Skip to main content

Codebuff

Local-file provider for the Codebuff AI coding agent (named manicode in its on-disk layout for historical reasons). Reads per-chat JSON files under ~/.config/manicode/, ~/.config/manicode-dev/, and ~/.config/manicode-staging/, and aggregates per-model token totals plus Codebuff credit spend. No network calls and no authentication.

At a glance​

  • Provider ID — codebuff
  • Detection — any of ~/.config/manicode/, ~/.config/manicode-dev/, ~/.config/manicode-staging/ exists, or a codebuff binary on PATH
  • Auth — local file
  • Type — coding agent
  • Tracks:
    • Total chats, chats today, chats in the last 7 days
    • Total messages
    • Total input, output, cache-read, and cache-write tokens
    • Codebuff credits used (a Codebuff-internal unit, not USD)
    • Per-model token totals with inferred upstream provider hint
    • Daily series for chats, tokens, and credits

Setup​

Auto-detection​

OpenUsage registers the provider when any of the three default channel roots exist or codebuff is on PATH. Auto-detection records the first matching root under data_dir. At fetch time every existing root is scanned and an additional path from the CODEBUFF_DATA_DIR environment variable is also walked when set.

Manual configuration​

{
  "accounts": [
    {
      "id": "codebuff",
      "provider": "codebuff",
      "extra": {
        "data_dir": "~/.config/manicode"
      }
    }
  ]
}

data_dir adds an additional channel root to the scan. The default roots and CODEBUFF_DATA_DIR are still honored; values are de-duped after path normalization, so listing the same directory twice is safe.

Data sources & how each metric is computed​

For each resolved channel root the provider walks <root>/projects/<project>/chats/<chatId>/chat-messages.json. Each file is a JSON array of message objects; only those with role: "assistant" and a recognizable token-usage block are counted.

Usage extraction (three-tier fallback)​

The provider looks for token counts in this order on each assistant message and stops at the first non-empty match:

  1. metadata.usage
  2. metadata.codebuff.usage
  3. metadata.runState.sessionState.mainAgentState.messageHistory[].providerOptions.usage — the first history entry with any non-zero token wins

Field mapping inside the chosen usage block:

Upstreamopenusage metric
input_tokenstotal_input_tokens
output_tokenstotal_output_tokens
cache_read_input_tokenstotal_cache_read
cache_creation_input_tokenstotal_cache_write
creditstotal_credits
modelRawModelID

When model is missing the per-model row falls back to the literal string codebuff-unknown. The upstream provider hint is inferred from the model name prefix: claude- → anthropic, gpt- or o1- → openai, gemini- → google, otherwise unknown.

Credits​

total_credits is emitted in Codebuff's internal credit unit, not in USD. The dashboard renders it as a separate row labelled Credits. There is no automatic conversion to dollars; the upstream JSON does not expose a rate.

Chat counts​

A "chat" is keyed by the tuple <channel>/<project>/<chatId>. Channel is the basename of the channel root (manicode, manicode-dev, manicode-staging, or the basename of a custom root). Project is the first path segment under <root>/projects/. chatId is the directory name immediately above the chat-messages.json file.

  • total_chats — distinct chat tuples observed
  • chats_today — chats with at least one assistant message timestamped on the current UTC day
  • chats_7d — chats with at least one assistant message in the last 7 days

Timestamp resolution​

The provider tries, in order: metadata.timestamp, message.timestamp, then a derivation from the chatId. A Codebuff chatId looks like 2025-12-14T10-00-00.000Z — an ISO-8601 timestamp with the time-portion's : separators replaced by -. The provider rebuilds the colons in the time half only, leaving the date half alone, then parses RFC 3339.

Per-message dedup​

Each assistant message is hashed by (input, output, cache_read, cache_write, ordinal) if it lacks a stable id. The dedup key is then prefixed with <channel>/<project>/<chatId> so the same chat on two channels does not collide. This is what keeps repeated polls from inflating totals when an in-progress chat is rewritten between snapshots.

Daily series​

DailySeries["sessions"] (chats by day), DailySeries["tokens"], and DailySeries["credits"] are populated.

What's NOT tracked​

  • USD cost. Codebuff bills in credits. The provider does not run a pricing lookup, so total_cost_usd is never emitted.

Caveats​

  • The on-disk directory name is manicode rather than codebuff for historical reasons. Both refer to the same product.
  • Three channels (manicode, manicode-dev, manicode-staging) are walked separately so dev / staging activity stays counted, but tagged by channel inside the chat tuple.
  • CODEBUFF_DATA_DIR is honored at fetch time; setting it does not require restarting the daemon for the next poll cycle.
  • The chatId time-parsing has to keep date-half - separators intact while restoring : in the time half. If you spot timestamps falling on the wrong day, file an issue with a sample chatId.

Troubleshooting​

  • Tile is empty — run a Codebuff chat so chat-messages.json lands under one of the channel roots. Confirm with openusage detect.
  • Credits show 0 despite running paid chats — the installed CLI version did not write credits into the usage block. Check ~/.config/manicode/projects/<project>/chats/<chatId>/chat-messages.json and grep for "credits":.
  • Custom data dir not being scanned — confirm echo $CODEBUFF_DATA_DIR returns the expected path in the same shell that launched the OpenUsage daemon. The default roots are also scanned in addition to the override.
  • Claude Code — sibling local-file coding-agent provider
  • OpenCode — sibling local-file coding-agent provider