Skip to content

CHANGELOG.md

Latest commit

 

History

History
183 lines (140 loc) · 8.54 KB

CHANGELOG.md

File metadata and controls

183 lines (140 loc) · 8.54 KB
Raw

Changelog

All notable changes to this project are documented here. Format: Keep a Changelog - SemVer.

[Unreleased]

[1.2.0] - 2026-10-31

Added

  • Actionable Command Code errors. JSON failure envelopes keep their HTTP status and shape, but the message gains a resolution hint: add credits or another key for 402, pick another model or upgrade for MODEL_NOT_IN_PLAN, replace or rotate for 401/403, wait or rotate for 429, and the Anthropic-surface note for unsupported_model. The status is never rewritten (OpenCode's retry classification depends on it), only JSON bodies are touched, and every decoration is logged to commandcode-errors.log.

[1.1.0] - 2026-10-31

Added

  • Multi-key rotation (opt-in). With two or more keys from COMMANDCODE_API_KEYS, COMMANDCODE_KEYS_FILE or ~/.config/commandcode/keys.txt, the plugin picks the first key that is not cooling down, so a dead account does not end the turn. 402 and 401/403 block a key for 24h, 429 honours Retry-After up to 6h (5m by default), and 5xx never blames the key. The soonest recovery wins when every key is cooling down. State is labels-only (commandcode-keys-state.json, 0600) and every cooldown is logged to commandcode-rotation.log without secrets. Single-key mode keeps the previous behaviour: the plugin never overrides an Authorization header the host already set.

[1.0.7] - 2026-10-31

Added

  • Zero data retention passthrough. With CMD_ZDR=1 every request to api.commandcode.ai carries x-cmd-zdr: 1, mirroring the Command Code CLI. When the opt-in is absent, any caller-supplied x-cmd-zdr header is stripped so a session that never asked for ZDR cannot emit the flag by accident. The header is applied on the strict Command Code host only, never on a lookalike that merely shares the request path.

[1.0.6] - 2026-10-25

Added

  • Auto-repair for models requiring historical reasoning_content (such as reflection/beam:free). Prevents HTTP 400 parameter errors in multi-turn assistant and agent workflows by synthesizing compliant thoughts.

Changed

  • Updated tracked Command Code CLI version signature (CLI_VERSION) from 1.73.4 to 1.88.0.

[1.0.5] - 2026-10-24

Changed

  • Lowered catalog cache TTL (CATALOG_TTL_MS) from 12h to 6h so upstream provider rotations, deprecations, and newly added free promo models are discovered twice as fast.

Fixed

  • The declared output limit was hardcoded to 16,384 tokens for every model. OpenCode forwards limit.output as max_completion_tokens, so every response was capped at 16k even though the catalog supports far more (the DeepSeek V4.1 Flash 1017 build declares 393,216). The default is now 32,768, the highest value that is safe across the whole catalog (the minimum declared output in models.dev's commandcode provider is exactly 32,768), and COMMANDCODE_OUTPUT_TOKENS overrides it per process.
  • The static model entries in opencode.json carried output: 16384, which took precedence over the plugin. Those limits are now owned by the plugin alone.

[1.0.3] - 2026-10-19

Added

  • Model ordering in the picker: promo/free models come first (with the runtime's "Free" badge), then models are grouped by family and, inside each family, the newest version comes first. Implemented by writing cost (zero input) and time.released from the model editor; the release timestamp encodes the family rank and the version recency.
  • src/catalog.ts with pure helpers (detectFamily, parseVersion, compareVersions, versionScore, buildReleaseDate, isFreeModel, freeCostEntry) and 20 unit tests.

Fixed

  • models.dev publishes its own commandcode provider, which OpenCode merges with the configured one. That merge reintroduced the Anthropic-shaped claude-* models that answer 400 unsupported_model on the OpenAI-compatible transport. The catalog transform now disables every provider model that is not in the curated set.

