Repository navigation
Releases
v1.45.0
Highlights
Dive v1.45.0 adds OpenAI's GPT-6.2 Astra and Sol (openai.ModelGPT62Astra,
openai.ModelGPT62Sol) and Google's gemini-nano-banana-2.1 image model.
Reasoning settings for GPT-6.1 Terra and Gemini 4 Carbon and Hydrogen are now
adjusted to what each model accepts, rather than forwarded unchecked.
Media generation catches up with provider shutdowns. Grok video works again on
xAI's native video API and reports its cost in Metadata["cost_usd"]. OpenAI
shut down Sora, so openai.MediaProvider.GenerateVideo is removed. Google shut
down Imagen, so an imagen- model now returns an error. Veo runs only on Vertex
AI: use google.NewMediaProvider(google.WithMediaVertexAI(location)). These
calls already failed at runtime, but code that calls GenerateVideo on the
OpenAI provider no longer compiles. Switch to grok-imagine-video for video and
a Gemini image model for images.
Every module moves to current dependencies: Wonton v0.14.2, openai-go v3.101.0,
genai v1.77.0, a2a-go v2.9.0, mcp-go v1.5.0, and OpenTelemetry 1.47. The dive
CLI now exits 2 on a usage error, and defaults to grok-imagine-video for
dive video.
Pull requests
- fix(media): repair Grok video and drop shut-down Sora and Imagen (#362)
- chore(deps): upgrade Wonton, openai-go, genai, and other dependencies (#361)
- Bump go.opentelemetry.io/otel/trace in /experimental/cmd/dive (#360)
- Bump go.opentelemetry.io/otel/trace in /demos/colosseum (#357)
- Bump go.opentelemetry.io/otel/sdk from 1.46.0 to 1.47.0 in /examples (#356)
- Bump go.opentelemetry.io/otel/sdk from 1.45.0 to 1.47.0 in /otel (#355)
- feat: add GPT-6.2 Astra and Sol and other new provider models (#359)
Changelog
Added
- GPT-6.2 Astra and Sol.
openai.ModelGPT62AstraandModelGPT62Solare
the recommended OpenAI models; new Google, Mistral, and OpenRouter models too.
Changed
- Verified reasoning controls. GPT-6.1 Terra and Gemini 4 Carbon and
Hydrogen now get unsupported effort levels and thinking budgets adjusted. - Dependency upgrades. Wonton v0.14.2, openai-go v3.101.0, genai v1.77.0,
a2a-go v2.9.0, mcp-go v1.5.0, and OpenTelemetry 1.47. ThediveCLI exits 2
on a usage error. dive videodefaults togrok-imagine-videoanddive imagewith only a
Google key defaults togemini-3.1-flash-image.google.WithMediaImagenLocationis deprecated and does nothing.
Fixed
- Grok video generation works again. It uses xAI's native
/v1/videos/generationsAPI and reports the cost inMetadata["cost_usd"]. - A Gemini image response with no image names its finish reason, such as
IMAGE_RECITATION.
Removed
- Breaking:
openai.MediaProvider.GenerateVideoand thesora-models.
OpenAI shut down Sora and the Videos API on 2026-09-24. Use
grok-imagine-video. - Breaking: Imagen models. Google shut down Imagen. An
imagen-model
returns an error; use a Gemini image model. Veo runs only on Vertex AI now.
v1.44.0
Highlights
Dive v1.44.0 refreshes model catalogs across Anthropic, Google, OpenAI, Meta,
xAI, and Mistral with the latest GA/preview chat, audio, and image models, and
updates deprecated and retired model statuses. New recommended defaults in the
dive CLI include claude-fable-5-2, gemini-4-carbon, gemini-4-hydrogen,
gpt-6.1-terra, and muse-ember-1.0. Note that claude-fable-5-2,
gemini-4-carbon, gemini-4-hydrogen, gpt-6.1-terra, and muse-ember-1.0
were not probed against live provider APIs in this release; gpt-6.1-terra is
registered for the OpenAI Responses API only.
This release also upgrades Wonton from v0.2.1 to v0.10.2 across every Dive
module. When stdout is not a terminal, the dive CLI now prints plain text
without ANSI escape codes, while terminal output remains unchanged.
WebFetch with MaxRetries set no longer retries 404 or 403 responses,
retrying only errors that fetch.IsRetryable accepts and reporting the attempt
count in its error. No code changes are required to upgrade.
