Distill skill hits "Prompt is too long" when processing many conversations

Überprüfe skills/distill/SKILL.md, mit Schwerpunkt auf den Prompts der Sub-Agents und dem Sammlungsschritt des Koordinators. Führe /distill für einen Stapel von Konversationen aus, einschließlich genügend Sitzungen, um das Kontextlimit zu reproduzieren. Erledigt ist die Aufgabe, wenn Sub-Agents direkt in das journal schreiben, nur kurze Zusammenfassungen zurückgeben und der vollständige Distillations- und Aktualisierungsablauf ohne einen Fehler "Prompt is too long" abgeschlossen wird.

Problem

I have 10-25 Claude Code sessions/day. Today, the /distill skill's coordinator session started hitting "Prompt is too long" during the end-of-day/memory-maintenance runs — every API call after sub-agents completed returned this error, making the session completely unresponsive.

In my case, sub-agents were returning ~25-37KB of JSON each. With 18+ conversations to process, the accumulated results in the coordinator's context (~630KB / ~164K tokens estimated) plus the system prompt exceeded the context limit.

Root cause

Claude and I did some debugging and our understanding of the current flow is:

1. Coordinator spawns sub-agents in parallel
2. Each sub-agent reads a conversation, extracts JSON, returns it to the coordinator
3. Coordinator collects all results, then writes to journal

Step 2 seems like the problem. All returned results accumulate in the coordinator's context. Claude Code context is append-only until compaction — there's no way to discard results between turns. The more conversations you have, the more result data piles up.

Fix attempted

Have sub-agents write directly to the journal via log_journal() instead of returning results to the coordinator. Each sub-agent returns only a short summary string (e.g., "Wrote 3 actions, 2 facts, 1 decision to journal.").

The coordinator then reads from the journal (get_recent_journal) when it needs the data for entity/state file updates.

Before: sub-agent → returns 35KB JSON → coordinator writes to journal
After: sub-agent → writes to journal directly → returns ~50 bytes

Tested

Patched locally and ran against 25 conversations without issue. Previously failed with 18 "Prompt is too long" results. All 25 conversations completed successfully — distillation, state file updates, entity updates, index maintenance finished without hitting context limits.

Proposed changes to skills/distill/SKILL.md

1. Sub-agent prompt: instead of "Extract and return as JSON", instruct to call log_journal() directly with topics distilled, distilled-fact, distilled-decision
2. Sub-agent prompt: "Return ONLY a short summary line" instead of "Return ONLY the JSON"
3. Step 3 (Collect): coordinator reads from journal instead of parsing returned JSON
4. Add note: "Sub-agents MUST write to journal themselves and return only a short summary"

Happy to open a PR with the full SKILL.md change if this approach looks good.