fix: complete goal resume and comment input follow-ups (#3361)

This commit is contained in:
Bohdan Triapitsyn
2026-09-05 10:53:34 +03:00
committed by GitHub
parent f3f844463b
commit 7c9fdd1ee5
9 changed files with 172 additions and 62 deletions
@@ -124,7 +124,10 @@ before touching the filesystem). Rationale: metadata rides every
The consecutive state is derived from the loaded message history, not
persisted, using `info.time.created` chronology rather than message IDs.
Summary messages are not agent turns; an ordinary completed assistant
turn naturally breaks the consecutive condition;
turn naturally breaks the consecutive condition. Explicit Resume grants
one new recovery attempt over the same transcript; the continuation
consumes that permission, so another truncation blocks again. Resume
does not bypass assistant errors or the token budget;
- otherwise, small-model audit of the objective + the last assistant turn
only — no conversation history and no continuation prompts
(`restrictToPreferredProvider`, session's own provider/model preferred):