Fix /loop false-positive stall detection

[androidand]

Proposal

Problem

The /loop command stalls too aggressively. A loop with noProgressLimit: 10 (default) and NoProgressSimilarityThreshold: 0.92 will stall when:

1. Iterations have tool calls but near-identical output (bigram similarity >= 0.92)
2. Output is genuinely different but the bigram similarity crosses 0.92 due to structural similarity (e.g., long file contents, boilerplate)

The current logic at packages/opencode/src/loop/loop.ts:548-569:

const limit = updated.info.noProgressLimit
const noToolCalls = result.toolCalls === 0
const nearIdentical =
  updated.lastOutput !== undefined &&
  similarity(result.output, updated.lastOutput) >= NoProgressSimilarityThreshold
const streak =
  noToolCalls && (updated.noProgressStreak === 0 || nearIdentical) ? updated.noProgressStreak + 1 : 0

This counts a streak only when noToolCalls is true AND either first iteration or near-identical output. The problem is:

1. Tool calls should reset the streak unconditionally — if the model is making tool calls, it's progressing, regardless of output similarity
2. The bigram similarity threshold (0.92) is too sensitive — outputs that differ in small but meaningful ways (e.g., a single word change in a long file) will still score 0.92+

Root cause

The streak increments only when noToolCalls is true. This means tool-call iterations already don't contribute to the stall streak. The real issue is likely:

• The similarity function is too aggressive at declaring outputs "near-identical"
• The default noProgressLimit: 10 is too low for complex tasks that produce structurally similar output across iterations
• The bigram approach doesn't account for semantic changes in long outputs

Fix

1. Calibrate the similarity threshold — increase NoProgressSimilarityThreshold from 0.92 to 0.96 or higher, or switch to a more robust similarity metric
2. Increase default noProgressLimit — consider raising the default from 10 to 15 or making it adaptive based on output length
3. Add a minimum output-change heuristic — if the output length changes by more than X%, consider it progress even if bigram similarity is high

Verification

• Add test cases for: loop with tool calls across iterations, loop with structurally similar output, loop with semantic changes in long output
• Ensure existing stall behavior is preserved for truly stalled loops (identical output, no tool calls)

Disposition (2026-09-18)

Shipped. Shipped: similarity threshold 0.96 + tests.

Tasks

• 1. Audit the current no-progress stall logic in packages/opencode/src/loop/loop.ts
• 2. Write failing test cases for false-positive stalls
• 3. Calibrate the similarity threshold and default noProgressLimit
• 4. Update the stall logic to be more robust
• 5. Add regression tests
• 6. Verify with real /loop usage

Plan changes

6 done

[androidand]

Archived 2026-09-18 — see the change's Disposition section.