feat: fetch GonkaGate models dynamically

AGENTS.md

@@ -19,6 +19,8 @@ Current honest state:
 
 - this repository contains a TypeScript CLI that writes OpenHuman's native
   GonkaGate custom-provider config shape
+- the CLI fetches `GET https://api.gonkagate.com/v1/models` with the user's API
+  key and uses that live response as the runtime model source of truth
 - the CLI preserves unrelated TOML config, creates one timestamped backup when
   replacing an existing config file, and verifies the written shape
 - the CLI accepts secrets only through hidden prompt, `GONKAGATE_API_KEY`, or
@@ -37,22 +39,23 @@ security flow changes, update this file immediately.
 - the stable GonkaGate provider id is `gonkagate`
 - the canonical GonkaGate base URL is `https://api.gonkagate.com/v1`
 - the current transport target is `/v1/chat/completions`
+- the runtime model catalog source is authenticated `/v1/models`
 - OpenHuman configuration is TOML-based and must preserve unrelated user config
 - `gp-...` secrets must never be printed
 - secrets must not be accepted through a plain `--api-key` flag
 - safe secret inputs are hidden prompt, `GONKAGATE_API_KEY`, and
   `--api-key-stdin`
-- setup success is based on local config verification and an optional direct
-  GonkaGate smoke check when a key is available, with OpenHuman sign-in reported
-  as a separate runtime gate
+- setup success is based on live model-catalog fetch, local config verification,
+  and a direct GonkaGate smoke check, with OpenHuman sign-in reported as a
+  separate runtime gate
 
 ## OpenHuman Integration Notes
 
 OpenHuman already has an OpenAI-compatible provider path, but the setup utility
 must write the current native custom-provider shape:
 
 - `cloud_providers` entry with slug `gonkagate`
-- workload provider strings such as `gonkagate:moonshotai/kimi-k2.6`
+- workload provider strings such as `gonkagate:<live-model-id>`
 - provider credentials under `provider:gonkagate`, profile `default`, entered
   through OpenHuman Settings -> AI unless a future supported OpenHuman RPC/API
   write path is added
@@ -66,8 +69,8 @@ The setup utility must still account for OpenHuman's internal tier model names:
 - `fast-v1`
 
 GonkaGate expects concrete model ids. The CLI routes workloads to concrete
-GonkaGate model ids and reports OpenHuman's active-session gate for custom
-providers when the user is signed out.
+GonkaGate model ids returned by `/v1/models` and reports OpenHuman's
+active-session gate for custom providers when the user is signed out.
 
 ## Change Discipline
 

CHANGELOG.md

@@ -1,5 +1,10 @@
 # Changelog
 
+## Unreleased
+
+- Fetch GonkaGate models dynamically from authenticated `/v1/models` and remove
+  runtime reliance on a hardcoded catalog.
+
 ## 0.1.0
 
 - Initial TypeScript CLI scaffold for `@gonkagate/openhuman-setup`.

README.md

@@ -26,9 +26,9 @@ GonkaGate configured as a native OpenAI-compatible custom provider without
 hand-editing OpenHuman TOML config.
 
 The CLI writes OpenHuman's native `cloud_providers` shape, routes OpenHuman
-agent workloads to curated GonkaGate model IDs, preserves unrelated config,
-creates a timestamped backup before replacing an existing config file, and
-verifies the written result.
+agent workloads to live GonkaGate model IDs from `/v1/models`, preserves
+unrelated config, creates a timestamped backup before replacing an existing
+config file, and verifies the written result.
 
 Current OpenHuman credential storage is not a safe standalone JSON write target.
 OpenHuman may store provider secrets in the OS keychain or encrypted JSON
@@ -54,13 +54,16 @@ You need:
 - Detects the effective OpenHuman `config.toml`, including
   `OPENHUMAN_WORKSPACE`, active user, legacy active workspace, staging root, and
   pre-login local config.
+- Fetches `GET https://api.gonkagate.com/v1/models` with the user's GonkaGate
+  API key and treats that response as the model source of truth.
 - Adds one GonkaGate `cloud_providers` entry with slug `gonkagate`.
 - Writes workload provider strings for chat, reasoning, agentic, coding, and
   memory/summarization paths.
 - Preserves unrelated TOML values and unrelated cloud providers.
 - Creates one timestamped backup before replacing an existing config file.
 - Verifies the local config shape after writing.
-- Runs a direct GonkaGate `/v1/chat/completions` smoke when a key is available.
+- Runs a direct GonkaGate `/v1/chat/completions` smoke with the selected live
+  model.
 - Reports OpenHuman sign-in and credential state separately.
 
