feat(cli): add CLI agent providers (Claude Code, Copilot, Kiro)

[buraksakalli]

Adds first-class support for driving evaluations through user-installed CLI agents (claude, copilot, kiro-cli) so users can run upskill against existing subscriptions instead of raw ANTHROPIC_API_KEY / OPENAI_API_KEY.

Architecture:

• New CliAgentExecutor peer to LocalFastAgentExecutor and RemoteFastAgentExecutor, dispatching to one of three CliAgentRunner implementations (ClaudeCodeRunner, CopilotCliRunner, KiroCliRunner) via a small registry.
• Per-model executor routing (executors/router.py): cli.
  models route to CliAgentExecutor; everything else keeps its existing local/jobs path. This lets one upskill eval call mix CLI agents and fast-agent models in a single benchmark.
• CliFastAgentAdapter satisfies the narrow SkillGenerator protocol upskill's generate.py uses, so skill generation, test generation, and refinement also work via a CLI provider — no API keys required.
• _agent_session() dispatcher in cli.py: pure-cli runs skip _fast_agent_context entirely; hybrid runs strip cli.* slots from fast-agent's model_references and override per phase.
• Subscription billing is preserved: ClaudeCodeRunner unsets ANTHROPIC_API_KEY / ANTHROPIC_AUTH_TOKEN / CLAUDE_CODE_USE_BEDROCK / CLAUDE_CODE_USE_VERTEX from the subprocess env by default and prints a one-shot stderr note when it does. Power users can override cli_providers..unset_env: [] to keep the vars (e.g. for Bedrock/Vertex routing).
• Smart defaults: when every -m is cli., executor: jobs auto-downgrades to local (CLI providers always run locally) and test_gen_model defaults to the first cli. eval model so users with non-cli config defaults don't need extra flags.

Testing:

• 48 new unit tests across runners, executor, router, env stripping, config, mixed-model orchestration, CLI help text.
• 4 end-to-end tests that drive the full stack against fake CLI binary fixtures (no real CLIs needed in CI).
• All 143 tests pass; ruff format/lint, ty, cpd all clean.

Docs:

• README "CLI Providers" section with provider table, quick-start, full config example, and subscription-billing notes.
• --model / --eval-model / --test-gen-model help text mentions the new cli. aliases.

[Copilot]

Pull request overview

Adds first-class support for running upskill evaluations (and generation flows) via user-installed CLI agent tools (claude, copilot, kiro-cli) using cli.<provider> model aliases, enabling subscription-backed runs without requiring Anthropic/OpenAI API keys. The PR introduces a per-model executor router so a single eval/benchmark can mix CLI-backed models and fast-agent models.

Changes:

• Introduces CliAgentExecutor + per-provider runners (Claude Code, Copilot CLI, Kiro CLI) and integrates them via a per-model executor router.
• Updates CLI orchestration to support pure-CLI and hybrid runs (CLI + fast-agent), including smart defaults/downgrades and test-gen routing.
• Adds extensive unit + E2E tests and updates documentation for CLI providers and configuration.

Reviewed changes

Copilot reviewed 28 out of 29 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
src/upskill/cli_agents/__init__.py	Exposes the CLI agents public API (CliProviderConfig, runners/protocols, defaults).
src/upskill/cli_agents/base.py	Defines shared CLI runner contracts, subprocess execution helpers, env stripping + warning behavior, and default provider configs.
src/upskill/cli_agents/claude_code.py	Implements Claude Code runner with JSON envelope parsing and token usage extraction.
src/upskill/cli_agents/copilot.py	Implements Copilot CLI runner via plain-text execution helper.
src/upskill/cli_agents/fast_agent_adapter.py	Adds an adapter so CLI executors can satisfy the generation/refinement protocol used by generate.py.
src/upskill/cli_agents/kiro.py	Implements Kiro CLI runner via plain-text execution helper.
src/upskill/cli_agents/protocols.py	Introduces a narrow SkillGenerator protocol for decoupling generation flows from fast-agent specifics.
src/upskill/config.py	Adds cli_providers config with default providers and a merge strategy for user overrides.
src/upskill/executors/cli_agent.py	Adds a new executor that materializes artifacts/workspace and dispatches one-shot requests to CLI runners.
src/upskill/executors/router.py	Adds per-model executor routing between CLI executor and fast-agent local/jobs executors.
src/upskill/model_resolution.py	Updates model resolution to default test-gen to a CLI model when all eval models are cli.*.
src/upskill/generate.py	Switches generation flows to depend on the new narrow generator protocol.
src/upskill/cli.py	Integrates CLI providers into eval, benchmark, and generate flows, adding session dispatching and per-model executor factories.
tests/test_cli_agent_executor.py	Unit tests for CliAgentExecutor prompt composition, artifact persistence, cancellation, and error paths.
tests/test_executor_router.py	Tests routing behavior for CLI vs fast-agent models and jobs/local selection.
tests/test_claude_code_runner.py	Unit tests for Claude runner JSON parsing, error handling, and argv formation.
tests/test_copilot_cli_runner.py	Unit tests for Copilot runner stdout parsing, error handling, and argv formation.
tests/test_kiro_cli_runner.py	Unit tests for Kiro runner argv formation (including opt-in flags) and error handling.
tests/test_cli_env_stripping.py	Tests env stripping and one-shot warnings, plus an end-to-end subprocess env capture for Claude.
tests/test_cli_agents_contract.py	Contract tests for config schema, defaults, merge behavior, and runner protocol conformance.
tests/test_cli_agents_e2e.py	End-to-end tests using fake CLI binaries to validate full router/executor/runner stack.
tests/test_mixed_model_eval.py	Validates mixed-model eval routing and pure-CLI test-gen skipping fast-agent context.
tests/test_cli_help.py	Ensures CLI help text mentions cli.<provider> model aliases.
tests/test_cli_generate_benchmark.py	Updates generate/benchmark tests to validate the new executor-factory plumbing.
tests/fixtures/fake_cli_agents/fake_claude.py	Fake Claude binary fixture for E2E tests (JSON envelope).
tests/fixtures/fake_cli_agents/fake_copilot.py	Fake Copilot binary fixture for E2E tests (plain text).
tests/fixtures/fake_cli_agents/fake_kiro.py	Fake Kiro binary fixture for E2E tests (plain text).
README.md	Documents CLI providers, configuration, billing/env stripping behavior, and caveats.
.gitignore	Ignores generated runs/ and skills/ output directories by default.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.