dsh-subagent-antigravity

dsh-subagent-antigravity

Route native DeepSeek Harness subagent_fork child sessions onto your Google AI Pro (Antigravity) subscription (daily-cloudcode-pa.googleapis.com) with zero API billing, real-time token streaming, full DSH tool calling, ~200 ms HTTPS quota checks, and a vertically centered right-side UI panel for live usage, model setting, effort level, and /subagent toggling.

parent session (any model route — e.g. deepseek-flash or google-vertex)
  └── subagent_fork → native DSH child session (header tab, live output, token meter, send_message)
        └── AntigravityLlmAdapter / Loopback Proxy (provider: "antigravity", dynamic model & effort)
              └── daily-cloudcode-pa.googleapis.com/v1internal:streamGenerateContent?alt=sse

Features

1. Native subagent_fork Delegation (No TUI / No PTY):

   • Automatically pins subagent_fork child sessions to antigravity/<selected-model> and <selected-effort> by wrapping ctx.subagents.startActivation.
   • Every delegated child is a first-class DSH session: it opens as a navigable tab, streams real SSE text and tool-call deltas as they arrive, supports send_message continuation and interrupt_agent, and reports exact usageMetadata to DSH's token meter.
   • Zero vendor preset tampering required: works out of the box with shipped standard, cordis, and ptc presets.

2. Direct In-Process LlmAdapter + Loopback OpenAI Proxy:

   • Registers AntigravityLlmAdapter directly on ctx.llm (provider: "antigravity") so child sessions require zero cordis.patch.yml / llm-pi-ai boilerplate.
   • Also starts a loopback OpenAI-compatible proxy on 127.0.0.1:8791/v1 (GET /v1/models, POST /v1/chat/completions, GET /health) for compatibility with llm-pi-ai or external local tools.
   • Full bidirectional tool calling: converts JSON Schemas to Gemini's strict uppercase Schema subset (collapsing anyOf/oneOf/allOf and ensuring ARRAY items), pairs parallel functionCall / functionResponse turns (separating inline tool-result images into a follow-up user turn), caches and replays Gemini 3.x cryptographic thoughtSignature tokens in ~/.dsh/antigravity/tool-signatures.json (bounded to 600 entries), and gracefully degrades foreign/unsigned parent history turns to readable text.

3. Instant HTTPS Plan Usage (~200 ms, Zero PTY):

   • Queries v1internal:fetchAvailableModels and v1internal:retrieveUserQuotaSummary directly over HTTPS using the CLI's OAuth token (~/.gemini/antigravity-cli/antigravity-oauth-token), refreshing tokens automatically in memory.
   • Exposed via the antigravity_usage model tool and GET /api/dsh-antigravity/usage (?force=1 bypasses the 5-minute cache).

4. Vertically Centered Right-Side UI Panel (shell.overlay + conversation.input.right):

   • Renders a compact capsule vertically centered on the right edge of the DSH UI (shell.overlay), featuring:
     • Simple Quota & /subagent Button: Shows the remaining 5-hour quota percentage for the active setting (amber under 25%, red under 8%, solid blue when /subagent is armed). Click to toggle /subagent in the active composer draft; Alt-click to force a live quota refresh.
     • Slim Side Arrow: Expands and collapses the floating settings popover panel.
   • Expanded Settings & Usage Popover:
     • Setting (Model & Usage Left): Choose between Gemini 3.8 Flash, Claude Sonnet 5.5, Claude Opus 5.5, and GPT-OSS 120B, each displaying its live remaining quota percentage, progress bar, and reset countdown.
     • Effort Level: Choose Low (0 budget), Med (4k), High (24k), or Max (32k) thinking token budget. Persisted to ~/.dsh/antigravity/settings.json via GET|POST /api/dsh-antigravity/settings and applied immediately to new subagent_fork activations.
     • Delegate with /subagent Toggle: Synced with the active conversation composer via a headless bridge in conversation.input.right.

Supported Models

Model ID	Family	Upstream	Context Window	Max Output	Thinking Budget	Quota Bucket
gemini-3.8-flash-high (default)	gemini-3.8-flash	gemini-3.8-flash-tiered	1,048,576	65,536	24,576	Gemini Flash/Pro
gemini-3.8-flash-medium	gemini-3.8-flash	gemini-3.8-flash-tiered	1,048,576	65,536	4,096	Gemini Flash/Pro
gemini-3.8-flash-low	gemini-3.8-flash	gemini-3.8-flash-tiered	1,048,576	65,536	0	Gemini Flash/Pro
claude-sonnet-5-5-high	claude-sonnet-5-5	claude-sonnet-5-5-high	1,000,000	128,000	adaptive	Claude / GPT-OSS
claude-sonnet-5-5-medium	claude-sonnet-5-5	claude-sonnet-5-5-medium	1,000,000	128,000	adaptive	Claude / GPT-OSS
claude-sonnet-5-5-low	claude-sonnet-5-5	claude-sonnet-5-5-low	1,000,000	128,000	adaptive	Claude / GPT-OSS
claude-opus-5-5-high	claude-opus-5-5	claude-opus-5-5-high	1,000,000	128,000	adaptive	Claude / GPT-OSS
claude-opus-5-5-medium	claude-opus-5-5	claude-opus-5-5-medium	1,000,000	128,000	adaptive	Claude / GPT-OSS
claude-opus-5-5-low	claude-opus-5-5	claude-opus-5-5-low	1,000,000	128,000	adaptive	Claude / GPT-OSS
gpt-oss-120b-medium	gpt-oss-120b	gpt-oss-120b-medium	131,072	65,536	adaptive	Claude / GPT-OSS