Every write to the per-sign-in Codex memory checked the account digest,
but the reads did not. A credential reconnected to another ChatGPT
account therefore tried the old account's working model first, sorted
by its refusals and trusted its model list, until the first write
rebound the entry, and the vision gate kept answering from the old
account in the meantime.
The client reads (working model, refusals, model list) now take the
credential and treat an entry learned under another account as empty.
The vision gate holds no account id, so the connect routes rebind the
entry and the disconnect routes drop it, and the gate reads the entry
as its process last bound it.
A fixed preference list goes stale at every retirement: when GPT-5.5
retired for ChatGPT sign-in on 2026-10-14, every model the shipped list
named was refused and Codex answered 503 until a release changed it. The
working model also sat in one slot for the whole process, so two sign-ins
on different plans overwrote each other's model and the vision gate read
whichever had answered last.
The client now reads the models the account is offered the way the Codex
CLI does (GET /backend-api/codex/models?client_version=..., same sign-in
as the responses call; shape from codex-rs/protocol/src/openai_models.rs)
and walks: CODEX_MODEL first; our preference order kept to what is
offered; the other offered models in the list's priority order, leaving
out hidden, specialist and frontier-tier entries; the rest of our order
last, so the walk never holds fewer models than before. With
CODEX_MODEL_FALLBACK_CHAIN set the order is the operator's and the list is
not read.
The list is a hint, never a gate. The read goes through safeFetch with a
5 s bound and a 4 MiB cap, is shared by concurrent requests, cached per
sign-in for an hour, and a failed read waits ten minutes. Any failure
leaves the static walk, one ai.codex.models_unreadable warning per sign-in
reports it, and each request that builds the walk annotates a pinned
codex.models.* shape.
The working model, refused models and the list are kept per stored
credential (a person's connection, or the operator-shared central one),
bounded by TTL and a 1000-entry LRU, with a digest of the account id so a
credential reconnected to another account starts empty. A model the list
offered that then answers model_not_found goes to the back of that
sign-in's line for six hours. When the list names a model's input
modalities, the Codex vision gate reads them; otherwise the model family
decides, as before.
Refs #1287
A gateway that refuses the reasoning field does so by its configuration,
so the lesson now holds for 24 hours instead of one. At the hour every
expiry cost one refused request, a few a day on a gateway that runs the
briefings unattended. Lessons read from a single reply keep the hour.
Refs #1279
OpenAI retired the 5.x line from Codex for ChatGPT sign-in in steps: GPT-5.2
and GPT-5.3-Codex on 2026-05-26, GPT-5.4 and GPT-5.4 mini on 2026-08-31,
and GPT-5.5 on 2026-10-14. The backend answers a retired model with 400
model_not_found and "not supported when using Codex with a ChatGPT
account".
Pin that exact refusal as a reason to try the next slug, pin that an
account refused the head of the default list walks on to the next rung
and caches it, and keep the long-retired models off the default list.
gpt-6.1-sol, a rung on the default list, answers reasoning "off"
(effort none) with 400 unsupported_value and lists low as its lowest
effort (seen 2026-10-24). The client must stay on the slug, retry with
low and remember it, not walk on as if the model were retired.
Background jobs ask for medium reasoning at most, and status cards ask for
none. A model that cannot be switched down (GLM behind a vLLM gateway, for
instance) ignores both, spends the whole room thinking and returns only
reasoning, so the daily briefing and the status cards fail as
reasoning_only and the chain pauses. The Coach on High works because it
leaves 12k tokens to think in.
The chain runner now marks every call of a job walk as unattended, in one
place rather than at each call site. The gateway and local clients give
such a call at least the High thinking room once its model has answered
with reasoning alone. The room has to ride on the #1267 repeat with none:
one call may send two requests, and a once-a-day job meets an expired
lesson every morning, so a room that waited for the unswitchable lesson
would never reach the first briefing of the day. It is a ceiling, not a
charge: a model that does stop at none never touches it. Calls a person is
reading, the Coach included, keep the room of their own level, and the
requested effort is unchanged.
The room calculation both clients shared is now one function.
Reported by @Tio61 in #1276.
A gateway or local entry can carry its own reasoning setting (#1126), sent
on every call that has no level of its own. An endpoint that refuses
reasoning_effort answered each of those calls with a 400 that was never
learned, because only the call's own level counted as sent: every status
note, every summary and every other level-less call failed, for as long as
the setting stood.
The refusal is now learned whichever level put the field on the wire, by
the same path as any refused field: the call is asked once more without it
inside its two-request budget, later calls send no field for that endpoint
and model, a model seen reasoning unasked keeps its thinking room, and one
ai.provider.reasoning_field_refused line says so per lesson instead of a
failure per call.
