Skip to content

Latest commit

 

History

History
384 lines (222 loc) · 23.5 KB

CHANGELOG.md

File metadata and controls

384 lines (222 loc) · 23.5 KB

[1.6.3] - 2026-02-22

Fixed

  • Gemini 3.1 Pro Multi-turn Tool Call Error - Fixed a bug where Gemini thinking models (gemini-3.1-pro-preview-customtools) would throw a 400 INVALID_ARGUMENT error (Function call is missing a thought_signature) during multi-turn tool execution.
  • Thought Signature Recovery - Added local cache recovery for thoughtSignature to prevent signature loss caused by the Vercel AI SDK stripping metadata when reconstructing conversation history.
  • Wrapped Request Injection - Correctly injects the recovered thoughtSignature into functionCall parts for Vercel AI SDK wrapped requests, ensuring seamless compatibility with OpenCode's execution loop.

Changelog

[1.6.0] - 2026-02-21

Added

  • Proactive Context Overflow Guard - Implemented a pre-flight token estimator for Claude models that prevents sending requests exceeding the ~200k token limit.
  • Auto-Compact Integration - Automatically triggers the OpenCode /compact command and notifies the user via toast when an overflow is detected, returning an actionable synthetic response.
  • Session ID Tracking - Captures and tracks OpenCode session IDs to enable programmatic command execution.

[1.5.11] - 2026-02-20

Fixed

  • Gemini model safety gate - Added hard allowlist for known working Gemini text models to prevent accidental use of stale/unsupported model IDs.
  • Legacy model redirects - Added compatibility redirects for common legacy names (gemini-pro, gemini-flash, preview/customtools variants) to supported working targets.

Changed

  • Validation scope - Gemini allowlist enforcement now applies only to pure Gemini text model paths, so Claude alias compatibility remains intact.
  • Resolver tests - Updated model resolver tests to assert unsupported legacy Gemini IDs are rejected with explicit errors.

Verified

  • Full isolated smoke run - typecheck, vitest (28 files / 899 tests), and full test:e2e:models all passed.
  • Model E2E summary - 17/17 models passed in isolated environment.

[1.5.10] - 2026-02-20

Added

  • Gemini 3.1 Pro endpoint integration - Merged PR #2 (aikazu/feat/update-gemini-3-1-pro) to align backend routing and model handling updates for Gemini 3.1 Pro flows.

Fixed

  • Typecheck stability - Fixed src/plugin/accounts-quota.test.ts imports and storage typing to remove invalid symbol imports and pass strict TypeScript verbatimModuleSyntax checks.

Verified

  • Gemini E2E model suite - Confirmed passing runs for Gemini categories with isolated repo copies:
    • gemini-cli: 4/4 passed (google/gemini-3-flash-preview, google/gemini-3.1-pro-preview, google/gemini-2.5-pro, google/gemini-2.5-flash)
    • antigravity-gemini: 3/3 passed (google/antigravity-gemini-3.1-pro-low, google/antigravity-gemini-3.1-pro-high, google/antigravity-gemini-3-flash)

[1.5.8] - 2026-02-19

Fixed

  • Sonnet fallback: retry claude-sonnet-4-6 automatically when claude-sonnet-4-6-thinking returns 404 from antigravity backend.
  • Windows E2E: run model/regression scripts with opencode.cmd and shell: true on win32 to avoid spawn ENOENT failures.

Changed

  • Model scripts: normalize thinking model runs to use base model IDs plus explicit --variant for Sonnet/Opus checks.

[1.5.7] - 2026-02-18

Changed

  • Models: migrated all Claude Sonnet references from 4.5 to 4.6 across runtime model config, resolver aliases, request handling, tests, and docs.
  • Thinking: kept Sonnet thinking variants aligned with claude-sonnet-4-6-thinking and antigravity model keys used by OpenCode configs.

