---
summary: "Qoder provider notes: separate credits and token planes, legacy SQLite versus current JSONL sources, source precedence, bounded reads, and first-party token gaps."
ids: [qoder, qodercn]
read_when:
  - Changing or debugging Qoder or Qoder CN usage, limits, source discovery or storage migration
  - Investigating Qoder CN usage that is missing, zero or stale in the widget
  - Touching the fork's token_monitor/qodercn.rs, providers/qodercn/paths.js or Qoder roots in clientSources.js
---

# Qoder provider

## Identity and ids

Qoder has two deliberately separate ids:

| Id | Role | Unit |
| --- | --- | --- |
| `qodercn` | opt-in tracked client for local Qoder CN usage | tokens and estimated API cost |
| `qoder` | AI Tool Limits provider for global or CN accounts | big-model credits |

Credits are quota units, not tokens. Never copy credit deltas into `qodercn` usage rows or token history.

## Data sources

| Data | Source |
| --- | --- |
| Current Qoder CN token usage | Claude-compatible JSONL transcripts under the configured `projects` directory |
| Legacy Qoder CN token usage | `SharedClientCache/cache/db/local.db` under the platform QoderCN app-data root |
| Credit quota and plan | Qoder's global or CN web API, authenticated with the separately configured `qoder` cookie |

Token Monitor's tokscale fork parses Qoder CN (`crates/tokscale-core/src/token_monitor/qodercn.rs` in Javis603/tokscale); upstream tokscale has no `qodercn` client, so the catalog marks it `forkOnly`. The parser reads both usage layouts because an upgraded machine can retain historical SQLite data while new sessions are written to JSONL. JSONL token semantics differ from Anthropic's envelope: `input_tokens` already includes `cache_read_input_tokens`, so the parser subtracts the cached subset before emitting uncached input. Model display names and routing tiers (`auto`, `ultimate`, …) are normalized there too; routing tiers are never priced from the model catalog.

Only rows with reported token counts enter usage. Current first-party, plan-billed rows can contain `credits` and `context_usage_ratio` while every token field is zero. The active context-window denominator is not reliably present in the observed session data and can vary per session, so a model-wide window must not be used to reconstruct tokens. Those rows remain absent from token totals; their credits belong to AI Tool Limits. BYOK/custom-model rows that contain measured token fields are counted normally.

## Source precedence

The JSONL projects directory resolves in this order:

1. `TOKEN_MONITOR_QODER_CN_PROJECTS_PATH`;
2. Qoder's own `QODERCN_CONFIG_DIR/projects`;
3. `~/.qoder-cn/projects`.

The legacy database uses `TOKEN_MONITOR_QODER_CN_DB_PATH` when set, then the platform default documented in the README. `src/shared/providers/qodercn/paths.js` resolves the same paths for watching, source health and the anchor fingerprint; it must stay in step with the fork's resolution.

A missing JSONL root is a valid empty source for a legacy-only install. Each source (the database and each transcript) is fingerprinted and its normalized rows are cached in `<tokscale cache dir>/token-monitor/qodercn.json`. Once traversal starts, directory, stat or stream failures, and the file, byte, row and line budgets, fail that whole source: the fork serves its last cached rows and logs a warning rather than publishing a partial read, and a database that is locked or unreadable is treated the same way. Do not skip an unreadable enumerated entry and publish the remaining rows as complete; a scan that returned less would feed a negative delta into the anchored watch tick.

The database path and projects directory are both part of the persisted-anchor fingerprint. Moving either source invalidates the anchor instead of combining today's data from one location with older periods captured from another.

## Credentials and transport

Local token usage needs no credential and stays on disk. The `qoder` limits provider stores its cookie through the shared credential boundary and uses the injected transport for the selected global or CN origin. The local `qodercn` client and the remote `qoder` account do not authenticate or identify each other.

## Invariants and known gaps

- Token history and credit quota remain separate units and surfaces.
- JSONL zero-token rows are not measured zero usage; they are uncountable with the currently persisted data.
- JSONL reads are bounded and fail closed after source discovery begins.
- The explicit Token Monitor projects override takes precedence over Qoder's config-root override.
- Output tokens cannot be recovered from credits or context occupancy.
- Legacy database sessions carry only `chat_session.project_name`, a bare name rather than a path, and project attribution needs a resolved workspace path, so those sessions have no project. JSONL sessions use the transcript `cwd`.
- Rows without a timestamp are stored at the epoch and fall outside any `--since` window.
- Qoder CN is excluded from WSL scans: under a Windows host the fork resolves the legacy database from the host's `APPDATA`, not from `--home`, so a WSL scan would recount host usage.
- The two storage generations have no verified cross-format message key; do not invent one from path or timestamp similarity.

## Verification

Run:

```bash
node --test tests/shared/collectorAnchorPersistence.test.js tests/shared/anchorSeed.test.js tests/shared/clientHealth.test.js tests/shared/wslUsage.test.js
```

The parser's own tests live in the fork: `cargo test -p tokscale-core token_monitor::qodercn`.

Real first-party Qoder CN installs are still needed to verify whether a future build begins persisting a trustworthy per-session active context window. Do not enable reconstruction from a model catalog or preference default alone.