GeezerVison

APM Review Panel: ship_with_followups - qualified draft publication only

DRAFT / NATIVE ACCEPTANCE HOLD. Actual Claude SessionStart, SubagentStart and UserPromptSubmit witnesses and native consent remain absent. Successful CI is not native, specification or release readiness. Keep PR3290 draft and issue3191 open.

Current head: 0d4ee8095f43c241995b30c5c3b83a3c4fe50f0b; integrated base: 53ccf06d11dd2d675e8ea8c3e2523a6d321c1b79.
This ancestry-preserving continuation integrates landed Codex behavior with the prior Claude foundation. It preserves separate native registrars, queues, ledgers, canonical source-plan/hash ownership and default hashing semantics.

Independent delta disposition

Seven genuine specialists reviewed the frozen initial merge tree: architecture, supply-chain, coverage, performance, DevX, documentation and authentication. Later folds were separately qualified at the exact published head. The genuine CEO stance is ship_with_followups for qualified draft publication, not permission to merge.

• Folded lazy Claude phase loading and existing both-target installed-CLI lifecycle coverage.
• Preserved both native/static catalogs and retained-hook conformance; fixed only the merged statement budget and wrapped-output assertion without weakening their contracts.
• The CEO rejected proposed loose projection from an opaque Codex bundle because it contradicts canonical suppression of all loose projections.
• Retained factual corrections: e69283b was the earlier Claude head, not Codex; remote CI was observed after publication, not before it. Original specialist/CEO returns remain preserved. No new panel ran for this metadata closeout.

Exact-head qualification

610 focused tests (including 68 quality cases), 31 static cases, six existing official Codex metadata/content cases, and 348 conformance cases passed; two conformance cases skipped. All ten local lint/architecture/strict guards passed.

Sixteen Claude/shared mutations produced 40 failures; 41 producer Codex/API mutants produced 46 test-call failures; two additional static disablings produced 15. All 81 distinct failed IDs passed after restoration. Five Codex failures are downstream contract exceptions, not literal assertions; none are setup/import errors or skips. API source bytes remained unchanged; in-memory mutations are disclosed.

Owner precision: five deterministically detected owners map to 13 executed current-head functional IDs. Actual completion status remains blocked, so canonical verify_completion skips terminal-owner enforcement. The separately disclosed advisory-status projection is an architecture-only probe, not an actual terminal gate or readiness pass.

Observed CI

Fresh exact-head REST reads confirm successful CI 39470640291, Spec conformance 39470640316 and Merge Gate 39470640367. CI job evidence includes both Linux shards/Coverage, Windows gates, lint, architecture and APM/lifecycle/binary smoke. CodeQL, Docs, NOTICE and eligibility workflows also succeeded.

Original yellow-CI completion, validation, publication and handoff receipts remain untouched. This green observation is additive; historical e69283b results and its needs_discussion stance are not relabelled.

Remaining boundaries

#3181 alone owns native runtime investigation under the existing consent boundary. Metadata responses and substituted Claude transport do not supply runtime witnesses. No new trust, activation, credential use or native action occurred.

Unproven replay optimization, cross-dependency failure isolation and unrelated raw-local missing-manifest MCP consistency remain outside this repair. Native acceptance itself is still an unmet original criterion, not deferred or waived.

Cumulative counts remain 6 outer / 3 Copilot rounds / 2 CI recoveries, under the authorized session-only budget override. This metadata closeout adds no iteration or recovery. GitHub reports MERGEABLE / BLOCKED; the PR remains draft/open, with no auto-merge, reviewer request, merge or issue closure.

Generated by autopilot-pr-merge-worker. This comment is AI-generated and may contain errors.

Published d23d1acca67e569352ae89022f0cc8fcb7e42f31 on this same draft PR, preserving 0d4ee809 ancestry and merging main 7e20e044.

The resumed repair shares native admission with batch skill preflight, preserves same-name personal loose skills, checks local locks against original source snapshots, validates raw local MCP declarations in disposable scratch, and keeps complete Claude audit verification out of loose-deployment replay. The command companion #3291 consumed this public foundation; no private producer patch was required.

