Skip to content

Dodaios is an open-source agent OS for your desktop: it runs your coding agents, browser and apps by voice or chat, on your own machine. Formerly Personal Jarvis.

fix(wiki): keep memory work on subscriptions, never the voice API key

rubenluetke10-beep150cb17b208a
diff --git a/jarvis/memory/wiki/consolidator.py b/jarvis/memory/wiki/consolidator.py
index e02dccf59b..3c2b398573 100644
--- a/jarvis/memory/wiki/consolidator.py
+++ b/jarvis/memory/wiki/consolidator.py
@@ -499,9 +499,9 @@ class Consolidator:
start_ns = time.time_ns()
from jarvis.memory.wiki.provider_chain import (
+ background_wiki_providers,
build_wiki_provider_chain,
complete_with_fallback,
- credential_ready_wiki_providers,
)
available = set(self._registry.available())
@@ -510,7 +510,7 @@ class Consolidator:
model_override=self._curator_cfg.model,
available=available,
credential_ready=(
- credential_ready_wiki_providers(
+ background_wiki_providers(
available=available,
config=self._root_cfg,
)
diff --git a/jarvis/memory/wiki/curator_llm.py b/jarvis/memory/wiki/curator_llm.py
index 8a08595221..88f948b518 100644
--- a/jarvis/memory/wiki/curator_llm.py
+++ b/jarvis/memory/wiki/curator_llm.py
@@ -533,9 +533,9 @@ class WikiCuratorLLM:
start_ns = time.time_ns()
from jarvis.memory.wiki.provider_chain import (
+ background_wiki_providers,
build_wiki_provider_chain,
complete_with_fallback,
- credential_ready_wiki_providers,
)
# Key-aware fallback (AP-22/23): cross to a reachable family instead of
@@ -546,7 +546,7 @@ class WikiCuratorLLM:
model_override=self._cfg.model,
available=available,
credential_ready=(
- credential_ready_wiki_providers(
+ background_wiki_providers(
available=available,
config=self._config,
)
diff --git a/jarvis/memory/wiki/extractor.py b/jarvis/memory/wiki/extractor.py
index 0f60ac79a5..e5a6c857a2 100644
--- a/jarvis/memory/wiki/extractor.py
+++ b/jarvis/memory/wiki/extractor.py
@@ -855,9 +855,9 @@ class ConversationFactExtractor:
start_ns = time.time_ns()
from jarvis.memory.wiki.provider_chain import (
+ background_wiki_providers,
build_wiki_provider_chain,
complete_with_fallback,
- credential_ready_wiki_providers,
)
# Key-aware fallback (AP-22/23): try the configured provider, then cross
@@ -869,7 +869,7 @@ class ConversationFactExtractor:
model_override=self._curator_cfg.model,
available=available,
credential_ready=(
- credential_ready_wiki_providers(
+ background_wiki_providers(
available=available,
config=self._root_cfg,
)
diff --git a/jarvis/memory/wiki/provider_chain.py b/jarvis/memory/wiki/provider_chain.py
index 310354877d..9abe364464 100644
--- a/jarvis/memory/wiki/provider_chain.py
+++ b/jarvis/memory/wiki/provider_chain.py
@@ -218,6 +218,32 @@ def credential_ready_wiki_providers(
return ready
+def background_wiki_providers(
+ *,
+ available: set[str] | frozenset[str],
+ config: Any,
+ is_subscription: Callable[[str], bool] | None = None,
+) -> set[str]:
+ """Providers wiki background work may bill: subscriptions before any key.
+
+ Memory curation is background work the user never waits on, so while any
+ subscription login is connected it runs ONLY on subscriptions plus keyless
+ local providers and never touches a per-token key (live 2026-09-29: an
+ OpenAI key added for voice paid ~$0.30 of wiki calls after a two-minute
+ chat). With no subscription connected the full credential-ready set stays
+ in play, so a single-key install keeps a working wiki (AP-21/AP-22).
