[brdloush]

Related to #615.

Problem

After an auto-compaction, ECA does not auto-compact again until the user turn ends. A long turn can fill the context a second time and end with Context window exceeded. Steering makes this common, because steered messages join the running turn.

Real case: 262k context, 75%. One compaction at ~196k, then 6 steered messages in the same turn, no second compaction, and finally Input length (267765) exceeds model's maximum context length (262144).

Cause: resume-after-compact! sets :auto-compacted? true on the chat-ctx, and every later tool-call round of the turn reuses it. So the threshold check and the context-overflow fallback stay blocked for the whole turn. The flag comes from #391 (stop a compact → overflow → compact loop). That goal is right, but it only needs to cover the request right after a compaction.

Changes

• Per-compaction guard. The resumed turn gets a :post-compact-guard* atom. It is released on the first usage update that shows the compacted context under the threshold. After that, the turn can compact again. If the compacted context is still too big, the guard stays on, so the ECA becomes unresponsive when tool output causes LLM context overflow #391 protection holds.
• autoCompactMaxPerTurn (default 3, global or per agent). Each compaction sends the whole context to the LLM, so a turn that keeps refilling it (e.g. re-reading a huge file) must not compact without end. When the limit is reached, the turn ends with the normal overflow error.
• The "Continue with the task" prefix no longer stacks up. Docs and schema are updated. A separate commit fixes a stale autoCompactPercentage: 85 in the default config docs (the default is 75).

Alternatives

• Remove the guard: when the compacted context is itself too big, every request would compact again, which brings back the ECA becomes unresponsive when tool output causes LLM context overflow #391 loop.
• Guard in the db, like :compacting?: it would need clearing at every turn end, including errors and stops. In chat-ctx it belongs to the turn by design.

Tests

• auto-compact-again-later-in-same-turn-test and context-overflow-later-in-same-turn-auto-compacts-test: a second compaction in the same turn (both fail on master), and no loop when the compacted context is still too big.
• auto-compact-limit-per-turn-test: stops at the limit, including autoCompactMaxPerTurn 1.
• context-overflow-auto-compact-guard-test now goes through the real resume-after-compact!.

A/B with the real server against a fake LLM: master overflows after one compaction. This PR compacts twice and finishes the task, and a longer task stops at 3 compactions.

• I added a entry in changelog under unreleased section.
• This is not an AI slop. (I have read, simplified and pinpointed the important parts myself, the human 😄)

🤖 Generated with Claude Code

[ericdallo]

thank you!

[brdloush]

The pleasure is all mine. I was just checking some long running goals my agent was doing yesterday evening and indeed it survived those multi-compactions-per-turn thanks to this patch. I guess I'll be able to go back to master branch soon 👍