Repository navigation
fix(provider): keep carried pieces apart in Chat, so a switched-from thought never runs into its answer - #48
Merged
Conversation
…thought never runs into its answer A model switch carries a finished thought over as text. Messages, Responses and Gemini give each carried piece its own block, but Chat's assistant content is one string, and its degraded encoder appended the pieces with nothing between them: the next model read "...in one sentence.Muse Spark answer 1". Found driving stage 28 in a terminal against two loopback models. Carried pieces now sit a blank line apart. The user chose that over a list of text parts, which some compatible gateways refuse on an assistant message.
benjipeng
added a commit
that referenced
this pull request
Sep 23, 2026
## What Closes stages 16 and 28. The user tested both in their terminal on 2026-09-23 against two loopback models: `/effort max` applied at the idle boundary and reached the next request as `max`; switching to another model carried the first model's finished thought across as text, a blank line from its answer (#48), with the notice once; switching back replayed the original reasoning in its own field. Both plans are deleted; the phase file and roadmap say complete. The one decision only stage 28's plan held, rejecting a second press to arm a switch, now sits beside MDL-1. Live-provider acceptance of degraded shapes stays marked unproven in the phase file. ## Verified Local: citations, frames, typos and the pre-commit gates. The request log from the user's run shows the efforts, wire ids and message shapes above. Documents only. CI not yet observed.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Stage 28 found in the terminal: after a model switch, Chat Completions glued a carried thought onto the answer with nothing between them (
...in one sentence.Muse Spark answer 1: ...), the concatenation the stage's plan meant to prevent. Messages, Responses and Gemini already keep each carried piece as its own block; Chat's content is one string, so the pieces now sit a blank line apart. PRV-3's degradation paragraph says so, with the list-of-text-parts alternative rejected: some compatible gateways refuse it on an assistant message.Verified
Local:
prv_3_chat_keeps_carried_pieces_apart_with_a_blank_line, which fails with the glued string on the old encoder; all provider crate tests, clippy, fmt, citations, typos. Found by driving the real binary through/effort max, a reasoning reply and/modelin a PTY against two loopback models. CI not yet observed.