[BUG] `.claude/settings.json` emits Cursor-formatted hooks, bare `${PLUGIN_ROOT}`, and duplicate entries after fresh install of `microsoft/azure-skills`

[aaronrogers]

Describe the bug

After a fresh install of microsoft/azure-skills into a project with a .claude/ marker dir, .claude/settings.json contains four distinct problems:

1. Cursor's ${CURSOR_PLUGIN_ROOT} variable inside Claude Code's config
2. Bare ${PLUGIN_ROOT} (no tool prefix) that Claude won't expand
3. Cursor's camelCase postToolUse event key emitted alongside Claude's PascalCase PostToolUse
4. Every hook entry written twice

Net effect: of the four PostToolUse entries written, only one is well-formed and will execute in Claude Code; the other three are dead config — three different failure modes — and the entire postToolUse (camelCase) block is unreachable in Claude.

To Reproduce

1. Create empty project: mkdir -p ~/temp/apm-repro && cd ~/temp/apm-repro
2. Initialize APM: apm init
3. Add marketplace: apm marketplace add github/awesome-copilot
4. Create Claude marker dir: mkdir .claude
5. Install: apm install microsoft/azure-skills/.github/plugins/azure-skills
6. Inspect: cat .claude/settings.json

Expected behavior

• .claude/settings.json should contain only Claude-formatted hooks under the PascalCase PostToolUse key.
• All paths should use ${CLAUDE_PLUGIN_ROOT} (which Claude expands) or fully-resolved literal paths — never ${CURSOR_PLUGIN_ROOT} or bare ${PLUGIN_ROOT}.
• Each unique hook should appear exactly once.
• Cursor-formatted hooks belong in .cursor/hooks.json, not in Claude's settings.

Environment

• OS: macOS 26.4.1 (Apple Silicon)
• Python Version: 3.14.0
• APM Version: 0.9.4 (latest, released 2026-04-27 ~6 hours before this report)
• VSCode Version: N/A (CLI behavior)

Logs

Full .claude/settings.json after the repro above (verbatim, 48 lines):

{
  "hooks": {
    "PostToolUse": [
      {
        "type": "command",
        "bash": "${PLUGIN_ROOT}/hooks/scripts/track-telemetry.sh",
        "powershell": "${PLUGIN_ROOT}/hooks/scripts/track-telemetry.ps1",
        "_apm_source": "azure-skills"
      },
      {
        "hooks": [
          {
            "type": "command",
            "command": "bash .claude/hooks/azure-skills/hooks/scripts/track-telemetry.sh"
          }
        ],
        "_apm_source": "azure-skills"
      },
      {
        "type": "command",
        "bash": "${PLUGIN_ROOT}/hooks/scripts/track-telemetry.sh",
        "powershell": "${PLUGIN_ROOT}/hooks/scripts/track-telemetry.ps1",
        "_apm_source": "azure-skills"
      },
      {
        "hooks": [
          {
            "type": "command",
            "command": "bash .claude/hooks/azure-skills/hooks/scripts/track-telemetry.sh"
          }
        ],
        "_apm_source": "azure-skills"
      }
    ],
    "postToolUse": [
      {
        "type": "command",
        "command": "bash ${CURSOR_PLUGIN_ROOT}/hooks/scripts/track-telemetry.sh",
        "_apm_source": "azure-skills"
      },
      {
        "type": "command",
        "command": "bash ${CURSOR_PLUGIN_ROOT}/hooks/scripts/track-telemetry.sh",
        "_apm_source": "azure-skills"
      }
    ]
  }
}

apm.lock.yaml deployed_files excerpt (showing duplicates from a single install):

deployed_files:
- .github/hooks/azure-skills-copilot-hooks.json
- .github/hooks/azure-skills-copilot-hooks.json
- .github/hooks/azure-skills-cursor-hooks.json
- .github/hooks/azure-skills-cursor-hooks.json
- .github/hooks/azure-skills-hooks.json
- .github/hooks/azure-skills-hooks.json
- .github/hooks/scripts/azure-skills/hooks/scripts/track-telemetry.sh
- .github/hooks/scripts/azure-skills/hooks/scripts/track-telemetry.sh