+ """
+ from jarvis.core.config import PROVIDER_SECRET_CANDIDATES
+
+ ready = credential_ready_wiki_providers(available=available, config=config)
+ probe = is_subscription or (lambda name: subscription_login_ready(name) is True)
+ subscriptions = {name for name in ready if probe(name)}
+ if not subscriptions:
+ return ready
+ keyless = {name for name in ready if name not in PROVIDER_SECRET_CANDIDATES}
+ return subscriptions | keyless
+
+
def build_wiki_provider_chain(
*,
primary: str,
@@ -617,6 +643,7 @@ async def complete_with_fallback(
__all__ = [
+ "background_wiki_providers",
"build_wiki_provider_chain",
"complete_with_fallback",
"credential_ready_wiki_providers",
diff --git a/jarvis/memory/wiki/recurate.py b/jarvis/memory/wiki/recurate.py
index 486ccf10c8..aebbbb5654 100644
--- a/jarvis/memory/wiki/recurate.py
+++ b/jarvis/memory/wiki/recurate.py
@@ -179,9 +179,9 @@ async def recurate_profile(
)
from jarvis.memory.wiki.provider_chain import (
+ background_wiki_providers,
build_wiki_provider_chain,
complete_with_fallback,
- credential_ready_wiki_providers,
)
# An injected registry (tests) skips the credential filter — the same
@@ -198,7 +198,7 @@ async def recurate_profile(
model_override=wiki_cfg.curator.model,
available=available,
credential_ready=(
- credential_ready_wiki_providers(available=available, config=config)
+ background_wiki_providers(available=available, config=config)
if credential_filter
else available
),
diff --git a/jarvis/memory/wiki/search_aliases.py b/jarvis/memory/wiki/search_aliases.py
index e7392e4793..dcf4e77626 100644
--- a/jarvis/memory/wiki/search_aliases.py
+++ b/jarvis/memory/wiki/search_aliases.py
@@ -262,9 +262,9 @@ async def generate_aliases(
from jarvis.core.protocols import BrainMessage, BrainRequest # noqa: PLC0415
from jarvis.memory.wiki.provider_chain import ( # noqa: PLC0415
+ background_wiki_providers,
build_wiki_provider_chain,
complete_with_fallback,
- credential_ready_wiki_providers,
)
# Same key-aware chain the curator uses: lead with the configured wiki
@@ -272,11 +272,15 @@ async def generate_aliases(
# credential (AP-22 — never a single-provider brick).
try:
available = set(registry.available())
+ curator = getattr(getattr(getattr(cfg, "memory", None), "wiki", None), "curator", None)
+ primary = str(getattr(curator, "provider", "") or "").strip() or str(
+ getattr(getattr(cfg, "brain", None), "primary", "") or ""
+ )
chain = build_wiki_provider_chain(
- primary=str(getattr(getattr(cfg, "brain", None), "primary", "") or ""),
+ primary=primary,
model_override="",
available=available,
- credential_ready=credential_ready_wiki_providers(
+ credential_ready=background_wiki_providers(
available=available, config=cfg
),
)
diff --git a/jarvis/plugins/brain/codex.py b/jarvis/plugins/brain/codex.py
index 208c1fe3b0..45f9c70550 100644
--- a/jarvis/plugins/brain/codex.py
+++ b/jarvis/plugins/brain/codex.py
@@ -751,7 +751,9 @@ class CodexBrain:
# OPENAI_API_KEY, so without this a key added for voice silently
# billed every memory pass on the most expensive model (live
# 2026-09-29: ~$0.30 of wiki calls after a two-minute voice chat).
- # The key stays the fallback when the subscription turn fails.
+ # A failed subscription turn raises instead of crossing to the key:
+ # background callers walk their own chain to the next subscription,
+ # and the maintainer wants the key spent only on the live voice call.