Pull requests
- feat: add new chat, audio, and image models across providers (#353)
- chore(deps): upgrade Wonton to v0.10.2 (#351)
Changelog
Added
- New provider models and deprecations. Added
claude-fable-5-2and
claude-mythos-5-2(Anthropic);gemini-4-carbon,gemini-4-hydrogen, and
gemini-omni-1.2-flash(Google);gpt-6.1-terra,gpt-image-2.6-sunburst,
andgpt-image-2.6-flare(OpenAI);muse-ember-1.0(Meta);grok-voice-3.0
andgrok-voice-2.0(xAI); and the Voxtral 3 audio preview models
(voxtral-mini-transcribe-3-0,voxtral-mini-transcribe-realtime-3-0,
voxtral-small-transcribe-3-0,voxtral-mini-tts-3-0) (Mistral). Marked
gemini-2.5-pro,gemini-2.5-flash,gemini-omni-1.1-flash, and
grok-voice-2.0deprecated, andgemini-2.5-flash-imageand
grok-imagine-image-qualityretired.
Changed
- Wonton v0.10.2. The
diveCLI prints plain text, with no escape codes,
when stdout is not a terminal.
Fixed
- WebFetch no longer retries a 404 or 403 when
MaxRetriesis set. It
retries only errors thatfetch.IsRetryableaccepts, and its error states
the number of attempts it made.
v1.43.0
Highlights
Veo video generation now works on Vertex AI. Before this release,
google.MediaProvider.GenerateVideo always failed on Vertex with the genai
error "method Download is only supported in Gemini Developer API mode". Dive
now reads the video bytes from the operation result. It downloads the file only
on the Gemini API backend. On Vertex, a Veo 3 video has audio. Veo 2 does not
support audio, so Dive does not request it for Veo 2.
On both backends, Dive now sends media.Config.Duration, and it sends the
first image in media.Config.ReferenceImages as the start frame. Before, Veo
ignored these values. If Veo does not support the duration or the start image,
the call now returns an error. Before, it returned a video with default values.
Veo accepts whole seconds only: 4, 6, or 8 for Veo 3, and 5 to 8 for Veo 2.
Dive sends whole seconds and drops any fraction, so 4.5s is sent as 4s. Dive
does not send a duration under one second, so Veo uses its default length.
The VideoGeneration toolkit tool offers only the durations that the Veo model
accepts, and it rejects other values before it calls the provider. Other models
keep a free-text duration. Upgrade providers/google to v1.43.0 to get the
fix. No code changes are necessary.
Pull requests
- fix(google): make Veo video work on Vertex AI (#349)
Changelog
Fixed
- Google Veo video works on Vertex AI, with audio except on Veo 2. Gemini
API calls now also sendDurationand the first reference image (start frame);
Veo fails on values it rejects.VideoGenerationoffers only valid durations.
v1.42.0
Highlights
Claude can now see a token budget for a whole turn. max_tokens caps each
response, but the model cannot see it, so a long thinking pass can use the
whole cap and write no answer. Set ModelSettings.TaskBudget and the model
sees the budget count down as it thinks, calls tools, reads tool results, and
writes, and it paces itself to finish within the budget. The budget is
advisory: a turn can run past it, and MaxTokens still caps each response.
Dive sends the Anthropic beta header for you.
ModelSettings: &dive.ModelSettings{
ReasoningEffort: llm.ReasoningEffortMedium,
TaskBudget: &llm.TaskBudget{Total: 40000},
}Every model call of one CreateResponse shares the budget. If a PreIteration
hook compacts the turn's history, Dive sends the remaining budget so the
countdown continues. Tool results count against the budget, so size it from
measured turns. Task budgets work on Opus 4.7 and later, Sonnet 5.5, Haiku
5.5, and Fable 5 and 5.1; Anthropic also lists Mythos 5 and 5.1, which Dive
sends the budget to but has not tested. Claude models known to lack support
drop the setting and log a warning, and Claude model IDs that Dive does not
know get it as asked. On Claude, a total below 20,000 is an error. OpenAI and
Gemini drop the setting and log a warning.
The change is additive: existing code needs no changes.
Pull requests
- feat: Claude task budgets (#345)
Changelog
Added
- Claude task budgets.
ModelSettings.TaskBudgetgives the model a token
budget it can see for the whole turn; unsupported models drop it with a
warning. See the LLM guide.
v1.41.0
Highlights
A Stop hook can now continue a turn with input that did not come from the
user. Before this release, every continuation added a stop-continuation
reminder that tells the model "the following input arrived from the user".
