Release 0.7.0

[deepfates]

The 0.7.0 release change. Nothing is tagged, published or merged; the owner tags v0.7.0 and runs mix hex.publish.

It is 0.7.0, not 0.6.1, under the semver ruling in decisions.md (a release that changes what a caller receives bumps the minor): cost is nil where it was ReqLLM's estimate (#253), an incomplete stream returns an error where it returned the text that had arrived (#255), the model recorded for a streamed call, and the Imp.Usage key for some clients, change (#255), and ReActV2's :last_request_note history entry carries empty tool_calls and tool_call_results (#260).

What changed

• mix.exs @version "0.7.0"; priv/public_api.json package_version (regenerated with mix imp.public_api; only that line changed); test/package_contract_test.exs asserts "0.7.0".
• {:imp, "~> 0.7"} in README, docs/getting-started/setting-up.md, docs/production.md, docs/coming-from-dspy.md, the five livebooks, and examples/deployment (mix.exs, README.md). The tree/v0.6.0 and blob/v0.6.0 links now point to v0.7.0; they 404 until the tag exists. README: "Imp 0.7 is experimental". The {:req, "~> 0.6"} lines are Req's version and are unchanged.
• CHANGELOG.md: ## Unreleased is ## 0.7.0 — 2026-09-28.
• RELEASE_NOTES.md: rewritten for 0.7.0: why minor, install, headline changes, Upgrading from 0.6 (five steps), Known limits.

Update: fact-check corrections

The fact-check ran the uncertain claims on v0.6.0 and 0.7.0, and commit "Release notes: apply the fact-check corrections" applies its text changes to RELEASE_NOTES.md, the 0.7.0 CHANGELOG section and one sentence of the Imp.Clients.ReqLLM moduledoc. No code changed.

• A streamed call records the configured model id without its provider prefix. ReqLLM 1.24 never passes a streamed response's reported model through, so the claim that it records the model the provider reported is gone. In 0.6.0 a real streamed call had no Imp.Usage entry at all.
• A failed stream's usage is recorded on the failed :model_response event only. Imp.Usage does not see streamed calls, which run in a separate process; this is a new Known limit.
• OpenRouter's charge: 0.6.0 read it only for models ReqLLM's catalog does not price.
• Effort translation: Anthropic and Google get a thinking budget; OpenAI, OpenRouter, Groq and xAI receive the effort as written. This replaces "maps or clamps", in the CHANGELOG and the moduledoc.
• Both tuple shapes are named, "Using that filler/text for history is Imp's" replaces the earlier wording, Upgrading 6 notes that a reloaded history holds %{tool_calls: []}, and the CHANGELOG note entry gives the conversion.

Checks after this commit: format exit 0, compile --warnings-as-errors exit 0, docs --warnings-as-errors exit 0, imp.public_api --check current, the contract tests plus test/req_llm_client_test.exs 207 tests, 0 failures (1 excluded). mix check was not re-run, since only docs text changed.

Update: #260 merged in

origin/main (99ef8cd, #260) is merged into this branch as a merge commit, with no rebase or force-push. Its two Unreleased > Changed entries sit at the end of 0.7.0's Changed section, word for word; there is no Unreleased heading left. The RELEASE_NOTES edits for #260 are in the same merge commit (c95a832).

• The why-minor paragraph now names four caller-visible changes, the note's history shape among them.
• Headline changes gain two bullets: the note renders as a user message with no invented reply (in the last request and on replay), and the written-mode replay of a step that recorded no call and no other output.
• Upgrading step 6: a host that stores returned ReActV2 histories and recognises the note by its exact shape must accept tool_calls and tool_call_results, or match the note by its key or text. A note stored by an earlier version keeps the input-only shape and, passed back, still renders with the "Not supplied" reply. It can be converted by adding tool_calls: %Imp.Adapter.Types.ToolCalls{tool_calls: []} and tool_call_results: [] to it; for a dumped history, Imp.History.load!/1, add the fields, and dump again. I checked the conversion with a throwaway test (since deleted), in native and written tool modes: an old-shape entry passed back as history: produces the filler in the request, and the converted entry does not. The dumped form carries __imp_type__ tags, which is why the step says to load and re-dump rather than edit the JSON.
• The "Not supplied for this conversation history message." filler: the release notes say it is Imp's own and that DSPy 3.2.1 renders a missing history output as None. I found no text in the notes, CHANGELOG, docs or lib that calls it DSPy's. Two nearby texts are not changed here: the comment above append_user_turn/2 in lib/imp/predict/react_v2.ex calls an input-only entry "a DSPy History entry", which names the structure, not the filler; and the message of commit 67297cd says "keep DSPy's rendering", which history can't change. The divergence itself is History entries with missing outputs: Imp invents a filler DSPy doesn't; native replay drops answers #261 and is not in this release.

