v0.11.0

New

• Add --web-search and --web-fetch to chat for Straitly's hosted tools. The
  provider runs them and returns a finished answer, so no tool loop is needed.
• Add a per-provider web tool map. Straitly is enabled; every other preset and custom
  endpoint fails with a clear error before any request. callm raw can still send a
  provider's own tool format.
• Report sources separately from the answer: text and streaming output list them on
  stderr, leaving stdout answer-only; --result-json gains an additive sources
  array; --json keeps the original annotations.

Changed

• Completion validation accepts web tool records only when the matching request
  enabled web tools and the answer finished. Other tool calls and a tool_calls
  finish reason still fail.
• The agent skill marks web tools as paid: agents must get explicit user confirmation
  before each use and must not enable them by default in scripts or pipelines.
• Release notes are now published from the tagged version's CHANGELOG.md section,
  and a release fails if that section is missing.

Cost and known limits

• Each search or fetch is billed on top of tokens, and the model chooses how many to
  run. --stats, --json and --result-json report the total the provider charged.
• Straitly did not apply --schema when web tools were on (three models tested). Ask
  for JSON in the prompt and use --validate-schema FILE; failed validations are billed.
• Sources are provider-reported and can include pages the answer did not use.

Maintenance

• Keep local development audit records out of version control and remove public
  documentation links to them. Preserve local copies for development use.

What's Changed

• Add Straitly hosted web search and fetch (v0.11.0) by @steamvogue in #1

New Contributors

• @steamvogue made their first contribution in #1

Full Changelog: v0.10.0...v0.11.0

v0.10.0

• Add regular-file prompt/system inputs, raw --body-file FILE|-, and a combined
  64 MiB request limit with a lower --max-input-bytes override. Raw bodies now
  require one valid JSON value and reject extra/conflicting body arguments.

• Add --schema provider constraints plus offline JSON Schema validation, and
  --validate-schema for local-only checks. Buffer and validate before publishing;
  block external schema references. Builds now use pinned Go modules for validation.
  OpenRouter schema requests require provider parameter support; local validation
  remains mandatory even after a successful API response.

• Add strict, versioned --result-json status/answer/model/finish/refusal/usage/timing
  envelopes, including request/completion/schema errors. Keep original --json.

• Add an optional Python mini harness with sequential manifests, atomic artifacts,
  hashes/journal/resume and one call per step. Test the harness against localhost
  mocks in CI; no automatic retries, repairs or model-controlled tools.

• Make live tests free-only: verify all catalog prices are zero before at most two
  explicit OpenRouter :free calls, with no paid fallback. Record a separate
  24-call free-model evaluation; leave production provider/model defaults unchanged.

• Remove the Go race runtime's repeated one-second exit delay from CLI fixture
  subprocesses while retaining detector settings, keeping the expanded CI suite
  within its existing 60-second timeout.

• Reject truncated, refused, filtered, tool-dependent and empty text completions.
  Add --strict for terminal-reason enforcement (including JSON), and
  --allow-empty for intentionally empty answers. Plain --json retains diagnostic
  envelope semantics. Partial streamed output can precede a failure.

• Preserve literal <think> content by default; add explicit --parse-think.
  Propagate answer/reasoning/statistics/raw/info/table/help/version write errors,
  including renderer short writes and flush failures.

• Include Anthropic cache reads and cache creation in processed-input statistics
  without adding cache categories again to compatible-provider totals.

• Add bounded model profiles for native Claude adaptive thinking/effort and known
  GPT-6 reasoning, sampling and token-limit parameters. Keep provider/model defaults;
  explicit caps remain unchanged. Legacy/manual and adaptive effort can consume
  different amounts of reasoning; evaluate the new behavior on your workload.

• Reject native Anthropic editor exports with native-adapter guidance; retain
  table and lossless JSON catalogs. Improve redacted missing-key diagnostics.

• Test supported Go 1.26/1.27 patches in CI; build releases with Go 1.27.2.
  Keep the Go 1.22 module language minimum.

• Correct provider-specific examples and add guarded JSON publication. Free-only
  live tests require valid answer/usage/finish metadata and zero reported cost.

• Validate repairs with unit/local mock tests and synthetic fixtures. Paid provider
  generation and editor imports were not executed; free-model evaluations do not
  establish compatibility with direct paid providers.

Download the archive for your platform and verify it with checksums.txt. The CLI is a standalone binary; the optional mini harness requires Python 3 on Linux/macOS.

