feat: fetch GonkaGate models dynamically

AGENTS.md

@@ -72,12 +72,11 @@ These audited assumptions are the current implementation contract:
 - the v1 durable secret target is user-level
   `settings.env.GONKAGATE_API_KEY`
 - after API-key collection, the installer must make a separate authenticated
-  `GET https://api.gonkagate.com/v1/models` request and confirm all three
-  supported models before rendering the picker or writing config
-- the managed Qwen provider catalog must include all three supported models:
-  - `qwen/qwen3-235b-a22b-instruct-2507-fp8`
-  - `moonshotai/Kimi-K2.6`
-  - `minimaxai/minimax-m2.7`
+  `GET https://api.gonkagate.com/v1/models` request before rendering the
+  picker or writing config
+- authenticated `/v1/models` is the source of truth for user-visible model ids,
+  names, `--model` validation, default selection, and managed provider writes
+- no checked-in GonkaGate model id list may become the runtime source of truth
 - `qwen auth` is removed in the audited baseline; status is an interactive
   `/doctor` concern rather than a standalone `qwen auth status` command
 - `modelProviders` is currently marked with `replace` merge semantics, so

CHANGELOG.md

@@ -2,11 +2,10 @@
 
 ## [0.2.1](https://github.com/GonkaGate/qwen-code-setup/compare/v0.2.0...v0.2.1) (2026-06-23)
 
-
 ### Bug Fixes
 
-* allow future Qwen Code versions ([676d44d](https://github.com/GonkaGate/qwen-code-setup/commit/676d44d7a0181f82c22b3c584652cd642879eca0))
-* allow future Qwen Code versions ([e084d4e](https://github.com/GonkaGate/qwen-code-setup/commit/e084d4e1a2f48f6434e6c5f84311bf2cc74b8794))
+- allow future Qwen Code versions ([676d44d](https://github.com/GonkaGate/qwen-code-setup/commit/676d44d7a0181f82c22b3c584652cd642879eca0))
+- allow future Qwen Code versions ([e084d4e](https://github.com/GonkaGate/qwen-code-setup/commit/e084d4e1a2f48f6434e6c5f84311bf2cc74b8794))
 
 ## [0.2.0](https://github.com/GonkaGate/qwen-code-setup/compare/v0.1.0...v0.2.0) (2026-06-16)
 
@@ -22,5 +21,5 @@
 - Add TypeScript, CI, release, package, docs, contract-test, and mirrored skill infrastructure.
 - Record the initial Qwen Code compatibility assumptions without claiming the installer runtime is implemented.
 - Expand the Qwen Code setup PRD with concrete v1 requirements for Qwen settings, secret storage, scope behavior, verification, and diagnostics.
-- Require support for all three current GonkaGate models and authenticated `/v1/models` discovery after API-key intake.
+- Require authenticated `/v1/models` discovery after API-key intake.
 - Implement the Qwen Code setup runtime with safe secret intake, authenticated model discovery, managed user/project settings writes, backups, rollback, dry-run, durable verification, current-session shadowing warnings, redacted human/JSON output, and fake-tested optional live verification.

README.md

@@ -44,7 +44,7 @@ The happy path is:
 1. Detects the local `qwen` binary and records its version.
 2. Collects or reuses `GONKAGATE_API_KEY` without printing the secret.
 3. Calls authenticated `GET https://api.gonkagate.com/v1/models`.
-4. Requires all supported GonkaGate models before writing config.
+4. Uses that live model response for defaulting, selection, and validation.
 5. Writes the GonkaGate provider into `modelProviders.openai[]`.
 6. Sets `security.auth.selectedType = "openai"` and `model.name`.
 7. Stores the durable key reference at `settings.env.GONKAGATE_API_KEY`.
@@ -56,12 +56,11 @@ settings. Use `--json` for machine-readable, redacted output.
 
 ## Supported Models
 
-The installer only offers models that are present in GonkaGate's authenticated
-model catalog:
-
-- `qwen/qwen3-235b-a22b-instruct-2507-fp8`
-- `moonshotai/Kimi-K2.6`
-- `minimaxai/minimax-m2.7`
+The installer has no checked-in user-facing model catalog. After collecting the
+GonkaGate API key, it calls authenticated
+`GET https://api.gonkagate.com/v1/models` and uses the returned model ids and
+names as the source of truth for the picker, `--model` validation, default
+selection, and Qwen Code provider writes.
 
