refactor(agent-core-v2): remove the legacy kosong request layer (#352)

## Requirement or Bug

Remove the legacy `packages/agent-core-v2/src/kosong` request layer.
Port the features that existed only there and fix the bugs it caused.
Stacked on #349.

## Bug Reproduction Steps

1. Configure a provider with `type = "openai"` and `env = {
OPENAI_API_KEY = "sk-..." }`, with no `apiKey` and no process env key.
2. Start a session and send a turn.
3. The auth check fails with `AuthTokenMissingError`, even though the
request path would find the key.

## Root Cause

20 of the 62 files under `src/kosong` still loaded at runtime through
`app/auth/*` and `app/kosongConfig/*`. kosong keeps its own
provider-definition map, and at runtime that map held only the pythinker
definitions, because the standard definitions never loaded. So
`resolveModelAuthMaterial` / `resolveModelForReady` (auth check) and
`envOverlay` (vendor `*_BASE_URL`) ignored provider-`env` values for
every non-pythinker provider. Real requests use `llm-adapter` and find
them. This is a fundamental fix: one provider-definition registry, the
live one.

## Code Changes

- Delete `src/kosong` (61 files, about 11k lines). The 19 importers now
import the same symbols from `#/llm-adapter/*` (types are identical
apart from the import path).
- Port features that existed only in the deleted copy, each with a test
that fails before and passes after:
- **OpenCode billing errors:** a 401/402/403 whose body says
"insufficient balance", "insufficient credit", "credits exhausted" or
"please recharge" is a provider error, not `provider.auth_error`.
Ordinary 401s stay auth errors.
- **DSML / Hermes tool calls:** tool calls that some models write as
text tags on the chat-completions stream are parsed into real tool
calls.
  - **`modelRecordProviderId`** moved to `llm-adapter/model/model.ts`.
- `apps/vis` imports
`@pymodel/agent-core-v2/llm-adapter/contract/tokens` instead of the
`kosong` subpath.
- `scripts/check-identity-freeze.mjs` drops the deleted kosong path.
- Test fix: an MCP registry test set the wrong home variable, so it did
not isolate the home directory. It now sets `PYTHINKER_CODE_HOME`.

The default-model fallback that also lived only in kosong is **not**
restored: since #323 the gateway tests require that `default_model` is
never rewritten. That is a product decision, tracked in #351.

## Behavior Changes and Affected Users

| Behavior | Before | After | Who relies on the old behavior | Escape
hatch |
|---|---|---|---|---|
| Vendor API key in a provider's `env` table | Auth check ignores it,
turn fails with `AuthTokenMissingError` | Key is found, turn runs |
Nobody (the old behavior was a bug) | n/a |
| Vendor `*_BASE_URL` in an `env` provider of non-pythinker type |
Ignored by `envOverlay` | Applied | Nobody (bug) | Remove the variable
from `env` |
| OpenCode 401/402/403 with a billing message | `provider.auth_error`
("not logged in") | Provider error with the billing message, not retried
| Clients that map `provider.auth_error` to a re-login prompt for this
case | None needed; the old message was wrong |
| Chat-completions stream with DSML/Hermes tool tags | Tags shown as
assistant text, no tool call | Parsed into tool calls. Text that could
start a tag is held back until it is known not to be a tag, and
`llm.streaming.finish` arrives after the stream ends | Nobody relies on
raw tags | None |
| `@pymodel/agent-core-v2/kosong/*` subpath import | Resolves | Gone |
`apps/vis` (updated in this PR); no other consumer found in the repo |
Import from `.../llm-adapter/*` |

Affected modules and coverage:
- `app/auth`: `test/app/auth/auth.test.ts` (provider-env key case, fails
on `main`).
- `app/kosongConfig/envOverlay`: new non-pythinker base-url case (fails
on `main`).
- `human/llm` openai format and stream: billing-error tests and DSML
parser/recovery tests.
- Full suites: agent-core-v2 6,491 pass, agent-gateway 1,407 pass,
vis-server 173 pass; `tsc` and `tsgo` clean; no-comments and
identity-freeze checks pass.

## Checklist

- [x] I have read the
[CONTRIBUTING](https://github.com/PyModel/pythinker-code/blob/main/CONTRIBUTING.md)
document.
- [x] I have linked a related issue (external PRs: issue must have a
maintainer's `/approve`).
- [x] I have added tests that prove my feature works.
- [x] The behavior-change table above is complete, and every removed
behavior or flipped default is named in the changeset and either has an
escape hatch or was explicitly approved by a maintainer in this PR.
- [x] Ran `gen-changesets` skill, or this PR needs no changeset.
- [x] Ran `gen-docs` skill, or this PR needs no doc update.

Author: elkaix

.changeset/dsml-tool-call-recovery.md

@@ -0,0 +1,6 @@
+---
+"@pymodel/pythinker-code": patch
+"@pymodel/pythinker-desktop": patch
+---
+
+Run tool calls that some models (such as DeepSeek) write as DSML or <tool_call> text instead of showing them as plain text.

.changeset/opencode-billing-error.md

@@ -0,0 +1,6 @@
+---
+"@pymodel/pythinker-code": patch
+"@pymodel/pythinker-desktop": patch
+---
+
+Report an insufficient-balance response from OpenAI-compatible providers as a billing error instead of an authentication error.