[1.5.6] - 2026-02-15

Fixed

  • Chore: remove .flowcrate/flowcrate.db from git tracking and add to gitignore to prevent sensitive local data leaks

CI/CD

  • Release: automated npm publishing pipeline via GitHub Actions

[1.5.5] - 2026-02-15

Fixed

  • Quota: wire quota check results to rateLimitResetTimes for accurate auth menu badges (ab276a7)

[1.5.4] - 2026-02-15

Added

  • Ui: show rate-limited family detail with countdown in auth menu badges (d5f935e)

[1.5.3] - 2026-02-15

Changed

  • Chore: remove temporary test output files (62ef7a4)
  • Chore: final cleanup of package references and comments (641f6fe)
  • Docs: update repo references in documentation (4758a6d)
  • Docs: consistency update for branding and strict quota explanation (2dbce88)
  • Docs: update branding, credits, and donation info (8f43b92)
  • Chore: prepare release 1.5.2 (47efecd)

Fixed

  • Fix: auto-clear verification flags after successful requests (80f25f2)

[1.5.2] - 2026-02-14

Added

  • Strict Soft Quota Locking - Accounts exceeding the soft quota threshold (confgurable, default 70%) are now strictly locked until their specific quota reset time, ignoring cache TTL. This prevents "leaky" usage where expiring cache allowed over-quota accounts to be reused before they were actually replenished.
  • Proxy Support - Added support for undici ProxyAgent, allowing the plugin to work behind corporate proxies via standard environment variables.
  • Oh-My-OpenCode Integration - integrated session recovery features from oh-my-opencode:
    • Tool Crash Recovery: Automatically fixes tool_result_missing errors when operations are cancelled.
    • Thinking Block Recovery: Fixes corrupted or out-of-order thinking blocks.
    • Strip Illegal Thinking: Automatically removes thinking blocks when switching to models that don't support them.
  • Interactive Quota Pause - The check command now pauses after displaying the quota table, giving users time to read the stats before the command exits or proceeds.

Changed

  • Quota Threshold Buffer - Default soft_quota_threshold_percent lowered from 90% to 70% to provide a safer 30% buffer against accidental overage and bans.

Fixed

  • Gemini 400 Bad Request - Fixed an issue where the x-goog-user-project header caused 400 errors for users without the Cloud Code API enabled on their project. This header is now stripped from requests.

[1.5.1] - 2026-02-11

Changed

  • Header Identity Alignment - ideType changed from IDE_UNSPECIFIED to ANTIGRAVITY and platform from PLATFORM_UNSPECIFIED to dynamic WINDOWS/MACOS (based on process.platform) across all header sources (getAntigravityHeaders, oauth.ts, project.ts). Now matches Antigravity Manager behavior

  • Gemini CLI Client-Metadata Header - Gemini CLI requests now include Client-Metadata header, aligning with actual gemini-cli behavior. Previously only Antigravity-style requests sent this header

  • Gemini CLI User-Agent Format - Updated from GeminiCLI/{ver}/{model} to GeminiCLI/{ver}/{model} ({platform}; {arch}) to match real gemini-cli UA strings. Version pool updated from 1.2.0/1.1.0/1.0.0 to 0.28.0/0.27.4/0.27.3 to align with actual release numbers

  • Randomized Headers Model-Aware - getRandomizedHeaders() now accepts an optional model parameter, embedding the actual model name in Gemini CLI User-Agent strings instead of a hardcoded default

  • Fingerprint Platform Alignment - Antigravity-style Client-Metadata platform now consistently matches the randomized User-Agent platform, fixing a potential mismatch where headers could disagree on reported platform

Removed

  • Linux Fingerprints - Removed linux/amd64 and linux/arm64 from ANTIGRAVITY_PLATFORMS and fingerprint generation. Linux users now masquerade as macOS (Antigravity does not support Linux as a native platform)

  • getAntigravityUserAgents() Function - Removed unused helper that had no callers in the codebase

  • X-Opencode-Tools-Debug Header - Removed debug telemetry header from outgoing requests

[1.5.0] - 2026-02-11

Added

  • Account Verification Flow - Auth login menu now supports verify and verify-all actions. When Antigravity returns a 403 with validation_required, the account is automatically disabled, marked with a verification URL, and cooled down. Users can verify accounts directly from the menu with a probe request to confirm resolution

  • Dynamic Antigravity Version - Plugin version is now fetched at startup from the Antigravity updater API, with a changelog-scrape fallback and a hardcoded last-resort. Eliminates stale "version no longer supported" errors after Antigravity updates

  • Storage V4 Schema - New storage version adds verificationRequired, verificationRequiredAt, verificationRequiredReason, verificationUrl, and fingerprintHistory fields per account. Full migration chain from v1/v2/v3 to v4

