Compaction returns total=0, causing infinite loop with max-steps guard #37871

Description

Environment
Version: 1.18.3
Model: big-pickle (via opencode provider)
Agent: Sisyphus - Ultraworker → Atlas - Plan Executor
Steps to Reproduce
Configure opencode with a small-context model (e.g., one with ~8K context window)
Ask an agent to work on a task that involves reading several large source files (>10K tokens total, e.g., multiple large Java service/controller files)
Observe the log
What Happens
The agent enters a permanent loop of compaction → refill → compaction:

text
Sisyphus - Ultraworker → reads files, context overflows
→ tail fallback: budget=8000, size=10000+, total=0
→ compaction (returns total=0)
→ Plan Executor → re-reads same files
→ tail fallback again...

In one observed instance:

239 loop iterations over 1+ hour
104 compaction calls
8 tail fallbacks, all with total=0
CPU pegged at 100% on one core, ~900MB memory
Session eventually reached step 157+ with no sign of stopping
Key log evidence
text
tail fallback budget=8000 size=14073 total=0
tail fallback budget=8000 size=10360 total=0
tail fallback budget=8000 size=10152 total=0
tail fallback budget=8000 size=10164 total=0
... (repeated 8 times, every few seconds)

The total=0 after every compaction means the compacted output was empty/useless, so the agent had no choice but to re-read files, re-triggering the overflow.

Expected Behavior
Circuit breaker on compaction failure: If compaction repeatedly returns total=0 (or total < some threshold like budget * 0.1), the agent should stop and report an error to the user instead of retrying indefinitely.

Max-steps guard: There should be a configurable or built-in limit on the total number of loop iterations per session (e.g., 50-100). After that, the session should abort with a clear error message.

Progressive backoff: If compaction fails, the agent could try trimming context more aggressively before compaction, or split into smaller steps with fewer files at a time.

Impact
Resource waste: 1+ hour of 100% CPU, ~900MB memory consumed for zero progress
User confusion: The agent looks like it's "working" but is actually stuck in a loop with no feedback
No recovery possible: Without external intervention (kill the process), the loop never terminates

Plugins

No response

OpenCode version

No response

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response