Refs #1279
A status run whose whole chain was benched for background work sent no
request, yet the caller opened the note's negative window as for any
failure and raised its failure streak. The streak doubles the window up to
two hours, so a note kept waiting long after its provider recovered.
AllProvidersFailedError now says when no hop sent anything (sentNothing:
every provider benched, or refused by the operator cost cap). The status
chokepoint then holds the note for the first window without touching the
streak (holdStatusNoteWhileBenched), and the caller's own failure write
lands inside that window and neither moves nor counts it. A run that
reached a provider counts as before.
Refs #1279
A chain whose first provider reasoned without answering and whose second
one answered billed only the answer: result.tokensUsed is the answering
call's alone, and the failed hop's spend, already reported on its
FallbackHop, was dropped by every caller that reconciles the day's ledger.
The raw runners now report tokensSpent on success too, the sum the failed
hops before the answer reported. The briefing, status, Coach (no-tools path
and every tool-loop round) and medication-extract reconciles charge it on
top of the answer. On failure, the Coach turn ledger's close keeps what the
failed chain reported spending instead of refunding everything, and the
medication-extract and document-chat routes charge it too, as the status
and briefing paths already did. Document indexing resolves a single provider
and already charged its failed call.
Refs #1279
The on-demand briefing (POST /api/insights/generate) runs through the
briefing chokepoint, which named its chain walk "job" for the background
share of the day's ceiling. Since background work leaves out a provider the
health ledger benched, a person pressing "Re-run analysis" got no request at
all: the call failed with a connection refusal that never happened, and a
refused credential was never reported back.
The chain runners now take the walk apart from the budget surface: a call
can be rationed as a job and still have a person waiting on it
(interactive). runBriefingCompletion takes the flag and the on-demand
route's calls (the first pass and the grounding retry) pass it, so they keep
the job share and walk a benched provider at the tail, as background-bench
promises for interactive walks. Every other job-surface caller is a worker
or is enqueued; the status, narrative and warm routes only read and queue,
and document reads and lab scans resolve a single provider without a chain.
Refs #1279
A reply that ran into the token ceiling with unlabelled text was read as
reasoning the server did not separate whenever the ceiling held room for
reasoning, which is every call that carries a level, Off included. A model
that does not reason at all (llama3.1) lost its own truncated answer: the
call failed as reasoning_only, the briefing never reached its JSON-shape
retry, the Coach named the wrong error, the failure counted toward the
bench, and the client learned to ask with "none" for an hour.
Such text is now reasoning only once the endpoint and model have shown that
they reason: a reasoning field or a think tag in an earlier reply, or a
lesson that says so (knownReasoner). A sighting on a request that asked to
reason speaks only for requests that ask too. Otherwise the text is the
answer, cut off, with finishReason "length".
A lone closing </think> now splits a reply only where the chat templates put
it, at the start of a line, or anywhere once the model is known to reason. An
ordinary answer that mentions the tag in a sentence stays whole.
Two expectations encoded the old rule and move: the Coach loop case whose
first reply is unlabelled template-opened text from a model never seen
reasoning, and the Local case that read a never-seen model's cut text as a
thought. The streamed think-block splitting moved into local-reasoning.ts.
Refs #1276
A gateway that refuses reasoning_effort outright (LiteLLM in front of
Ollama: "does not support parameters: ['reasoning_effort']") in front of a
model that reasons without answering sent requests until something outside
stopped it. The answerless rule learned to ask with "none", the refusal of
that "none" was learned as "send no field" over it, the next request came
back answerless again, and each lesson undid the other.
Each call of the Local and the OpenAI-compatible gateway clients now carries
one request budget through every repeat (request-hops.ts): at most two
requests, after which the call ends with the failure it has. A structural
guard holds both clients to sending only through the budget and to handing
it to every repeat.
The refusal is now learned as what it is: an endpoint that refuses the
field, in front of a model already seen reasoning unasked, keeps both facts
(FIELD_REFUSED_REASONS_REGARDLESS), so it is never asked with "none" again
and keeps its thinking room. An answerless reply from an endpoint that
refuses the field is final. When the refused "none" was the repeat of an
answerless reply, the call fails as reasoning_only, which is what it ran
into. The refusal handling of both clients lives in wire-refusal.ts.
A status run that hits its bound now aborts the call it gave up on instead
of leaving it running detached.
Refs #1279
The health ledger already counted every failed call and paused a provider
that kept failing, with a pause that grows. A paused provider only moved to
the tail of the chain, though, and the tail is still walked: a chain whose
only entry answered every call with reasoning and no answer was called by
every background job regardless.
A walk for background work (surface "job") now leaves out every provider
the ledger has benched until its window lifts; the local floor only once
it is paused, past a brief backoff. One ai.chain.background_paused line
per window says so. A person's own call still reaches the provider at the
tail and reports its own error. The ledger is in Postgres, so the bench
holds across workers and restarts.