.changeset/provider-env-api-key-auth.md

@@ -0,0 +1,6 @@
+---
+"@pymodel/pythinker-code": patch
+"@pymodel/pythinker-desktop": patch
+---
+
+Accept a vendor API key or base URL set in a provider's env table (for example ANTHROPIC_API_KEY or ANTHROPIC_BASE_URL) instead of ignoring it.

apps/pythinker-code/src/cli/sub/provider.ts

@@ -406,7 +406,7 @@ export async function handleCatalogAdd(
 
   // Always restore `[thinking]` from what was there before — including
   // `undefined`. Persisting `enabled: false` when the user never set it would
-  // make `resolveThinkingEffort` (agent-core-v2/src/kosong/model/thinking.ts) treat
+  // make `resolveThinkingEffort` (agent-core-v2/src/llm-adapter/model/thinking.ts) treat
   // it as an explicit "off" request and silently disable thinking, even for
   // thinking-capable models.
   config.thinking = previousThinking;

apps/vis/server/src/lib/context-projector.ts

@@ -10,7 +10,7 @@ import {
   selectCompactionUserMessages,
   selectRecentUserMessages,
 } from '@pymodel/agent-core-v2/agent/contextMemory/compactionHandoff';
-import { estimateTokensForMessages } from '@pymodel/agent-core-v2/kosong/contract/tokens';
+import { estimateTokensForMessages } from '@pymodel/agent-core-v2/llm-adapter/contract/tokens';
 import { renderToolResultForModel } from '@pymodel/agent-core-v2/agent/contextMemory/toolResultRender';
 import type {
   ContentPart,
@@ -678,7 +678,7 @@ const MICRO_MIN_CONTENT_TOKENS = 100;
 
 /** Replicates the engine's per-char token weighting exactly, over the same
  *  `text` + `think` parts its gate counts. The engine
- *  (`packages/agent-core-v2/src/kosong/contract/tokens.ts`) sums per-part
+ *  (`packages/agent-core-v2/src/llm-adapter/contract/tokens.ts`) sums per-part
  *  estimates, each
  *  `estimateTokens(s) = Math.ceil(asciiCount / 4) + nonAsciiCount` (ASCII ~4
  *  chars/token, every non-ASCII/CJK code point a full token); other part types

apps/vis/server/test/lib/context-projector.test.ts

@@ -1,6 +1,6 @@
 // apps/vis/server/test/lib/context-projector.test.ts
 import { describe, it, expect, afterEach } from 'vitest';
-import { estimateTokensForMessages } from '@pymodel/agent-core-v2/kosong/contract/tokens';
+import { estimateTokensForMessages } from '@pymodel/agent-core-v2/llm-adapter/contract/tokens';
 import { buildCompactionContinuationText } from '@pymodel/agent-core-v2/agent/contextMemory/compactionHandoff';
 import { buildSessionFixture } from '../fixtures/build';
 import { projectContext } from '../../src/lib/context-projector';

packages/agent-core-v2/src/app/auth/auth.ts

@@ -1,7 +1,7 @@
 import { createDecorator, type ServiceIdentifier } from '#/_base/di/instantiation';
 import { Error2 } from '#/_base/errors/errors';
 
-import type { OAuthRef } from '#/kosong/provider/provider';
+import type { OAuthRef } from '#/llm-adapter/provider/provider';
 
 import { AuthErrors } from './errors';
 

packages/agent-core-v2/src/app/auth/authService.ts

@@ -14,15 +14,15 @@ import { type ILogger, ILogService } from '#/_base/log/log';
 import { IBootstrapService } from '#/app/bootstrap/bootstrap';
 import { IConfigService } from '#/app/config/config';
 import { LifecycleScope } from '#/app/scopes';
-import { IModelService, type ModelRecord } from '#/kosong/model/model';
+import { IModelService, type ModelRecord } from '#/llm-adapter/model/model';
 import {
   effectiveModelConfig,
   nonEmpty,
   providerNameFromFlatModel,
   resolveModelAuthMaterial,
   resolveModelForReady,
-} from '#/kosong/model/modelAuth';
-import { IProviderService, type OAuthRef } from '#/kosong/provider/provider';
+} from '#/llm-adapter/model/model-auth';
+import { IProviderService, type OAuthRef } from '#/llm-adapter/provider/provider';
 
 import {
   AuthModelNotResolvedError,

packages/agent-core-v2/src/app/auth/authStatusService.ts

@@ -1,8 +1,8 @@
 import { ScopeActivation, registerScopedService } from '#/_base/di/scope';
 import { LifecycleScope } from '#/app/scopes';
-import { IModelService } from '#/kosong/model/model';
-import { resolveModelForReady } from '#/kosong/model/modelAuth';
-import { IProviderService } from '#/kosong/provider/provider';
+import { IModelService } from '#/llm-adapter/model/model';
+import { resolveModelForReady } from '#/llm-adapter/model/model-auth';
+import { IProviderService } from '#/llm-adapter/provider/provider';
 
 import type { AuthSummary } from './authStatus';
 import { IAuthStatusService } from './authStatus';

packages/agent-core-v2/src/app/auth/configSection.ts

@@ -19,7 +19,7 @@ import {
   transformPlainObject,
 } from '#/app/config/toml';
 import { type AssertExact, type Equal } from '#/_base/utils/typeEquality';
-import type { OAuthRef } from '#/kosong/provider/provider';
+import type { OAuthRef } from '#/llm-adapter/provider/provider';
 
 export const SERVICES_SECTION = 'services';