Additional context

Plugin and marketplace

• Plugin: microsoft/azure-skills/.github/plugins/azure-skills @ e98cf5c5ee6a34ff5085a0a74693e3f224b55017
• Marketplace: github/awesome-copilot

Findings broken down

1. Cursor's variable leaked into Claude's config (settings.json lines 38, 43)

"command": "bash ${CURSOR_PLUGIN_ROOT}/hooks/scripts/track-telemetry.sh"

Claude Code does not expand ${CURSOR_PLUGIN_ROOT}. These hook commands would fail at execution time — though they're moot because they live under the postToolUse (camelCase) key that Claude never reads (see finding 3).

2. Bare ${PLUGIN_ROOT} and unrecognized keys (settings.json lines 6, 7, 21, 22)

"bash": "${PLUGIN_ROOT}/hooks/scripts/track-telemetry.sh",
"powershell": "${PLUGIN_ROOT}/hooks/scripts/track-telemetry.ps1",

Three distinct variable conventions appear in the same file (${PLUGIN_ROOT}, ${CURSOR_PLUGIN_ROOT}, and the literal .claude/hooks/... paths on lines 14, 29). Per src/apm_cli/integration/hook_integrator.py (lines 40, 267, 298), the canonical source convention APM expects is ${CLAUDE_PLUGIN_ROOT} — none of the three variants in the output match.

Additionally, bash / powershell keys at the same level as type: command don't conform to Claude Code's hook schema. This looks like a Cursor cross-platform-shell convention emitted into Claude's file.

3. Both PostToolUse and postToolUse keys present

Claude uses PascalCase event names (PostToolUse). Cursor uses camelCase (postToolUse). Both keys appear in .claude/settings.json. Claude only reads the PascalCase one, so the camelCase block is dead — and it's where the ${CURSOR_PLUGIN_ROOT} leak lives. The Cursor-formatted block should never have been written into .claude/settings.json at all.

4. Every hook entry written twice

After de-duplication, there are only two unique entries in PostToolUse and one unique entry in postToolUse, each appearing twice. The lockfile (logs section above) shows the same pattern in deployed_files. This appears distinct from the duplicate-hook bug fixed in #708 — that one was about re-running apm install appending entries. These duplicates appear after a single install, suggesting a different code path is iterating over hook sources twice (possibly: once in an "all tools" pass and once in a Claude-specific pass).

Hypothesis

src/apm_cli/integration/hook_integrator.py knows about ${CLAUDE_PLUGIN_ROOT} as the canonical source variable (line 40 comment: "resolved relative to package root, rewritten for target"). The bug appears to be in the emit path for the Claude target:

• Claude target receives entries from multiple source-format passes (Cursor format + Claude format + raw Copilot format) without filtering to Claude's schema.
• The "rewrite for target" step doesn't appear to run on at least some of those entries — ${CURSOR_PLUGIN_ROOT} and ${PLUGIN_ROOT} should both have been normalized before emission.

I haven't traced the exact code path; the symptoms are visible in the output without that.

Related

• Re-running apm install duplicates hook entries in settings.json #708 (closed 2026-04-15) — duplicates on re-install. Different mechanism but same symptom class. Worth re-checking the test coverage there to see if the fix only covered the re-install path.

Why it matters

microsoft/azure-skills is a Microsoft-published plugin in the awesome-copilot marketplace, so this affects every Claude Code user installing Azure agent skills via APM. The visible failure mode is "the hook silently doesn't run" — Claude reads the only well-formed entry, ignores the malformed duplicates, and never touches the camelCase block. End users may not realize their telemetry hooks aren't firing.

