For your second question: No LLM calls. Context Mode uses algorithmic processing — FTS5 indexing with BM25 ranking and Porter stemming. Raw output gets chunked and indexed in a SQLite database inside the sandbox, and only the relevant snippets matching your intent are returned to context. It's purely deterministic text processing, no model inference involved.
Really cool. A tangential task that seems to be coming up more and more is masking sensitive data in these calls for security and privacy. Is that something you considered as a feature?
The SQLite database is ephemeral — stored in the OS temp directory (/tmp/context-mode-{pid}.db) and scoped to the session process. Nothing persists after the session ends. For sensitive data masking specifically: right now the raw data never leaves the sandbox (it stays in the subprocess or the temp SQLite store), and only stdout summaries enter the conversation. But a dedicated redaction layer (regex-based PII stripping before indexing) is an interesting idea worth exploring. Would be a clean addition to the execute pipeline.
Yes — the database is tied to the MCP server process, so it's created fresh on each claude launch and lost when you exit; resuming a session starts a new process with a new empty database.
The BM25+FTS5 approach without LLM calls is the right call - deterministic, no added latency, no extra token spend on compression itself.
The tradeoff I want to understand better: how does it handle cases where the relevant signal is in the "low-ranked" 310 KB, but you just haven't formed the query that would surface it yet? The compression is necessarily lossy - is there a raw mode fallback for when the summarized context produces unexpected downstream results?
Also curious about the token count methodology - are you measuring Claude's tokenizer specifically, or a proxy?
On lossy compression and the "unsurfaced signal" problem:
Nothing is thrown away. The full output is indexed into a persistent SQLite FTS5 store — the 310 KB stays in the knowledge base, only the search results enter context. If the first query misses something, you (or the model) can call search(queries: ["different angle", "another term"]) as many times as needed against the same indexed data. The vocabulary of distinctive terms is returned with every intent-search result specifically to help form better follow-up queries.
The fallback chain: if intent-scoped search returns nothing, it splits the intent into individual words and ranks by match count. If that still misses, batch_execute has a three-tier fallback — source-scoped search → boosted search with section titles → global search across all indexed content.
There's no explicit "raw mode" toggle, but if you omit the intent parameter, execute returns the full stdout directly (smart-truncated at 60% head / 40% tail if it exceeds the buffer). So the escape hatch is: don't pass intent, get raw output.
On token counting:
It's a bytes/4 estimate using Buffer.byteLength() (UTF-8), not an actual tokenizer. Marked as "estimated (~)" in stats output. It's a rough proxy — Claude's tokenizer would give slightly different numbers — but directionally accurate for measuring relative savings. The percentage reduction (e.g., "98%") is measured in bytes, not tokens, comparing raw output size vs. what actually enters the conversation context.
Hey! Thank you for your comment! You can actually use an MCP on this basis, but I haven't tested it yet. I'll look into it as soon as possible. Your feedback is valuable.
And when you say only returns summaries, does this mean there is LLM model calls happening in the sandbox?
The SQLite database is ephemeral — stored in the OS temp directory (/tmp/context-mode-{pid}.db) and scoped to the session process. Nothing persists after the session ends. For sensitive data masking specifically: right now the raw data never leaves the sandbox (it stays in the subprocess or the temp SQLite store), and only stdout summaries enter the conversation. But a dedicated redaction layer (regex-based PII stripping before indexing) is an interesting idea worth exploring. Would be a clean addition to the execute pipeline.
Does that mean that if I exit claude code and then later resume the session, the database is already lost? When exactly does the session end?
The tradeoff I want to understand better: how does it handle cases where the relevant signal is in the "low-ranked" 310 KB, but you just haven't formed the query that would surface it yet? The compression is necessarily lossy - is there a raw mode fallback for when the summarized context produces unexpected downstream results?
Also curious about the token count methodology - are you measuring Claude's tokenizer specifically, or a proxy?
--
On lossy compression and the "unsurfaced signal" problem:
Nothing is thrown away. The full output is indexed into a persistent SQLite FTS5 store — the 310 KB stays in the knowledge base, only the search results enter context. If the first query misses something, you (or the model) can call search(queries: ["different angle", "another term"]) as many times as needed against the same indexed data. The vocabulary of distinctive terms is returned with every intent-search result specifically to help form better follow-up queries.
The fallback chain: if intent-scoped search returns nothing, it splits the intent into individual words and ranks by match count. If that still misses, batch_execute has a three-tier fallback — source-scoped search → boosted search with section titles → global search across all indexed content.
There's no explicit "raw mode" toggle, but if you omit the intent parameter, execute returns the full stdout directly (smart-truncated at 60% head / 40% tail if it exceeds the buffer). So the escape hatch is: don't pass intent, get raw output.
On token counting:
It's a bytes/4 estimate using Buffer.byteLength() (UTF-8), not an actual tokenizer. Marked as "estimated (~)" in stats output. It's a rough proxy — Claude's tokenizer would give slightly different numbers — but directionally accurate for measuring relative savings. The percentage reduction (e.g., "98%") is measured in bytes, not tokens, comparing raw output size vs. what actually enters the conversation context.
Codex CLI:
Or in ~/.codex/config.toml: opencode:In opencode.json:
We haven't tested yet — would love to hear if anyone tries it!