0.7.0 — 2026-07-25

Release Notes

Added

• Kimi (Moonshot AI) static-key provider. New kimi provider pinned to
  https://api.moonshot.ai/v1 with broker-owned key discovery from
  MOONSHOT_API_KEY / KIMI_API_KEY. Catalog covers Kimi K2.7 Code
  (default), K2.7 Code Highspeed, K3 (1M context), and K2.6 — all
  OpenAI-compatible with tool calling and reasoning support. Use models as
  kimi/<model-id>. /effort exposes exact documented reasoning modes per
  model: K3 supports low / high / max (default max) via the
  top-level reasoning_effort field; K2.6/K2.5 support only a thinking
  on/off toggle (thinking.type); the K2.7 Code family always thinks with
  no effort control. Unsupported levels are rejected at mutation time —
  never silently swallowed by the API.

• Typed, brokered provider authentication and routing. synaps login now
  discovers typed OAuth, cloud, and static-key targets from provider
  descriptors. New broker-owned flows cover Grok/xAI, GitHub Copilot, Google
  Gemini CLI / Code Assist, Azure OpenAI, Amazon Bedrock, and Google Vertex AI,
  while preserving Anthropic and OpenAI Codex. Provider-qualified identities
  survive login, persistence, model selection, catalog refresh, runtime routing,
  and status output without exposing long-lived credentials to model runtimes.

• Exact model capabilities and live effort controls. /settings reuses the
  model picker and /effort exposes only modes authorized for the active exact
  model. OpenAI Codex keeps seven source-controlled ChatGPT OAuth model rows and
  distinct xhigh, max, and logical ultra modes. Claude Fable 5 adds wire
  max and logical ultracode; UltraCode lowers to Anthropic xhigh plus
  bounded orchestration doctrine rather than sending a made-up wire value.

• Typed modular prompt orchestration. Source-controlled manifests compose
  exact provider/model/role prompt modules, path-confine local files, persist
  provenance, reload transactionally, and install runtime-owned delegation
  policy. Worker authorization validates exact models, roles, write scopes,
  grants, concurrency, and recursion before credentials, network, billing,
  threads, or worker-state side effects.

• Reactive subagent lifecycle tools and completion gates. Foregrounds can
  start, poll, steer, inspect, collect, resume, and explicitly reconcile
  workers. Completion waits for required reconciliation. Manifestless
  OpenRouter sessions have an exact three-model worker authority for DeepSeek
  V4 Pro, GLM 5.2, and Kimi K2.7 Code; other manifestless providers remain
  foreground-only unless explicitly authorized.

• Brokered dynamic catalogs and provider-native streaming. Catalog and
  streaming support now covers xAI Responses, GitHub Copilot descriptors,
  Gemini Code Assist, Azure OpenAI, Bedrock ConverseStream, Vertex SSE, OpenAI
  Codex, and Anthropic. Broker requests pin hosts and paths, reject redirects,
  and retain credentials inside the broker boundary.

Changed

• Model identity is provider-qualified end to end; model switching now replaces
  manifestless orchestration authority atomically and is refused while workers
  still require collection or reconciliation.
• OpenAI-compatible schema translation recursively sanitizes tool schemas and
  preserves provider-specific call identifiers, while Gemini function
  responses use Code Assist-compatible object payloads and replay thought
  signatures.
• Direct Codex transport errors include provider hosts and full cause chains,
  retry only bounded transient send failures, and use idle read timeouts so
  healthy long-running streams can continue.

Fixed

• Reactive subagent wake and garbage collection. Finished workers no longer
  disappear before collection, and expired unreconciled workers become bounded,
  transport-free tombstones. Tombstones release steering/shutdown channels,
  receiver/thread state, tool logs, and resumable conversation state while
  retaining a UTF-8-safe terminal-output tail and idempotent
  collection/reconciliation path. Reconciled tombstones are removed on the next
  reaper pass, eliminating the completion-gate ⇄ GC deadlock.
• Gemini Code Assist continuation now honors rate-limit reset hints, serializes
  function responses correctly, and removes a public-API alias that Code Assist
  rejects.
• TUI model and effort changes fail closed on unsupported provider/model/mode
  combinations, stale capability generations, active streams, and
  unreconciled-worker races without persisting partial state.

Testing

• Added zero-network OAuth, broker, cloud, catalog, routing, reasoning,
  orchestration, provider-confinement, and streaming harnesses; golden request
  bodies for OpenAI and Anthropic; and an external tombstone harness proving
  bounded allocation release and real completion-gate recovery.
• Live certification artifacts exercised Grok 4.5 fan-out, GLM 5.2 recovery,
  and a heterogeneous DeepSeek → GLM → Kimi workflow. These are validation
  evidence, not guarantees of upstream entitlement or availability.

Fixed (existing development line)

• Reactive subagent wake (Phase 1) — subagents no longer silently vanish
  when the parent turn ends before they complete. cleanup_finished() now
  retains uncollected finished handles for 15 minutes (TTL reaper); a
  finalize_subagent hook fires exactly once at each subagent thread exit
  (outside catch_unwind), stamps finished_at, and pushes a
  subagent_completion event to the parent's EventQueue so the idle parent
  wakes and can call subagent_collect. Kill-switch:
  SYNAPS_DISABLE_SUBAGENT_WAKE=1. Covers start, resume, panic, timeout, and
  tokio-build-failure exit paths. Encodes the live regression as integration
  test I1 (parent_wakes_after_turn_end_multi_subagent). See
  IMPLEMENTATION-PLAN.md in the workspace for full design rationale.

Install synaps 0.7.0

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/HaseebKhalid1507/SynapsCLI/releases/download/v0.7.0/synaps-installer.sh | sh

Download synaps 0.7.0

File	Platform	Checksum
synaps-aarch64-apple-darwin.tar.gz	Apple Silicon macOS	checksum
synaps-x86_64-apple-darwin.tar.gz	Intel macOS	checksum
synaps-aarch64-unknown-linux-gnu.tar.gz	ARM64 Linux	checksum
synaps-x86_64-unknown-linux-gnu.tar.gz	x64 Linux	checksum
synaps-x86_64-unknown-linux-musl.tar.gz	x64 MUSL Linux	checksum