fix(athena): parallelize council launches and gate handoff actions

2026-02-18 19:26:20 +01:00
parent ca199ee85b
commit 4764d65db1
9 changed files with 433 additions and 411 deletions
@@ -0,0 +1,119 @@
+export const HEALTH_TEMPLATE = `# Health Command
+
+## Purpose
+
+Ping each configured model provider with a tiny request and report latency and connectivity status. Useful for debugging provider connectivity, authentication failures, and rate limiting.
+
+---
+
+# PHASE 1: DISCOVER CONFIGURED PROVIDERS
+
+Read the OpenCode configuration to enumerate the providers the user has set up.
+
+Configuration locations to check, in order:
+1. Project: \`.opencode/opencode.json\` or \`.opencode/opencode.jsonc\`
+2. User: \`~/.config/opencode/opencode.json\` or \`~/.config/opencode/opencode.jsonc\`
+
+Use the Read tool to load whichever file exists. If both exist, the project file takes precedence but you should still check the user file for additional providers.
+
+Look for the top-level \`provider\` object. Each key under it is a provider id (for example \`anthropic\`, \`openai\`, \`google\`, \`openrouter\`, \`groq\`, \`deepseek\`, \`xai\`, \`mistral\`, \`cohere\`, \`opencode\`, etc.). Each value can have:
+- \`options.baseURL\` - custom endpoint (use this when present)
+- \`options.apiKey\` - API key (NEVER print this in output)
+- \`models\` - object whose keys are model ids
+
+If no \`provider\` block exists, inform the user that no providers are configured in opencode.json and stop.
+
+---
+
+# PHASE 2: RESOLVE ENDPOINT FOR EACH PROVIDER
+
+For each discovered provider, determine the health check endpoint. Prefer the cheapest read-only endpoint that still validates auth.
+
+| Provider id            | Default base URL                              | Health endpoint suffix |
+|------------------------|-----------------------------------------------|------------------------|
+| anthropic              | https://api.anthropic.com                     | /v1/models             |
+| openai                 | https://api.openai.com                        | /v1/models             |
+| google                 | https://generativelanguage.googleapis.com     | /v1beta/models         |
+| openrouter             | https://openrouter.ai/api                     | /v1/models             |
+| groq                   | https://api.groq.com/openai                   | /v1/models             |
+| deepseek               | https://api.deepseek.com                      | /v1/models             |
+| xai                    | https://api.x.ai                              | /v1/models             |
+| mistral                | https://api.mistral.ai                        | /v1/models             |
+| cohere                 | https://api.cohere.com                        | /v1/models             |
+| together               | https://api.together.xyz                      | /v1/models             |
+| fireworks              | https://api.fireworks.ai/inference            | /v1/models             |
+| perplexity             | https://api.perplexity.ai                     | /v1/models             |
+| opencode               | https://opencode.ai                           | /                      |
+
+Rules:
+- If the provider config has \`options.baseURL\`, use that instead of the default.
+- If the provider id is not in the table above, default to \`<baseURL>/v1/models\` if \`baseURL\` is set, otherwise mark the provider as \`unknown\` and skip the network call.
+
+---
+
+# PHASE 3: PING EACH PROVIDER
+
+For each provider, run a single curl call via the Bash tool. Use the smallest possible request (HTTP GET, no body) and a 5 second timeout.
+
+Construct the auth header per provider id:
+- \`anthropic\`: \`-H "x-api-key: $KEY" -H "anthropic-version: 2023-06-01"\`
+- \`google\`: append \`?key=$KEY\` to the URL, no Authorization header
+- everything else: \`-H "Authorization: Bearer $KEY"\`
+
+Read the API key from the appropriate environment variable when the config does not contain \`options.apiKey\`. Common env var names:
+- \`ANTHROPIC_API_KEY\`, \`OPENAI_API_KEY\`, \`GOOGLE_API_KEY\` / \`GEMINI_API_KEY\`, \`OPENROUTER_API_KEY\`, \`GROQ_API_KEY\`, \`DEEPSEEK_API_KEY\`, \`XAI_API_KEY\`, \`MISTRAL_API_KEY\`, \`COHERE_API_KEY\`, \`TOGETHER_API_KEY\`, \`FIREWORKS_API_KEY\`, \`PERPLEXITY_API_KEY\`
+
+Bash command template (do NOT echo the key):
+
+\`\`\`
+curl -sS -o /dev/null \\
+  --max-time 5 \\
+  -w "%{http_code} %{time_total}" \\
+  -H "Authorization: Bearer $API_KEY" \\
+  "$ENDPOINT"
+\`\`\`
+
+The output is two whitespace-separated values: HTTP status code and total time in seconds.
+
+Interpret results:
+- HTTP \`200\`-\`299\`: status \`ok\`
+- HTTP \`401\` / \`403\`: status \`auth_failed\` (reachable, key is bad or missing)
+- HTTP \`429\`: status \`rate_limited\`
+- HTTP \`5xx\`: status \`server_error\`
+- curl exit non-zero or empty status: status \`unreachable\`
+
+Convert latency to milliseconds (multiply by 1000, round to nearest integer).
+
+CRITICAL SAFETY RULES:
+- Never include the API key value in the printed output.
+- Never write the API key to a file or to the terminal. Pass it through environment variables only.
+- Do not call any endpoint that costs tokens (chat completions, messages, etc). Only the listed read-only endpoints.
+
+---
+
+# PHASE 4: REPORT RESULTS
+
+Print a single Markdown table with one row per configured provider:
+
+\`\`\`
+| Provider | Endpoint | Status | HTTP | Latency |
+|----------|----------|--------|------|---------|
+| anthropic | https://api.anthropic.com/v1/models | ok | 200 | 142 ms |
+| openai | https://api.openai.com/v1/models | auth_failed | 401 | 89 ms |
+| google | https://generativelanguage.googleapis.com/v1beta/models | unreachable | - | - |
+\`\`\`
+
+After the table, add a short \`Summary\` line: \`<N> ok / <N> failing / <N> total\`.
+
+If any provider returned \`auth_failed\`, \`server_error\`, or \`unreachable\`, add a \`Suggested next steps\` section with one bullet per failing provider explaining the most likely cause (missing env var, wrong base URL, network firewall, expired key).
+
+---
+
+# IMPORTANT CONSTRAINTS
+
+- Read-only command: do not modify any file.
+- Do not call delegate_task, ralph_loop, or any background agent.
+- Do not retry failed pings - one attempt per provider is enough.
+- Total wall time budget: under 30 seconds.
+- The user must be able to run /health from any directory; tolerate missing project config gracefully.
+`