deploy(vera): Sonnet 5 on the Claude subscription + an alarm for the silent fallback

.github/workflows/ci.yml

@@ -21,6 +21,15 @@ jobs:
           assert m["archetype"]["soul"].strip(), "archetype soul present"
           print("manifest ok:", m["id"], "-", len(m["plugins"]), "members")
           PY
+      - name: Watchdog check logic
+        # The attribution rule in check_model_fallback.py and the verdict rules in
+        # check_oauth_health.py decide whether #alerts gets woken — a bug there is
+        # either a silent degrade nobody hears about or a pager that cries wolf, and
+        # neither shows up in a smoke run against a healthy container. Stdlib
+        # unittest: CI here is python3 + PyYAML, and a watchdog suite earning a
+        # dependency install is the wrong trade.
+        run: python3 -m unittest discover tests -v
+
       - name: README pins match bundle manifest and Dockerfile
         # Prevents the member-table and base-image prose from drifting out of sync
         # with the canonical pins in protoagent.bundle.yaml and the Dockerfile FROM line.

Dockerfile

@@ -9,7 +9,7 @@
 # protoAgent core forward on the same image roll — core and member bumps are
 # decoupled. Bump this deliberately (and re-verify), keeping it in step with the
 # manifest's `verified_against`. Tag format is bare semver (no `v` prefix).
-FROM ghcr.io/protolabsai/protoagent:0.137.1
+FROM ghcr.io/protolabsai/protoagent:0.145.0
 
 USER root
 
@@ -27,7 +27,7 @@ RUN curl -fsSL https://deb.nodesource.com/setup_22.x | bash - \
 # Bake the bundle members at their RELEASE TAGS (both public — no build secrets).
 # The tags mirror protoagent.bundle.yaml's pins; bump both together (the manifest
 # is the source of truth, this bake is its image form).
-ARG GITHUB_PLUGIN_REF=v0.3.0
+ARG GITHUB_PLUGIN_REF=v0.5.0
 RUN git clone --depth 1 --branch "${GITHUB_PLUGIN_REF}" \
       https://github.com/protoLabsAI/github-plugin.git /opt/protoagent/plugins/github \
     && rm -rf /opt/protoagent/plugins/github/.git

README.md

@@ -14,7 +14,7 @@ python -m server plugin install https://github.com/protoLabsAI/qaEngineer
 | Member | Pin | Role |
 |---|---|---|
 | `workflows` (builtin) | core | the recipe engine the review panels run on |
-| [github-plugin](https://github.com/protoLabsAI/github-plugin) | v0.3.0 | the verdict surface — formal Review API tools with CI-terminal + self-review guards inside the tools |
+| [github-plugin](https://github.com/protoLabsAI/github-plugin) | v0.5.0 | the verdict surface — formal Review API tools with CI-terminal + self-review guards inside the tools |
 | [pr-reviewer-plugin](https://github.com/protoLabsAI/pr-reviewer-plugin) | v0.35.0 | the machinery — webhook chokepoint, structural trigger, panel dispatch, evidence grounding, convergence, approve-on-green sweep, on-demand summon, telemetry + eval |
 
 Persona: [`SOUL.md`](./SOUL.md) (Vera — verdict system, three-layer verification, 80% bar,
@@ -84,7 +84,44 @@ plus prior-round context. In practice that's **~37k tokens per finder**.
 - **No aggressive TPM cap.** A free tier at 12,000 tokens/minute rejects a single finder
   outright.
 - Routed through a gateway alias, so swapping models is a gateway edit rather than a code
-  change. Model settings are host-scoped (ADR 0047).
+  change. Model settings are host-scoped (ADR 0047) — **or** a native subscription lane
+  (see below), which is not a gateway edit at all.
+
+**The reference host's lane (since 2026-08-23):** `protolabs/smart` on the gateway
+(`model.provider: openai`), with `routing.fallback_models: ["protolabs/cloud"]` behind
+it. She ran a native Claude subscription (`anthropic-oauth` / `claude-sonnet-5`) from
+2026-08-21 to 08-23 and it reviewed well — ~25% faster, same grounding rate — but a
+five-finder panel across a fleet's worth of repos **exhausted the subscription's 7-day
+quota in under two days** (`anthropic-ratelimit-unified-7d-status: rejected`,
+`retry-after: 356343` = 4.1 days). A reviewer that cannot review for four days is not a
+reviewer, so the gateway is the sustainable primary and a subscription is the treat.
+Set `model.name` and `model.provider` in the SAME config POST either way: they are one
+decision (protoAgent#2623).
+
+**The old native-OAuth shape, for reference:** Native OAuth **bypasses the gateway entirely** (ADR 0097) — the fallback alias is
+reachable only because protoAgent#2571 lets a namespaced slot name opt out of the native
+provider, so `model.api_base` + a gateway key must stay set even though the primary never
+uses them. Set `model.name` and `model.provider` in the SAME config POST: they are one
+decision (protoAgent#2623), and a native provider paired with a `protolabs/*` name is
+rejected on every call.
+
+**A fallback is SILENT.** protoAgent wires langchain's `ModelFallbackMiddleware` raw, and
+that middleware swallows the primary's exception with no log, no counter and no event —
+found on core 0.144.0 and filed as protoAgent#2956. So a dead subscription doesn't break
+the review; it quietly changes which model writes the verdict. That is what
+`scripts/check_model_fallback.py` exists to catch, by inference from gateway metrics:
+protoAgent-UA traffic arriving at the gateway from Vera's container *is* a fallback,
+because her primary never goes there.
+
+**#2956 is FIXED in core 0.145.0**, which the pins above now carry:
+`ObservableModelFallbackMiddleware` logs a WARNING and publishes a `model.fallback` bus
+event (ADR 0039). The inference script is therefore scheduled for deletion — but not
+yet, and the distinction matters: **pinning a version is not running it.** Vera rolls on
+watchtower after a merge, so between the pin landing and the roll completing she is on
+the old core with no event at all. Retire the script once the running instance reports
+0.145.0 *and* the event has been seen firing; deleting the inference before its
+replacement is observed working would leave the silent-degrade window covered by
+neither.
 
