Thread root agent's system_prompt into sub-agent decomposition prompts

[acdonaire]

Problem

DecompositionEngine's parallel/sequential/hybrid prompt templates are
hardcoded in English and don't inherit the root agent's system_prompt.
In AgentMode.AUTO/SUB_AGENTS, this can leak English-language reasoning
into generated subtasks regardless of the agent's configured language —
found while testing PR #150's Spanish keyword support with a real
multi-agent run.

Solution

Thread system_prompt through context (root agent -> SubAgentRouter.route
-> DecompositionEngine.decompose) and use it to add an explicit language
instruction to the LLM decomposition prompt. No-op when system_prompt
is unset.

Validation

• 137/137 existing tests pass unmodified
• Verified with a mock llm_client: prompt sent to the model now includes
  the language instruction when system_prompt is set, and is byte-identical
  to before when it isn't
• No behavior change for callers that don't set system_prompt

🤖 Generated with Claude Code

https://claude.ai/code/session_01JGxRqBhgH19YxWzGyrXqJr

[ctrl-gaurav]

Merged into main — thank you. Finding this while testing #150 with a real multi-agent run is exactly how it should have been found; it isn't visible from the unit tests.

Your commits landed with your authorship intact:

• e3c1c35 Thread root agent's system_prompt into sub-agent decomposition prompts
• c11aca9 Propagate root agent's system_prompt to spawned sub-agents

Why this shows as closed rather than merged: the commits were replayed onto main rather than merged as-is, which changed their hashes, so GitHub can't link them back automatically. The author on both is still you, so they count as your contributions.

The code went in unchanged. Two things I checked specifically, because both would have been quiet failures:

• parent is in scope at the point you use it in _run_real_sub_agent (parent = self.parent_agent, line 565) — I flagged it while reading the diff, since the signature is (self, subtask, config), and it's fine.
• All four .format() call sites on the templates pass language_note, and those templates aren't formatted anywhere else, so no KeyError is reachable.

Your "no-op when unset" claim held up: I asserted the decomposition prompt is byte-identical with no system_prompt, with context={}, and with an unrelated context key, across all four strategy branches.

One thing needed fixing before it could land (8285873): self.config.system_prompt in agent_orchestration.py pushed the mypy ratchet over its baseline — "AgentOrchestrationMixin" has no attribute "config" went 10 → 12, and scripts/mypy_ratchet.py fails the build on any increase. Rather than re-record the baseline, I declared config: AgentConfig in the mixin's existing TYPE_CHECKING block, which is what AgentStreamingMixin already does and which resolves all 12 rather than just the 2 you added. Worth running python scripts/mypy_ratchet.py locally before opening a PR — it's cheap and it's a required check.

I also added tests/unit/test_system_prompt_threading.py (12 cases, 5 of which fail against the pre-merge tree) covering the byte-identical no-op path, the note reaching all four strategies, and the 200-character truncation — that truncation is a good call and worth having pinned, since without it a long persona would crowd the task out of the prompt.

One small comment tidy in orchestrator.py (1d54274): the (PR #151) reference was removed, since shipped source here describes behaviour rather than the process that produced it. The sentence reads the same without it.

[ctrl-gaurav]

Landed on main as e3c1c35 and c11aca9. Closing — see the review comment above for what changed on top and why this reads as closed rather than merged.