if (
api_key
and self._structured_prompts
@@ -762,25 +764,14 @@ class CodexBrain:
"CodexBrain.complete: structured request on the ChatGPT subscription (model=%s)",
self._cli_model or "default",
)
- sub_emitted = False
- try:
- stream = (
- self._complete_via_cli(req)
- if self._subscription_text_only
- else self._complete_via_app_server(req)
- )
- async for delta in stream:
- if delta.content:
- sub_emitted = True
- yield delta
- return
- except Exception as exc: # noqa: BLE001 — re-raised once output was emitted
- if sub_emitted:
- raise
- log.warning(
- "CodexBrain: subscription path failed (%s) — using the API key",
- type(exc).__name__,
- )
+ stream = (
+ self._complete_via_cli(req)
+ if self._subscription_text_only
+ else self._complete_via_app_server(req)
+ )
+ async for delta in stream:
+ yield delta
+ return
if api_key:
log.info("CodexBrain.complete: API-key path (model=%s)", self._model)
client = self._ensure_client(api_key)
diff --git a/tests/unit/brain/test_codex_brain.py b/tests/unit/brain/test_codex_brain.py
index 8182959886..3465ea58ec 100644
--- a/tests/unit/brain/test_codex_brain.py
+++ b/tests/unit/brain/test_codex_brain.py
@@ -152,7 +152,7 @@ async def test_structured_request_prefers_subscription_over_paid_key(
@pytest.mark.asyncio
-async def test_structured_request_falls_back_to_key_when_subscription_fails(
+async def test_structured_request_never_crosses_to_the_paid_key(
monkeypatch: pytest.MonkeyPatch,
) -> None:
api_calls: list[str] = []
@@ -166,10 +166,11 @@ async def test_structured_request_falls_back_to_key_when_subscription_fails(
brain = CodexBrain(structured_prompts=True)
req = BrainRequest(messages=(BrainMessage(role="user", content="extract facts"),))
- text = "".join([d.content or "" async for d in brain.complete(req)])
+ with pytest.raises(RuntimeError, match="app server down"):
+ async for _delta in brain.complete(req):
+ pass
- assert text == "from-api"
- assert api_calls == ["api"]
+ assert api_calls == []
@pytest.mark.asyncio
diff --git a/tests/unit/memory/wiki/test_curator_llm.py b/tests/unit/memory/wiki/test_curator_llm.py
index 38210ebf9b..6a0ffd6b5c 100644
--- a/tests/unit/memory/wiki/test_curator_llm.py
+++ b/tests/unit/memory/wiki/test_curator_llm.py
@@ -764,6 +764,11 @@ async def test_propose_updates_uses_default_registry_if_none_injected(
"jarvis.memory.wiki.provider_chain.credential_ready_wiki_providers",
lambda *, available, **_kwargs: set(available),
)
+ # This host's real CLI logins must not reorder the chain under test.
+ monkeypatch.setattr(
+ "jarvis.memory.wiki.provider_chain.subscription_login_ready",
+ lambda _provider, **_kwargs: None,
+ )
llm = WikiCuratorLLM(
config=cfg,
diff --git a/tests/unit/memory/wiki/test_provider_chain.py b/tests/unit/memory/wiki/test_provider_chain.py
index b7019a1b6a..52b11cc476 100644
--- a/tests/unit/memory/wiki/test_provider_chain.py
+++ b/tests/unit/memory/wiki/test_provider_chain.py
@@ -173,6 +173,40 @@ def test_credential_probe_uses_core_portable_storage_and_keeps_oauth(
assert ready == {"nvidia", "future-oauth"}
+def test_background_work_skips_paid_keys_while_a_subscription_is_connected(
+ monkeypatch: pytest.MonkeyPatch,
+) -> None:
+ import jarvis.memory.wiki.provider_chain as chain_module
+
+ monkeypatch.setattr(
+ chain_module,
+ "credential_ready_wiki_providers",
+ lambda *, available, config: {"openai", "gemini", "codex", "ollama"},
+ )
+ ready = chain_module.background_wiki_providers(
+ available={"openai", "gemini", "codex", "ollama"},
+ config=object(),
+ is_subscription=lambda name: name == "codex",
+ )
+ assert ready == {"codex", "ollama"} # subscription + keyless local only
+
+
+def test_background_work_keeps_keys_when_no_subscription_exists(
+ monkeypatch: pytest.MonkeyPatch,
+) -> None:
+ import jarvis.memory.wiki.provider_chain as chain_module
+
+ monkeypatch.setattr(
+ chain_module,
+ "credential_ready_wiki_providers",
+ lambda *, available, config: {"openai"},
+ )
+ ready = chain_module.background_wiki_providers(
+ available={"openai"}, config=object(), is_subscription=lambda _name: False
+ )
+ assert ready == {"openai"} # a single-key install keeps a working wiki
+
+
# --- the fallback loop -------------------------------------------------------