The status, briefing and document-index paths refunded a failed call in
full, so background spend on replies without an answer never reached the
day's token ceiling. A failed chain now carries the tokens its hops
reported (tokensSpent) and those paths charge them.
The two clients share one answerless step (askAgainUnreasoned), which
keeps both inside the source size ratchet; its baselines go down.
Refs #1279
The #1267 rule asks a reply that reasoned without answering once more with
reasoning_effort "none", and asks every later call that way for an hour.
That is what Gemma 4 on Ollama needs. gpt-oss behind vLLM accepts only
low / medium / high and ignores "none", and a GLM chat template reads an
effort it does not know as its maximum, so for those models the repeat and
every later "none" were wasted requests or longer thoughts.
When a reply sent as "none" comes back without an answer as well, the
Local and gateway clients now learn that too: later calls go out as the
call asks, with the thinking room #1276 keeps, and an answerless reply is
final. One call costs two requests at most and one after that, and the
log gets one line for each lesson instead of one per call.
A failed answerless call now carries what it cost on its error, and a
repeat that answers reports both requests, so the day's token ledger can
charge them. The learned lessons are pinned on globalThis, so the Coach's
route handlers and the worker's jobs share one copy per process instead
of learning the same thing twice.
Refs #1279
A gateway or local server in front of a model that always reasons (GLM
behind vLLM, for one) accepts reasoning_effort "none" and ignores it. Two
things went wrong with such a model.
Once it answered a round with reasoning alone, the #1267 lesson asked every
later call with "none", and the token ceiling followed the wire: "none" meant
no thinking room, so a Coach round got its bare 600-token answer budget. The
model spent that thinking and stopped at finish_reason "length" before it
reached a tool call or an answer, for the hour the lesson holds. A turn set
to Off had the same ceiling from the start.
Where the server did not separate the reasoning (no reasoning parser, or one
that falls back to content), the reply's content was the thought itself: an
inline <think> block, the text before a lone </think> when the chat template
opened the block in the prompt, or unlabelled text cut off at the ceiling.
The gateway client passed all of it through as the answer; the Local client
split a leading <think> block, and only for callers that asked to reason.
On these wires the ceiling now follows the call's level instead of the wire
effort: Off keeps the low room, a call sent as "none" keeps its own level's
room, and a call without a level gets the low room once the model has shown
it reasons unasked. It is a ceiling, not a charge; a model that does switch
off never touches it. Both clients now take reasoning out of content for
every caller: a leading or a lone closing think tag, and a reply that filled
a ceiling sized for thinking with unlabelled text and no reasoning field
beside it. What comes out goes to the reasoning channel when the caller asked
for it and nowhere else. A reply left without an answer takes the #1267 path
and ends as reasoning_only, never as the thought.
The loop test drives a Coach turn through the gateway client onto an
upstream that answers like the reported wire capture, with and without a
reasoning parser.
Refs #1276
Extracting lab values from a stored document still sent the whole text
in one call. A multi-page report produced more JSON than one call's
output ceiling, the answer was cut mid-fact, the parse failed twice, and
the read came back as 422 "Couldn't read the document. Try a clearer
copy." The lab scan had already learned to read the same report a page
or a part at a time, but each surface carried its own copy of the read,
so the document path never got it. It also ran under the shared 60 s
timeout, dropped every fact past 120 (or every fact when one did not
match the schema), and charged a flat reservation.
The read now lives once, in lib/documents/ocr-read.ts and ocr-plan.ts:
the per-read limits, the corrective retry, the salvage of a cut
answer's complete rows, the wire re-check and budget reservation before
each read, and the planning of a PDF a page at a time and text in parts
at line boundaries. The lab scan and the document extraction (the
extract route, its background run and the lab auto-staging) both read
through it, and the background run plans its reads on the same deadline
as a lab scan. The per-read budget moves to a shared metered budget.
The document path validates the answer fact by fact, keeps the lab
scan's row cap (400), and says why nothing came back: 422
documents.inbound.outputLimit when the reads stopped at the output
limit, 429 when the budget refused them, extractFailed only for answers
that were not the schema. A run whose reads were cut or not reached
reports `incomplete` on its result.
A guard holds the class: the output ceiling is set only in the core,
both surfaces read through runOcrReads, both runs share the deadline,
and a lab read and a document read go out under the same limits.
Refs #1257
The ChatGPT Codex backend and the OpenAI API serve gpt-6-sol, gpt-6-luna,
gpt-6-astra and gpt-6.1-sol, but the reasoning, Chat Completions and vision
gates only knew the gpt-5 line. A gpt-6 model got no reasoning control and
no real "off", went out on the legacy wire with max_tokens and a
temperature that reasoning models refuse, and was denied image input.
