Engine sidecar floods the node log with the raw model event stream

[AbirAbbas]

What happens

With the coding engine enabled, the supervised sidecar's entire stdout — the raw
per-token/per-tool event stream — is relayed into the node's own log file with a
[pro-engine] prefix.

Measured on a single trivial build (add one function + one test) against a local
control plane:

• ~/.agentfield/logs/swe-planner.log reached 6.1 MB in ~30 minutes
• of the first 2,144 lines, 1,987 were message.part.delta events (~93%)
• individual lines run to several KB each, since tool events embed full file
  contents, diffs and reasoning traces

Why it matters

• af logs swe-planner and AgentField Desktop's log view are unreadable — the
  handful of lines an operator actually needs (startup, engine acknowledgement,
  dispatch, outcome) are buried under streaming deltas.
• Log growth is unbounded and proportional to token throughput, not to events
  worth recording. A long-running node on a busy repo will fill a disk.
• The same content is already delivered to the control plane as structured DAG
  nodes (turns, tool calls, notes), so this is duplicate storage of data that
  has a better home.

Suggested fix

Filter what the supervisor relays rather than passing the child's stdout
through verbatim. Roughly, in increasing order of effort:

1. Drop message.part.delta (and other high-frequency streaming event types)
   from the relay entirely — they carry no operator value once the turn's
   result is recorded.
2. Relay only lifecycle lines: serve/registration, per-task start and outcome,
   errors and warnings.
3. Put the full stream behind a debug switch (e.g. SWE_PRO_LOG_STREAM=1) for
   when someone genuinely needs to debug the engine.

Environment

• SWE-AF main @ 4237beb, engine enabled by default via af install
• AgentField v0.1.122
• Linux, OpenRouter-only configuration