feat: add Amplifier configuration skill

[Sam Schillace (ramparte)]

Problem

Amplifier CLI configuration spans root-session provider selection, spawned-session routing, bundle composition, and local/project/global scope overrides. Users need one reliable workflow that identifies the controlling plane and current effective configuration before changing anything.

Solution

Package an inline, user-invocable /amplifier-config skill that:

• inspects effective configuration and authoritative current sources;
• separates root provider selection from spawned routing;
• proposes the smallest safe change with provenance, blast radius, and rollback;
• applies only authorized changes and verifies observed outcomes.

Why app-cli

The skill depends on CLI-specific configuration behavior and app-cli:cli-expert. Packaging it with app-cli keeps its guidance version-locked to the installed CLI and its version-matched expert.

Files changed

• amplifier_app_cli/data/skills/amplifier-config/SKILL.md — packaged inline configuration workflow.
• tests/test_merge_utils.py — package-relative discovery and skill contract coverage.
• tests/test_skills_cli.py — packaged slash-command discovery and invocation coverage.

Verification

• uv run pytest -q: 1470 passed, 1 skipped, 13 deselected, 1 xfailed.
• Built wheel includes amplifier_app_cli/data/skills/amplifier-config/SKILL.md.
• Fresh-process pilot smoke verified slash discovery and the root-vs-spawned provider distinction.
• Focused packaged-skill checks: 3 passed.

Breaking changes

None.

[Brian Krabach (bkrabach)]

Thanks for this, Sam Schillace (@ramparte) — the underlying signal is real: users need a repeatable way to inspect effective Amplifier CLI configuration (root provider selection vs. spawned-session routing vs. bundle composition vs. scope precedence) instead of guessing from memory or stale docs. That's exactly the kind of friction an inline /amplifier-config skill should absorb.

Per our review policy for community PRs, here's the due diligence we ran before merging:

Accuracy check against this week's changes

The three "landmines" we specifically checked this skill against (all merged to main after this branch was cut):

• fix(provider): wizard honors default=None — never force-writes choice/boolean values (in_memory defect) #284 (provider wizard: None-default fields render "leave unset" and omit the key)
• amplifier-module-provider-openai#69 (model-gated ConfigFields, e.g. prompt_cache_retention/text_verbosity/enable_reasoning_context, plus did-you-mean on unknown keys)
• feat(provider): reuse-or-separate credential flow for multi-instance providers #281 (multi-instance credential reuse-or-separate flow)

None of these are hardcoded anywhere in SKILL.md. The skill is written at the right altitude: it classifies configuration planes (root selection, spawned routing, bundle composition, scope precedence, runtime session state, skill discovery, provider-module fields) and requires live inspection + delegation to app-cli:cli-expert for anything version-sensitive, rather than freezing today's field names or wizard behavior into the doc. That design choice is exactly what keeps a config-help skill from going stale the next time a provider module or wizard ships a change — so there was nothing to revise here. We also confirmed the scope names (global/project/local persisted, session runtime-only) and the ${VAR} env-placeholder convention it references match current lib/settings.py / provider_config_utils.py, and that app-cli:cli-expert is the real, existing agent reference used elsewhere in this repo (agents/cli-expert.md, behaviors/cli-expertise.yaml).

Wiring verdict: confirmed discoverable, not inert

amplifier_app_cli/data/skills/<name>/SKILL.md is auto-appended to tool-skills's config by _ensure_default_skills_dirs() (runtime/config.py) — no separate index/registration step, same mechanism that already ships goalify, goal-batch, and ten-lane-highway. The new test_packaged_amplifier_config_discovery_and_invocation test in tests/test_skills_cli.py exercises this end-to-end (reads the packaged dir, parses frontmatter, confirms /amplifier-config <args> dispatches to load_skill), and test_amplifier_config_skill_contract in tests/test_merge_utils.py pins the frontmatter contract (name/user-invocable/license/version, no context/model_role/allowed-tools, $ARGUMENTS and app-cli:cli-expert present in the body). Both pass.

Test results

• Full suite on the PR branch as submitted (2 commits behind main): 1471 passed, 1 skipped, 13 deselected, 1 xfailed.
• Rebased onto current main (picks up chore(context): remove legacy scaffolding shown to cause over-delegation (-27% tokens, quality preserved) #283/fix(provider): wizard honors default=None — never force-writes choice/boolean values (in_memory defect) #284) to get an apples-to-apples baseline: 1493 passed, 1 skipped, 13 deselected, 1 xfailed — i.e. main's 1491 + this PR's 2 new tests, zero regressions. (The 1471 vs. 1491 baseline gap before rebasing is fully explained by tests/test_provider_config_unset_fields.py, added by fix(provider): wizard honors default=None — never force-writes choice/boolean values (in_memory defect) #284, which this branch predates.)

Disposition

Accurate and wired as submitted — no revisions needed. Since this PR comes from a fork, we can't push a rebase to your branch, so we're merging the original commit as-is via squash (repo has no branch protection issue, but the ruleset requires 1 approving review — using admin-merge at team direction to land it, since CI is 9/9 green and content review is complete). Credit stays with you as author; thanks for surfacing the gap.