  • saveAccountsReplace - New destructive-write storage function that replaces the entire accounts file without merging, preventing deleted accounts from being resurrected by concurrent reads

  • setAccountEnabled / Account Toggling - New account management methods: setAccountEnabled(), markAccountVerificationRequired(), clearAccountVerificationRequired(), removeAccountByIndex()

  • Secure File Permissions - Credential storage files are now created with mode 0600 (owner read/write only). Existing files with overly permissive modes are tightened on load

  • opencode.jsonc Support - Configure models flow now detects and prefers existing opencode.jsonc files. JSONC parsing strips comments and trailing commas before JSON.parse

  • Header Contract Tests - New src/constants.test.ts validates header shapes, randomization behavior, and optional header fields for both Antigravity and Gemini CLI styles

Changed

  • Unified Gemini Routing - Gemini quota fallback between Antigravity and Gemini CLI pools is now always enabled for Gemini models. The quota_fallback config flag is deprecated and ignored (backward-compatible, no breakage)

  • cli_first Honored in Routing - resolveHeaderRoutingDecision() centralizes routing logic and properly respects cli_first for unsuffixed Gemini models

  • Fingerprint Headers Simplified - buildFingerprintHeaders() now returns only User-Agent. Removed X-Goog-QuotaUser, X-Client-Device-Id, X-Goog-Api-Client, and Client-Metadata from outgoing content requests to align with Antigravity Manager behavior

  • Client Metadata Reduced - Fingerprint client metadata trimmed to ideType, platform, pluginType only. Removed osVersion, arch, sqmId

  • Gemini CLI User-Agent Format - Updated from google-genai-sdk/... to GeminiCLI/... format

  • Search Model - Changed from gemini-2.0-flash to gemini-2.5-flash for improved search result quality

  • Deterministic Search Generation - Search requests now use temperature: 0 and topP: 1 instead of thinking config

  • OAuth Headers Dynamic - oauth.ts and project.ts now use getAntigravityHeaders() instead of static constants, removing stale X-Goog-Api-Client from token/project calls

Fixed

  • #410: Strip x-goog-user-project header for ALL header styles, not just Antigravity. This header caused 403 errors on Daily/Prod endpoints when the user's GCP project lacked Cloud Code API
  • #370 / #336: Account deletion now persists correctly. Root cause: saveAccounts() merged deleted accounts back from disk. Fixed by introducing saveAccountsReplace() for destructive writes and syncing in-memory state immediately
  • #381: Disabled accounts no longer selected via sticky index. getCurrentAccountForFamily() now skips disabled accounts and advances the active index
  • #384: google_search tool no longer returns empty citations when using gemini-3-flash. Search model switched to gemini-2.5-flash
  • #377: Configure models flow now respects existing opencode.jsonc files instead of creating duplicate opencode.json
  • Excessive Disk Writes - Fixed project context auth updates causing 3000+ writes/sec during streaming. Changed from reference equality to value comparison on auth tokens and added throttled saves. Prevents SSD wear on macOS
  • Fingerprint Alignment - Force-regenerated fingerprints to match current Antigravity Manager behavior, fixing ideType and stripping stale client metadata fields

Removed

  • Extra Outgoing Headers - X-Goog-Api-Client, Client-Metadata, X-Goog-QuotaUser, X-Client-Device-Id no longer sent on content requests
  • Fingerprint Metadata Fields - osVersion, arch, sqmId removed from fingerprint client metadata
  • updateFingerprintVersion Helper - Removed from accounts module (fingerprint version rewriting no longer needed)

Documentation

  • AGENTS.md expanded with detailed architecture, code style, and fingerprint system documentation
  • README.md, CONFIGURATION.md, MULTI-ACCOUNT.md updated to reflect deprecated quota_fallback and automatic Gemini pool fallback behavior
  • antigravity.schema.json marks quota_fallback as deprecated/ignored

[1.4.5] - 2026-02-05

Added

  • Configure Models Menu Action - Auth login menu now includes a "Configure models" action that writes plugin model definitions directly into opencode.json, making setup easier for new users

  • cli_first Config Option - New configuration option to route Gemini models to Gemini CLI quota first, useful for users who want to preserve Antigravity quota for Claude models

  • toast_scope Configuration - Control toast visibility per session with toast_scope: "root_only" to suppress toasts in subagent sessions