The three gates now read one classifier, openai-model-family.ts, so a new
generation is taught in one place. Every GPT from 5.1 on, the gpt-6 line
included, is the family that takes effort "none"; every GPT from 5 on reads
images. The upstream catalog (codex-rs/models-manager/models.json) lists
image input for every gpt-6 slug and the same reasoning presets as the
gpt-5.6 line; the presets omit "none" for gpt-5.5 too, which accepts it.
The Codex slug ladder moves into codex-slug-chain.ts and gains gpt-6-sol,
gpt-6.1-sol and gpt-6-luna. gpt-5.6-terra stays first as the slug verified
against the backend; the catalog names gpt-6-sol as its upgrade, and a
rotation retires a whole line at once, so the gpt-6 rungs follow before
terra's own siblings. The ladder test now holds every rung to a recognised
reasoning family with a real off and image input. The OpenAI presets offer
gpt-6.1-sol and gpt-6-luna.
The ChatGPT Codex backend now rejects every slug from gpt-5.5 down, so a
deploy without a CODEX_MODEL_FALLBACK_CHAIN override walked the whole
default ladder into "all configured Codex slugs were rejected". The
default now starts at gpt-5.6-terra, with its luna and sol siblings and
gpt-5.5 as the last rung.
The gpt-6 line stays out of the default: the reasoning, sampling and
vision gates do not recognise that family yet, and a test now holds every
default slug to a family the reasoning gate knows.
next stays on main's 16.4.2 (the RCE and Server Actions advisories) with its two companions; release's newer dev tools and its own additions are kept. The lockfile is release's, re-resolved. pt-BR, which main does not have, takes the one key main added.
Add pt-BR to the locale list and everything that hangs off it: the
client loader, Intl and date-fns mapping, the reply-language footers of
the Coach and insights prompts (English body, pt-BR footer, the same
path Korean and Dutch take), the citation-coverage normative words, the
no-key fallback copy, the AUDIT-C standard-drink line, country labels,
period search, CSV yes/no cells and the Telegram bare-word aliases.
The locale enums in the extract, managed-profile and record-settings
schemas now derive from the locale list instead of repeating it, and
the five notification and login-alert routes share one isLocale guard
instead of hand-rolled chains, so a later locale is one edit there.
A language tag is matched exactly and case-insensitively before its
primary subtag, so pt-BR and pt-PT browsers land on the Portuguese
bundle rather than English, and AUDIT-C no longer cuts a region tag to
two letters.
PHQ-9, GAD-7 and WHO-5 are not marked validated for pt-BR: the wording
has not been read against the published Brazilian forms, so readers get
the English items plus the existing note.
The regenerate button aborted its inline POST after a fixed 45 s, so on a
slow local model every regenerate was cut while the server still had
minutes left on the person's setting. It now uses clientAbortMs over the
route's sequential model calls, never less per call than the briefing's
own 180 s default, because the published value cannot say whether the
person set it.
A save of the about-me self-context derives the clarifying questions
inline and ran under the 15 s request default, which reported a failed
save while the server went on to commit it. All six save sites now take
their abort from the same setting. Every other browser request that waits
on a model either has no client timeout or only queues the work.
Refs #1267
Ollama serves thinking models such as Gemma 4 on /v1/chat/completions with
the thought in a separate `reasoning` field. A long briefing can come back
with empty `content`: the thought used the whole token budget, or the
answer itself was parsed into `reasoning`. The background briefing always
asks to reason, so every briefing on such a model ended in the empty-reply
sentinel, which the chain logged as status 0 / cause "network" in the same
second the server logged 200, and the read path showed as a timeout.
The Local client and the OpenAI-compatible gateway now treat a reply with
reasoning and no answer as a lesson: the endpoint and model are remembered
in the reasoning dialect cache as answering only unreasoned, the call is
asked once more with reasoning_effort "none", and later calls ask that way
until the lesson expires. A test drives the real client, safeFetch and the
pinned dispatcher against a loopback server that answers the way Ollama
serves Gemma 4, and reproduces the reported log line without the fix.
Failed hops now carry what the call ran into: `failure` (timeout, aborted,
refused, reset, socket, parse, empty_response, reasoning_only, ...) and a
code-shaped `code` on ai.chain.link_failed, closed vocabulary only. The
cause only says "timeout" for a deadline, and the briefing failure marker
records the kind so the empty state suggests a longer timeout only when
time ran out.
Refs #1267
The text layer went to the model as the authority for digits. It is not
one: a scanned PDF often carries an invisible OCR layer with its own
mistakes, a subset font can map to the wrong characters, and a crafted
PDF can carry invisible text that differs from what the reviewer sees.
The prompt now calls it a hint only, makes the image authoritative, forbids
adding a value the image does not show, and asks for a low value
confidence where the two disagree. The layer is fenced with the document
text markers, which it cannot close, and so is text mode's OCR'd text.
