feat: fetch GonkaGate models dynamically

AGENTS.md

@@ -32,13 +32,11 @@ Current honest state:
   sandbox XDG config tree so project-scope verification matches the local
   resolver on `@kilocode/cli >=7.2.0` installs, and that temporary sandbox now
   stages outside the repository with cleanup after verification
-- the stock public build now ships `moonshotai/Kimi-K2.6` as the recommended
-  validated curated default, with installer-managed
+- the installer now fetches `GET https://api.gonkagate.com/v1/models` after
+  safe API-key intake and uses that live response as the model source of truth
+- the managed GonkaGate provider config writes every fetched chat-completions
+  model into Kilo's OpenCode-style `/models` picker with installer-managed
   `limit.output = 8192` for Kilo `7.2.0` compatibility
-- the managed GonkaGate provider config now writes the validated
-  chat-completions catalog, currently Kimi K2.6, Qwen3 235B A22B Instruct
-  2507 FP8, and MiniMax M2.7, so Kilo's OpenCode-style `/models` picker can
-  switch between validated GonkaGate models
 - Kilo detection now accepts `@kilocode/cli >=7.2.0` without pre-blocking
   future Kilo releases, while the audited compatibility baseline remains
   `@kilocode/cli@7.2.0`
@@ -59,12 +57,12 @@ The intended happy path is:
 
 1. user runs `npx @gonkagate/kilo-setup`
 2. installer validates local `kilo` or fallback `kilocode`
-3. installer offers curated validated Kilo model choices
-4. installer auto-selects `project` inside a git repository or `user`
+3. installer collects a GonkaGate key through a hidden prompt,
+   `GONKAGATE_API_KEY`, or `--api-key-stdin`
+4. installer fetches live Kilo model choices from GonkaGate `/v1/models`
+5. installer auto-selects `project` inside a git repository or `user`
    otherwise, and only asks on interactive reruns when the previous
    installer-managed scope differs from the new recommendation
-5. installer collects a GonkaGate key through a hidden prompt,
-   `GONKAGATE_API_KEY`, or `--api-key-stdin`
 6. installer writes the minimum safe Kilo config layers
 7. for `project` installs, installer reports Kilo global UI-model cache risk
    and can clear the current cached model on request
@@ -134,10 +132,8 @@ These are implementation facts today:
 - `src/install/` contains shipped Kilo detection, path resolution, secret
   intake, managed config writes, rollback, effective-config verification, and
   orchestration
-- `src/constants/models.ts` now exposes Kimi K2.6 as the recommended validated
-  production default plus the validated Qwen3 235B A22B Instruct 2507 FP8
-  and MiniMax M2.7 catalog entries, each with installer-managed
-  `limit.output = 8192` in the written Kilo provider config
+- `src/install/model-catalog.ts` fetches and parses GonkaGate `/v1/models` as
+  the installer model source of truth
 - `docs/specs/kilo-setup-prd/spec.md` is the copied Kilo setup PRD
 - `docs/release-readiness.md` records the current production-readiness audit
 - `.github/workflows/release-please.yml` and `.github/workflows/publish.yml`

CHANGELOG.md

@@ -2,6 +2,9 @@
 
 ## Unreleased
 
+- Fetched GonkaGate model availability from authenticated `GET /v1/models`
+  after safe API-key intake, using the live response for the model picker,
+  config writes, and `--model` validation.
 - Updated managed GonkaGate `limit.context` values to match the current
   deployed windows: `240000` for Kimi/Qwen and `180000` for MiniMax.
 - Wrote the validated GonkaGate chat-completions catalog into the managed

README.md

@@ -72,13 +72,10 @@ Current public baseline:
 - package: `@gonkagate/kilo-setup`
 - primary command after setup: `kilo`
 - current transport target: `chat/completions`
-- curated default:
-  `moonshotai/Kimi-K2.6`
-- managed validated catalog for Kilo's OpenCode-style `/models` picker:
-  `moonshotai/Kimi-K2.6`,
-  `qwen/qwen3-235b-a22b-instruct-2507-fp8`, and
-  `minimaxai/minimax-m2.7`
-- installer-managed `limit.output = 8192` for Kilo compatibility
+- model source of truth: `GET https://api.gonkagate.com/v1/models` after safe
+  API-key intake
+- Kilo's OpenCode-style `/models` picker receives every fetched GonkaGate model
+- installer-managed `limit.output = 8192` for each fetched model entry
 - no native Windows production claim yet
 