 ## Target OpenHuman Contract
@@ -72,8 +75,7 @@ You need:
 - OpenHuman provider shape: `cloud_providers` entry plus workload provider
   strings
 - OpenHuman credential key: `provider:gonkagate`, profile `default`
-- Recommended model: `moonshotai/kimi-k2.6`
-- Additional validated model: `qwen/qwen3-235b-a22b-instruct-2507-fp8`
+- Model catalog: authenticated `GET /v1/models` with the user's API key
 
 ## Safe Secret Handling
 
@@ -104,27 +106,24 @@ printf '%s' "$GONKAGATE_API_KEY" | npx @gonkagate/openhuman-setup --api-key-stdi
 
 ## Model Overrides
 
-Defaults:
-
-- reasoning, agentic, coding: `gonkagate:moonshotai/kimi-k2.6`
-- chat, memory/summarization: `gonkagate:qwen/qwen3-235b-a22b-instruct-2507-fp8`
-
-Use curated model keys only:
+By default, the CLI uses the first model returned by GonkaGate `/v1/models` for
+all OpenHuman workloads. Override values must be IDs returned by that live
+response:
 
 ```bash
 npx @gonkagate/openhuman-setup \
-  --reasoning-model kimi-k2.6 \
-  --agentic-model kimi-k2.6 \
-  --coding-model kimi-k2.6 \
-  --summarization-model qwen3-235b-a22b-instruct-2507-fp8
+  --reasoning-model provider/model-id \
+  --agentic-model provider/model-id \
+  --coding-model provider/model-id \
+  --summarization-model provider/model-id
 ```
 
 ## What This Tool Does Not Do
 
 - It does not install OpenHuman.
 - It does not write the old `api_url` + `api_key` + `model_routes` setup shape.
 - It does not accept arbitrary custom base URLs.
-- It does not accept arbitrary raw model IDs.
+- It does not use repository-hardcoded GonkaGate model catalogs.
 - It does not print or persist `gp-...` secrets outside approved OpenHuman
   credential handling.
 - It does not directly mutate OpenHuman `auth-profiles.json`.
@@ -148,7 +147,7 @@ The CLI can be exercised against a temporary OpenHuman workspace with:
 
 ```bash
 npm run build
-OPENHUMAN_WORKSPACE="$(mktemp -d)" node bin/gonkagate-openhuman.js --json --yes
+GONKAGATE_API_KEY=gp-... OPENHUMAN_WORKSPACE="$(mktemp -d)" node bin/gonkagate-openhuman.js --json --yes
 ```
 
 The product requirements for the implementation live in

docs/how-it-works.md

@@ -4,10 +4,11 @@ The utility:
 
 1. detect the active OpenHuman configuration location
 2. collect a GonkaGate API key through a safe input path
-3. preserve unrelated OpenHuman TOML config
-4. configure GonkaGate as an OpenHuman `cloud_providers` entry
-5. write workload provider strings for OpenHuman task tiers
-6. verify the effective setup with a real GonkaGate/OpenHuman smoke check
+3. fetch `GET https://api.gonkagate.com/v1/models` with Bearer auth
+4. preserve unrelated OpenHuman TOML config
+5. configure GonkaGate as an OpenHuman `cloud_providers` entry
+6. write workload provider strings using only fetched live model ids
+7. verify the effective setup with a real GonkaGate/OpenHuman smoke check
 
 Direct OpenHuman credential-store mutation is intentionally skipped. Current
 OpenHuman writes provider secrets through runtime keychain or encrypted JSON
@@ -25,13 +26,16 @@ label = "GonkaGate"
 endpoint = "https://api.gonkagate.com/v1"
 auth_style = "bearer"
 
-chat_provider = "gonkagate:qwen/qwen3-235b-a22b-instruct-2507-fp8"
-reasoning_provider = "gonkagate:moonshotai/kimi-k2.6"
-agentic_provider = "gonkagate:moonshotai/kimi-k2.6"
-coding_provider = "gonkagate:moonshotai/kimi-k2.6"
-memory_provider = "gonkagate:qwen/qwen3-235b-a22b-instruct-2507-fp8"
+chat_provider = "gonkagate:<live-model-id>"
+reasoning_provider = "gonkagate:<live-model-id>"
+agentic_provider = "gonkagate:<live-model-id>"
+coding_provider = "gonkagate:<live-model-id>"
+memory_provider = "gonkagate:<live-model-id>"
 ```
 
+The concrete model id comes from the authenticated `/v1/models` response, not a
+repository catalog.
+
 The implementation must also handle OpenHuman tier model names such as
 `reasoning-v1`, `agentic-v1`, `coding-v1`, `summarization-v1`, and `fast-v1`.
 
@@ -42,5 +46,5 @@ JSON output includes separate machine-readable states for:
 - config write and backup path
 - credential presence or manual key action
 - local config verification
-- direct GonkaGate smoke check when a key is available
+- direct GonkaGate smoke check
 - OpenHuman app-session presence

docs/security.md

@@ -10,6 +10,7 @@ close to the existing GonkaGate setup tools.
 - the CLI must not accept a plain `--api-key` flag
 - support safe inputs only: hidden prompt, `GONKAGATE_API_KEY`, and
   `--api-key-stdin`
+- use the key to fetch the live GonkaGate `/v1/models` catalog with Bearer auth
 - preserve unrelated OpenHuman configuration
 - create backups before managed rewrites
 - redact secret-bearing diagnostics and error messages

docs/specs/openhuman-setup-prd/fact-check.md

@@ -45,11 +45,11 @@ label = "GonkaGate"
 endpoint = "https://api.gonkagate.com/v1"
 auth_style = "bearer"
 
-chat_provider = "gonkagate:qwen/qwen3-235b-a22b-instruct-2507-fp8"
-reasoning_provider = "gonkagate:moonshotai/kimi-k2.6"
-agentic_provider = "gonkagate:moonshotai/kimi-k2.6"
-coding_provider = "gonkagate:moonshotai/kimi-k2.6"
-memory_provider = "gonkagate:qwen/qwen3-235b-a22b-instruct-2507-fp8"
+chat_provider = "gonkagate:<live-model-id>"
+reasoning_provider = "gonkagate:<live-model-id>"
+agentic_provider = "gonkagate:<live-model-id>"
+coding_provider = "gonkagate:<live-model-id>"
+memory_provider = "gonkagate:<live-model-id>"
 ```
 
 The matching credential key is:
@@ -104,6 +104,7 @@ provider:gonkagate / default
 ## Implementation Consequences
 
 - Use `cloud_providers` and workload provider fields as the main write path.
+- Fetch concrete GonkaGate model IDs from authenticated `/v1/models`.
 - Do not write `api_url = "https://api.gonkagate.com/v1"` for v1. That is the
   old direct-provider mental model and risks confusing backend API resolution.
 - Do not rely on `model_routes` as the primary tier-routing mechanism.

docs/specs/openhuman-setup-prd/implementation-plan.md

@@ -28,7 +28,7 @@ Done means:
   workload fields, and `provider:gonkagate` credentials.
 - Do not use the legacy `api_url` + `api_key` + `model_routes` path as the v1
   write target.
-- Do not add arbitrary base URLs or raw model ids.
+- Do not add arbitrary base URLs or model ids absent from live `/v1/models`.
 - Do not install or patch OpenHuman.
 - Do not print, log, or place `gp-...` secrets in argv, output, repository
   files, or backup filenames.

docs/specs/openhuman-setup-prd/release-proof.md

@@ -5,8 +5,9 @@ Date: 2026-06-21
 ## Automated Proof
 
 - `npm run ci` passes.
-- Temp-workspace CLI smoke writes OpenHuman native GonkaGate config:
-  `OPENHUMAN_WORKSPACE=/tmp/openhuman-setup-smoke-codex node bin/gonkagate-openhuman.js --json --yes`.
+- Temp-workspace CLI smoke writes OpenHuman native GonkaGate config with a
+  mocked authenticated `/v1/models` response:
+  `OPENHUMAN_WORKSPACE=/tmp/openhuman-setup-smoke-codex GONKAGATE_API_KEY=gp-... node bin/gonkagate-openhuman.js --json --yes`.
 - The smoke-written config contains one `cloud_providers` entry with slug
   `gonkagate` and workload provider strings for chat, reasoning, agentic,
   coding, and memory.
@@ -17,7 +18,7 @@ Not completed in this environment.
 
 Current blocker:
 
-- `GONKAGATE_API_KEY` is not set.
+- A live `GONKAGATE_API_KEY` is not set.
 - OpenHuman discovery resolves to pre-login config:
   `/Users/daniil/.openhuman/users/local/config.toml`.
 - OpenHuman session inspection reports `missing`.
@@ -35,8 +36,8 @@ and a real GonkaGate key entered through OpenHuman Settings -> AI.
   pre-login `users/local/config.toml`.
 - T2 Pure TOML Merge: fixture tests cover fresh config, unrelated provider
   preservation, GonkaGate replacement, and idempotency.
-- T3 Model Selection: tests cover defaults, global `--model`, per-workload
-  overrides, and curated-key rejection.
+- T3 Model Selection: tests cover live `/v1/models` parsing, first-live-model
+  defaults, global `--model`, per-workload overrides, and live-id rejection.
 - T4 Secret Intake And Redaction: tests cover `GONKAGATE_API_KEY`,
   `--api-key-stdin`, plain `--api-key` rejection, and `gp-...` redaction.
 - T5 Credential Storage: current safe outcome is no direct `auth-profiles.json`