[github-actions]

Triage decision

accept

Proposed labels

theme/portability
area/multi-target
area/cli
type/bug
status/accepted
priority/high

Milestone

0.9.4

Suggested next action

Trace the Claude target emit path in src/apm_cli/integration/hook_integrator.py to filter hook entries to Claude's schema only and ensure the variable rewrite (${PLUGIN_ROOT} and ${CURSOR_PLUGIN_ROOT} -> ${CLAUDE_PLUGIN_ROOT}) runs consistently before writing .claude/settings.json.

Suggested issue comment

Thanks for this report, and especially for the level of detail -- the verbatim `.claude/settings.json` output, the `apm.lock.yaml` excerpt, the `hook_integrator.py` line references, and the cross-reference to #708 make this immediately actionable. Really appreciated.

Confirmed bug. The four failure modes you identified are all real and distinct:

1. `${CURSOR_PLUGIN_ROOT}` written into Claude Code's config (Claude only expands `${CLAUDE_PLUGIN_ROOT}`)
2. Bare `${PLUGIN_ROOT}` and Cursor-specific `bash`/`powershell` keys in `.claude/settings.json`
3. Both `PostToolUse` (Claude PascalCase) and `postToolUse` (Cursor camelCase) keys written to the same file
4. Every hook entry duplicated on a single fresh install (a different code path from the re-install duplication fixed in #708)

The root cause appears to be in the Claude target emit path in `src/apm_cli/integration/hook_integrator.py`: entries from multiple source-format passes (Cursor, Claude, raw Copilot) are merged into `.claude/settings.json` without filtering to Claude's schema, and the variable rewrite step does not run consistently before emission.

This is accepted as a high-priority bug. `microsoft/azure-skills` is a Microsoft-published package in the awesome-copilot marketplace; this silently breaks hooks for all Claude Code users installing it via APM.

If you would like to contribute a fix, the relevant entry point is the Claude target's hook emit and variable-rewrite logic in `hook_integrator.py`. A PR that ensures only Claude-schema-formatted entries reach `.claude/settings.json` -- with `${CLAUDE_PLUGIN_ROOT}` as the sole path variable -- would be a great contribution and would be reviewed promptly. Otherwise, this will be picked up in the next patch.

Per-lens notes (collapsed)

DevX UX Expert -- User-Need Reviewer

The user task is concrete and routine: install a Microsoft-published plugin via apm install and have its hooks fire in Claude Code. The bug violates the core "install adds, never silently mutates" mental model in the worst possible way -- apm install exits successfully with no diagnostic output, yet .claude/settings.json contains three different variable conventions, Cursor-specific schema keys, and duplicate entries. The failure mode is entirely invisible: Claude reads the one well-formed entry, silently ignores the malformed duplicates, and the camelCase block is unreachable. The surface is apm install -> .claude/settings.json integrity, touching area/multi-target (Cursor-format entries contaminating Claude's config) and area/cli (install command correctness).

Supply Chain Security Expert -- Risk-Surface Reviewer

No risk-surface implication for theme/security or theme/governance. The bug involves hook variable confusion and schema mismatch between targets, but does not touch lockfile integrity, credential handling, path traversal, marketplace name resolution, MCP trust, or signing. The duplicate entries in deployed_files are a data-correctness issue. The ${CURSOR_PLUGIN_ROOT} and ${PLUGIN_ROOT} variables that fail to expand could theoretically be exploited if a malicious package plants environment variables with those names, but this is outside the current threat model scope for this issue. The primary concern is that telemetry/audit hooks silently fail -- a DX consequence, not a supply-chain attack vector.

OSS Growth Hacker -- Contributor-Tone Reviewer

Author aaronrogers has author_association: "NONE" -- an external user with no prior recorded contribution to microsoft/apm. The report quality is exceptional: reproduction steps, verbatim logs, source file line references, a plausible root cause hypothesis, a cross-reference to a related closed issue (#708), and a business impact statement. This is the kind of bug report that, if met with a warm and specific reply, converts a user into a repeat contributor. The suggested comment thanks the report quality explicitly, confirms the bug is accepted with priority, and plants an open invitation to contribute the fix. The "reviewed promptly" signal is important: external contributors who file high-quality bugs need fast acknowledgement to maintain engagement.