 ## Shortest Start Path
@@ -95,15 +92,16 @@ The installer will:
 
 1. detect `kilo`, or fall back to `kilocode`
 2. verify local Kilo is at least `@kilocode/cli >=7.2.0`
-3. show the curated model choice
-4. write the validated GonkaGate provider model catalog for Kilo's `/models`
+3. collect your API key through a hidden prompt
+4. fetch the live GonkaGate model catalog from `/v1/models`
+5. show the live model choices
+6. write the fetched GonkaGate provider model catalog for Kilo's `/models`
    picker
-5. choose the recommended scope automatically:
+7. choose the recommended scope automatically:
    - inside a git repository: `project`
    - outside a repository: `user`
-6. collect your API key through a hidden prompt
-7. write the managed config and verify the result
-8. return you to normal `kilo` usage
+8. write the managed config and verify the result
+9. return you to normal `kilo` usage
 
 On interactive reruns, the installer asks about scope only if the last
 installer-managed scope differs from the new recommendation.
@@ -205,18 +203,14 @@ This repository intentionally stays narrow today:
 
 - minimum accepted Kilo floor: `@kilocode/cli >=7.2.0`
 - current transport target: `chat/completions`
-- current curated default:
-  `moonshotai/Kimi-K2.6`
-- validated GonkaGate models exposed to Kilo's `/models` picker:
-  `moonshotai/Kimi-K2.6`,
-  `qwen/qwen3-235b-a22b-instruct-2507-fp8`, and
-  `minimaxai/minimax-m2.7`
+- model availability comes from authenticated `GET /v1/models`
+- if no explicit model is selected, setup uses the API-provided default when
+  present, otherwise the first returned model
 - real-path Kilo verification is not the production default
 - native Windows production support is not claimed yet
 - future Kilo releases are not pre-blocked by version, but observed
   compatibility breaks still need fixes
-- unvalidated extra models and new flows are not implied just because this
-  package exists
+- new transport flows are not implied just because this package exists
 
 The shipped runtime treats effective Kilo config as the real success gate. It
 uses the local resolver as the durable verifier and keeps the XDG-isolated

docs/README.md

@@ -22,9 +22,7 @@ Current contract documents:
 - [`specs/kilo-setup-prd/spec.md`](./specs/kilo-setup-prd/spec.md): copied Kilo
   setup PRD from the planning repository
 
-This repository ships the Kilo installer runtime with a validated curated
-GonkaGate chat-completions catalog, Kimi K2.6 as the recommended default,
-MiniMax M2.7 and Qwen3 235B A22B Instruct 2507 FP8 as additional validated
-catalog entries, installer-managed `limit.output = 8192` for Kilo
-compatibility, and a minimum accepted Kilo floor of `@kilocode/cli >=7.2.0`
-without a preset upper version bound.
+This repository ships the Kilo installer runtime with authenticated GonkaGate
+`GET /v1/models` discovery as the model source of truth, installer-managed
+`limit.output = 8192` for each fetched Kilo model entry, and a minimum accepted
+Kilo floor of `@kilocode/cli >=7.2.0` without a preset upper version bound.

