[Birfy]

Closes #884

--agent now folds case and -/_ the same way algorithm names, parameter names, and --kind already do. init --agent claude-code and algorithms run … --agent openai-compatible write the canonical agent.ref (claude_code, openai_compatible) instead of rejecting a spelling doctor / install already print.

Kind mismatch checks still run on the folded name (--agent claude-code on kind=text is still refused as a WorkspaceAgent, #456). Install-only aliases that are not a spelling fold (deepseek-harness, claudecode, open-code) still hint; grok-bot still explains it has no worker CLI.

Verification

• pytest -q tests/test_init_honor_agent.py tests/test_changelog_fragments.py tests/test_init_default_agent.py tests/test_workspace_agent_hint.py tests/test_cli_json_usage_errors.py tests/test_cli_algorithms_run.py tests/test_install_host_aliases.py passed
• live init --agent claude-code / --agent openai-compatible write the canonical refs

[Birfy]

Verdict: PASS
Tip SHA: fbe66d8
Ran: pytest -q tests/test_init_honor_agent.py tests/test_changelog_fragments.py tests/test_init_default_agent.py tests/test_workspace_agent_hint.py tests/test_cli_json_usage_errors.py tests/test_cli_algorithms_run.py tests/test_install_host_aliases.py at the tip → passed; live init --agent claude-code / --agent openai-compatible write the canonical refs; grok-bot and kind=text --agent claude-code still refuse. risk_gate.py from main → T2 (91 changed lines / 3 files).
Not run: full suite (CI covers 3.9–3.13).

Correctness. --agent now folds case and -/_ the same way algorithm names, parameter names, and --kind already do, so a spelling copied from doctor / install (claude-code, openai-compatible, CLAUDE_CODE) writes the canonical agent.ref. Kind mismatch still runs on the folded name (no silent swap, #456). Install-only aliases that are not a spelling fold (deepseek-harness, claudecode, open-code) still hint; grok-bot still explains it has no worker CLI.

Risk. T2 by size/path (CLI validation only; no evolve-loop change).

[Birfy]

Verdict: PASS
Tip SHA: fbe66d8
Ran: pytest -q tests/test_init_honor_agent.py tests/test_changelog_fragments.py tests/test_init_default_agent.py tests/test_workspace_agent_hint.py tests/test_cli_json_usage_errors.py tests/test_cli_algorithms_run.py tests/test_install_host_aliases.py at the tip → passed (exit 0). risk_gate.py from main → T2 (91 changed lines / 3 files).
Not run: full suite (CI covers 3.9–3.13).

Correctness. --agent now folds case and -/_ onto a built-in short ref (claude-code / CLAUDE_CODE → claude_code, openai-compatible → openai_compatible) in write_starter_spec, so both init and algorithms run write the canonical agent.ref instead of rejecting a spelling doctor / install already print. Kind-mismatch checks still run on the folded name (--agent claude-code on kind=text is still refused as a WorkspaceAgent, #456). Install-only aliases that are not a spelling fold (deepseek-harness, claudecode, open-code) still hint at the worker; grok-bot still explains it has no worker CLI; module:attribute refs and unrelated names like bogus are unchanged.

Design. fold_agent_ref matches the existing algorithm-name / parameter-name / --kind fold (strip + casefold + hyphen-to-underscore, accept only when the result is exactly one known ref). Applying it once in the shared writer is the right layer — no second parser path to drift. The leftover did-you-mean branch in init_agent_unknown for hyphen/case aliases is now dead for this call site (fold happens first) and remains a safe fallback if another caller forgets to fold.

Tests / docs. New cases cover hyphen/case aliases on init, algorithms run, the #456 kind-mismatch gate, fold_agent_ref itself (including leaving deepseek-harness and module:attribute alone), and changelog fragment 884-agent-aliases.md. Existing install-host hint tests still pass.

Risk. T2 by size (package CLI validation only; 91 lines / 3 files, under the 250-line cap). No evolve-loop or installer change. AGENT_PAUSE=false; will squash-merge when required CI is green.