[Birfy]

Closes #882

What changed

--kind and spec kind now behave like algorithm and parameter names: a value that matches exactly one of text, skill_dir, skill_library, agent_dir, agent_code, plugin, policy_slot once case and -/_ are ignored resolves to it (skill-library → skill_library, SKILL_DIR → skill_dir). A remaining close typo appends ; did you mean 'skill_library'?. Unrelated names like nope still get the full list and no guess.

algorithms list/run, init, MCP list_algorithms / algorithm_spec, and compose share resolve_kind. Written specs use the canonical name.

Verification at the tip

Tip SHA: ef39c17
Ran:

• pytest -q tests/test_evolvespec.py tests/test_cli_algorithms_bare.py tests/test_cli_algorithms_run.py tests/test_guess_kind_skill_library.py tests/test_cli_json_usage_errors.py tests/test_mcp_algorithm_spec.py tests/test_algorithms.py tests/test_changelog_fragments.py → passed
  Not run:
• full suite — CI

[Birfy]

Verdict: PASS
Tip SHA: ef39c17
Ran: pytest -q tests/test_evolvespec.py tests/test_cli_algorithms_bare.py tests/test_cli_algorithms_run.py tests/test_guess_kind_skill_library.py tests/test_cli_json_usage_errors.py tests/test_mcp_algorithm_spec.py tests/test_algorithms.py tests/test_changelog_fragments.py at the tip → passed (exit 0). risk_gate.py from main → T2 (185 changed lines / 8 files).
Not run: full suite (CI covers 3.9–3.13).

Correctness. --kind and spec kind now fold case and -/_ onto the canonical KINDS entry (skill-library / SKILL_LIBRARY → skill_library), so algorithms list, init, algorithms run, MCP list_algorithms / algorithm_spec, and compose accept the same aliases algorithm and parameter names already do. Written specs store the canonical kind (so skill_library seeding and layout still run). A near-miss typo such as skil_library ends with did you mean 'skill_library'?; unrelated nope still lists every kind and does not guess. argparse choices= on init / run --kind is kept (unknown values stay invalid choice) after a type=fold_kind conversion.

Risk. T2 by size (package CLI/spec validation, under the 250-line cap). No evolve-loop or installer change. AGENT_PAUSE=false; will squash-merge when required CI is green (or arm --auto while it runs).

[Birfy]

Verdict: PASS
Tip SHA: ef39c17
Ran:

• .github/scripts/risk_gate.py --repo ThinkFlowLab/agentdescent --pr 883 → T2 (185 changed lines / 8 files)
• pytest -q tests/test_evolvespec.py tests/test_cli_algorithms_bare.py tests/test_cli_algorithms_run.py tests/test_guess_kind_skill_library.py tests/test_cli_json_usage_errors.py tests/test_mcp_algorithm_spec.py tests/test_algorithms.py tests/test_changelog_fragments.py → passed
  Not run:
• full suite — CI

fold_kind / resolve_kind match the existing algorithm-name and parameter-name path: case and -/_ fold to a unique KINDS entry (skill-library / SKILL_LIBRARY → skill_library), and a leftover close typo (skil_library, difflib cutoff 0.75) appends ; did you mean 'skill_library'? without guessing at unrelated names like nope.

The shared helper is wired through the surfaces #882 named: algorithms list/run (including the free-text group-level --kind), init / write_starter_spec, MCP list_algorithms and algorithm_spec (the latter via algorithms_run_payload), EvolveSpec.__post_init__ so later kind in alg.kinds checks see the canonical name, and compose. Written specs keep skill_library, which the init and algorithms run tests cover. The changelog fragment is present.

One non-blocking note: init --kind and algorithms run --kind still go through argparse type=fold_kind, choices=KINDS. Aliases resolve, but a near-miss like init --kind skil_library still gets argparse's invalid choice rather than the did you mean suffix that algorithms list --kind skil_library and resolve_kind produce. That's acceptable — choices= keeps the valid list in --help — and is not a merge blocker.

Risk: T2 (under the 250-line cap). AGENT_PAUSE=false.