chore: dogfood APM by migrating primitives from .github/ to .apm/

[danielmeppiel]

Motivation

Eat our own dogfood: microsoft/apm now uses APM to manage its own AI primitives. This PR:

1. Validates fix: deploy root project .apm/ primitives without a sub-package stub #715 by exercising apm install --target copilot end-to-end on the real repo with 26 primitives.
2. Unblocks a future gh-aw automation PR that will declare microsoft/apm as a dependency in CI workflows so apm-action can pack/restore the primitives bundle (see dogfood-layout-spec.md sections 3-4 for the consumer-side install path).
3. Establishes the source-of-truth model: .apm/<type>/ is authored; .github/<type>/ is the deployed/regenerated artifact that stays committed for raw Copilot CLI users without APM and for PR review visibility.

Migration matrix

Type	Count	Source (authored)	Target (deployed)
Skills	8	.apm/skills/<name>/SKILL.md	.github/skills/<name>/SKILL.md
Agents	10	.apm/agents/<name>.agent.md	.github/agents/<name>.agent.md
Instructions	8	.apm/instructions/<name>.instructions.md	.github/instructions/<name>.instructions.md

Note on the 8th instruction: .apm/instructions/python.md was previously dead content (the integrator globs *.instructions.md only — instruction_integrator.py:34-42). It is renamed to python.instructions.md so it now deploys properly.

Deleted as legacy (pre-dated the skill/agent primitive model, no source/test/workflow references):

• .github/prompts/*.prompt.md (7 files)
• .github/chatmodes/oss-architect.chatmode.md (1 file)

Pack verification approach

apm pack from microsoft/apm's own root is a no-op today (pack_bundle only walks dependency deployed_files, not lockfile.local_deployed_files). That is fine for the gh-aw flow: gh-aw declares microsoft/apm as a dependency in a fresh CI workspace, runs apm install (which routes microsoft/apm through FreshDependencySource -> integrators -> .github/<type>/), then runs apm pack — at which point the deployed files ARE in the dependency's deployed_files and ARE bundled. See bundle/packer.py:122-135 and dogfood-layout-spec.md section 3.

Byte-equality regression evidence

After apm install --target copilot --verbose against the migrated .apm/:

$ diff -ruN .dogfood-baseline/skills        .github/skills        && echo "[+] skills OK"
[+] skills OK

$ diff -ruN .dogfood-baseline/agents        .github/agents        && echo "[+] agents OK"
[+] agents OK

$ diff -ruN .dogfood-baseline/instructions  .github/instructions
diff -ruN .dogfood-baseline/instructions/python.instructions.md .github/instructions/python.instructions.md
--- .dogfood-baseline/instructions/python.instructions.md   1970-01-01 01:00:00
+++ .github/instructions/python.instructions.md             2026-04-21 23:45:24
@@ -0,0 +1,8 @@
+---
+description: Python development guidelines
+applyTo: '**/*.py'
+---
+
+Use type hints for all function parameters and return values.
+Follow PEP 8 style guidelines.
+Write comprehensive docstrings.

All 7 pre-existing .github/instructions/*.instructions.md files round-trip byte-identical. The single delta is the newly-resurrected python.instructions.md (renamed from the dead python.md), which is the documented spec outcome (section 7).

Lockfile shape

$ grep -E "^local_deployed_files:|^dependencies:" apm.lock.yaml
dependencies: []
local_deployed_files:

26 entries under local_deployed_files (8 agents are wrong, sorry: 10 agents + 8 instructions + 8 skills = 26). No synthetic dependency. SHA-256 hashes recorded under local_deployed_file_hashes for collision/staleness detection.

Test evidence

uv run pytest tests/unit tests/test_console.py: 4770 passed, 1 deselected. The single deselection (test_user_scope_skips_workspace_runtimes) is a pre-existing failure on main (verified by re-running with my branch stashed) that is unrelated to this migration — it asserts MCP scope filtering behavior. Out of scope for this PR.

