imposter-who

name	imposter-who
description	Probe the SmoothLlmImposter router for the resolved upstream model behind the current harness/agent by issuing a non-streaming `--who?` request. Use when the user asks "which model am I really hitting", "who am I talking to", "what is the imposter target for a model", "am I being rerouted", or `/imposter-who`. Also use when verifying an imposter mapping (e.g. confirming `gpt-5.5` is rewritten to `glm-5.2`) without spending upstream tokens. Works around the streaming-harness limitation that silently bypasses the in-band `--who?` switch (LADR-05).
allowed-tools	["Bash(.agents/skills/imposter-who/scripts/imposter-who.sh:*)","Bash(${CLAUDE_PLUGIN_ROOT}/.agents/skills/imposter-who/scripts/imposter-who.sh:*)","Bash(curl:*)"]
models	{"claude":"haiku","copilot":"gpt-5.4-mini","codex":"gpt-5.4-mini"}

Imposter Who

Discover which upstream model the SmoothLlmImposter router resolves an inbound model to, by sending a non-streaming --who? probe directly to the router. The router short-circuits with a synthetic reply naming the inbound model, the resolved target (or passthrough), the auth scheme, and the session identity — with zero upstream HTTP calls.

Why a skill instead of just typing --who?

Agent harnesses (Codex, Claude Code) stream chat requests by default ("stream": true). The router intentionally refuses to synthesize --who? replies for streaming requests (HLD 010, LADR-05), so a --who? typed into the harness forwards to the real upstream, which then answers it as an ordinary user turn — the probe is silently bypassed. This skill issues a separate non-streaming curl to the router, so the short-circuit fires and the real routing decision is reported.

Workflow

This is a single-shot probe. Run the script once, relay the one-line reply, stop. Do not enumerate other models, sweep dialects, or re-probe unless the user names a specific model.

Important — what the router matches on. The SmoothLlmImposter router rewrites on the inbound (From) model — the model name the harness is configured to send to the router (e.g. claude-opus-4-7, gpt-5.5). The probe's --model argument takes that From value, NOT your own executing model. When the router is configured From: claude-opus-4-7 → To: minimax-m3, the agent running through the imposter is minimax-m3 — that is the resolved upstream, not the inbound model. Probing --model minimax-m3 will always return Passthrough: minimax-m3 because no From mapping starts from the To model. The script cannot auto-detect the harness's inbound model from the environment; only the user (or the harness config) knows it.

1. Determine the From model and dialect to probe, in this order of precedence:
   • If the user named a model in their request, use that model as the From. That is the common case — the user usually names the model they configured their harness to send (e.g. /imposter-who gpt-5.5 or "am I really hitting gpt-5.5?").
   • If the user did not name a model, DO NOT fall back to probing your own executing model (that is the To, and probing it returns a meaningless Passthrough). Ask the user once, briefly, which inbound model their harness is configured to send (e.g. "which model is your harness set to send — e.g. gpt-5.5, claude-opus-4-7?"). Do not run the script until you have a From model.
   • Pick the dialect from the From model name: gpt-* (and most non-Claude chat models) → OpenAI; claude-* → Anthropic. The script auto-detects dialect from OPENAI_BASE_URL / ANTHROPIC_BASE_URL; pass --dialect only when the model name is ambiguous.
2. Run the probe once:

.agents/skills/imposter-who/scripts/imposter-who.sh --model gpt-5.5

For an Anthropic model:

.agents/skills/imposter-who/scripts/imposter-who.sh --dialect anthropic --model claude-sonnet-4-6

3. Relay the single-line output verbatim to the user. It has the shape:

Imposter: gpt-5.5 -> glm-5.2 (auth: Bearer, session: null)

• Imposter: in -> out — the inbound model is rewritten to out upstream.
• Passthrough: inbound — no mapping matched; the request reaches the real provider unchanged.
• auth: — the auth scheme the router will use upstream (Bearer / ApiKey / none / caller-passthrough).
• session: — resolved session identity, or null.

Do not loop over candidate models, probe both dialects, or build a summary table unless the user explicitly asks for a sweep. The default invocation is one curl, one line, one answer.

Script

scripts/imposter-who.sh — takes --dialect, --model, --session, --base-url. Derives the dialect from OPENAI_BASE_URL / ANTHROPIC_BASE_URL when not forced. No API key is required: the router resolves upstream credentials from its own config, not the inbound request.

Exit codes: 0 synthetic reply printed; 1 router did not short-circuit (feature disabled or unexpected response); 2 env/curl failure.

• --port N overrides the port in the resolved base URL (it replaces the :<port> authority in OPENAI_BASE_URL / ANTHROPIC_BASE_URL). Useful when the router runs on a non-default port or when you want to point at a different router instance without re-setting the env vars.

Notes

• The probe never reaches the upstream provider — it is answered by the router itself, so it costs zero upstream tokens and works even when upstream credits are exhausted.
• --session id attaches a caller session id (body session_id for OpenAI, session_id header for Anthropic) so a prior imposter-newsession mapping reflects in the session: field of the reply.
• The model passed to --model should match what the harness actually sends; the router matches on the configured From entries (HLD 007).