Repository navigation
wiring up jinja input marking #28249
Description
Note: this issue is assigned, contributors should not take it
Jinja engine already support "input marking" to avoid special token injections in user messages, but the rest of the code base was not ready for it.
TODO:
- add
mtmd_tokenize_from_partsfor fine-control per-segment tokenization (for example, one segment can be marked asparse_special, while other doesn't) - Refactor
chat.cppto output parts instead of a single string - Wire it up to server
Related issue: #24382
Reacted by Quentin McGaw
Reacted by Donny
Activity
ogwhic commentedon Oct 12, 2026
Is there a workaround for this until the implementation lands?
Ran into it when trying to edit chat templates with an agent, each time the agent read the file and started reasoning about it it'd write the chat template tags and break. Injection problem too from a security perspective.
Edit: This is a slightly separate problem. The model itself writing the tags as text, which may be unavoidable unless the distinction can be made between "special token with id" != "text matching ASCII representation of token with id". (edit edit: use markdown emphasis instead of XML delimiters, oh the irony)
qdm12 commentedon Oct 12, 2026
I did tell my ai to write a proxy server pointing it to this github issie, to fix the tag issue and it does work since then. It figured out a bunch of things and I honestly didn't care much on what was wrong as long as it worked, anyway it's a temp fix. Maybe point your agent to implement something like that?
yudagua commentedon Oct 21, 2026