Out of scope (NOT in this PR)

• .github/copilot-instructions.md migration — tracked separately by apm compile should emit .github/copilot-instructions.md (dogfood gap) #792.
• .apm/AGENTS.md, .github/AGENTS.md, .github/CODEOWNERS, .github/ISSUE_TEMPLATE/, .github/aw/, .github/pull_request_template.md, .github/release.yml, .github/workflows/ — repo infrastructure, not APM-managed primitives.
• Extending pack_bundle to walk lockfile.local_deployed_files for self-publishable repos. The gh-aw consumer flow does not need it (see "Pack verification" above). File a follow-up if a future use case requires apm pack from microsoft/apm's own root.
• Fixing pre-existing non-ASCII characters in some migrated .instructions.md files (em dashes, en dashes, curly quotes). Pre-existing per spec section 7; out of scope per repo code-change rules.

Note on #792's "gitignore" suggestion

#792 floats the idea of gitignoring the deployed .github/<type>/ artifacts. This PR explicitly does not adopt that fork. The .github/<type>/ deployed/compiled output should always remain committed because:

1. Raw Copilot CLI users without APM installed still need primitive discovery — Copilot reads from .github/skills/, .github/agents/, .github/instructions/ directly. Gitignoring breaks the contributor-onboarding zero-tools path.
2. PR reviewers need to see deployed output in the diff to catch link-resolver rewrites, target-format drift, and integrator-introduced bugs. Hiding the artifact hides regressions.
3. CI staleness guard is the correct enforcement mechanism, not gitignore. A simple workflow that runs apm install --target copilot and asserts no diff catches drift without compromising reviewability.

I'll restate this on #792 as a comment.

Related issues

• fix: deploy root project .apm/ primitives without a sub-package stub #715 (self-install dogfood)
• apm compile should emit .github/copilot-instructions.md (dogfood gap) #792 (copilot-instructions.md migration; this PR is a stepping stone)
• [BUG] the .github/ directory as a signal for auto-detecting AI agent isn't ideal #805 (gh-aw consumer flow)

Reviewer checklist

• Confirm apm install --target copilot runs cleanly from a fresh checkout
• Confirm .github/<type>/ is byte-identical to pre-migration content (modulo the documented python.instructions.md resurrection)
• Confirm apm.lock.yaml local_deployed_files shape
• Spot-check one skill, one agent, one instruction file content survived intact

[Copilot]

Pull request overview

This PR dogfoods APM in microsoft/apm by migrating the repo's own AI primitives into .apm/ as the authored source of truth, while keeping .github/{skills,agents,instructions}/ as committed deployed artifacts regenerated by apm install --target copilot.

Changes:

• Add root apm.yml and apm.lock.yaml to track locally deployed primitives via local_deployed_files (+ hashes).
• Introduce authored primitives under .apm/ (agents, skills, instructions) and add the newly-deployed .github/instructions/python.instructions.md.
• Remove legacy .github/prompts/*.prompt.md and .github/chatmodes/oss-architect.chatmode.md, and document the migration in CHANGELOG.md.

Show a summary per file

File	Description
apm.yml	Adds a root manifest to make the repo an APM project with no external deps.
apm.lock.yaml	Records locally deployed .github/{agents,instructions,skills} artifacts + hashes for drift detection.
CHANGELOG.md	Adds an Unreleased entry describing the dogfood migration and legacy deletions.
.github/instructions/python.instructions.md	Adds the resurrected Python Copilot instructions file in deployed form.
.github/prompts/*.prompt.md	Removes legacy prompt files no longer used by the repo's primitive model.
.github/chatmodes/oss-architect.chatmode.md	Removes a legacy chatmode artifact.
.apm/agents/*.agent.md	Adds authored agent primitives that deploy into .github/agents/.
.apm/skills/**/SKILL.md	Adds authored skills that deploy into .github/skills/.
.apm/instructions/*.instructions.md	Adds authored instructions that deploy into .github/instructions/.

Copilot's findings

• Files reviewed: 37/38 changed files
• Comments generated: 1

[Copilot]

CHANGELOG entry does not follow the repo's Keep a Changelog rules: entries should be one line per PR and end with a PR reference like "(#NNN)". These bullets are multi-sentence and missing the required PR number suffix; please reformat accordingly (and keep code references in backticks).

[danielmeppiel]

Addressed in 5ad9d89 — reformatted both entries to one line per PR with the (#823) suffix and backticked code refs, per .github/instructions/changelog.instructions.md. Thanks @copilot-pull-request-reviewer for catching this.