Full Changelog: v0.9.0...v0.10.0

v0.9.0

Full Changelog: v0.8.0...v0.9.0

v0.8.0

Full Changelog: v0.7.1...v0.8.0

v0.7.1

Full Changelog: v0.7.0...v0.7.1

v0.7.0

Full Changelog: v0.6.0...v0.7.0

callm v0.6.0

Kimi Code subscribers can now select the coding API with --kimi and KIMI_API_KEY. The preset uses https://api.kimi.com/coding/v1 and defaults to kimi-for-coding.

export KIMI_API_KEY="your-kimi-code-key"
callm --kimi "Review this Go function"

• Supports chat, SSE streaming, models, info, and raw requests through the OpenAI-compatible client, with existing key, URL, model, and timeout overrides.
• Alias change: --kimi previously selected Moonshot. Use --ms or --moonshot with MOONSHOT_API_KEY for the separate pay-as-you-go Moonshot API.
• Subscription statistics show server-returned usage and cost only; no per-call subscription price is inferred.
• Updated CLI help, README, examples, and bundled agent skill.

Validation: local mock tests, go vet, ShellCheck, and actionlint passed. GitHub CI passed, including the Linux amd64 race detector and builds for Linux/macOS amd64 and arm64, and Windows amd64. Live subscription access was not tested.

callm v0.5.0

• All API requests now identify the project with the default User-Agent
  CallM (Call-LLM; +https://github.com/steamvogue/callm).
• Added --user-agent to chat, models, info, and raw commands, with
  CALLM_USER_AGENT as the environment/configuration-file default override.
  Explicit flags win; an empty environment value keeps the project default.
  --user-agent "" omits the header; control characters are rejected.
• Updated CLI help, README examples, and agent skills.

Validation: local mock HTTP tests cover defaults, overrides, omission, and invalid values across all API commands and both streaming protocols; catalog pagination is also checked. Existing regression probes and go vet passed.

callm v0.4.0

• Added OrcaRouter with --orca, ORCA_API_KEY, base URL
  https://api.orcarouter.ai/v1, and default model orcarouter/auto.
  --or remains OpenRouter. All existing timeout defaults and overrides apply.
• Supports chat, SSE streaming, models, info, raw requests, and --orca --claude.
  --effort passes OrcaRouter’s reasoning_effort; --thinking-budget rejects
  with guidance to use effort instead. Model capabilities remain provider-dependent.
• OrcaRouter --stats requests billed cost through X-OrcaRouter-Include-Cost
  and reads usage.cost_usd, including streaming usage.
• Updated CLI help, README, agent skill, and optional live-test provider list.

Validation: local mock API tests, 91 regression probe scenarios, go vet, actionlint, and ShellCheck passed. No billed OrcaRouter inference was performed.

callm v0.3.0

This release fixes stalled calls, dropped input, silently ignored options, and
misleading success results across chat, streaming, model discovery, and raw requests.

Timeouts and input

• All API calls default to a 300-second total timeout. --timeout accepts
  seconds or durations (600, 10m); 0 disables it.
• Added --header-timeout and streaming chat --idle-timeout, both inheriting
  --timeout unless explicitly overridden. Zero disables each selected limit.
• Added independent --stdin-timeout (default 300 seconds) and --no-stdin.
  Pipes are read through EOF on all platforms, including delayed producers;
  cancellation interrupts both input and HTTP waits.
• Bound input files/stdin/buffered responses to 64 MiB and SSE events to 8 MiB.

Correctness and provider support

• Parse full SSE events, reject malformed or truncated streams, and stop Anthropic
  streams immediately at message_stop. HTTP and API errors return failure.
• Honor configured clients and endpoint overrides consistently. Follow Anthropic
  model catalog pagination under one deadline and detect repeated cursors.
• Update Claude defaults to claude-sonnet-4-6 directly and
  anthropic/claude-sonnet-4.6 through Straitly/OpenRouter.
• Convert Anthropic images correctly, preserve top-p, validate thinking budgets,
  and preserve explicit token caps. Unsupported JSON-object requests fail explicitly.
• Apply reasoning display controls to both output modes. Preserve original provider
  JSON and large numbers; --json --stats retains separate stderr statistics.
• Request streaming usage for --stats; handle string costs and unknown usage,
  prices, and context sizes without reporting missing data as zero.
• Accept provider flags before subcommands and honor --stream=false. Reject
  malformed numbers, conflicting flags, and unsupported provider combinations.
  Normalize token limits for OpenAI o-series models.

Compatibility changes

• Missing provider keys no longer fall back to another provider's credentials.
  Use the selected provider's key, CALLM_API_KEY, or an explicit key override.
  Cross-origin redirects are rejected before credentials reach the destination.
• Streaming and reasoning display default on only when stdout is a terminal.
  Reasoning goes to stderr; final answers go to stdout.
• Automation relying on incomplete streams, silently dropped settings, or API
  errors returning success must now handle nonzero exit codes.

Documentation and validation

• Synchronize CLI/subcommand help, README, examples, and the callm agent skill with
  current options, aliases, defaults, and configuration precedence. Add a README/help
  consistency test and documented maintenance rules.
• Add mock regression coverage and 91 local CLI scenarios; improve live-test
  assertions and expose make test-race, enabled in Linux amd64 CI.
• Release archives cover Linux amd64/arm64, macOS amd64/arm64, and Windows amd64,
  with SHA256 checksums. Mock validation does not establish live provider generation
  quality, billing, or account-specific model availability.