  • Soft Quota Protection - Skip accounts over 90% usage threshold to prevent Google penalties, with configurable soft_quota_threshold_percent and wait/retry behavior

  • Gemini CLI Quota Management - Enhanced quota display with dual quota pool support (Antigravity + Gemini CLI)

  • OPENCODE_CONFIG_DIR Environment Variable - Custom config location support for non-standard setups

  • quota_refresh_interval_minutes - Background quota cache refresh (default 15 minutes)

  • soft_quota_cache_ttl_minutes - Cache freshness control for soft quota checks

Changed

  • Model Naming and Routing - Documented antigravity-prefixed model names and automatic mapping to CLI preview names (e.g., antigravity-gemini-3-flash → gemini-3-flash-preview)

  • Antigravity-First Quota Strategy - Exhausts Antigravity quota across ALL accounts before falling back to Gemini CLI quota (previously per-account)

  • Quota Routing Respects cli_first - Fallback behavior updated to respect cli_first preference

  • Config Directory Resolution - Now prioritizes OPENCODE_CONFIG_DIR environment variable

  • Enhanced Debug Logging - Process ID included for better traceability across concurrent sessions

  • Improved Quota Group Resolution - More consistent quota management with resolveQuotaGroup function

Fixed

  • #337: Skip disabled accounts in proactive token refresh
  • #233: Skip sandbox endpoints for Gemini CLI models (fixes 404/403 cascade)
  • Windows Config Auto-Migration: Automatically migrates config from %APPDATA%\opencode\ to ~/.config/opencode/
  • Root Session Detection: Reset isChildSession flag correctly for root sessions
  • Stale Quota Cache: Prevent spin loop on stale quota cache
  • Quota Group Default: Fix quota group selection defaulting to gemini-pro when model is null

Removed

  • Fingerprint Headers for Gemini CLI - Removed fingerprint headers from Gemini CLI model requests to align with official behavior
  • web_search Configuration Leftovers - Cleaned up remaining web_search config remnants from schema

Documentation

  • Updated README with model configuration options and simplified setup instructions
  • Updated MODEL-VARIANTS.md with Antigravity model names and configuration guidance
  • Updated CONFIGURATION.md to clarify quota_fallback behavior across accounts
  • Updated MULTI-ACCOUNT.md with dual quota pool and fallback flow details

[1.3.2] - 2026-01-27

Added

  • Quota check and account management in auth login - Added new --quota and --manage options to the auth login command for checking account quota status and managing accounts directly from the CLI (#284)

  • Request timing jitter - Added configurable random delay to requests to reduce detection patterns and improve rate limit resilience. Requests now include small random timing variations

  • Header randomization for fingerprint diversity - Headers are now randomized to create more diverse fingerprints, reducing the likelihood of requests being grouped and rate-limited together

  • Per-account fingerprint persistence - Fingerprints are now persisted per-account in storage, allowing consistent identity across sessions and enabling fingerprint history tracking

    • Added fingerprint restore operations to AccountManager
    • Extended per-account fingerprint history for better tracking
    • Fingerprint now shown in debug output

  • Scheduling mode configuration - Added new scheduling modes including cache-first mode that prioritizes accounts with cached tokens, reducing authentication overhead

  • Failure count TTL expiration - Account failure counts now expire after a configurable time period, allowing accounts to naturally recover from temporary issues

  • Exponential backoff for 503/529 errors - Implemented exponential backoff with jitter for capacity-related errors, matching behavior of Antigravity-Manager

Changed

  • Increased MODEL_CAPACITY backoff to 45s with jitter - Extended the base backoff time for model capacity errors from previous values to 45 seconds, with added jitter to prevent thundering herd issues

  • Regenerate fingerprint after capacity retry exhaustion - When all capacity retries are exhausted, the fingerprint is now regenerated to potentially get assigned to a different backend partition

  • Enhanced duration parsing for Go format - Improved parsing of duration strings to handle Go-style duration formats (e.g., 1h30m) used in some API responses

Fixed

  • Prevent toast spam for rate limit warnings - Added 5-second debounce for rate limit warning toasts to prevent notification flooding when multiple requests hit rate limits simultaneously (#286)

  • getEnabledAccounts now treats undefined as enabled - Fixed issue where accounts without an explicit enabled field were incorrectly filtered out. Accounts now default to enabled when the field is undefined

  • Show correct position in account toast for enabled accounts - Fixed the account position indicator in toast notifications to only count enabled accounts, showing accurate position like "Account 2/5" instead of including disabled accounts

  • Filter disabled accounts in all selection methods - Ensured disabled accounts are properly excluded from all account selection strategies (round-robin, least-used, random, etc.)