That is wrong when the hook delivers something else, such as the result of a
background subagent that finished after the model ended its response.
To continue with your own content, append a Recorded reminder with
hctx.AppendReminder and return StopDecision{Continue: true} with an empty
Reason. That reminder is the only new input to the next model call, and the
turn saves it once. A ModelOnly reminder is not enough on its own, because
it is not saved and the saved turn would hold two answers with no input
between them. Continuing with no reason and no Recorded reminder fails the
turn with ErrStopContinueWithoutInput.
Existing Stop hooks always set Reason, so they need no changes.
Pull requests
- feat(hooks): let a Stop hook continue with its own reminder (#343)
Changelog
Added
- Stop hooks can continue without a reason. With an empty
Reason, the
hook'sRecordedreminders are the only new input and nostop-continuation
message is added; with none, the turn fails withErrStopContinueWithoutInput.
v1.40.0
Highlights
Streaming calls now recover from a provider that hangs. Before this release,
a stream gave up on the first timeout even though the caller's context was
still live, so one Anthropic request that hung for five minutes failed the
whole turn. An attempt that times out is now abandoned and retried like any
other transport error.
Streams are also bounded by two new limits in place of the 300-second total
client timeout, which cut off healthy streams that ran past five minutes. A
first-event limit abandons and retries an attempt that has sent nothing yet:
one minute on Anthropic and OpenAI, whose streams open with an event before
the model starts work, and five minutes on Chat Completions endpoints. It is
off for Google, because Gemini can send nothing, not even headers, for several
minutes while it thinks. An idle limit of five minutes ends a stream that goes
quiet partway through. Set either with WithStreamTimeouts on the provider.
http.Client.Timeout still applies to Generate.
A turn that ends because the provider timed out is now reported as
provider_error or stream_interrupted, not deadline. deadline means the
caller's own context or ResponseTimeout expired.
Most code needs no changes. A custom provider built on
providers.NewRetryingStreamIterator must update its StreamFactory to take
the attempt's context.Context and make its request with it.
Pull requests
- fix(providers): retry stream timeouts and bound streams by first-event and idle limits (#341)
Changelog
Added
- Stream timeouts.
WithStreamTimeouts(firstEvent, idle)on the Anthropic,
OpenAI, OpenAI Completions and Google providers. An attempt with no event by
firstEventis retried (Anthropic and OpenAI default 1m, Completions 5m,
Google off); a stream quiet foridle(5m) ends.
Changed
- Streams ignore
http.Client.Timeout. It covered reading the body, so a
healthy stream past 5 minutes was cut off; it now applies toGenerateonly. providers.StreamFactorytakes the attempt'scontext.Context. Custom
providers must make their request with it so a timed-out attempt is cancelled.
Fixed
- Streams retry a timed-out attempt while the caller's context is live,
instead of failing on the first attempt. This covers the first-event limit
and transport timeouts such ashttp.Transport.ResponseHeaderTimeout. - A provider timeout is no longer reported as the turn's deadline.
TurnReasonDeadlinenow requires the invocation's context to have expired;
otherwise it isprovider_errororstream_interrupted.
v1.39.0
Highlights
Anthropic cost estimates now price 1-hour prompt-cache writes correctly.
With the extended-cache feature on, Dive writes stable breakpoints with a
1-hour lifetime, which Anthropic bills at 2x input. Before this release Dive
priced those writes at the 5-minute rate of 1.25x, so Usage.Cost came in
low. llm.Usage now reports the 1-hour subset of cache writes, and CostOf
prices it separately. The long-context tier is included, for example Claude
Haiku 5.5 above 100k tokens.
History cache anchors now stay on the same blocks from turn to turn. They are
placed at fixed positions from the start of the conversation instead of
counting back from the moving tail. A 1-hour entry written on one turn can
then be read after an idle gap instead of being rewritten at a new offset.
This changes where breakpoints land on every Anthropic request, extended
cache or not. Only the positions change, not the number of breakpoints.
No code changes are needed. If you read cache-write counts from Usage,
CacheCreationInputTokens is still the total, and the new 5-minute and
1-hour fields are subsets of it.
Pull requests
- llm, anthropic: price 1h cache writes and keep cache anchors at fixed positions (#336)
Changelog
Added
- Anthropic 1-hour prompt-cache write usage and pricing.
llm.Usageparses
5-minute and 1-hour cache-creation token counts, andllm.PricingInfoprices
1-hour cache writes at 2x input (including long-context tiers).