Each page's layer is capped while it is collected, and a read shares its
text budget out per page, cut at a line and marked, so a long first page
no longer pushes the later ones out.
A further read now first books what the reads so far cost beyond their
reservation, then its own ceiling, both against the cap of the owner the
route judged the first read by, so the counter the next read is judged
against never lags. The wire is re-checked before every read, the first
included. A re-check or reservation that itself fails stops the scan with
the pages read rather than failing it, and a scan stopped before any row
answers with the stop's reason (budget, or the wire's own refusal) instead
of an unreadable page.
A scan sent the whole report to one call under an 8000-token ceiling and
validated the reply against a 100-row envelope. A long multi-panel report
lost everything either way: the reply stopped mid-row and failed to parse,
which the person saw as "try a clearer photo", or it held more than 100
rows and the schema refused all of them. One malformed row refused the lot
as well. Reasoning models made it worse, because the call sent no
reasoning setting and OpenAI, Codex and Claude's 5 generation reason by
default against the same output allowance.
A PDF of more than one page is now rendered and read a page per call on
every provider (up to 20 pages), as many reads as the run's deadline
allows, and the pages are merged in order with a reading printed twice
listed once. A photo, a one-page PDF on a provider that reads PDFs, and
OCR'd text up to 4000 characters stay one read; longer text is split at
line boundaries. Each read carries an explicit reasoning off, so a model
that can stop reasoning does and one that cannot gets its minimum budget
on top of the answer's. A reply cut at the output limit is detected from
the finish reason (Codex's incomplete stream included), keeps the rows
that closed before the cut and is named in the result instead of retried;
when no read produced a row the run fails as labs.ocr.outputLimit. A
provider failure after the first page keeps the pages already read.
Rows are validated one by one, the row cap is 400 for the scan and the
commit, and a report past it keeps its first 400 and says how many it
had. The run settles what the provider reported it spent rather than the
fixed reservation, so a page-by-page read still counts against the day's
budget in full.
Two callers refreshing one Codex token raced: the loser heard
refresh_token_reused, read it as a dead grant and parked the connection
before the winner wrote its rotated tokens, leaving fresh tokens beside
an expired status. Refreshes now share one flight per process and queue
on the per-credential advisory lock across processes, re-read the row
inside it, use a token a peer already rotated, write the rotated tokens
compare-and-set on the spent one together with connected, and re-read
after a refusal before parking. refresh_token_reused no longer counts as
a dead grant, and neither does a bare 403 without an OAuth error code:
an edge challenge answers a valid grant so, and it keeps the ledger's
timed cooldown instead of parking until the next sign-in.
Output of `node scripts/strip-planning-tags.mjs --write`: 3,157 comment
lines in 1,308 files lose the work-item identifier at their head. Release
versions and issue references stay; nothing outside comment text
changes, which the script checks per file against a comment-free
printout. healthlog/no-planning-tags is enabled at error level with this
commit, so the tree it lands on is clean.
Adds messages/nl.json (AI-initial, informal je/jij register, full key
parity with en.json) and registers nl everywhere the locale list lives:
config, loaders, Intl/date-fns mapping, zod enums, OpenAPI spec, scripts,
the managed-profile picker and the notification locale guards.
Dutch follows the Korean pattern for the AI surfaces (English body plus a
Dutch reply-language footer, no safety-contract matrix) and gets its own
banks in the outbound safety screen (dose, risk, causal, continuation),
the prose-grounding and citation-coverage tables, the reference-range
prose and the no-key fallback copy, each with Dutch test fixtures.
PHQ-9 / GAD-7 / WHO-5 are not marked validated for nl yet: the Dutch
wording has not been read against the published forms, so Dutch readers
get the English items plus the existing note until it has.
The AI settings page posted to /api/consent/ai/web on every mount and the
server minted an ai_full receipt for it, so an account that had agreed to
nothing sent its health data to the operator's provider from background
jobs. Native builds synced a receipt the same way on every launch.
- Both grant routes now require `confirm: true`; anything else writes
nothing and answers 400 consent.ai.confirmRequired.
- Remove the mount-time heal. A consent card names the operator's service
and the data that would go there, with one Agree button, in Settings > AI
and on the Insights overview while AI is closed by the missing consent
alone.
- Publish the standing consent as `ai.consent` on /api/auth/me (status,
since, recipients); the settings card renders active-since and the
withdraw control from it.
- Existing receipts stay valid.
- Guard that no client grants outside a click handler, route tests for the
confirmation, and an integration test that a fresh operator-key account
gets no background AI text until it agrees.
