Regression in 1.14.48: context compaction breaks sessions with large MCP tool sets #27631

Description

Starting with v1.14.48, opencode's context compaction behavior regresses for sessions with large MCP tool definitions. Sessions that worked correctly on v1.14.47 become unusable — context usage jumps erratically (3% to 68% in a single turn), compaction fails or produces incorrect results, and the session enters a degraded state where every turn re-sends the full bloated context.

v1.14.47 and earlier handle this correctly: context stays steady at ~61% per turn with auto-compaction working as expected.

Reproduction

Environment:

• Provider: Z.AI coding plan (GLM-5-Turbo, 200K context)
• MCP tools: 206 tools x 2 transports (SSE + stdio) = 412 tool definitions (approx 123K tokens overhead per turn)
• OS: macOS 26.4.1, Apple Silicon (M1 Air 8GB)
• Install method: Homebrew

Steps to reproduce:

1. Install opencode v1.14.48 (or v1.14.49 / v1.14.50 — all affected)
2. Configure MCP servers with a large tool set (~100+ tools)
3. Start a session with any model on a provider with a moderate context window (200K tokens or less)
4. Send 2-3 messages and observe context usage

Expected behavior (v1.14.47):

• First turn: ~123K tokens (61% of 200K) — steady
• Subsequent turns: ~400 tokens (cache hit) or ~123K (cache miss) — predictable
• Compaction triggers at the configured threshold and works correctly
• Context percentage stays stable between turns

Actual behavior (v1.14.48+):

• First turn: ~3% context usage
• Second turn: jumps to 68-90%+ context usage
• Context percentage fluctuates erratically between turns
• Session degrades progressively, becoming slower and less reliable
• Eventually becomes unusable without DB intervention

Workaround

Downgrade to v1.14.47:

# Download from GitHub releases
curl -L -o /tmp/opencode-1.14.47.zip \
  https://github.com/anomalyco/opencode/releases/download/v1.14.47/opencode-darwin-arm64.zip
unzip /tmp/opencode-1.14.47.zip -d /tmp/opencode-1.14.47/

# Backup current binary
cp /opt/homebrew/bin/opencode /opt/homebrew/bin/opencode.bak

# Replace
cp /tmp/opencode-1.14.47/opencode /opt/homebrew/bin/opencode

# Prevent brew from overwriting on next upgrade
brew pin opencode

After downgrading:

• Context stays at steady ~61% per turn
• Compaction works correctly
• Session remains stable

Version Information

• Broken: v1.14.48 (May 11), v1.14.49 (May 13), v1.14.50 (May 14)
• Working: v1.14.47 (May 11)
• Both v1.14.47 and v1.14.48 are dated May 11 — the regression was introduced in the same day's release

Additional Data

Token breakdown per turn (measured with reserved=8000, 200K context):

The per-turn token cost is identical between versions. The regression is in how compaction manages the context window, not in the initial token calculation.

Related Issues

• fix(provider): classify "tokens in request more than max tokens allowed" as context overflow #27519 — Z.AI overflow error not in OVERFLOW_PATTERNS (separate but related: the missing pattern means even when compaction does trigger, the overflow error is not caught)
• Session permanently stuck after auto-compaction: post-compaction auto-trigger fires tool_use without tool_result error and is non-recoverable #27594 — compaction re-trigger bug (post-compaction auto-fire causes orphaned tool_use blocks)