[Birfy]

Closes #892

status, watch, show, apply, cancel, resume and diagnose now take latest (or last, any case) in place of a run id and resolve it to the newest run in the store, so agentdescent demo && agentdescent show latest works without copying the timestamped id.

• A real run directory literally named latest/last still wins.
• An empty store raises a clear no runs under … yet error (also via --json), instead of no run 'latest'.
• Resolution happens in cli.main only; runstore.py is untouched.
• show/watch/diagnose help mentions latest.

Tests: tests/test_latest_run_alias.py (newest wins, real dir wins, empty store, other ids untouched); -k "cli or help or docs or json" suite passes locally.

[Birfy]

Self-review at tip 2c51e76

Verdict: PASS

• Alias resolution is confined to cli.main for the seven run-id commands; audit/--run-id flags and runstore.get are unchanged.
• A real latest/last run dir is preferred; empty store errors through the existing RunStoreError path (exit 2, --json object).
• New tests cover newest-wins, literal-dir-wins, empty store and untouched ids; local -k "cli or help or docs or json" suite green.
• risk_gate: T2 (83 lines / 3 files).

[Birfy]

Verdict: PASS

Tip SHA: 2c51e76
Ran: tests/test_latest_run_alias.py (4 passed); -k "cli or help or json or runstore" plus tests/test_docs_links.py on the tip — all pass except tests/test_plugin_host_cli.py::test_present_host_passes, which fails identically on main in this environment (a gateway key ARK_API_KEY is set on the review box), so unrelated. risk_gate.py → T2 (83 lines / 3 files).
Not run: full suite; CI matrix 3.9–3.13 still in progress at review time.

Correctness

• _resolve_latest_run is gated to the seven run-id commands by fn identity, so no other run_id-shaped argument (e.g. on bench/algorithms) is rewritten. Good.
• runstore.list_runs returns newest first via reverse-sorted dir names, and run ids are timestamp-prefixed, so runs[0] is the newest. Unreadable/status-less dirs are skipped there too, so latest can't land on a half-created dir.
• A real directory named latest/last wins via store_path(store) (which normalises ~, runstore: a literal '~' in --store or $AGENTDESCENT_HOME is only half expanded (runs created, list says none) #402), and the empty-store error is a RunStoreError, so it flows through the existing --json error object (cli: run-store and archive errors ignore --json (status/show/apply/cancel/resume/diagnose/portfolio/audit/safety) #658) — the test checks exactly that and that nothing gets created.
• Resolution sits inside the try, after parsing and after json_errors is reset, which is the right place.

Design fit: keeping it in cli.main and leaving runstore untouched keeps the MCP/library API exact-id only, which is the safer default. Case-insensitive match is fine since generated ids are never those words.

Tests/docs: cover newest-wins, real-dir-wins, empty store, and untouched other ids; help text updated on show/watch/diagnose; changelog fragment present.

Nits (non-blocking): status/apply/cancel/resume help could mention latest too for discoverability; with a default store and a run being created concurrently, latest is naturally racy, which is acceptable for an interactive alias.