  • Robust handling for capacity/5xx errors - Implemented comprehensive retry logic for model capacity and server errors, achieving parity with Antigravity-Manager's behavior

    • Reordered parsing logic to prioritize capacity checks
    • Fixed loop retry logic to prevent state pollution
    • Added capacity retry limit to prevent infinite loops (#263)

  • Fixed @opencode-ai/plugin dependency location - Moved @opencode-ai/plugin from devDependencies to dependencies section, fixing runtime errors when the plugin was installed without dev dependencies

Removed

  • Removed deprecated web_search configuration - The deprecated web_search.default_mode and web_search.grounding_threshold configuration options have been fully removed. Use the google_search tool instead (introduced in 1.3.1)

[1.3.1] - 2026-01-21

Added

  • New google_search tool for web search - Implements Google Search grounding as a callable tool that the model can invoke explicitly
    • Makes separate API calls with only { googleSearch: {} } tool, avoiding Gemini API limitation where grounding tools cannot be combined with function declarations
    • Returns formatted markdown with search results, sources with URLs, and search queries used
    • Supports optional URL analysis via urlContext when URLs are provided
    • Configurable thinking mode (deep vs fast) for search operations
    • Uses gemini-3-flash model for fast, cost-effective search operations

Changed

  • Upgraded to Zod v4 and adjusted schema generation for compatibility
  • Deprecated web_search config - The web_search.default_mode and web_search.grounding_threshold config options are now deprecated. Google Search is now implemented as a dedicated tool rather than automatic grounding injection

Fixed

  • keep_thinking=true now works without debug mode - Fixed Claude multi-turn conversations failing with "Failed to process error response" when keep_thinking=true after tool calls, unless debug mode was enabled

    • Root cause: filterContentArray trusted any signature >= 50 chars for last assistant messages, but Claude returns its own signatures that Antigravity doesn't recognize
    • Fix: Now verifies signatures against our cache via isOurCachedSignature() before passing through. Foreign/missing signatures get replaced with SKIP_THOUGHT_SIGNATURE sentinel
    • Why debug worked: Debug mode injects synthetic thinking with no signature, triggering sentinel injection correctly

  • Fixed tool calls failing for tools with no parameters - Tools like hive_plan_read, hive_status, and hive_feature_list that have no required parameters would fail with Zod validation error state.input: expected record, received undefined

    • Root cause: When Claude calls a tool with no parameters, it returns functionCall without an args field. The response transformation only processed parts where functionCall.args was defined, leaving args as undefined
    • Fix: Changed condition to handle all functionCall parts, defaulting args to {} when missing, ensuring opencode's state.input always receives a valid record

  • Auth headers aligned with official Gemini CLI - Updated authentication headers to match the official Antigravity/Gemini CLI behavior, reducing "account ineligible" errors and potential bans (#178)

    • GEMINI_CLI_HEADERS["User-Agent"]: 9.15.1 → 10.3.0
    • GEMINI_CLI_HEADERS["X-Goog-Api-Client"]: gl-node/22.17.0 → gl-node/22.18.0
    • ANTIGRAVITY_HEADERS["User-Agent"]: Updated to full Chrome/Electron user agent string
    • Token exchange now includes Accept, Accept-Encoding, User-Agent, X-Goog-Api-Client headers
    • Userinfo fetch now includes User-Agent, X-Goog-Api-Client headers
    • fetchProjectID now uses centralized constants instead of hardcoded strings

  • quiet_mode now properly suppresses all toast notifications - Fixed quiet_mode: true in antigravity.json not suppressing "Status dialog dismissed" and other toast notifications (#207)

    • Root cause: The showToast helper function didn't check quietMode, and only some call sites had manual !quietMode && guards
    • Fix: Moved quietMode check inside showToast helper so all toasts are automatically suppressed when quiet_mode: true

Removed

  • Removed automatic googleSearch injection - Previously attempted to inject { googleSearch: {} } into all Gemini requests, which never worked due to API limitations. Now uses the explicit tool approach instead

[1.3.0] - Previous Release

See releases for previous versions.