 ## Known Qwen Code Baseline
 

docs/architecture-decisions.md

@@ -32,37 +32,27 @@ The initial expected fields are:
 Activation is expected through:
 
 - `security.auth.selectedType = "openai"`
-- `model.name = <curated GonkaGate model id>`
+- `model.name = <live GonkaGate model id>`
 
 This decision was revalidated against `@qwen-code/qwen-code` `0.18.0` on
 June 12, 2026. The provider surface remains valid, but `modelProviders` is
 marked with `replace` merge semantics, so project-scope writes must stay gated
 until later executable compatibility proof confirms they cannot hide
 user-managed providers.
 
-## Decision 3: Combine Curated Support With Authenticated Model Discovery
+## Decision 3: Use Authenticated Model Discovery As Runtime Catalog
 
-The setup flow should expose the three supported GonkaGate models, not an
-arbitrary model id box, and it must confirm availability through GonkaGate after
-API-key intake.
+The setup flow exposes the live GonkaGate model catalog returned for the user's
+API key, not a checked-in model id list.
 
 Reasons:
 
-- predictable public UX
-- support and troubleshooting clarity
-- compatibility metadata can be attached to each model
-- validation proof can gate end-user exposure
-- the authenticated `/v1/models` response proves the supplied API key can
-  actually use the required models
-
-The v1 required model set is:
-
-- `qwen/qwen3-235b-a22b-instruct-2507-fp8`
-- `moonshotai/Kimi-K2.6`
-- `minimaxai/minimax-m2.7`
-
-All three must be written into `modelProviders.openai[]`; `model.name` only
-selects the default.
+- adding or removing a GonkaGate network model should not require a repository
+  change
+- authenticated `/v1/models` response proves the supplied API key can see the
+  model before writes
+- the same live ids drive picker, `--model` validation, default selection, and
+  `modelProviders.openai[]`
 
 ## Decision 4: Use User Settings `env` For The Durable Secret
 

docs/how-it-works.md

@@ -31,10 +31,11 @@ The runtime:
 3. Parse and record the local Qwen Code version.
 4. Collect the GonkaGate API key through a safe input.
 5. Fetch `https://api.gonkagate.com/v1/models` with that key.
-6. Confirm all three supported GonkaGate models are available.
-7. Choose the setup default from the supported model picker.
-8. Merges managed GonkaGate settings into Qwen Code user settings, including all
-   three models in `modelProviders.openai[]`.
+6. Parse and dedupe the live GonkaGate model catalog.
+7. Choose the setup default from the first fetched model or the live model
+   picker.
+8. Merges managed GonkaGate settings into Qwen Code user settings, including
+   the fetched models in `modelProviders.openai[]`.
 9. Persist the secret in user-level `settings.env.GONKAGATE_API_KEY`.
 10. Verifies the durable Qwen Code outcome from locally inspectable evidence.
 11. Reports current-session shadowing when process or project env overrides the

docs/model-validation.md

@@ -1,62 +1,41 @@
 # Model Validation
 
-This document records the validation gate for curated GonkaGate models exposed
-through `@gonkagate/qwen-code-setup`.
+This document records the live model discovery rule for
+`@gonkagate/qwen-code-setup`.
 
-## Current Supported Models
+## Source Of Truth
 
-The v1 installer must support all three GonkaGate models currently available
-for this setup flow:
+The installer has no checked-in user-facing GonkaGate model id list. After
+collecting the GonkaGate API key, it must make authenticated
+`GET https://api.gonkagate.com/v1/models` with Bearer auth.
 
-- `qwen/qwen3-235b-a22b-instruct-2507-fp8`
-- `moonshotai/Kimi-K2.6`
-- `minimaxai/minimax-m2.7`
+The OpenAI-compatible response shape is expected to contain at least:
 
-The static registry records each model with:
+```json
+{
+  "data": [{ "id": "provider/model", "name": "Optional display name" }]
+}
+```
 
-- a stable curated key
-- the GonkaGate model id
-- a display label
-- `status: "validated"`
-- validation evidence date
-- Qwen Code compatibility notes for the audited `0.18.0` baseline
-- optional Qwen `generationConfig` fragments
+The installer deduplicates returned ids, rejects malformed or empty responses
+with `validated_models_unavailable`, and uses the remaining live models for:
 