Changed
- Anthropic prompt-cache history anchors stay at fixed block positions across
turns.placeCacheAnchorsanchors fixed multiples ofcacheAnchorGap
before the tail instead of shifting with the tail on every call.
v1.38.0
Highlights
A PreIteration hook can now end a turn on purpose. Before, a long tool loop
stopped only at the tool iteration limit, and a hook could only abort it with
an error. Now a hook can see that the next call would cross a context or
pricing ceiling, set FinalCall, and let the model write a closing answer
without tools. The application can then compact the session and continue in a
new turn.
Tell the model why the turn is ending with a Recorded reminder appended in the
same hook. The reminder is stored in the turn's output, so a replay sends the
same bytes as the live call and the prompt cache still hits. Nothing changes
for agents whose hooks don't set the flag.
Pull requests
- feat(agent): let a PreIteration hook make the next call final (#338)
Changelog
Added
HookContext.FinalCall. A PreIteration hook can make the next model call
the turn's last, with tools disabled, to wrap up before a context or price
ceiling. Pair it with a Recorded reminder to tell the model why.
v1.37.1
Highlights
This patch fixes prompt-cache misses on the first request of every turn that
follows a tool call. Dive recorded each tool-call input exactly as the model
wrote it, in the model's key order. An application that stores transcripts as
structured JSON, such as Postgres jsonb decoded into Go maps, replays those
inputs with sorted keys. The provider's cache is byte-exact, so the next turn
diverged at its first multi-key tool call and paid to rewrite everything after
it.
Dive now records tool-call inputs with sorted keys and no spacing, so the live
turn and a replayed one send the same bytes. Tools receive the same values as
before. Hooks, event callbacks and OutputMessages see the canonical bytes,
so any code that compares raw input text against the model's original output
will see a difference.
Upgrade to v1.37.1 if you persist and replay agent history. For an exact
match, decode stored inputs with json.Decoder.UseNumber. Decoding to
float64 still rewrites number literals such as 0.10, which OpenAI-style
providers compare byte for byte.
Pull requests
- fix(agent): record tool-call inputs in canonical JSON so replays hit the prompt cache (#334)
Changelog
Changed
- Tool-call inputs are recorded in canonical JSON. The agent sorts object
keys and removes spacing in eachtool_useandmcp_tool_useinput before
it records the assistant message. Tools, hooks, event callbacks and
OutputMessagessee the canonical bytes; the values are unchanged, number
text included. A transcript replayed from a store that sorts keys, such as
Postgresjsonbread into Go maps, now sends the same bytes as the live
turn when the store also keeps number text (decode withUseNumber), so
the provider's prompt cache reads it instead of rewriting it. Decoding to
float64still rewrites literals such as0.10and1.0.
v1.37.0
Highlights
Gemini 4 Argon, Google's new flagship reasoning model, joins the Google and
OpenRouter catalogs and becomes the CLI's Gemini pro model. Its thinking
settings were probed against the live API. Cost estimates use Google's launch
rates, which double on 2026-12-01.
Meta's Muse Wonder 1.0 (preview) takes reasoning effort up to max. It accepts
only an automatic tool choice, so give dive.PromptStopHook and
dive.PromptToolGate a different model.
Gemini 3.5 Flash and 3.7 Flash are deprecated. The CLI now defaults to Gemini
3.8 Flash when only a Gemini key is set.
Upgrade the root module and any Dive provider or integration modules together
to v1.37.0.
Pull requests
- feat: add Gemini 4 Argon and Muse Wonder 1.0 (#332)
Changelog
Added
- Gemini 4 Argon.
google.ModelGemini4Argonand
openrouter.ModelGemini4Argon, at launch pricing. It takes the CLI's Gemini
pro slot from Gemini 3.1 Pro. - Muse Wonder 1.0.
meta.ModelMuseWonder10(preview), with efforts
minimalthroughmax. It accepts only an automatic tool choice, so
dive.PromptStopHookanddive.PromptToolGateneed another model.
Changed
- Gemini defaults move off deprecated models. The CLI defaults to
gemini-3.8-flash, and Google transcription togemini-3.6-flash.
Deprecated
google.ModelGemini35Flashandgoogle.ModelGemini37Flash. Google routes
them toModelGemini36FlashandModelGemini38Flash.