#480 fix: silently strip from_* tool calls before dispatch

[hoblin]

Summary

Implements the fix for #480 based on the research-spike findings recorded in the comment on the issue.

The spike (PR #485) confirmed that spurious from_* tool calls — from_shell-runner, from_zero-width-sleuth, from_melete_goal, etc. — are LLM hallucinations, not internal leaks. The model emits them as legitimate tool_use blocks in the raw Anthropic response, complete with fabricated input schemas, despite the from_* naming convention from PR #445 + system-prompt sister block + a Level 2 memory explicitly documenting the anti-pattern. Three reproductions in ~3 minutes of a single deliberately-constructed session.

What this PR does

A small pure filter — Aoide::PhantomCallFilter.call(hash) → hash — drops every block where type == "tool_use" && name.start_with?("from_") from response["content"] before the rest of the main loop sees it. The filter has no AR, no events, no I/O — pure hash transformation, easy to spec in isolation.

It runs at the top of Events::Subscribers::LLMResponseHandler#emit, immediately after log_raw_response so the dev-only diagnostic log from PR #485 keeps observing the pre-filter API response. This matters: we want to keep measuring hallucination rate after the fix lands, and decide later whether a feedback loop (synthesised tool_result, Mneme nudge) is worth adding.

After the strip, the existing handler logic carries the rest of the work unchanged:

• normalize_tool_uses operates on the filtered content → no from_* enters dispatch.
• extract_text is unaffected by the strip → text always survives.
• The if tool_uses.any? … elsif session.may_response_complete? … branch handles the empty case naturally → from_*-only responses land at clean response_complete!.

Three edge cases the user explicitly asked to cover

Response shape	Behaviour
Only from_* tool calls	No dispatch, no failed tool_result, session → idle (clean end of turn)
from_* + legitimate tool calls	Only legitimate calls dispatch; session → executing; one tool_call message persisted
Text + from_* tool calls	Text persisted as agent_message; no dispatch; session → idle

All three are asserted in spec/lib/events/subscribers/llm_response_handler_spec.rb. The filter itself has 13 unit specs against fixture payloads (including symbol-keyed content, missing content key, non-array content, malformed blocks).

Diagnostic continuity

log/aoide.log keeps logging the raw API response and raw tool_use blocks pre-filter. Post-fix, dispatching tool=from_* lines should disappear, but from_* blocks may still appear in raw tool_use blocks: — that's the signal that the model is still hallucinating. If frequency stays high we'll know without re-instrumenting.

Test plan

• bundle exec rspec spec/lib/aoide/phantom_call_filter_spec.rb — 13 examples, 0 failures
• bundle exec rspec spec/lib/events/subscribers/llm_response_handler_spec.rb — 16 examples (3 new edge cases), 0 failures
• bundle exec rspec spec/jobs/drain_job_spec.rb — 11 examples, 0 failures (sanity on upstream emitter)
• bundle exec standardrb — clean on changed files
• bundle exec reek lib/aoide/phantom_call_filter.rb lib/events/subscribers/llm_response_handler.rb — 2 pre-existing FeatureEnvy warnings remain, none introduced
• Smoke test on dev brain — restart on this branch, reproduce the parallel-subagent scenario, confirm dispatching tool=from_* lines no longer appear and no UnknownToolError lands in the conversation. (Doing right after this PR opens.)

Closes #480

[bonk-moltbot]

🐭✅ From spike to fix — measure, then cut.

PhantomCallFilter is exactly the right shape: pure function, hash in → hash out, no AR, no events, no I/O. Immutable — .merge instead of mutating content in-place, original response untouched. 13 unit specs cover every edge: symbol keys, nil content, bare prefix, non-hash blocks, multiple phantoms, mixed real+phantom.

Placement is surgical: after log_raw_response (diagnostic data preserved pre-filter) but before normalize_tool_uses (no from_* enters dispatch). The logging spec updated accordingly — raw blocks still show from_*, but no dispatching tool=from_* lines anymore.

Three edge cases from the ticket explicitly covered in integration specs:

• from_*-only → clean idle, no failed tool_result, no wasted tokens
• from_* + legitimate → only legitimate dispatches, one tool_call persisted
• text + from_* → text kept as agent_message, session idles cleanly

The filter keeps hallucination rate observable (raw blocks in aoide.log) while eliminating the symptom (UnknownToolError, wasted round-trips). Future feedback loop (synthesised tool_result, Mneme nudge) remains a clean extension point.