Three levels (frugal, standard, full) with an operator ceiling and default
and a per-account choice within them, resolved in the capability layer for
the record and published on GET /api/auth/me as ai.usage: the level in
effect, the stored choice, the bounds, why (user, default, admin_cap), the
three options with whether each is allowed and an estimate of tokens per
day scaled from the record's last seven full days of coach_usage, and the
measured mean. A delegate sees the level, never the owner's spend.
PUT /api/auth/me/ai-usage stores the account's own choice or null for the
default, refusing a level above the ceiling with 422 ai.usage.aboveCeiling
and meta.maxLevel. The new supplementLabel capability needs standard or
above; below it the capability is unavailable with the new usage_level
reason, refused with 403 ai.usage.levelTooLow and meta.requiredLevel.
Refs #1217
The nightly briefing warm completed every night and read "all-failed",
because one run event held every account's annotations and the last
account won. Each candidate now runs in its own wide event and the run
reports every outcome through `reportJobRun`, so ok, dead sign-in,
withheld and failed accounts show side by side.
A dead Codex refresh token threw a bare error with no status, so the chain
read it as a network failure: the health ledger recorded hard_failed, no
"reconnect" hint appeared, and every night tried the same token. The
refresh now throws `CodexRefreshError` with a 401 for a dead grant
(invalid_grant / invalid_refresh_token on a 400 included), as do a missing
or undecryptable stored credential. The ledger records auth_failed, the
chain marks the primary credential expired, and the nightly pass reports
the account as auth_failed without queuing a retry.
A briefing the outbound screen withheld (a causal claim) failed outright,
and the 45-minute provider retry then asked the same question eight times
a night. The generator now makes one bounded repair that names the broken
rule and keeps the repaired payload only if it passes the same parse,
grounding and screen; a withheld outcome queues no retry.
A provider that keeps failing is paused: from five consecutive hard
failures the skip window starts at 30 minutes and doubles per further
failure, capped at a day. A paused provider is walked last and tried again
when the window lifts; one success resumes it. The operator card shows how
many accounts it is paused for and until when. After a 5xx from a
non-canonical OpenAI-compatible endpoint the client asks the free model
listing, and a configured model the endpoint no longer offers is named in
the error, on the hop line and on the card.
Also:
- The tiered context's live day band read `measured_at AT TIME ZONE tz`,
which treats the stored UTC as local time; a far-from-UTC reader's
evening readings landed on the next day. It converts from UTC first now.
- Google Health intraday heart rate re-reads two hours per sync instead of
a day; daily summaries and sleep keep the day (#1023).
- The first filtered daily read that answers 400 and triggers the
snake_case retry is recorded as a note, not an error (#1023).
- The single derived-score read goes through the derived cache with
stale-while-revalidate, like the batch beside it.
Own-plan turns get 165k tokens, 200 s and 16 rounds, operator turns 55k,
120 s and 8; the final-answer reserve, the concurrent-turn slot and the
step cap move with them.
Codex requests carry a per-conversation prompt_cache_key and a stable
session id, api.openai.com gets the same key, Anthropic usage counts
cache reads and writes in the gross, and local servers report cached
prompt tokens, so the turn budget discounts cached input on every
provider.
The Coach loop's final round sent toolChoice "none" without the tool
definitions. With tool_use/tool_result blocks in the history Anthropic
refuses such a request (400), and OpenAI Chat Completions refuses
tool_choice without tools, so every budget, time, cap and no-progress stop
ended in an error on those wires. The final round now carries the tools
with toolChoice "none", which every wire accepts, thinking included.
Codex kept reasoning items as provider state even when the request asked
for no encrypted reasoning; replaying such an item under store:false is a
400. Items are kept only when the request included
reasoning.encrypted_content and the item carries it.
The reasoning dialect cache learned from any 4xx that mentioned
reasoning or thinking (auth, rate limits, model names such as
deepseek-reasoner) and from replay errors, and kept the downgrade for the
life of the process. Learning now requires an HTTP 400 that refuses a
sent parameter, ignores replay errors and the model name, and a learned
downgrade expires after an hour.
providerState may be read only in the clients, the contract and the Coach tool loop, never on a line that logs, annotates or persists, and has no schema column.
The mock reports a fixed summary, title and provider state for calls that ask for reasoning. The runner passes reasoning and onReasoning to every candidate unchanged and adds the reasoning token and summary counts to the wide event, never the text or the state.
OpenAI sends reasoning_effort for reasoning models on rounds without tools, off mapped to each family's lowest value. Gateways send reasoning_effort, OpenRouter its reasoning object, and the reasoning details or thinking blocks a gateway returns go back on the next round's assistant message. Local servers send reasoning_effort and read reasoning_content, reasoning and a leading think block, live on the stream without letting it reach the answer.
The thinking budget rides on top of the answer budget where reasoning counts against it. A refused reasoning field is retried once without it and remembered per endpoint and model, before the JSON dialect check. A call's own level wins over the stored chain setting, which still applies when the call has none.
