Skip to content

A plugin names the adapter's files as a deployment does - #108

Merged
haizhongzheng merged 8 commits into
Infini-AI-Lab:mainfrom
ruiling-smartbear:deploy-matches-training
Nov 20, 2026
Merged

haizhongzheng merged 8 commits into
Infini-AI-Lab:mainfrom
ruiling-smartbear:deploy-matches-training

Conversation

@ruiling-smartbear

@ruiling-smartbear ruiling-smartbear commented Nov 18, 2026 •

Copy link
Copy Markdown
Contributor

Change

Roadmap §5 (#37). A hermit deploy workspace's context says where the adapter's files are; a hermit convert plugin did not.

  • Claude Code plugins: a SessionStart hook prints the installed folder and the same file list.
  • Codex, Pi, Hermes plugins: the files are listed under unnamed_files.
  • A test that training deploys an adapter exactly as hermit deploy does, on every host.

Validation

tests/core/test_deploy.py, tests/core/test_convert.py; claude -p --plugin-dir with the converted example names the folder and files.

Review

  • Tests and documentation cover the affected behavior, where applicable.
  • Commands, defaults, and public interfaces are consistent.
  • Shared logs and examples contain no credentials or private data.
  • Documentation-only: not applicable.
  • Independent review (post-PR audit): findings fixed.

Maintainer update

  • Merged main (after the task-environment refactor, Task/environment refactor: task directories, Harbor as environments, every agent Hermit's (#112-#116) #117–Remove the forwards of moved modules #121): one conflict, CHANGELOG.md (both entries kept). The PR's code merged unchanged; its tests still reach hermit.agent.acting.workspace._deploy and mini_swe_agent.CONFIG_FILE, which the refactor kept.
  • Parity test tightened (tests/core/test_deploy.py): the with-tools test let a whole tool-server file differ, so training losing Hermes's skills settings passed. Now every other file must match byte for byte, the tool-server file must equal hermit deploy's once the tools' entry is removed (OpenCode, Hermes, mini-swe-agent), and each host's tool handoff (workspace server or the session's own) is asserted on its own. Mutants (Hermes skills dropped; harness_tools dropped) fail the new test; the old one missed the first.
  • Tests: test_deploy.py + test_convert.py with mini-swe-agent 2.4.6 installed: 85 passed, 1 skipped (all 7 hosts ran). pytest --changed --base origin/main locally: 1784 passed, 11 failed; all 11 fail the same way on a clean main checkout (a stale hermit install in that environment), none in this PR's files.
  • Independent review (/dual-review, two rounds, this session and Codex): the only fix was the test above. Left, recorded: a .env's first line in the list (deploy's context already lists it); the hook's output past Claude Code's 10,000-character cap (the folder note comes first); unconnected files inside skills/ or agents/ not listed (deploy's filter); nested .git or a linked directory listed differently from what is embedded; non-UTF-8 locale writes (as every plugin file before); Windows without Git Bash (documented); a converted role's fresh context not seeing the folder note; reconversion replacing hooks/ wholesale (as agents/ and skills/); no hook for an adapter with only its context.

ruiling-smartbear and others added 8 commits November 18, 2026 04:50
…loys as hermit deploy does

A Claude Code plugin prints, from a SessionStart hook, where its embedded adapter is and the files no export
connects, the list a deployment's context gives. Codex, Pi and Hermes plugins, which cannot say where they are
installed, list those files under unnamed_files. A test holds that training deploys an adapter as hermit deploy
does, file for file on every host, the tools' server entry aside.
…Windows wording; tighter tests

A Hermes plugin's register appends the folder note and the list of other files to its system prompt section,
within its 4,000 characters. The Claude Code README mentions the hook only when there is one, and the docs say
the hook needs a POSIX shell. Tests: a reconversion removes a hook no longer needed; the deployment records
differ only in the tools' entries.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… taken out

The test let a whole tool-server file differ between training and `hermit deploy`, so training losing
Hermes's skills settings passed, and whether a host's tools come from the workspace or the session's own
server was checked only when a file differed. Now every other file must match byte for byte, the
tool-server file must match once the tools' entry is removed (OpenCode, Hermes, mini-swe-agent), and
each host's tool handoff is asserted on its own.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@haizhongzheng
haizhongzheng merged commit 42fbd08 into Infini-AI-Lab:main Nov 20, 2026
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants