multi-agent-skill-silent-phase-compression

Subagent outputs land verbatim in the orchestrator's chat history as Agent tool results, accumulating 200k+ tokens across phases. Under this context pressure the orchestrator rationalizes compressing mandatory phases like reflection and debate into the judge's integration.

Check the process log for headings like 'Phase 4-5 (Compressed)' with summaries instead of per-reviewer outputs, look for a missing 'Phases run:' manifest in the report header, and re-run the skipped phases to see if net-new findings surface.

Use file-based state passing: have each subagent Write its full output to a state file on disk and return only the path, a 100-word summary, and a one-line top finding. This drops orchestrator cost from ~215k to ~3k tokens for a 5-reviewer panel.

No. Each subagent has its own context window, but their full outputs still return verbatim into the orchestrator's chat history. The pressure accumulates in the orchestrator, not the subagents, which is what triggers compression.

Only when the phase has an explicit skip condition in the spec and that condition is met, such as skipping targeted verification when zero unresolved disputes remain. Mandatory phases like reflection, debate, and blind finals must never be silently compressed.

Audit the state directory for one file per reviewer-phase cell, confirm Agent tool results are capped near 150 tokens, check the report header lists all mandatory phases with no COMPRESSED warning, and stress test with 7+ reviewers.