Up to 4.5 a thinking budget (with the interleaved-thinking header on tool rounds), from 4.6 adaptive thinking with an effort and summarized display. From the 5 generation thinking cannot be switched off, so off becomes the lowest effort. Temperature is never sent from 4.7 on, nor with thinking, which also removes a 400 on the newer model presets.
Thinking blocks travel to the next round unchanged, signature included; a tool round another provider started runs without thinking. Every text block is joined instead of only the first. The JSON instruction goes only to callers that ask for JSON. Unknown model names learn the dialect and the sampling rule from a refusal.
A call that carries a reasoning level sends reasoning {effort, summary} with include [reasoning.encrypted_content]; off is effort none. Summary titles are reported as soon as the bold first line closes, the text when the part is done. Reasoning items come back as provider state and are re-inserted before the next round's message and calls, only for the slug that made them.
A 400 that names a reasoning parameter is answered with the lesser request the backend lists (no summary, the nearest supported effort, no reasoning), remembered per slug. Calls without a level keep reasoning null and include [].
The provider-chain summary marked every entry available, so the Coach's
provider select offered entries without a key or sign-in. Choosing one
saved it to the front of the chain, the resolver skipped it, and the
select jumped back without a word.
Availability now comes from the presence probe, which names every chain
type whose credential is present. After a switch the picker reads the
chain back and says so when the server still answers with another
provider.
An endpoint that does not know reasoning_effort answers with an unknown
parameter error, which the JSON-mode probe read as a refusal of
response_format and switched that base URL to plain prompts.
Reasoning models on Ollama and similar servers think by default and can
spend the whole max_tokens budget before the visible answer starts, so
the call comes back empty or times out. The Local and OpenAI-compatible
chain entries now carry an optional reasoning setting: Default sends
nothing, Off sends reasoning_effort "none", Low / Medium / High send
their own value. It is stored on the entry in aiProviderChain, so a
chain can mix a thinking model with one that does not know the field.
- Every exported resolver binds the entry's value onto the client, the
connection test included, keyed on the instance type so the operator
provider, OpenAI, Anthropic and Codex never carry it.
- PATCH /api/user/ai-provider accepts localReasoningEffort and
compatReasoningEffort and GET reports them. A provider absent from
the chain holds its value on a disabled entry; restating Default on a
chain that was never customised writes nothing.
- PUT /api/insights/provider-chain accepts reasoningEffort on local and
openai-compatible entries, refuses it on any other, and keeps the
stored value when the body leaves it out. Its schema is now shared
with the OpenAPI registry.
- A Default request body is byte-identical to the one sent before.
Refs #1126
The response-timeout setting reached the briefing, the status notes and
the Coach, which threaded it by hand. Document reads, document
extraction, lab OCR and lab staging called the provider without a
timeout, so every client fell back to its own 60 s literal and a slow
self-hosted model was cut off however high the setting stood.
The setting now rides on the provider instance. resolveProvider,
resolveProviderChain and resolveProviderForTest stamp the record's
aiResponseTimeoutSeconds on everything they return, and each wire
client reads its ceiling through callTimeoutMs: the setting, else the
surface's own timeoutMs, else 60 s. A call site no longer has to pass
it, so a new one cannot forget it. The connection test runs under the
same ceiling a read does.
The Coach nudge tick and the reaction line keep their short ceilings
(timeoutPolicy: "surface-ceiling"): the tick runs accounts one after
another under a shared budget, the reaction line's claim lease is
shorter than the largest setting, and both have deterministic copy
that stands in.
/api/auth/me publishes the effective value as
ai.provider.responseTimeoutMs, so a client can size its own request
timeout from it instead of guessing.
Refs #1090
Making `=true` grant nothing would stop a single-user self-host's local
model on a patch update. It is deprecated for one release cycle instead,
in a narrowed and safe form: it still grants a private host, but only to
the operator's own configurations (the instance-wide admin provider and
AI settings saved on an admin account), and only through the pinned
operator-approved dispatcher, so redirects stay forbidden and metadata,
link-local and unspecified addresses stay refused. A non-admin account's
own base URL needs an exact origin in AI_PRIVATE_ORIGINS.
Provider resolution passes the owning account's role to the client
(`operatorTrusted`); the save route and the connection test apply the
same rule to the account being configured. The boot summary now says the
setting is deprecated and names the replacement, and the admin AI
settings page shows the exact AI_PRIVATE_ORIGINS line for the private
origins saved on the instance (every account's Local and gateway URL and
the admin base URL, scheme, host and port only), so the operator can copy
it and drop the old setting.
ALLOW_LOCAL_AI_PRIVATE_HOSTS=true let any user of the instance point the
Local provider or the OpenAI-compatible gateway at any private address
the server could reach, the metadata range included, and a granted host
was dialled with no connect-time pin at all.
