Description

When context_budget_tokens is set too tight (e.g. 18K with a 12K+ system prompt), the compaction logic enters a multi-turn loop where every single turn triggers compaction, compaction summaries grow the context further, and eventually the API returns 400 (context too large).

Reproduction

Config: context_budget_tokens = 18000, compaction_threshold = 0.65, auto_budget = false, provider = openai
Run 7+ turns with tool calls.

Observed Behavior

1. Turn 5: cached_tokens=15371, threshold=11700 → should_compact=true — compaction M0: Project bootstrap — workspace and crate skeleton #1 fires: 40 messages → summary_tokens=3104. After: cached_tokens=12329 (still > threshold).
2. Turn 6: should_compact=true again — compaction M1: Ollama chat loop — interactive CLI with LLM #2 fires: 2 messages → summary_tokens=2755. After: cached_tokens=12217 (still > threshold).
3. Turn 7: compaction M2: Skills system — SKILL.md loading and shell execution #3 fires — 400 Bad Request from API (context too large).

The root problem: compaction summaries are injected back as system messages. With a very tight budget, the system prompt + injected summaries alone exceed the threshold, so every turn triggers compaction even when there are almost no messages left to compact.

Secondary bug: compaction #2 produced a 2755-token summary for only 2 messages — the summary is larger than what it replaced, making the context worse.

Debug dump evidence

Request #3 (failed): 3 system messages (18630 + 2059 + 13957 chars), tool output 15799 chars, max_tokens=4096. Total context clearly exceeds 18K token budget.

Expected Behavior

• After compaction, if cached_tokens is still above threshold, emit WARN "context compaction could not reduce usage below threshold (compacted N messages, still at M/B tokens)" and stop attempting further compaction for that session — surface Stopping: context window is nearly full to user.
• OR: add a post-compaction cooldown: skip should_compact() for the next N turns after a successful compaction.
• Compaction summary should be bounded — if summary_tokens > freed_tokens, it was counterproductive; log WARN and don't apply.

Severity

Medium — only affects users with extremely tight context_budget_tokens settings (well below the default). With auto_budget = true (default), this doesn't occur. Manual tight budgets hit this edge case.