-Exactly one registry entry is marked as the recommended default for `--yes`.
-For the current set, that default is
-`qwen/qwen3-235b-a22b-instruct-2507-fp8`.
+- the interactive picker
+- `--model` validation
+- the `--yes` default, which is the first fetched model
+- `modelProviders.openai[]` writes
+- `model.name`
+- install-state managed model metadata
 
-After collecting the GonkaGate API key, the installer must make a separate
-authenticated request to `https://api.gonkagate.com/v1/models` and confirm that
-all three ids are available before showing the picker or writing Qwen Code
-settings.
+## Runtime Rule
 
-If any required id is missing, setup must fail with
-`required_models_unavailable` instead of writing a partial provider catalog.
+Only models returned by authenticated `/v1/models` for the user's key should
+appear in the picker or be accepted by `--model`.
 
-## Minimum Validation Proof
+The Qwen Code provider catalog under `modelProviders.openai[]` must include the
+live fetched model ids for that run, not a static subset. The selected
+model controls `model.name`.
 
-Before a model can be added to or removed from the supported set, record proof
-for:
-
-- GonkaGate `/v1/models` availability for an authenticated `gp-...` key
-- Qwen Code model provider selection
-- non-streaming chat completion
-- streaming chat completion if Qwen Code uses streaming for normal operation
-- tool/function calling behavior if Qwen Code requires it for agentic tasks
-- context-window assumptions
-- timeout/retry behavior
-- any provider-specific generation settings
-
-## Registry And Runtime Rule
-
-Only models with `status: "validated"` and present in the authenticated
-`/v1/models` response should appear in the end-user picker.
-
-The Qwen Code provider catalog under `modelProviders.openai[]` must include all
-three supported models, not only the selected default. The selected default
-controls `model.name`.
-
-Unsupported curated model keys and arbitrary raw model ids must fail before any
-runtime work. Extra GonkaGate `/v1/models` entries remain ignored until this
-registry and the public docs are explicitly updated.
+Adding or removing a model in the GonkaGate network must not require a
+repository change. No checked-in GonkaGate model id list may become the runtime
+source of truth.

docs/security.md

@@ -25,45 +25,33 @@ Allowed secret inputs:
 
 ## Qwen Code Secret Storage
 
-Qwen Code provider entries refer to an `envKey`. For v1, the durable secret
-target is user-level `settings.env.GONKAGATE_API_KEY` inside the active Qwen
-settings file:
+Qwen Code provider entries refer to `envKey`.
+
+For v1, the durable secret target is user-level
+`settings.env.GONKAGATE_API_KEY` inside the active Qwen settings file:
 
 - `~/.qwen/settings.json`, or
 - `<QWEN_HOME>/settings.json` when Qwen Code is configured to use `QWEN_HOME`
 
-The installer uses this target because Qwen Code always treats `settings.env`
-as a fallback environment source, while a separate user `.env` can be skipped
-when a trusted project `.env` exists.
-
 The secret must never be written to project `.qwen/settings.json`, project
 `.env`, or shell profiles.
 
 ## Model Discovery Request
 
-After collecting the key, the installer must make an authenticated
-`GET https://api.gonkagate.com/v1/models` request to confirm the three supported
-models are available.
+After collecting the key, the installer must make authenticated
+`GET https://api.gonkagate.com/v1/models` and use that live response as the
+user-visible model catalog.
 
-Security rules for this request:
+Security rules for the request:
 
-- pass the key only through an `Authorization: Bearer ...` header
-- never print the header or raw key
+- pass the key only through the `Authorization: Bearer ...` header
+- never print the raw header key
 - do not include the raw response in user-facing errors
-- report missing model ids through redacted diagnostics such as
-  `required_models_unavailable`
+- report invalid, empty, or unavailable live model catalogs through redacted
+  diagnostics such as `validated_models_unavailable`
 
 ## Diagnostics
 
-Do not print raw config or environment dumps if they can contain substituted
-secrets.
-
-If a verification command exposes secret-bearing output, capture it internally
-only and render a redacted summary.
-
-## Backups And Rollback
-
-Existing managed targets are backed up under
-`~/.gonkagate/qwen-code/backups/` before replacement. Project settings backups
-also stay under that user-level backup root. Failed writes or failed durable
-verification roll back prior managed writes where the filesystem permits it.
+Do not print raw config or environment dumps that contain substituted secrets.
+If verification output may contain secrets, capture it internally and render
+only a redacted summary.