 **Reviews are not cheap.** One structural review is nine LLM steps and 5–9 minutes of
 wall clock. On a hosted frontier model that's roughly $0.12–0.15 each; on local inference
@@ -164,7 +201,7 @@ from the payload):
 ## Deploying Vera (the reference host)
 
 This repo doubles as Vera's image source: `Dockerfile` = stock protoAgent (**pinned
-base** — `protoagent:0.137.1`, in step with the manifest's `verified_against`; bump
+base** — `protoagent:0.145.0`, in step with the manifest's `verified_against`; bump
 deliberately so a member-pin bump can't drag the core forward on the same roll) +
 node/`clawpatch` + the bundle members baked at their manifest pins +
 `deploy/vera.langgraph-config.yaml` (seed, not force) + `SOUL.md`.
@@ -191,6 +228,31 @@ static check can't see a *running* instance whose live config drifted;
 `scripts/check_card_drift.py` is the runtime half — point it at the (tailnet-only) card
 from the ava fleet cron: `python3 scripts/check_card_drift.py` (exit 1 on drift).
 
+### The watchdogs
+
+Four checks, all run from the ava fleet cron through `scripts/vera-watchdog.sh`, which is
+the piece that makes a failure LOUD (a Discord `#alerts` post, with exit 1 = verdict and
+exit 2 = unreachable kept as distinct alarms). Every one of them exists because Vera fails
+*quietly* — a starved panel, a drifted card, a swapped model underneath a verdict.
+
+| Mode | Check | Asks |
+|---|---|---|
+| `health` | `check_review_health.py` | is the gate still producing verdicts? (growth in unreviewed/exhausted, completion rate) |
+| `drift` | `check_card_drift.py` | does the live card still match the seed? |
+| `fallback` | `check_model_fallback.py` | did she silently answer from her fallback model? (gateway-metrics inference — protoAgent#2956) |
+| `oauth` | `check_oauth_health.py` | is the subscription credential still signed in, refreshable, and coherent with `model.name`? |
+
+The wrapper runs **installed copies** in `~/.local/bin`, not `scripts/*.py` — this repo is
+also the deploy source, so a branch switch would silently disarm a guard that lived inside
+it. Re-`install` them after changing a script (there is no auto-update) — the install lines
+are in the wrapper's header. That includes **`scripts/vera_api.py`**, the shared
+operator-API helper: Python puts a script's own directory on `sys.path[0]`, so a flat
+module installed alongside imports cleanly — and a check installed *without* it dies on
+ImportError. `python3 -m unittest discover tests` covers the attribution and verdict rules
+(CI runs it). It reads `DISCORD_WEBHOOK_ALERTS` from `infisical run`, so an
+expired Infisical session downgrades every alarm to a log line nobody reads; that is worth
+checking whenever the alerts channel goes quiet for a suspiciously long time.
+
 ## Other orgs
 
 Nothing here is protoLabs-specific except the pins and the seed: the bundle installs into

protoagent.bundle.yaml

@@ -25,11 +25,11 @@ description: >-
   own work; never posts a blocking verdict against pending CI.
 
 # The core version this pin set was last verified against (ADR 0049 rule 2).
-verified_against: 0.137.1
+verified_against: 0.145.0
 
 plugins:
   - { id: workflows, builtin: true }   # the recipe engine (code-review panels run through it)
-  - { id: github, url: https://github.com/protoLabsAI/github-plugin, ref: v0.3.0 }
+  - { id: github, url: https://github.com/protoLabsAI/github-plugin, ref: v0.5.0 }
   - { id: pr-reviewer, url: https://github.com/protoLabsAI/pr-reviewer-plugin, ref: v0.35.0 }
 
 enabled: [workflows, github, pr-reviewer]   # suggested turn-on list (applied to plugins.enabled)