Evidence: 424 functional and 120 static passes on the merged working tree; full conformance 517 passed / 2 skipped; 46 exact-head functional cases bind all 8 detected owners; full current lint mirror and strict guards pass. Four process-memory guard mutations produced 10 intended failures with zero errors before restoration.

Current CI, spec conformance, and Merge Gate are successful at this head. The original lifecycle job completed 394 tests successfully, with 13 skips, then hit its 10-minute job deadline. Its one isolated retry passed in 511.37 seconds / 9:05 total job time. No timeout, selection or source was changed. The original cancellation remains recorded; historical 6/3/2 is preserved, with cumulative CI recoveries now 3.

Acceptance remains blocked, not ready to merge. Normal native consent for the isolated owned fixture session is still absent. Actual Claude Code 2.1.291 SessionStart, SubagentStart and UserPromptSubmit execution, distinct effective roots/resources, and runtime lifecycle-state witnesses must follow that consent through the designated runtime owner. No native probe, trust bypass or activation occurred here. Metadata, substituted transport, conformance and green CI do not replace those witnesses. Keep this PR draft and #3191 open; final qualification/merge stays with the coordinating maintainer.

Generated by autopilot-pr-merge-worker. This comment is AI-generated and may contain errors.

Native Claude check (10 minutes)

This is the merge gate for #3181: Claude Code itself must load the plugin APM registered and run its namespaced command, with ${CLAUDE_PLUGIN_ROOT} resolving to the intact plugin. An agent cannot pass Claude's trust prompt, so this needs one human run. Everything happens in a throwaway HOME, so your real ~/.claude and ~/.apm are not touched.

Requirements: macOS or Linux, git, node + npm, uv, python3, and a Claude login (or ANTHROPIC_API_KEY). The fixture is the #3291 plugin tests/fixtures/native_claude_commands: its manifest selects one command, selected/status.md, which runs node "${CLAUDE_PLUGIN_ROOT}/scripts/context.mjs", which reads a bundled resource file.

1. Isolated tools (about 3 minutes)

export CHECK="$(mktemp -d /tmp/apm-native-claude-check.XXXXXX)"
export SHA="$(git ls-remote https://github.com/microsoft/apm refs/heads/danielmeppiel-claude-command-qualification | cut -f1)"
echo "check dir: $CHECK  stack head: $SHA"

# Claude Code 2.1.291 is the only version APM's native route is qualified for.
npm install --silent --prefix "$CHECK/npm" @anthropic-ai/[email protected]
# APM from the #3290 + #3291 stack head.
if ! { uv venv -q "$CHECK/apm-venv" && uv pip install -q --python "$CHECK/apm-venv/bin/python" "git+https://github.com/microsoft/apm@$SHA"; }; then
  # Fallback if uv cannot resolve from your index: run the stack source with
  # the .venv of an existing microsoft/apm checkout.
  export APM_REPO="$HOME/Repos/awd-cli"
  git -C "$APM_REPO" fetch -q origin danielmeppiel-claude-command-qualification
  git -C "$APM_REPO" worktree add -q --detach "$CHECK/apm-src" "$SHA"
  mkdir -p "$CHECK/apm-venv/bin"
  printf '#!/bin/sh\nPYTHONPATH="%s/apm-src/src" exec "%s/.venv/bin/python" -m apm_cli.entrypoint "$@"\n' "$CHECK" "$APM_REPO" > "$CHECK/apm-venv/bin/apm"
  chmod +x "$CHECK/apm-venv/bin/apm"
fi
"$CHECK/apm-venv/bin/apm" --version   # the version line ends with the first 9 chars of $SHA

2. Isolated HOME and CLAUDE_CONFIG_DIR

export REAL_HOME="$HOME"
export HOME="$CHECK/home"
export CLAUDE_CONFIG_DIR="$CHECK/home/.claude"
export PATH="$CHECK/npm/node_modules/.bin:$CHECK/apm-venv/bin:$PATH"
mkdir -p "$HOME/.apm" "$CHECK/work"
claude --version     # expect: 2.1.291 (Claude Code)

