Skip to content

ASE 2.0.7

Choose a tag to compare

@rse rse released this 04 Nov 13:24
· 129 commits to master since this release

2.0.7 (2026-11-04)

  • FEATURE [code, docs]: Model Selection of Agents
    The new definition option -M|--model of all agents (also model of the definition files and
    of the sections default and per agent of a fleet) selects the model of the agent tool, either
    for all tasks, or as <size>=<model> pairs per Effort: size of the task (like
    XS=haiku,*=sonnet), while the new hook model of a definition can override it per task.

  • FEATURE [code, docs]: Partial Staging Skill
    The new skill ase-repo-stage stages only the working copy changes belonging to a topic, the
    last CHANGELOG.md entry (--entry), the last turn (--turn), or the session (--session),
    splitting mixed hunks, checks that the staged state builds on its own (optionally in a temporary
    worktree), and optionally chains into ase-repo-changelog and ase-repo-commit.

  • IMPROVEMENT [code, docs]: Verification Levels for Edit Skills
    The boolean option --verify|-v of ase-code-edit and ase-spec-edit now takes a level:
    none (former default), light (new default: lint, build, and regression tests resp. SpecBook
    lint), or full (additionally ad-hoc manual tests), also steering the grilling questions. Both
    skills now fix and re-verify for at most ten rounds and list the remaining failures.

  • IMPROVEMENT [code, docs]: ChangeLog Entries in File Style
    The skill ase-repo-changelog now follows the style of the existing entries instead of a fixed
    60-80 character summary, keeps each entry within about 5 lines, writes the section via Edit
    (so its changes show up as a diff), and triggers on prompts like "add a changelog entry".

  • IMPROVEMENT [code]: Free-Text Replies in Next-Step Dialogs
    The next-step dialogs of ase-task-implement, ase-task-integrate, ase-task-grill,
    ase-task-reboot, and ase-task-condense now accept a free-text reply as an implicit DONE:
    the skill finishes, and the text is then handled as a regular request, instead of the dialog
    being re-opened with an "Invalid option selected!" error.

  • IMPROVEMENT [code]: Commit Message Skill Activation
    The frontmatter description of ase-repo-commit now lists trigger phrases (like "write" or
    "propose" a "commit message"), so the skill is activated under more circumstances.

  • IMPROVEMENT [code]: find(1) Fallback for Source Resolution
    ase-code-lint now resolves directory and wildcard source references with find via Bash
    (newly allowed) if the Glob tool is not available, to support foreign LLMs.

  • BUGFIX [code]: Gitea Issue Dependencies Content-Type
    The Gitea task store plugin (ase-task-store-plugin-gitea.ts) now sends the issue dependency
    create and remove requests with Content-Type: application/json, fixing the HTTP 422 errors
    against Forgejo/Gitea.

  • BUGFIX [code]: Silent parallel sub-agents
    Forbid any output while concurrently dispatched sub-agents report back one by one, as Claude
    Code runs them in the background despite run_in_background: false, so skills like
    ase-code-lint and ase-docs-proofread no longer print progress messages per returned batch
    (ase-control.md).

  • BUGFIX [code]: Skill Started Message First
    ase-skill.md now requires the "skill started" message as the very first output, before any
    tool call (including ase_getopt) and also for skills invoked via the Skill tool, so that
    "do not output anything" instructions of later steps no longer suppress or defer it.

  • BUGFIX [code]: Synopsis Band Continuation Indentation
    The continuation lines of the skill synopsis in the hint band are now indented by 2 instead of
    4 spaces, so they align with the command after the prompt marker.

  • UPDATE [infr]: NPM dependencies
    Updated NPM dependencies.