docs/gonkagate-x-kilo.md

@@ -129,16 +129,13 @@ Right now the public default is deliberately small:
 - base URL: `https://api.gonkagate.com/v1`
 - transport: `chat/completions`
 - minimum Kilo floor: `@kilocode/cli >=7.2.0`
-- recommended validated model: `moonshotai/Kimi-K2.6`
-- validated catalog exposed to Kilo's OpenCode-style `/models` picker:
-  `moonshotai/Kimi-K2.6`,
-  `qwen/qwen3-235b-a22b-instruct-2507-fp8`, and
-  `minimaxai/minimax-m2.7`
-- managed limits: `limit.output = 8192` for all validated entries;
-  `limit.context = 240000` for Kimi/Qwen and `180000` for MiniMax
-
-We are treating model support as a curated list, not as a vague "it probably
-works" promise.
+- model source of truth: authenticated `GET /v1/models`
+- catalog exposed to Kilo's OpenCode-style `/models` picker: every fetched
+  GonkaGate model id
+- managed limit: `limit.output = 8192` for each fetched model entry
+
+We are treating model availability as GonkaGate API truth, not as a checked-in
+package catalog.
 
 ## What we are not claiming
 

docs/how-it-works.md

@@ -29,48 +29,46 @@ Current public limit:
 - the published contract keeps a minimum Kilo floor of
   `@kilocode/cli >=7.2.0`, `chat/completions`, and non-Windows production
   claims
-- the curated default is
-  `moonshotai/Kimi-K2.6` with `limit.context = 240000` and
-  `limit.output = 8192`
-- the written provider config includes the validated chat-completions model
-  catalog, currently `moonshotai/Kimi-K2.6`,
-  `qwen/qwen3-235b-a22b-instruct-2507-fp8`, and
-  `minimaxai/minimax-m2.7`, so Kilo's OpenCode-style `/models` picker can
-  switch between validated GonkaGate models
-- broader model claims still require additional proof before models should be
-  added to the validated catalog
+- model availability comes from authenticated `GET /v1/models` after safe
+  API-key intake
+- the written provider config includes every fetched GonkaGate
+  chat-completions model so Kilo's OpenCode-style `/models` picker can switch
+  between the live GonkaGate options
+- each fetched model entry gets installer-managed `limit.output = 8192` for
+  Kilo `7.2.0` compatibility
 
 ## Install Flow
 
 1. Check that `kilo` is available, or fall back to `kilocode`.
 2. Verify the minimum accepted Kilo floor: `@kilocode/cli >=7.2.0`.
-3. Resolve the curated model choice and scope.
-4. Use the recommended scope automatically in the default interactive flow:
-   - `project` inside a git repository
-   - `user` outside a repository
-5. On interactive reruns, ask about scope only when the previous
-   installer-managed scope differs from the new recommendation.
-6. Accept a GonkaGate API key through:
+3. Accept a GonkaGate API key through:
    - a hidden interactive prompt
    - `GONKAGATE_API_KEY`
    - `--api-key-stdin`
-7. Save the secret only under `~/.gonkagate/kilo/api-key`.
-8. Write or update the user-level provider definition, including the validated
+4. Fetch the live GonkaGate model catalog from `GET /v1/models`.
+5. Resolve the live model choice and scope.
+6. Use the recommended scope automatically in the default interactive flow:
+   - `project` inside a git repository
+   - `user` outside a repository
+7. On interactive reruns, ask about scope only when the previous
+   installer-managed scope differs from the new recommendation.
+8. Save the secret only under `~/.gonkagate/kilo/api-key`.
+9. Write or update the user-level provider definition, including the fetched
    GonkaGate model catalog for Kilo's `/models` picker.
-9. When `project` scope is chosen, write only activation settings into
-   `.kilo/kilo.jsonc`.
-10. On rerun, remove only installer-owned stale GonkaGate activation from the
+10. When `project` scope is chosen, write only activation settings into
+    `.kilo/kilo.jsonc`.
+11. On rerun, remove only installer-owned stale GonkaGate activation from the
     old location and preserve unrelated Kilo config.
-11. Verify the durable intended Kilo outcome with the local resolver and use
+12. Verify the durable intended Kilo outcome with the local resolver and use
     the XDG-isolated oracle as a compatibility check.
-12. If `KILO_CONFIG_CONTENT` is active, or the installer is running inside an
+13. If `KILO_CONFIG_CONTENT` is active, or the installer is running inside an
     active `kilo` terminal session with runtime config overrides, verify the
     current session separately.
-13. For `project` installs, surface Kilo global UI-model cache notices and
+14. For `project` installs, surface Kilo global UI-model cache notices and
     optionally clear the current cached model when the user requests it.
-14. Report redacted blockers or mismatches instead of printing raw resolved
+15. Report redacted blockers or mismatches instead of printing raw resolved
     config.
-15. Finish by sending the user back to plain `kilo`.
+16. Finish by sending the user back to plain `kilo`.
 
