@mmmbuto/nexuscrew 0.9.23 → 0.9.25

package/CHANGELOG.md

@@ -2,6 +2,106 @@
 
 All notable changes to NexusCrew are tracked here.
 
+## 0.9.25 - 2026-09-13
+
+- **Compatibility note**: codex-vl >= 0.154.0-vl.3 inside a bound Fleet cell
+  requires NexusCrew >= 0.9.25 for the `nc_*` tools; NexusCrew 0.9.24 only
+  reads the legacy identity variables, which vl.3 removes from MCP children on
+  verified bindings. Standalone (legacy) cells keep working with both.
+
+- **MCP identity reads the verified binding metadata first (fail-closed).**
+  `nexuscrew mcp` now evaluates the `NEXUSCREW_VERIFIED_*` metadata delivered
+  by codex-vl >= 0.154.0-vl.3 launchers before any legacy source: version,
+  owner instance, cell, channel and the authority introspection must all be
+  coherent (source `verified-env`, verified identity in `nc_identity`), and any
+  incoherence fails closed with `NEXUSCREW_MCP_IDENTITY_VERIFIED_ENV_INVALID`
+  instead of falling back to the legacy variables. Without the metadata the
+  legacy path is unchanged.
+- **`CODEX_APP_SERVER_IDENTITY_REQUIRED=1` is exported only in authority
+  mode.** The fleet launcher now exports the flag for codex-vl cells only when
+  the fleet identity mode is `authority` and the authority is constructible;
+  in legacy mode it exports an explicit `0` and names the reason in a
+  structured log line. Explicit user overrides (`engine.env`, `cfg.env`) still win, with a
+  warning when they force `1` outside authority mode.
+- **Standalone cells (`CODEX_APP_SERVER_IDENTITY_REQUIRED=0`) no longer
+  receive the identity channel.** The launcher created fd 3:4 and
+  `NEXUSCREW_IDENTITY_FD` even when it had resolved the flag to `0`, so the
+  child started an identity handshake against a legacy lease server that
+  answered `revoked` and the cell died at startup. The channel decision now
+  comes from the same resolution that produces the flag.
+- **A configured but unavailable fleet identity authority now refuses
+  protected cell launches instead of silently falling back to standalone.**
+  The launch fails with `IDENTITY_AUTHORITY_UNAVAILABLE` naming the reason
+  (never a value), and `nexuscrew doctor` reports the same state.
+- **Upgrade NexusCrew before codex-vl**: `codex-vl >= 0.154.0-vl.3` with
+  `NexusCrew <= 0.9.24` fails closed in Fleet cells (the old launcher forces
+  identity without an authority).
+- **`nexuscrew identity provision`.** New command that generates the two
+  distinct authority credentials (32 random bytes each, files `0600` inside a
+  `0700` directory under `~/.nexuscrew/identity-authority/`), never overwrites
+  existing credentials without `--force`, activates
+  `fleet.identity.mode = "authority"` in the config (or prints the exact keys
+  with `--no-write-config`). `loadConfig` reads the credential files
+  automatically and `nexuscrew doctor` shows the identity mode and authority
+  constructibility (no values).
+- **MCP identity remediation rewritten.** `nc_identity` and `doctor` no longer
+  recommend allowlisting `TMUX`/`TMUX_PANE`/`NEXUSCREW_MCP_SESSION` in the MCP
+  server `env_vars`: on a shared verified spawn that allowlist fails the launch
+  (the names are reserved). The two valid paths are a verified session
+  (codex-vl >= 0.154.0-vl.3 + NexusCrew >= 0.9.25 with provisioned authority)
+  or a legacy session inheriting the three variables.
+
+- **Ollama Cloud catalog refreshed against the official measured sample of
+  2026-09-12.** Bare-name `gemma4` and `qwen3.5` enter the catalog (verified
+  alive with a 200 on `/v1/chat/completions`; official measured sample:
+  22 live ids / 8 dead variants), with context and capabilities taken from
+  the channel cards (`gemma4` 128K text+image, `qwen3.5` 256K text+image).
+  The 404 variants (`gemma4:12b`/`:e4b`, `qwen3.5:27b`/`:35b`/`:122b`, bare
+  `mistral-large-3`, `nemotron-3-super:120b`, `nemotron-3-nano:4b`) and
+  community ids stay out on purpose. `glm-5.3` and `deepseek-v4.1-flash`
+  were already listed and are not duplicated.
+- **Fixture-based liveness guard for the Ollama Cloud models.** New test
+  (`tests/fleet-ollama-liveness.test.js`): every catalog id must belong to
+  the official measured sample; when upstream retires an id, the test goes
+  red naming it as "to be verified". No network calls in tests; the real
+  periodic check remains a separate item.
+- **OpenCode Go catalog refreshed signal-only (401 CreditsError on
+  2026-09-12).** The Go plan could not generate (credits exhausted, listing
+  OK): the additions are wire signals, not end-to-end measurements.
+  `glm-5.3`, `glm-5.3-flash`, `longcat-2.0`, `deepseek-v4.1-flash` and
+  `hy4-preview` enter all three wires; `qwen3.8-flash` enters Messages/Chat
+  ONLY (Responses rejected, measured). New limits declared by models.dev
+  (1M context; `hy4-preview` 1024000), marked "declared, not measured".
+  Retired `glm-5`, `kimi-k2.5`, `minimax-m2.5` and `qwen3.5-plus` are
+  removed from the lists, LIMITS and the catalog, with a test guard against
+  re-introduction. `hy3` unchanged (diverging source on its output limit:
+  to be re-checked), default `deepseek-v4-flash` unchanged, VISION
+  unchanged (only the PNG-measured vision-exp id).
+
+## 0.9.24 — 2026-09-12
+
+- **Verifica dell'identità legata alla challenge, con claims completi.** La
+  verifica online di un identity proof (metodo `nexuscrew/identity/verify`, sul
+  canale di lease per cella) ora riceve dall'authority i claims che la challenge
+  ha effettivamente emesso, e il risultato è legato alla challenge stessa: una
+  risposta che non corrisponde alla challenge emessa viene rifiutata
+  (`challenge_mismatch`) invece di essere accettata. Il `nonce` torna fra i
+  claims perché il lato client lo confronta con la propria challenge al momento
+  del commit, senza fallback su claim esterni.
+
+- **Il canale serve più verifiche per connessione.** Dopo una verifica
+  conclusa il relay azzera il riferimento pendente: una seconda verifica sulla
+  stessa connessione (reconnect, resume, nuova incarnazione) non trovava più il
+  canale libero e falliva pur essendo il canale sano. Corretto, con e2e che
+  copre verifica ripetuta e challenge successiva.
+
+- **Requisito di compatibilità — la coppia versioni conta.** `codex-vl` **≥
+  0.154.0-vl.3** richiede **NexusCrew ≥ 0.9.24**: il daemon `codex-vl` con
+  identità obbligatoria e canale di verifica assente o inutilizzabile **non si
+  avvia** (fail-closed) e non può verificare l'identità contro un server che non
+  espone `verify`. Con NexusCrew 0.9.23 una cella gestita resta quindi
+  inutilizzabile; aggiornate il server prima del daemon.
+
 ## 0.9.23 — 2026-09-11
 
 - **Frontend bundle ricostruito.** Il banner «new version available /