests are well-structured and load-bearing.

_{Reviews (1): Last reviewed commit: "fix: collapse child_initial_input into a..." | Re-trigger Greptile}