v0.40.0

Release Notes — v0.40.0

Minor: A built-in catalogue of model facts

ogcode now knows the context window, prices (cache reads and writes included)
and image support of the current Anthropic and OpenAI models and the major
open-weight families — DeepSeek, Qwen, Kimi, GLM, MiniMax, Llama and Muse,
Mistral, Gemma, gpt-oss, Nemotron and Phi — checked against each vendor's docs
on 2026-09-28. The facts follow the model, not the provider: it is recognised
under any host's name for it (anthropic/claude-sonnet-5:batch on OpenRouter,
glm-5.3-flash:cloud on Ollama, llama-3.3-70b-versatile on Groq), including
a model you add by hand. A name that means different models on different hosts,
like deepseek-r1, is left unknown rather than guessed.

• Windows. Compaction and the context meter use the smallest window
  anything knows: the catalogue's, the host's own listing, or one learned from
  an overflow. A host that serves less than the vendor (MiniMax M3 at 512K)
  wins; one that claims more (OpenRouter's 1M for Claude Sonnet 4.5) does not.
  A model running on your own Ollama is left to what the instance reports, since
  it runs at num_ctx, not the model's maximum. Settings → Models shows each
  model's window next to its price, and a context meter beside the token pill
  shows how full the last request was and where compaction happens.
• Prices. OpenRouter's own per-model prices are now read, so a :free
  variant shows as free and a hosted model at OpenRouter's rate. Elsewhere the
  vendor's price applies, and local Ollama, Ollama Cloud and OGX show none, since
  none of them bills per token. ogcode run prices cache traffic at the
  published cache rates.
• OpenAI requests. GPT-5.4 and later take tools on Chat Completions only
  with reasoning off, so agent steps to them send reasoning_effort: "none";
  utility calls send each model's lowest effort, reasoning models get no
  temperature, and max_completion_tokens replaces max_tokens. Models whose
  tools need the Responses API (GPT-6 Astra, the pro and Codex models) are no
  longer offered. Kimi's current models and Claude Opus 4.7+ get no temperature
  on any host.
• Defaults. New installs start on Claude Sonnet 5, GPT-6 Sol and, on
  OpenRouter, anthropic/claude-sonnet-5, with the current generation enabled
  (Claude Fable 5.1, Opus 5.5, Sonnet 5, Haiku 4.5; GPT-6 Sol and Luna, GPT-5.6
  Sol and Terra). Retired models are gone from the lists; the ones other hosts
  still serve stay catalogued.

Minor: Token totals count every token, and every call

Three accounting fixes, found by checking live usage from Groq, Gemini,
OpenRouter, Ollama Cloud and a local Ollama against what ogcode recorded.

• Gemini's thinking is counted. Its OpenAI-compatible endpoint bills thinking
  as output but reports it only in total_tokens. The gap between that and
  prompt + completion now goes to output and reasoning; a live gemini-2.5-flash
  step had recorded 24 output tokens where 263 were billed. When the gap instead
  equals the cached count, the server reported the cache outside
  prompt_tokens, so input is kept rather than clamped to zero.
• One definition of a total. Every total — per step, per session, utility,
  ogcode run's total, the token pill — is now input + cache read + cache
  write + output (TokenCounts.Consumed), the figure providers report as
  total_tokens. v0.35.0 had left cache reads out of session totals, which
  showed a well-cached step that processed 3,267 tokens as 25 and left the pill's
  rows not adding up to its total. Cache reads are processed and billed on every
  step that sends them, so they count. The pill now marks reasoning and utility
  as parts of the rows above them.
• Work outside the step loop is charged to its session. Task and
  memory-recall sub-agents ran in ephemeral sessions that were deleted with
  their tokens, and the deep-search pipeline, the per-turn memory summary and
  plan auto-naming discarded their usage. The loop now stamps its session on the
  context: a sub-agent's run is folded into the session that spawned it before
  deletion, and the one-shot calls charge it directly, all in the utility
  subtotal. The Notes rewrite and transform calls have no session and are still
  not counted.

