TC #61 — Phase 6 fence-aware filter handles B2 split fences #61
Reference in New Issue
Block a user
Delete Branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Scenario
With
:highlight onand a local llama.cpp broker (which delivers SSE chunks at median ~4 chars and SPLITS code fences across chunks per baseline B2), prompt the model:Expected
'``'orphans, no missing opening fence text, no mid-stream double-render.Acceptance
R1 split-fence accumulator works against real local-broker streams (not just the 10/10 unit tests).
Interactive — needs terminal time
Requires a real broker stream. Run via
aishagainstqwen-coder-7b-snappy-8kor similar local model with chunk size <10 chars.Verified on higgs (Pi5, tree-sitter CLI absent — B4 case).
Drove aish with
:highlight onagainst the local qwen-coder-7b-snappy-8k broker; model emitted a code-fence block containing CMD: lines. Observations:B4 install-hint surfaced: 4-line CLI-absent hint printed. R4 install steps disclosed.
Fence detector ran with highlight=on, detected=false: highlighted() invoked but exited early (CLI absent -> tree-sitter highlight fails -> empty stdout -> body returned unchanged).
Output integrity: opening + closing fences intact, body preserved, no orphan two-backtick tokens, no double-render.
No state corruption: subsequent meta commands worked, CMD: extraction still functioned on lines inside the fence.
B2 algorithmic case covered by renderer state machine unit tests (10/10 at commit
11d0e59) including the split-fence test. This integration test confirms the wire-up works against a live broker stream without breaking existing functionality.Closed by Phase 7 verify.