Python Architect -- Architecture Reviewer

Not activated -- the issue is a well-scoped bug in hook_integrator.py's Claude target emit path; no cross-cutting design decision, schema change, or new primitive is required before code can land safely.

Doc Writer -- Documentation Reviewer

Not activated -- this is a bug fix restoring already-documented correct behavior (${CLAUDE_PLUGIN_ROOT} as the canonical hook variable for Claude Code); no new user-facing documentation is implied.

APM CEO -- Triage Arbiter

Accept with priority/high. microsoft/azure-skills is a Microsoft-published package; silent hook failure for every Claude Code user installing it via APM is a high-visibility correctness regression filed the same day as the 0.9.4 release. Community trust requires a prompt, specific reply and a fast-path fix. The primary theme is theme/portability -- this is fundamentally a multi-target contamination failure where the Cursor-format pass bleeds into the Claude-target output. No design decision is required; the fix is in the existing emit path. Milestone is 0.9.4 (current open patch milestone) given the scope of impact. The Growth Hacker's note about contributor retention is taken: the suggested comment explicitly invites the author to contribute the fix.

{
  "decision": "accept",
  "decision_detail": "",
  "theme": "theme/portability",
  "areas": ["area/multi-target", "area/cli"],
  "type": "type/bug",
  "status": "status/accepted",
  "priority": "priority/high",
  "preserved_labels": [],
  "milestone": "0.9.4",
  "next_action": "Trace the Claude target emit path in src/apm_cli/integration/hook_integrator.py to filter hook entries to Claude's schema only and ensure the variable rewrite runs consistently before writing .claude/settings.json.",
  "comment_markdown": "Thanks for this report, and especially for the level of detail -- the verbatim `.claude/settings.json` output, the `apm.lock.yaml` excerpt, the `hook_integrator.py` line references, and the cross-reference to #708 make this immediately actionable. Really appreciated.\n\nConfirmed bug. The four failure modes you identified are all real and distinct:\n\n1. `${CURSOR_PLUGIN_ROOT}` written into Claude Code's config (Claude only expands `${CLAUDE_PLUGIN_ROOT}`)\n2. Bare `${PLUGIN_ROOT}` and Cursor-specific `bash`/`powershell` keys in `.claude/settings.json`\n3. Both `PostToolUse` (Claude PascalCase) and `postToolUse` (Cursor camelCase) keys written to the same file\n4. Every hook entry duplicated on a single fresh install (a different code path from the re-install duplication fixed in #708)\n\nThe root cause appears to be in the Claude target emit path in `src/apm_cli/integration/hook_integrator.py`: entries from multiple source-format passes (Cursor, Claude, raw Copilot) are merged into `.claude/settings.json` without filtering to Claude's schema, and the variable rewrite step does not run consistently before emission.\n\nThis is accepted as a high-priority bug. `microsoft/azure-skills` is a Microsoft-published package in the awesome-copilot marketplace; this silently breaks hooks for all Claude Code users installing it via APM.\n\nIf you would like to contribute a fix, the relevant entry point is the Claude target's hook emit and variable-rewrite logic in `hook_integrator.py`. A PR that ensures only Claude-schema-formatted entries reach `.claude/settings.json` -- with `${CLAUDE_PLUGIN_ROOT}` as the sole path variable -- would be a great contribution and would be reviewed promptly. Otherwise, this will be picked up in the next patch."
}

---

Triage status: agentic proposal pending human ratification.
Silence is approval. Maintainers can:

• Override any label or milestone above by editing it directly --
  human edits are authoritative and will not be reverted on
  subsequent runs.
• Re-trigger triage by applying the status/needs-triage label, or
  by removing status/triaged to enroll the issue in the next
  daily sweep.

Posted by the Triage Panel workflow. See .apm/skills/apm-triage-panel for the panel skill.

Generated by Triage Panel for issue #1007 · ● 381.7K · ◷