Judgement calls in the CHANGELOG

1. Two Streamed ReqLLM calls record the usage and cost the provider reported #255 entries moved from Fixed to Changed and are now marked "Breaking:" with a migration, as Release 0.6.0 (draft: version pending owner) #246 did for unmarked breaking changes:
   • a stream that did not complete ends in {:error, %Imp.LMError{}} instead of a completion of the text that arrived;
   • the model a streamed call records. The Streamed ReqLLM calls record the usage and cost the provider reported #255 entry said only that the Imp.Usage key changes. In 0.6.0 a ReqLLM stream carried no usage, so it had no Imp.Usage entry at all; the change a host can see is the req_llm metadata model on the :model_response event (openai:gpt-test becomes gpt-test, or the model the provider reported). The entry now says that.
2. An addition to that entry, which no merged entry mentioned: provider_name/1 now goes through model_identity/1, so a non-streamed call to a client built from a string-keyed spec map or a {provider, opts} / {provider, model, opts} tuple records its provider, where 0.6.0's provider_name/1 matched only an atom-keyed :provider map or a "provider:model" string and gave nil. Its Imp.Usage key goes from "m" to "openai/m". At main I checked the keys with a throwaway test ({:openai, id: "m"}, %{"provider" => "openai", "id" => "m"} and "openai:m" each give ["openai/m"]) and that Imp.req_llm/2 keeps the spec as given. The 0.6.0 side comes from reading git show v0.6.0:lib/imp/clients/req_llm.ex; I did not run it.
3. The cost entry adds that estimated_cost is nil for a streamed call, as the Imp.Core.LMResponse and Imp.LM moduledocs say.

Judgement calls in RELEASE_NOTES

4. Carried-forward Known limits, each checked against main. None of the modules they describe changed between v0.6.0 and main (git diff --stat v0.6.0..origin/main touches only req_llm.ex, core.ex, lm.ex, saving.ex, streaming/execution.ex, exceptions.ex and a benchmark task), and mix.lock is unchanged:
   • mint/Finch: mix.lock holds mint 1.10.1 and Finch 0.23.0. Hex's newest Finch is still 0.23.0, and Close abandoned HTTP/1 connections after request errors sneako/finch#397 is still open and unmerged. The limit stays word for word.
   • The flat pair divergence, checkpoints as sensitive state, ReActV2 prose that quotes field names, ReqLLM's warning log with headers, the Bearer length in command output, and the GEPA full-validation checkpoint all stay.
   • The Install section's cowlib/Cowboy and mint paragraphs are still true: cowboy 2.19.0 and cowlib 2.20.0 are the newest on Hex, mint 1.11.0 is the newest mint, and mix.exs still declares {:mint, "~> 1.8"}.
5. Two new Known limits:
   • The usage map (:usage on the event, Imp.Usage, get_lm_usage/1) is ReqLLM's, so for a non-streamed call its :cost/:total_cost are the catalog estimate (ReqLLM.Usage.Cost.merge/3 puts them there, and sanitize_usage/1 keeps them). A model call's cost is the provider's reported charge; the catalog price is estimated_cost #253 changed the cost field but not this map, so a host reading usage[:cost] still gets the estimate.
   • A provider error sent inside a stream has status nil and retryable: true (stream_failure/2 passes true; Imp.LMError's moduledoc gives the reason).
6. Upgrading step 5 (spend totals grow) is a behaviour change, not a breaking one: 0.6.0 recorded no usage for a ReqLLM stream, and now records usage for streams, including failed ones.

Dependabot

GitHub reports open alerts for accelerate and nltk. They come from benchmarks/requirements-ifbench-parity.txt, the Python environment for a benchmark, not the Hex package. mix hex.build --unpack of this branch contains no benchmarks/ directory and no .txt file (top level: CHANGELOG.md, LICENSE, NOTICE, README.md, RELEASE_NOTES.md, assets, docs, examples, hex_metadata.config, lib, ...). This PR does not change that file.

Checks on this branch (merge commit c95a832)

• mix format --check-formatted: exit 0
• mix compile --warnings-as-errors --force: exit 0 (427 files)
• mix docs --warnings-as-errors: exit 0
• mix dialyzer.check: "Total errors: 142, Skipped: 142, Unnecessary Skips: 0", "done (passed successfully)"
• mix imp.public_api --check: "Public API manifest is current: priv/public_api.json"
• documentation_contract, documented_paths, public_api_manifest, public_surface, package_contract, deployment_reference, livebook_contract: 106 tests, 0 failures (1 excluded)
• MIX_ENV=test mix package.check: "clean-room package proof passed: tmp/package-clean-room"
• mix check (1-minute load 1.7, no other test VM): exit 0, "59 doctests, 9 properties, 3562 tests, 0 failures, 13 skipped (221 excluded)"

No test was added, so there was nothing to falsify by reverting. The release only moves version strings and documents changes that are already merged.