Skip to content

[bug]: bl skill init / skill remove ignore --dry-run and perform real changes #211

Description

@walle0624

Environment

  • CLI: bl 1.26.0 (npm install; upgraded from 1.3.2 via bl update today)
  • Skill: 1.26.0 (bailian-cli, installed by bl skill init)
  • Node: v22.22.2
  • OS: darwin 25.6.0 (macOS)
  • Region: cn

Summary

--dry-run is listed under "Global Flags" in bl skill init --help, bl skill remove --help (and add / update), but the whole skill command group ignores it: bl skill init --dry-run performed a full real installation (10 skills fanned out into 56 agent directories), and bl skill remove --name <x> --dry-run really deleted the skill. Exit code 0 in both cases.

Reproduce

# 1. "Preview" what init would do -> actually installs everything
bl skill init --dry-run

# 2. "Preview" a removal -> actually removes it
bl skill remove --name bailian-web-search --dry-run

Expected

With --dry-run, nothing is written or deleted; the command prints the plan (skills, target agent directories, link paths). Alternatively, if dry-run is not supported for these commands, the flag is rejected with a USAGE error and not advertised in their --help.

Actual

bl skill init --dry-run printed:

{
  "status": "success",
  "skills": [
    "bailian-cli", "bailian-finetune", "bailian-gen", "bailian-managed-agent",
    "bailian-protocol", "bailian-sandbox", "bailian-web-search",
    "bailian-docs-llm-wiki", "bailian-model-recommend", "bailian-train-deploy"
  ]
}

and on disk, all within the same second:

  • ~/.bailian/skills/<10 skills>/ were created (~33 MB, mostly bailian-docs-llm-wiki) and ~/.bailian/skills/skill-lock.json was written.
  • Symlinks were fanned out into 56 detected agent skill directories (~/.claude/skills, ~/.agents/skills, ~/.codex/skills, ~/.gemini/skills, ...).
  • The pre-existing real directory ~/.agents/skills/bailian-cli (installed from GitHub via npx skills add modelstudioai/cli and refreshed by bl update a few minutes earlier) was replaced by a symlink to ~/.bailian/skills/bailian-cli. I did not find a backup of the replaced directory.
  • ~/.bailian/telemetry.jsonl recorded a plain {"command":"skill init", ..., "params":{}} entry with no dry-run marker.

bl skill remove --name bailian-web-search --dry-run printed:

{
  "skills": [
    { "name": "bailian-web-search", "status": "removed", "removedLinks": 56 }
  ]
}

and the canonical directory plus all 56 links were really gone; a subsequent real bl skill remove --name bailian-web-search reported "reason": "not installed".

Likely cause

At HEAD, packages/commands/src/commands/skill/{init,add,remove,update,list}.ts never read ctx.settings.dryRun (there is no occurrence of dryRun anywhere under skill/). init.ts calls installSkillWithFanout(...) unconditionally and remove.ts calls unlinkSkillFromAgents(...) / removeSkillDir(...) unconditionally, while the global --dry-run flag is still rendered in every bl skill ... --help.

Why it matters

  • --dry-run is the documented way for agents to preview a mutating command before asking the user for consent (bailian-protocol). Here it silently performs the mutation, so an agent that follows the protocol ends up installing 10 skills into every detected agent app on the machine, which also changes default routing in those agents (e.g. bailian-gen / bailian-web-search claim generation / search requests by default).
  • bl update (npm path) and the automatic update check both run bl skill init afterwards (see also Allow opting out of auto-update and of the automatic bl skill init after updates #200), so there is currently no non-mutating way to see what a skill install/update is going to write.

Suggested fix

Either honor dryRun in the skill group (print planned installs/removals with target agent directories and link paths, write nothing), or make these commands reject --dry-run with a USAGE error and drop it from their --help so it cannot be mistaken for a safe preview.

Already tried

  • bl update done today; skill version (1.26.0) aligned with CLI (1.26.0)
  • bl auth status OK (not relevant: skill commands are [No Auth])
  • Different network / region: not applicable, purely local operation

Notes

  • Frequency: Always (reproduced with both skill init and skill remove)
  • Invoked via: Agent (Claude Code)
  • Registry: https://bailian-wiki.oss-cn-hangzhou.aliyuncs.com/skills

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions