Skip to content

Releases

v0.7.2

Choose a tag to compare

@github-actions github-actions released this 12 Nov 10:20

Fixed

  • Preserve existing environment variables during CLI tools apply. Applying Hermes settings now uses upsertEnvFileKey to update OPENAI_API_KEY without overwriting the rest of the .env file, preventing accidental erasure of platform bot tokens (Discord, Telegram) and credentials.

v0.7.1

Choose a tag to compare

@github-actions github-actions released this 12 Nov 09:47

Added

  • Dynamic model limits and 1M Qwen support. Model token limits now automatically synchronize from upstream accounts and propagate across configured CLI tools. Models such as qwen4-max are recognized with 1,000,000 tokens context window and 128,000 max output tokens.
  • Multi-path Hermes configuration sync. The CLI tools API now automatically discovers, backs up, and applies configuration to both standard ~/.hermes/ and custom $HERMES_HOME (e.g. ~/AI/hermes-stack/config/).

Changed

  • CLI tools configuration and reload automation. Applying CLI tool configurations (Hermes, Claude Code, Codex, etc.) automatically ensures model definitions, provider endpoints, and context windows are synchronized.
  • Model heuristics and upstream limit inheritance. Improved model pattern matching to prevent version heuristics collisions (e.g. Claude 4.5 vs Claude 5) and allow external models to seamlessly inherit policy token limits.

v0.7.0-muse

Choose a tag to compare

@github-actions github-actions released this 08 Nov 09:34

See CHANGELOG.md for details.

v0.7.0

Choose a tag to compare

@github-actions github-actions released this 08 Nov 09:04

Added

  • Bulk delete account groups. Each provider subgroup (gorouter.app, tabitoken.com, ...) now has a trash icon in its header that deletes every account in that group in one click, after a confirmation dialog showing the count and domain.

Changed

  • Codex config re-renders after Apply. The selected model now appears in the input field immediately after a successful apply, instead of only after reopening the Codex detail view.

v0.6.9

Choose a tag to compare

@github-actions github-actions released this 08 Nov 06:02

Changed

  • Update modal now has a single one-click Install button. The secondary "Go to Download" link that opened the GitHub releases page for manual download was removed. The self-update flow (download tarball → verify sha256 → swap binary → restart service) is the only path now.

v0.6.8

Choose a tag to compare

@github-actions github-actions released this 07 Oct 17:18

Fixed

  • The Antigravity verification link could verify the wrong Google identity. Google returns the VALIDATION_REQUIRED page with an empty authuser parameter, so the browser completed the owner check on its default session: the operator saw "Authentication successful" while the flagged account kept failing with 403. The stored link now pins authuser to the account email, so Google's account chooser targets the identity that is actually flagged.
  • Antigravity plan tags were wrong for paid accounts. Cards ran Antigravity through the Kiro subscription badge, which falls back to "Free" when subscriptionType is empty — mislabeling accounts whose loadCodeAssist reports a paid tier such as standard-tier. Both account payloads now expose antigravityTier; the card and the detail view badge the real tier, and an unknown tier renders no badge instead of a guessed one.

v0.6.7

Choose a tag to compare

@github-actions github-actions released this 06 Oct 18:50

Fixed

  • Self-update on systemd never claimed its new version. The dashboard's update wrapper is a child of the very process the restart kills, so the updater died mid-restart and never reached the post-restart version.json commit: the restarted service served the new binary while /health reported the old version, --check kept offering the same update, and update.state stayed running:true. The version claim is now committed before the restart, the previous claim is kept beside it, and a rollback restores it. Sandbox test T12 kills the script at its own restart step and asserts the new claim survived.
  • The update health check probed the wrong port. The spawn never passed OMNIPROXY_PORT, so update.sh defaulted to 8080 while a prefix install listens elsewhere (20131 on hitokiri): a perfectly good swap would fail the check and roll itself back. The endpoint now passes the configured listen port.

v0.6.6

Choose a tag to compare

@github-actions github-actions released this 06 Oct 18:21

Fixed

  • One-click update on a prefix install wrote the new binary where nothing executes it. The update endpoint hardcoded OMNIPROXY_LAYOUT=repo, the layout of the launchd dev box. On a prefix install (systemd user unit, ~/.omniproxy-user) update.sh therefore placed the new binary at ~/.omniproxy-user/omniproxy and the new web/ at ~/.omniproxy-user/web — one level above the bin/ subtree the service actually runs — then restarted the service, which came up on the untouched old binary. Nothing announced the failure: the dashboard reported success, /health kept the old version, and update.state stayed running:true because the wrapper that records the exit code died with the old process. The endpoint now derives the layout from where the running binary lives (installed web/ or scripts/ beside it with version.json one level up → prefix; version.json beside it → repo), and unit tests pin all three shapes.

v0.6.5

Choose a tag to compare

@github-actions github-actions released this 06 Oct 17:25

Fixed

  • Antigravity login no longer triggers Google's phone-verification step. The authorize request used the v1 endpoint (/o/oauth2/auth) plus a PKCE code_challenge — a shape no Cloud Code Assist desktop client sends, and one Google's consent screen answers with an SMS challenge for this client. Side-by-side with a working third-party router showed the difference: the v2 authorize endpoint and no PKCE at all, the client secret authenticating the exchange instead. OmniProxy now sends exactly that shape (/o/oauth2/v2/auth, no code_challenge/code_verifier), and a test pins it — endpoint, absent PKCE params, required params — so a regression cannot silently bring the verification step back.
  • Cancelling an Antigravity login right after starting it could panic the process. The callback server goroutine read session.server at run time while abort() nils that field on cancel, so the two could race into Serve on a nil *http.Server. The new shape test tripped it on its first run (start then deferred cancel); the server is now captured in a local before the goroutine spawns.

v0.6.4

Choose a tag to compare

@github-actions github-actions released this 06 Oct 05:07

Fixed

  • Applying the Hermes CLI-tool settings no longer writes an unparseable config.yaml. The generated provider block quotes model IDs through yamlQuoteIfNeeded, which only quoted on :#@/?&= and spaces. A model ID beginning with [ fell through bare — and reseller upstreams hand back exactly that shape, so [opencode]deepseek-v4-flash was discovered from a live account and emitted as an unquoted mapping key. In YAML a leading [ opens a flow sequence, so the parser read [opencode] as an array, choked on the trailing deepseek-v4-flash:, and reported the error against the enclosing models: block — pointing the operator several screens away from the line that was actually broken:
    cannot parse config.yaml: while parsing a block mapping, line 16, column 1:
      models:
    did not find expected key
    
    The failure was silent rather than loud, which is what made it costly: Hermes catches the parse error, logs a warning, and falls back to its last known-good config, so the gateway stayed up while serving a stale base_url and model list. Pressing Apply again regenerated the same broken file, so the operator's fix never took hold and nothing pointed at the proxy as the writer. Any leading YAML indicator (`-?:,[]{}#&*!|>%@``) is now quoted. The rule is deliberately broad rather than a list of observed prefixes — over-quoting a scalar is always safe, because the parser strips the quotes and yields the identical value.