3. Install the fixture plugin globally with APM

cat > "$HOME/.apm/apm.yml" <<EOF
name: native-claude-check
version: 1.0.0
targets: [claude]
dependencies:
  apm:
    - git: microsoft/apm
      ref: $SHA
      path: tests/fixtures/native_claude_commands
EOF
(cd "$CHECK/work" && apm install -g --target claude --only apm)

Expected output includes:

1 native Claude plugin(s) registered; existing activation preserved. New plugins
are disabled. Review and enable them in Claude; restart or reload plugins there.

claude plugin list --json
ls "$CLAUDE_CONFIG_DIR/commands" 2>&1    # expect: No such file or directory (no loose copies)

Expected: exactly one entry, "id": "apm-command-context@apm-claude", "scope": "user", "enabled": false, with readFromFolder ending in .apm/apm_modules/microsoft/apm/tests/fixtures/native_claude_commands.

4. Your consent: enable the plugin

APM never enables plugins. You do that:

claude plugin enable apm-command-context@apm-claude --scope user

5. Run the namespaced command in a real session

cd "$CHECK/work"
APM_COMMAND_PROBE_MARKER="$CHECK/witness.jsonl" claude

In Claude:

1. Accept the folder trust prompt and log in if asked (the config is fresh).
2. Type /apm-command-context:. The menu must offer status and must NOT offer unselected.
3. Run /apm-command-context:status qualified-3181. Approve the node command if Claude asks.
4. Claude prints JSON like this. The root is your $CHECK path, so it will differ:

{"root":"<CHECK>/home/.claude/plugins/cache/apm-claude/apm-command-context/1.0.0","resource":"APM_NATIVE_COMMAND_CONTEXT","argument":"qualified-3181"}

5. Type /exit.

6. Verify the witness (proof of success)

python3 - "$CHECK" <<'PY'
import json, pathlib, sys
check = pathlib.Path(sys.argv[1]).resolve()
rows = [json.loads(l) for l in (check / "witness.jsonl").read_text().splitlines() if l.strip()]
assert len(rows) == 1, rows
row = rows[0]
assert row["resource"] == "APM_NATIVE_COMMAND_CONTEXT", row
assert row["argument"] == "qualified-3181", row
root = pathlib.Path(row["root"]).resolve()
assert root.is_relative_to(check / "home"), root
assert (root / "scripts/context.mjs").is_file() and (root / "resources/identity.txt").is_file(), root
print("[+] NATIVE CLAUDE CHECK PASSED")
print("    root:", root)
PY

Success means the script prints [+] NATIVE CLAUDE CHECK PASSED, Claude showed the same JSON, and the menu had status but no unselected. Paste the two printed lines plus the claude plugin list --json output into this PR.

Failure looks like one of these: /apm-command-context:status is missing from the menu, unselected is offered, the JSON has an empty or wrong resource, root is outside $CHECK/home, or the witness file is missing. If Claude printed the JSON but no witness file exists, paste the JSON anyway; the env var may not reach the command shell.

Heads-up from the agent dry run: claude plugin details apm-command-context@apm-claude reports Skills (0). A plain plugin with no APM involved shows the same thing when its command is declared through a manifest commands path, while default commands/ files are counted. That could be a details blind spot or a real loading gap. This step settles it.

7. Clean up

(cd "$CHECK/work" && apm uninstall -g microsoft/apm/tests/fixtures/native_claude_commands)
claude plugin list --json      # expect: []
export HOME="$REAL_HOME"; unset CLAUDE_CONFIG_DIR
[ -n "${APM_REPO:-}" ] && git -C "$APM_REPO" worktree remove --force "$CHECK/apm-src"
rm -rf "$CHECK"

Then open a new shell so PATH is clean again.

What the agent already verified at stack head 73a2550d0 (no trust, enable, or session): with real Claude 2.1.291 and an isolated HOME and CLAUDE_CONFIG_DIR, step 3 registered apm-command-context@apm-claude disabled at user scope, created no loose commands, and step 7 removed the registration and the empty lock. On this machine the uv pip install path failed to resolve dependencies, and the step 1 fallback worked.

Generated by autopilot-pr-merge-worker. This comment is AI-generated and may contain errors.

