heimdall

Run a heimdall watch: tell the user, with evidence, which model has been serving this session and whether it silently swapped.

Do this:

1. Ground truth — served model per turn. Run the ledger script:

   bash "${CLAUDE_PLUGIN_ROOT}/scripts/model-ledger.sh"
   

   It reads the session transcript's message.model (stamped by the inference server, not the client) and prints per-model turn counts, token usage, the current served model, and any mid-session swaps. If it reports no rows yet, say so plainly — the field populates once turns have completed.

2. Routing context. Run the provider script:

   bash "${CLAUDE_PLUGIN_ROOT}/scripts/which-provider.sh"
   

   It shows provider (Anthropic-direct / Bedrock / Vertex), auth mode, network path, and the nearest CDN edge. Make clear this locates the user, not the model — the edge airport code says nothing about which model answered.

   If the user asks for more than "which model" — token/cache/cost detail, tier, advisor model, or why a cache miss happened — run the extended read:

   bash "${CLAUDE_PLUGIN_ROOT}/scripts/session-inspect.sh"
   

   It surfaces advisorModel, service_tier, speed, inference_geo, token totals, cache hit rate, and cache-miss reasons with re-charged token counts. Call out actionable findings — e.g. a large tools_changed / messages_changed miss means mid-session tool/MCP toggles or history edits busted the prompt cache; fewer of those → higher cache hit → lower cost.

3. Report. Lead with the answer: the current served model, and whether it changed this session. If a swap happened, explain the mechanism honestly: each model has its own usage bucket; when the primary (per model in settings) is exhausted or under high demand, the configured fallbackModel serves instead so the session keeps working — the swap is why the served model changed. Then give the two levers: /model to pin/switch, and /usage (or /usage-credits) to see per-model budget and add credits.

Rules:

• Never guess the model. Only report what message.model shows. If the transcript has no served-model rows yet, say that instead of inferring from latency, config, or a UI banner — those are not evidence.
• No secrets: the scripts never print API keys or tokens.
• Keep it tight: the headline is "you are on X (changed/unchanged this session)."

Optional: the plugin also ships a passive model-drift-net Stop hook that surfaces a one-line notice the moment a swap happens, so drift is caught automatically without running this command.