BEFORE — commit 95cb068c10a6cf0b33eed0a0cab3b298763fe852
Environment: Linux x64; Node 26.13.1; real T3 server + CLI; codex-cli 0.167.2.
Fresh isolated HOME and T3 base directory; no copied database; no provider credentials.
All paths below refer to synthetic fixtures, normalized as $SCRATCH; host identifier normalized as $HOST.

Seed: instance codex, setupMode=existing, homePath="", shadowHomePath=$SCRATCH/shadow,
instance environment CODEX_HOME=$SCRATCH/custom-codex; initially disabled.
Default and custom config files differ only in local endpoint (/default versus /custom).
Seeded usage: default 111 input + 10 output; custom 22222 input + 10 output.

1. Start real server: node apps/server/src/bin.ts serve --base-dir $SCRATCH/t3 --port 39856 --host 127.0.0.1
Filesystem readlink after startup:
{
  "sqlite": "$SCRATCH/home/.codex/sqlite",
  "config.toml": "$SCRATCH/home/.codex/config.toml",
  "shell_snapshots": "$SCRATCH/home/.codex/shell_snapshots",
  "cache": "$SCRATCH/home/.codex/cache",
  "logs": "$SCRATCH/home/.codex/logs",
  "mcp-oauth-locks": "$SCRATCH/home/.codex/mcp-oauth-locks",
  "worktrees": "$SCRATCH/home/.codex/worktrees",
  "thread-writer-locks": "$SCRATCH/home/.codex/thread-writer-locks",
  "plugins": "$SCRATCH/home/.codex/plugins",
  "sessions": "$SCRATCH/home/.codex/sessions",
  "archived_sessions": "$SCRATCH/home/.codex/archived_sessions",
  "skills": "$SCRATCH/home/.codex/skills"
}

2. Authenticated real WebSocket RPC server.getConfig (snapshot captured again after the turn):
{
  "instanceId": "codex",
  "continuation": {
    "groupKey": "codex:home:$SCRATCH/home/.codex"
  }
}

3. Authenticated real WebSocket RPC server.getUsageSummary with awaitRefresh=true:
{
  "sinceDay": "2026-11-13",
  "untilDay": "2026-11-13",
  "timeZone": "UTC",
  "buckets": [
    {
      "day": "2026-11-13",
      "provider": "codex",
      "model": "gpt-6.1-sol",
      "sourcePath": "$SCRATCH/home/.codex/sessions",
      "totals": {
        "uncachedInputTokens": 111,
        "cachedInputTokens": 0,
        "cacheCreationTokens": 0,
        "outputTokens": 10,
        "reasoningTokens": 0
      },
      "costUsd": 0.000322,
      "cacheSavingsUsd": 0,
      "categoryCostUsd": {
        "input": 0.000222,
        "cacheRead": 0,
        "cacheWrite": 0,
        "output": 0.0001
      },
      "costSource": "modelPriced",
      "records": 1,
      "unpricedRecords": 0,
      "unpricedTokens": 0,
      "sessions": 1
    }
  ],
  "sources": [
    {
      "fingerprint": {
        "hostId": "$HOST",
        "provider": "codex",
        "resolvedHomePath": "$SCRATCH/home/.codex/sessions",
        "volumeId": "64769:1469680"
      },
      "status": "ok",
      "scannedFiles": 1,
      "skippedFiles": 0,
      "malformedRecords": 0,
      "distinctSessions": 1,
      "message": null
    }
  ]
}

4. Enable same instance via server.updateSettings; refresh provider; add fresh project through CLI.
node apps/server/src/bin.ts thread start $SCRATCH/project "Say hello." --provider codex --model gpt-6.1-sol --json --wait --base-dir $SCRATCH/t3
Logging wrapper execs /usr/local/bin/codex unchanged. Subprocess launches:
[
  {
    "args": [
      "app-server",
      "-c",
      "thread_unload_delay_secs=0"
    ],
    "CODEX_HOME": "$SCRATCH/shadow",
    "sessions": "$SCRATCH/home/.codex/sessions",
    "config": "model_provider = \"evidence\"\nmodel = \"gpt-6.1-sol\"\n[model_providers.evidence]\nname = \"Synthetic evidence endpoint\"\nbase_url = \"http://127.0.0.1:39958/default\"\nwire_api = \"responses\"\nrequires_openai_auth = false\n",
    "thread": null
  },
  {
    "args": [
      "app-server",
      "-c",
      "thread_unload_delay_secs=0"
    ],
    "CODEX_HOME": "$SCRATCH/shadow",
    "sessions": "$SCRATCH/home/.codex/sessions",
    "config": "model_provider = \"evidence\"\nmodel = \"gpt-6.1-sol\"\n[model_providers.evidence]\nname = \"Synthetic evidence endpoint\"\nbase_url = \"http://127.0.0.1:39958/default\"\nwire_api = \"responses\"\nrequires_openai_auth = false\n",
    "thread": null
  },
  {
    "args": [
      "app-server",
      "-c",
      "thread_unload_delay_secs=0"
    ],
    "CODEX_HOME": "$SCRATCH/shadow",
    "sessions": "$SCRATCH/home/.codex/sessions",
    "config": "model_provider = \"evidence\"\nmodel = \"gpt-6.1-sol\"\n[model_providers.evidence]\nname = \"Synthetic evidence endpoint\"\nbase_url = \"http://127.0.0.1:39958/default\"\nwire_api = \"responses\"\nrequires_openai_auth = false\n",
    "thread": null
  }
]

Observed local HTTP request during this turn:
[
  {
    "method": "POST",
    "path": "/default/responses",
    "model": "gpt-6.1-sol"
  }
]

Actual CLI result (exit 1; intentional HTTP 400 from local endpoint, on both commits):
{"threadId":"mcp:acf30100-e53f-49b3-a0df-94b9faff40c4","runId":"run:thread:mcp%3Aacf30100-e53f-49b3-a0df-94b9faff40c4:ordinal:1","status":"failed","reply":null,"error":"{\"error\": {\"message\": \"Evidence endpoint reached; no paid model request is made.\", \"type\": \"invalid_request_error\", \"code\": \"evidence_capture\"}}"}
Thread mcp:acf30100-e53f-49b3-a0df-94b9faff40c4 stopped with status failed: {"error": {"message": "Evidence endpoint reached; no paid model request is made.", "type": "invalid_request_error", "code": "evidence_capture"}}


Interpretation: endpoint path identifies the config actually consumed by real Codex after the T3 adapter opens a session and starts a turn. This is not a completed inference or an OAuth/renewal check.