AI_PRIVATE_ORIGINS grants a private AI endpoint with the exact-origin
grammar NOTIFICATION_PRIVATE_ORIGINS and NIGHTSCOUT_PRIVATE_ORIGINS use.
A granted call goes through the pinned operator-approved dispatcher:
redirects forbidden, and metadata, link-local and unspecified answers
dropped at dial time even for a granted name. Everything else stays on
the public pin. The Local client (buffered and streaming) and the
gateway and admin-key tags of the OpenAI client all take their policy
from one function.
The host-list form of ALLOW_LOCAL_AI_PRIVATE_HOSTS keeps working with the
same pinning; a never-grantable literal or a host:port entry is refused
with a one-time warning. Its "true" form grants nothing any more and is
reported in the boot readiness summary with the replacement.
A failed lab-report scan no longer puts the endpoint's response body into
the wide event; status and model stay. The response to the caller never
carried it.
The status chokepoint, the comprehensive briefing, reaction lines and
AI-composed nudges now ask `aiEgressRefusal` about exactly the chain they
are about to send to, instead of each pairing the operator-key check
with a surface receipt lookup of its own. The consent kinds come from the
capability table, so the wire and the published capability answer from
the same rule.
The document summary, its catch-up pass and the index backfill stop on a
closed switch or module before anything is picked, and leave "no
provider" and "consent needed" to the document pick, which re-checks the
wire for the provider it chose. A closed switch still only heals PENDING
to NONE; a withheld pick records UNAVAILABLE. The per-pick document
consent helper is removed, since nothing calls it any more.
The auto-read toggle mints the shared `AI_EXTRACTION_CONSENT_KIND`, and
the web grant helper goes back to minting the master grant only.
Refs #975
Every AI job now asks its capability for the user before it builds a
snapshot, and skips with a `.skipped` annotation naming the reason: the
nightly pregenerate (briefing and statusText, each half on its own), the
status batch and the per-metric crons, the on-demand status worker, the
narrative cron, Coach nudges and their AI composer, the Coach memory
refresh, reaction lines, workout paragraphs, and the document summary,
catch-up and index jobs. Seven of these read no assistant switch at all
before, so the master switch now stops them too.
The cron candidate queries drop people who switched AI analysis off in
SQL, and no longer read `disableCoach` outside the Coach: hiding the
Coach hides the Coach. Coach nudges use the one presence probe, so a
central Codex account is no longer skipped; a missing consent still gets
the deterministic template. The arrival spine asks only the
provider-free half (switches and modules) before it enqueues, keeping
its module graph free of provider code.
Document reading needs an extraction receipt (ai_extraction or ai_full)
for any external provider; the auto-read toggle is the trigger, no longer
a consent shortcut, so a revoked receipt wins over a toggle left on. Lab
report scans get the same rule for a whole chain.
Two structural guards: every file that calls a provider is named with
the reason its call is admitted, and the chokepoints re-check a
capability; every AI job references its capability gate. An integration
matrix runs each worker under the refusing states with a provider spy.
Refs #975
probeProviderChain describes the chain that would serve a record
without decrypting, building a client or refreshing a token: which
entries hold a credential, in chain order, whether each reads images,
the in-browser OCR opt-in, and where the credential comes from. It
applies the provider-work authority, appends the central Codex only
behind the opt-in, and falls back to the legacy provider only for an
empty chain. hasAnyConfiguredProvider and resolveProviderAvailability
now answer through it.
Background insights can no longer use up the day before somebody asks a question; a fifth of the daily limit is held back for what a person does by hand.
The background ceiling applied only when the operator funded the chain, so on a
chain the operator does not pay for the automatic generators and the chat shared
one undivided ceiling and the first could still take all of it. That is the
configuration the recommended remedy for the original lockout produces — put the
personal Codex account first — so the guarantee was missing on exactly the setup
it was written for.
The ceiling is now expressed as what it is meant to protect. Background work may
run until the day's spend would leave less than a fixed reserve for whoever is
waiting on an answer, on both cost owners. A reserve and a quota look alike at
the boundary and mean opposite things in the middle: a quota stops useful
generation on a day nobody was ever going to interrupt, while a reserve costs it
nothing until the moment it would take the last of the day. An account
generating 1.45 million tokens of insights on its own plan keeps every one of
them and the chat still holds 400 000.
INTERACTIVE_RESERVE_SHARE replaces JOB_SURFACE_SHARE. The old name described a
quota over the jobs rather than a floor under the person waiting, and a name
that states the opposite of the intent is what put the wrong comparison in this
ledger twice already.
The fraction is a default an operator will want to change. The setting and its
documentation are separate work.
Marc-André Bombeck