Bounded delegation: cap in-flight workers by shared host memory, not only coordinator capacity #238

What: Bounded delegation should tell a coordinator to size parallel workers by the shared host's free memory when all children run on one machine.
Why: On 2026-10-19 00:07 Warsaw, a Chromeria backlog coordinator ran about 20 delegated workers on bestia (31 GB), most running test suites; systemd-oomd killed the shared Codex scope (664 processes) and eight runs were lost and had to be restarted.
So what: An agent edits skills/operations/references/bounded-delegation.md ("Prove and size the workflow") with one sentence; no user decision needed.

Outcome

Coordinators check shared-host memory and other running agents before each parallel wave.

Problem Statement

Current guidance: "Bound in-flight work by the coordinator's capacity to validate and integrate, not the host's maximum worker count." It names coordinator capacity only; nothing prompts checking host memory or other coordinators sharing the host. Evidence: bestia load average 31, available memory near zero; journalctl --user: t3code-codex-754631c5.scope: systemd-oomd killed 664 process(es).

Solution

Add to "Prove and size the workflow": when children share one host, also bound concurrency by that host's free memory and the agents already running there; check before each wave and shrink after a resource kill.

Elon record

• Requirements and who asked: the user's 2026-10-18 backlog request ran many parallel workers; this run's OOM exposed the gap.
• Deleted: a new tool or scheduler; one sentence suffices.
• Bottleneck: shared host memory, not coordinator review capacity.
• Checked myself: read bounded-delegation.md in full and the bestia journal entries above.

Acceptance criteria

• The sentence lands in bounded-delegation.md without duplicating it elsewhere.

Non-goals

• Per-model worker limits, automation or a scheduler.

Blockers

• None.

Required proof

• The edited reference reads correctly in context; real-world effect stays unverified until a later parallel run.

Filed by the Chromeria backlog-clearing coordinator (Claude Opus 5.5 in T3 Code).