Minor: Live previews at their own hostname, served only when published

A service the agent starts is now previewed at its own origin —
http://3000.preview.localhost:<server-port>/ — instead of under
/preview/3000/, so an app that boots off its own location (Next.js and
friends) starts normally. Old /preview/<port>/ links redirect there.

A preview hostname now serves only ports that were published: the agent
publishes a port by writing its preview URL in its reply (it takes effect as
soon as that step ends, not when the turn does), and you can add or remove one
on the Preview page. Everything else answers 403, and ogcode's own port is never
served. A loopback port is often private precisely because it is loopback-only,
and a preview link is something you share, so the proxy no longer opens every
port on the machine. Any other name under the preview domain answers 404 rather
than reaching ogcode, so a preview wildcard exposed more widely than the main
address never exposes ogcode itself.

Previewed apps behave as they do on their own: ogcode's CORS headers no longer
leak onto their responses (or swallow their preflights), they get their own
storage and cookies inside the Preview page, redirects to their loopback address
land back on the preview, and the scheme behind a TLS-terminating proxy is passed
on. An open preview no longer reloads every ten seconds.

*.localhost names resolve to the browser's own machine, so the default works
from a browser on the server (or through ssh -L). To open previews from
another machine, set OGCODE_PREVIEW_DOMAIN to a dedicated wildcard DNS name that
points at the server; the Preview page and the agent say so when they detect the
mismatch.

Minor: A shared in-flight budget for provider requests

Provider requests are now capped process-wide, so a project index that fans out
across many documents can no longer put its whole wave on the endpoint at once.
The failure this prevents is not slowness but refusal: an index starts a session
per batch, each a full agent turn of two or more requests, and the burst can trip
the endpoint's rate limiter — which the retry path then answers with further
retries, so the burst amplifies the very throttling it caused. One shared ceiling
turns the wave into a queue that drains at the rate the endpoint accepts.

The ceiling is OGCODE_PROVIDER_MAX_CONCURRENT (default 8, minimum 2). Two of
those slots are reserved for interactive turns: the index and your own turns
share one endpoint and one rate limit, so without a reserve a background refresh
could hold every slot and leave your next message waiting behind it. The cap sits
on the provider send path for both OpenAI and Anthropic, so it is independent of
the indexer's own concurrency setting.

Minor: A ledger of what you have spent

Every agent step and every utility call now writes one row to a ledger, so the
spend of every project, session and model can be totalled in one view. Settings
→ Usage shows it over Today, 7 days, 30 days or all time, for this workspace
(its task worktrees included), for all projects, or for one project opened from
the list.

The ledger keeps two very different numbers apart. Billed is what per-token
providers charged, at their own price. Plan value is what a flat plan or a
local model did, priced at the vendor's list price — the bill the plan spared
you, not a charge. A daily chart shows the shape of the spend, folding to weeks
on the longer ranges.

Prices are not stored with a row; they are applied when the ledger is read, from
the provider's listing or the built-in catalogue, so a corrected price reaches
history instead of being frozen into it. Each workspace's usage from before the
ledger existed is folded in once, and a row outlives the session it came from.

• The endpoint is recorded, not just the provider. A provider slot pointed
  somewhere else — the OpenAI slot at Z.ai, the Anthropic slot at DeepSeek —
  bills at neither the slot's name nor the model's, so the row keeps the host it
  actually called and is priced and named by it. Work whose provider was never
  recorded is inferred from what serves the model today, and says so; work that
  cannot be priced is never counted as a charge.
• Rows are keyed, not appended blindly. An agent step is filed under its
  assistant message id, so the same step recorded twice overwrites rather than
  doubles, and a utility call is one row of its own. Sessions deleted with their
  tokens leave their spend behind.
• One source for session cost too. The token pill's cost and ogcode run's
  figures read from the same pricing, so a session's total and the ledger agree.