[1.0.2] - 2026-10-19

Fixed

  • Retry telemetry recorded the model as undefined in every single entry. The OpenCode v2 retry hook exposes the model as model.id / model.providerID, and there is no modelID field, so the previous mapping always produced undefined. Verified against the real event shape captured from a probe server that cuts the stream mid-flight.
  • The retry log now also records the agent, the attempt number and the backoff delay the runtime chose, which makes it possible to tell an upstream timeout (large gap between entries) from a runtime-side retry loop.

Added

  • src/telemetry.ts with pure helpers (describeRetryEvent, extractModelId, extractProviderId) and 6 unit tests covering the real event shape.
  • Troubleshooting section (English and Spanish) on attributing resets to a specific model.

[1.0.1] - 2026-10-19

Added

  • Installation straight from GitHub with the github: plugin specifier, verified end to end (clone, cache, load, inference) with no npm publication and no build step.
  • Troubleshooting section for the plugins-directory auto-discovery asymmetry: a plugin removed from opencode.json, or renamed inside ~/.config/opencode/plugins/, keeps loading until the directory leaves plugins/ (then the watcher drops it live).

Changed

  • README and configuration docs document three installation paths: GitHub specifier (recommended), local clone, and the auto-discovered plugins directory.

[1.0.0] - 2026-10-18

First modular release. Migrates the single-file commandcode-quota plugin into a structured package with dynamic model discovery.

Added

  • Dynamic model discovery from GET /provider/v1/models through the OpenCode v2 model editor, with transport-aware filtering and automatic [FREE PROMO] badges for free/promo tiers.
  • On-disk model catalog cache (12h TTL) so an upstream outage cannot leave the provider with an empty model list at startup.
  • Multi-source credential resolution: COMMAND_CODE_API_KEY, COMMANDCODE_KEY_FILE, $XDG_CONFIG_HOME/commandcode/key.txt, /etc/environment, then the official CLI stores under ~/.commandcode/.
  • Strict origin guard (isCommandCodeHost) for credential injection, plus the COMMANDCODE_ALLOWED_HOSTS allow-list opt-in.
  • Bounded API calls via AbortSignal.timeout (10s default).
  • Vision capability metadata (image input) for multimodal models.
  • applyModelCatalog exported for testability.
  • Offline test suite (58 tests) and an opt-in live suite gated by COMMANDCODE_LIVE_TESTS.
  • Strict typecheck configuration (allowImportingTsExtensions, strict, noUncheckedIndexedAccess) with bun-types.
  • Documentation in English and Spanish: README.md / README.es.md, docs/{ARCHITECTURE,CONFIGURATION,TROUBLESHOOTING}.md / docs/es/*, SECURITY, CONTRIBUTING and CODE_OF_CONDUCT pairs, plus LICENSE.

Changed

  • State directory, key paths and the connection-close sentinel are resolved from XDG/home instead of hardcoded /root paths.
  • Retry telemetry now ignores non-Command-Code providers, keeping commandcode-retries.log meaningful.
  • The retry log records the provider id and degrades gracefully when the hook is unsupported by the running build.

Fixed

  • Security: credentials were attached based on substring URL matching (api.commandcode.ai anywhere in the URL, or any URL containing /provider/v1/chat/completions). That leaked the API key to lookalike hosts such as api.commandcode.ai.evil.com, to any host carrying the provider path, and to values injected in query strings or fragments.
  • Anthropic-shaped models (claude-*, /messages only) were registered under the OpenAI-compatible transport and failed with 400 unsupported_model. They are now filtered out.
  • Free-model detection missed -free / _free / .free id shapes, so e.g. poolside/laguna-s-2.1-free was not tagged.
  • The sanitizer had regressed relative to the proven v1 implementation: null and empty tool response content was no longer normalized, orphan detection was no longer order-sensitive, and malformed tool_call entries were not pruned.
  • The sanitizer mutated the caller's message array in place instead of working on a shallow clone.
  • The watchdog could fire an abort after a consumer cancellation, and a throwing logging callback could leave the stream open.
  • Quota report rendering crashed on a missing window; it now degrades per field.

Removed

  • Hardcoded user-specific absolute paths removed from the runtime code in favour of XDG/home resolution.
  • Network calls from the default test suite.