fix(spawn): register system-prompt factory for sub-sessions so prefix-placement hooks can compose

[Brian Krabach (bkrabach)]

Summary

Delegated sub-agent sessions were paying an uncached token tax on every single LLM request, for the entire life of the sub-session.

session_spawner.py injected the agent's system instruction as a STATIC message via add_message() and never registered a system-prompt factory.

The skills-visibility hook defaults to placement="prefix", which works by WRAPPING the context module's registered system-prompt factory so the skills index rides the provider's stable cached prefix. At tool-skills/hooks.py:216-221 it explicitly REFUSES to wrap when no factory is registered — correctly, because in context-simple a registered factory takes precedence over static system messages (context-simple/__init__.py:268-271), so wrapping a nonexistent factory would silently DROP the agent's instruction.

Result: every delegated sub-agent hit the fallback path and re-injected the entire skills index as a fresh per-request message on every provider:request, outside the cached prefix, plus emitted a confusing warning blaming the context module.

The context module was never at fault. context-simple implements set_system_prompt_factory at __init__.py:195. amplifier-foundation already does this correctly in both its session paths (_prepared.py:661,664 and :874,875). app-cli's session_spawner — the path that actually runs every delegate/Task sub-agent — was the only gap.

The Fix

amplifier_app_cli/session_spawner.py:788 — prefer factory registration, fall back to add_message, mirroring the foundation _prepared.py spawn path. The instruction content is byte-identical to before (system_instruction is already resolved at that point; the factory is a closure returning that resolved string, bound to a local to avoid late-binding). No re-expansion per request — deliberately out of scope.

Verification Evidence

Test suite, run by the reviewer independently:

• baseline on main: 1297 passed, 1 skipped, 13 deselected, 1 xfailed
• on this branch: 1301 passed, 1 skipped, 13 deselected, 1 xfailed
• delta: +4, exactly the new tests. No regressions.

End-to-end mechanism proof driving the REAL context-simple module and the REAL SkillsVisibilityHook (108-skill synthetic catalog):

BEFORE FIX  (session_spawner add_message only)
  prefix placement achieved : False
  hook action               : inject_context
  per-request injection size: 2,218 chars
  skills index IN system msg: False  (system msg 43 chars)
  agent instruction present : True
  WARNINGS                  : 1
    ! visibility.placement='prefix' (the default) but the context module offers no system-prompt factory surface...

AFTER FIX   (session_spawner registers factory)
  prefix placement achieved : True
  hook action               : continue
  per-request injection size: 0 chars
  skills index IN system msg: True  (system msg 2,263 chars)
  agent instruction present : True
  WARNINGS                  : 0

Critically, agent instruction present: True in BOTH cases — the fix does not drop the system instruction, which is precisely the hazard the hook's refusal-to-wrap guard exists to prevent.

Note on magnitude: The proof used 108 synthetic skills with short descriptions (~2.2KB). A real catalog carries ~19,766 bytes of skill descriptions (~5k tokens), so the real-world per-request saving is substantially larger. The mechanism is what the proof establishes; the magnitude scales with the actual catalog.

Lint/typecheck:

• ruff check on both changed files → All checks passed (baseline and post-change)
• pyright session_spawner.py → 3 errors, all pre-existing and unrelated (lines 486, 577, 578), identical before and after
• Pre-existing ruff format debt in test_session_spawner.py present in baseline too

Also checked for the same bug elsewhere in the package: The only other add_message({"role": "system"...})-shaped call site is the resume/transcript-restore path (~line 1298), which replays a saved transcript message-by-message. Different semantics, not the same bug, deliberately left unchanged.