 ## Why User-Level Provider Ownership
 

docs/release-readiness.md

@@ -17,15 +17,11 @@ docs, and tests for these facts:
 - Kilo command contract: `kilo` first, `kilocode` as fallback
 - minimum Kilo compatibility floor: `@kilocode/cli >=7.2.0`
 - current GonkaGate transport claim: `chat/completions`
-- curated public default:
-  `moonshotai/Kimi-K2.6`
-- validated provider catalog exposed to Kilo's OpenCode-style `/models` picker:
-  `moonshotai/Kimi-K2.6`,
-  `qwen/qwen3-235b-a22b-instruct-2507-fp8`, and
-  `minimaxai/minimax-m2.7`
-- curated model limits for the validated catalog:
-  `limit.output = 8192` for all validated entries; `limit.context = 240000`
-  for Kimi/Qwen and `180000` for MiniMax
+- model availability source: authenticated `GET /v1/models`
+- provider catalog exposed to Kilo's OpenCode-style `/models` picker: every
+  fetched GonkaGate model id
+- model limit written for Kilo compatibility: `limit.output = 8192` for each
+  fetched model entry
 - managed secret path: `~/.gonkagate/kilo/api-key`
 - project scope stays secret-free and still depends on a compatible
   user-level `provider.gonkagate` definition on each machine
@@ -34,21 +30,16 @@ docs, and tests for these facts:
 
 ## External Evidence Captured
 
-The current model-limit contract is backed by product direction plus Gonka
-deployment metadata checked on 2026-06-23:
+The current model contract is backed by product direction plus Kilo
+compatibility evidence:
 
-- Gonka deployment model args list Kimi K2.6 and Qwen3 235B A22B Instruct 2507
-  FP8 with `--max-model-len 240000`
-- Gonka deployment model args list `minimaxai/minimax-m2.7` with
-  `--max-model-len 180000`
-- the model card documents OpenAI-compatible chat completions access through
-  Moonshot's API
-- the API docs describe the supported request path as `POST /v1/chat/completions`
-- the package writes `limit.output = 8192` for validated catalog entries as
-  the installer-managed Kilo compatibility clamp because Kilo `7.2.0` requires
-  a numeric output limit in custom model config
-- the package keeps the same installer-managed `limit.output = 8192`
-  compatibility clamp for every validated catalog entry
+- model availability is resolved from GonkaGate `GET /v1/models` after safe
+  API-key intake
+- the API docs describe the supported request path as
+  `POST /v1/chat/completions`
+- the package writes `limit.output = 8192` for fetched catalog entries as the
+  installer-managed Kilo compatibility clamp because Kilo `7.2.0` requires a
+  numeric output limit in custom model config
 - npm registry metadata checked on 2026-04-29 showed `@kilocode/cli` patch
   releases in the `7.2.x` line, including `7.2.14`, with both `kilo` and
   `kilocode` binaries still exposed by the wrapper package
@@ -62,22 +53,21 @@ The repository must continue to avoid these claims:
 - no claim that GonkaGate `responses` transport works today
 - no claim that real-path Kilo verification is the production default
 - no claim that native Windows production support is proven
-- no claim that every future Kilo behavior or live GonkaGate model is
-  proven just because the current curated default is shipped
+- no claim that every future Kilo behavior or live GonkaGate model is proven
+  beyond fetching and writing the catalog shape
 
 ## Remaining Follow-Up Items
 
 These items still benefit from fresh proof or human review outside this
 repository-only pass:
 
-- fresh live GonkaGate/Kilo smoke evidence when the selected public model or
-  Kilo baseline changes
+- fresh live GonkaGate/Kilo smoke evidence when Kilo's model config
+  requirements or transport behavior changes
 - native Windows oracle-safety proof on a real runner or equivalent native VM
 - product/legal approval for public publishing under
   `@gonkagate/kilo-setup`
 
