codeburn/docs/providers/mistral-vibe.md

Mistral Vibe

Mistral Vibe CLI.

• Source: src/providers/mistral-vibe.ts
• Loading: eager (src/providers/index.ts)
• Test: tests/providers/mistral-vibe.test.ts

Where it reads from

$VIBE_HOME/logs/session/ when VIBE_HOME is set, otherwise ~/.vibe/logs/session/.

Storage format

Legacy Vibe 2.x stores each session as a directory:

• meta.json contains session metadata, cumulative token totals, active model config, model prices, timestamps, working directory, and available tools.
• messages.jsonl contains non-system messages and assistant tool_calls.

Subagent traces are stored under a parent session's agents/ folder with the same meta.json / messages.jsonl shape, so CodeBurn scans those one level down as separate sessions.

Vibe 2.25's Unified Harness (still the layout in 2.26) stores every session under unified/<session_id>/. A subagent is its own sibling session there, linked by parent_session_id; its usage is not included in the parent's. Its meta.json has no stats, and there is no messages.jsonl. CURRENT selects a committed generation containing projection-state.json, runtime-state.json, and a manifest. The projection's history can be stored in content-addressed chunks/ files.

CodeBurn reads snapshot.session.tokenUsage and the cumulative usage envelopes in journal/*.jsonl. It supports both projection_advanced and projection_delta records. These are recovery journals: Vibe retains only two segments after checkpointing. Summing action_result records alone loses older usage, while summing every nested usage object double-counts copies of a call. CodeBurn takes deltas between cumulative envelopes and recovers the older prefix from the snapshot. Each later envelope step is one call with its exact usage (the same numbers as that call's action_result.result.usage). The rotated prefix is split evenly into as many calls as runtime-state.json lists succeeded completion actions, across its recorded turns. Call counts and session totals stay exact; per-call values inside that prefix are an even split.

Caching

Legacy Vibe local logs do not expose cache-read/cache-write token fields, so CodeBurn reports cache token counts as 0. When meta.json.stats.session_cost is present, CodeBurn uses that session total instead of re-estimating from prompt/completion token prices because it is the best cache-aware cost signal available in the local log shape.

Unified inputTokens includes cachedInputTokens. CodeBurn subtracts cached input from ordinary input and records it as cache-read usage, then estimates cost using the pricing catalog. Unified sessions persist no cost of their own (Vibe's session_cost is computed in memory only), so these costs are marked estimated. It does not treat contextUsage (the latest context window) as cumulative spend.

The model is not recorded per call. runtime-state.json session_metadata.active_model is only a pin and is null for an unpinned session, which runs the configured default. CodeBurn then reads the top-level active_model from $VIBE_HOME/config.toml at parse time, falling back to mistral-medium-3.5 (Vibe's built-in default, priced at Vibe's own rates of $1.5/M input, $0.15/M cached, $7.5/M output). Changing config.toml later does not reprice sessions already cached, and mid-session model switches are not recoverable.

The source cache fingerprints CURRENT, metadata, and the bounded journal segments. Appending usage invalidates the cache even before Vibe publishes a new generation. Generation contents and chunks are immutable once published.

Deduplication

Legacy message IDs, or the session ID for an aggregate fallback. Unified usage uses the session ID, projection sequence, turn ID, and call index. Repeated journal sequences and cumulative totals are not counted again.

Quirks

• Usage is cumulative per session. Vibe does not write per-assistant-message token usage into messages.jsonl; token counts come from meta.json.stats.session_prompt_tokens and session_completion_tokens. CodeBurn splits assistant-message tools into their user turns for classification and distributes the cumulative token/cost totals across those assistant calls so session totals remain unchanged.
• Cost prefers Vibe's own session total. meta.json.stats.session_cost is used first. If it is missing, meta.json.stats.input_price_per_million and output_price_per_million are used with the active model config as a fallback. LiteLLM pricing is only used when Vibe provides no price data.
• Project names come from metadata. Discovery uses meta.json.environment.working_directory and falls back to the session directory name if that field is missing.
• Unified tool names carry their group (file_system.bash, ui.ask_user_question). CodeBurn drops the group before mapping, and a turn's tools are attached to its first call only.
• Tool calls come from messages. Assistant tool_calls[*].function.name is normalized to the standard CodeBurn names (bash to Bash, search_replace to Edit, etc.). Bash commands are extracted from function.arguments.command.

When fixing a bug here

1. Identify the harness first. Legacy sessions require meta.json and messages.jsonl; Unified sessions use CURRENT, generations, chunks, and journals.
2. Check legacy meta.json.stats or Unified projection session.tokenUsage, never contextUsage.
3. Test journal rotation and appends against a warm CodeBurn cache, as well as legacy compatibility. See tests/fixtures/mistral-vibe-unified.md for fixture provenance.