PR #460 introduced fullwidth pipe characters (｜) in DSML tool call formatting
to improve parsing robustness, but models exposed to these fullwidth pipes in
system prompts exhibit significantly higher rates of tool output hallucinations.
Reverting to halfwidth pipes (|) drastically reduces tokenizer/perplexity-driven
hallucinations while retaining the existing confusable-hardening in the parser.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The closing tag format was <｜/DSML｜tag> but must be </｜DSML｜tag>.
The scanner's closing-tag detection checks text[1] == '/', so the
slash must come immediately after '<', before the first full-width
pipe (U+FF5C). Tags like <｜/DSML｜tool_calls> would not set
closing=true and would not match any tool markup name.
Files fixed:
- internal/toolcall/tool_prompt.go: all closing tags
- internal/promptcompat/prompt_build_test.go: 1 test expectation
Replace all strings.ToLower usage with ASCII case-insensitive matching
(hasASCIIPrefixFoldAt, indexASCIIFold, hasDSMLPrefix) to prevent slice
bounds errors when Unicode characters change byte length after case
folding (e.g., Turkish İ U+0130 → i + combining dot: 2 bytes → 3 bytes).
Root cause: code created a strings.ToLower(text) copy, found byte
positions in that copy, then used those positions to slice the
original text — byte offsets that were valid in the lowercased copy
became out-of-bounds in the original when case folding changed byte
lengths.
Files changed:
- toolcalls_scan.go: remove 5 lower usages, add hasDSMLPrefix
- toolcalls_parse_markup.go: remove 3 lower usages, add indexASCIIFold
- toolcalls_markup.go: SanitizeLooseCDATA lower removal
- toolcalls_parse.go: updateCDATAStateForStrip lower removal
- tool_prompt.go: align DSML pipe characters with tool call spec
- tool_prompt_test.go: fix pre-existing test character mismatch
Track byte sizes of inline-uploaded files during PreprocessInlineFileInputs and convert them to conservative token estimates (bytes/3). RefFileTokens is threaded through StandardRequest into all OpenAI chat/responses usage builders so returned prompt_tokens/input_tokens reflect the full upstream context cost including attached files.
Extract the compacted-context prompt string into a single function in
promptcompat and add a [context note] block to the injected file wrapper
so the model knows the attached history is compressed context, not new
instructions.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