 Those items do not change the current package contract: the publishable surface
-remains the minimum Kilo `7.2.0` floor without an upper version bound, the
-validated GonkaGate chat-completions catalog with Kimi as the recommended
-default, Qwen3 and MiniMax as additional validated entries, and the current
-non-Windows verification policy documented above.
+remains the minimum Kilo `7.2.0` floor without an upper version bound,
+authenticated GonkaGate `/v1/models` discovery for the chat-completions
+catalog, and the current non-Windows verification policy documented above.

docs/security.md

@@ -54,6 +54,7 @@ proven.
   that sandbox up after verification
 - do not claim real-path Kilo verification as the production default
 
-The current runtime already rejects plain `--api-key`, keeps stock public runs
-blocked before writes when no validated curated model exists, and redacts
-secret-bearing text on user-facing error paths.
+The current runtime already rejects plain `--api-key`, fetches the model
+catalog only after safe API-key intake, blocks before writes when `/v1/models`
+does not return usable models, and redacts secret-bearing text on user-facing
+error paths.

docs/specs/kilo-setup-prd/compatibility-spike-notes.md

@@ -2,6 +2,9 @@
 
 Date: April 14, 2026
 
+Current model availability comes from authenticated GonkaGate `GET /v1/models`;
+model ids below are historical evidence examples, not the runtime catalog.
+
 These notes preserve redacted evidence for the GonkaGate Kilo setup PRD. They do
 not contain raw `kilo debug config` output or real secrets.
 

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

@@ -1,5 +1,10 @@
 # Implementation Plan: GonkaGate Kilo Setup
 
+Current repository behavior supersedes the draft model-catalog tasks below:
+the installer now fetches authenticated `GET /v1/models` after safe API-key
+intake and uses that live response for model choices, config writes, and
+`--model` validation.
+
 ## Overview
 
 This plan breaks the draft PRD into ordered, verifiable implementation tasks.
@@ -317,7 +322,7 @@ diagnostics for the user and for machine-readable JSON output.
 
 - [ ] Detect `KILO_CONFIG`, `KILO_CONFIG_DIR`, home-level directory, and
       `KILO_CONFIG_CONTENT` blockers.
-- [ ] Detect provider shape mismatch, missing curated model entry, provider
+- [ ] Detect provider shape mismatch, missing selected model entry, provider
       allow/deny, provider whitelist/blacklist, and secret-binding provenance
       mismatch.
 - [ ] Treat `disabled_providers` as stronger than `enabled_providers` when both
@@ -422,14 +427,13 @@ sandbox and installer-owned targets.
 ## Task 12: Enable Interactive Setup
 
 **Description:** Replace scaffold-only execution with an interactive install
-flow that detects Kilo, shows the installed version, offers curated model
-choices, recommends scope, accepts a hidden secret, writes config, verifies, and
+flow that detects Kilo, shows the installed version, fetches live model choices from `/v1/models`, recommends scope, accepts a hidden secret, writes config, verifies, and
 ends by sending the user back to plain `kilo`.
 
 **Acceptance criteria:**
 
 - [ ] Interactive flow shows the selected Kilo command and version.
-- [ ] Model picker shows only validated curated choices.
+- [ ] Model picker shows fetched `/v1/models` choices.
 - [ ] Project scope is recommended inside a git repository and user scope
       outside a git repository.
 - [ ] Prompt text explains Kilo-specific scope effects in user language.
@@ -626,8 +630,7 @@ acceptance criteria before claiming production installer support.
 
 ## Open Questions
 
-- What is the proven numeric `limit.output` for
-  `qwen/qwen3-235b-a22b-instruct-2507-fp8`?
+- What numeric `limit.output` policy should fetched model entries use?
 - Does the live GonkaGate/Kilo `chat/completions` smoke pass tool-call behavior
   for the selected model?
 - Will product/legal approve `@gonkagate/kilo-setup` as the public package