APM Review Panel (delta): needs_discussion - revived, waiting on the native Claude check

Daniel decided on 2026-11-12 to revive #3290 and #3291. This run brought them up to date with main. Code, lint and CI are green. Merging still needs the #3181 proof that Claude Code actually loads and runs the command; agents cannot pass Claude's trust prompt. Daniel's 10-minute script: #3290 (comment)

What changed in this run

• Merged current main (a14ee1521, 145 commits) into fix(claude): prepare intact user plugins through native registration #3290. Two conflicts:
  • commands/uninstall/cli.py: kept main's guarded _persist_uninstall_lockfile (symlink refusal, claim binding, profile-scoped updates). It now writes an empty last-package lock instead of deleting it right away. _persist_lock_then_resync_native deletes that lock only after Claude native cleanup succeeds, so an interrupted removal can still retry.
  • integration/mcp_config_view.py: switched to main's fix(mcp): deploy MCP servers from local-path Claude plugin dependencies #3540 behavior, which reads the install-synthesized apm_modules/_local manifest, warns when it is missing, and defers under a frozen cold cache. Removed the PR's scratch-snapshot path. Moved the marketplace-integrations-claude-ownership guard and its mutation cases onto main's redirect, and aligned one unit test with main's semantics. The publisher source is still never written.
• Extracted _queue_native_claude so integrate_package_primitives stays within C901 50 after the merge.
• Restacked test(claude): protect native command context and lifecycle #3291 cleanly: its diff against fix(claude): prepare intact user plugins through native registration #3290 is still exactly its 8 test, fixture and doc files.

Folded in this run

• (panel) python-architect: _persist_lock_then_resync_native now takes deploy_root, modules_dir and scope explicitly instead of **kwargs -- resolved in ea99d5f.
• (panel) test-coverage-expert: added unit traps for the deferred empty-lock branches and for deleting the lock only after native resync converges -- resolved in ea99d5f.

supply-chain-security-expert had no findings: the empty-lock window is symlink-checked and cannot make a frozen install pass vacuously, and reading apm_modules/_local is stronger than parsing a writable source copy.

Deferred (out-of-scope follow-ups)

• (panel) Explain why claude plugin details reports Skills (0) for the fixture, then lock the answer in with an assertion. Scope boundary: this needs an authenticated, trusted Claude session, which is exactly what the manual check covers.
• (panel) Turn the persist-then-native-resync ordering into a shared native-adapter lifecycle hook. Scope boundary: cross-harness design under [EPIC] Native plugin installation and lifecycle across all APM targets #3094.

Regression-trap evidence (mutation-break gate)

• test_empty_uninstall_lock_is_kept_only_when_removal_is_deferred and test_empty_uninstall_lock_is_removed_only_after_native_resync_converges: disabling deferral gave 3 failures; deleting the lock before resync gave 2 failures; the guard was restored.

Native evidence so far (agent, no trust, enable or session)

With real Claude Code 2.1.291, an isolated HOME and CLAUDE_CONFIG_DIR, and stack head 73a2550d0:

• apm install -g registered apm-command-context@apm-claude at user scope, disabled, with readFromFolder pointing at the intact apm_modules plugin and no loose commands.
• apm uninstall -g removed the registration and the empty lock.
• claude plugin details reports Skills (0). A plain non-APM plugin shows the same for a manifest custom-path commands entry. Daniel's check settles whether that is a display gap or a loading gap.

Lint contract

The full CI-mirror Lint job is clean at the exact head: ruff check, ruff format, the YAML, file-length and relative_to guards, pylint R0801, auth signals, and architecture boundaries.

Tests

• Related suites: 9346 passed. The only failures were 60 test_unix_install_ownership.py cases that flaked under -n 6 load; that file passes 376/376 alone on both this branch and main.
• Fold re-run (uninstall, commands, Claude lifecycle, receipts, owner mutations): 2116 passed.
• test(claude): protect native command context and lifecycle #3291: 29 passed.
• Owner-touch gate: 7 owners detected; the blocked completion verified.

CI

All green at ea99d5fab: CI, Spec conformance, Merge Gate, and CodeQL. That took 0 CI fix iterations.

Mergeability status

No merge, auto-merge, reviewer request, assignment or issue closure.

Generated by autopilot-pr-merge-worker. This comment is AI-generated and may contain errors.