io.github.NORTHTEKDevs/lossless-context-mcp

Forensics: what did the agent see of this file, and when? Every content version the flight recorder observed (SHA-256 + git blob SHA-1, first/last seen, capture source, sessions), plus what else was in context around a focus moment. The evidence-backed answer to "why did the agent do that?" — query history instead of trusting self-report.

Issue an HMAC-SHA256-signed context receipt: every file/view the model was shown via this ledger, the SHA-256 of each content version, git binding (blob SHA-1 per version, repo HEAD at issue time), delivery kinds, token totals, and an explicit coverage statement. With include_sweep, also attests native-tool reads captured by the transcript-sweep hooks. The auditable answer to "what did the AI see when it did this?". Signs with LOSSLESS_RECEIPT_KEY or the shared trust key file, so it verifies with the same key as trust-mcp receipts. Verify later with verify_context_receipt.

Where did this session’s file-read tokens go? Totals, per-repo breakdown, heaviest files, dedup savings, and a USD estimate (LOSSLESS_PRICE_PER_MTOK, default $3/MTok input). Counted with a real tokenizer on exactly what this server sent.

The air-traffic radar for concurrent agents on this machine: which agent processes are active (from the coordination plane the guard hooks maintain), what they have been editing, which files have cross-session activity, and where edit conflicts are brewing. Advisory visibility only — sessions without the hooks installed are invisible.

Build a fan-out context pack: the stable, hot files (ranked across sessions from the flight-recorder history) rendered as one deterministic block to embed VERBATIM in a custom agent-type's SYSTEM PROMPT. Siblings in a fan-out then hit the prompt cache on that shared prefix instead of each re-reading the same files (measured on a real review fan-out: 46.55% cheaper; per-task injection measured WORSE — see README). Files matching secret patterns are never packed.

Return a cheap structural outline of a file: declaration lines (functions, classes, types, methods) with line numbers, bodies elided. Use this to navigate a large/unknown file before reading specific parts with read_file.

Read a text file through the context ledger. The FIRST read of a file/view (or the first after a context compaction) returns full content. A later read of an UNCHANGED view returns a short "reuse what you have" marker. A later read of a CHANGED view returns a unified DIFF to apply to the copy you already have. Optionally read just one symbol (function/class by name) or a line range instead of the whole file. All dedup is lossless: it only diffs/withholds content it can prove you still have. Every read is metered (see context_stats) and attested in the signed context receipt. Pass force_full:true for the whole content regardless.

Read a working set of text files in one call, each through the same lossless ledger as read_file (full on first contact, unchanged-marker or diff on re-reads). One call for N files instead of N calls. Per-file errors are reported inline without failing the batch.

Re-emit the working set after a compaction (or /clear) destroyed it. Serves CURRENT disk content of the requested files — by default the top of the pre-compaction manifest the sweep hook wrote — through the lossless ledger, annotating any file that changed since the model last saw it. Budget-capped so a restore cannot blow the fresh context; files over budget are listed, not silently dropped.

Verify a context receipt + signature pair against the local receipt key (timing-safe, canonicalized so JSON field order does not matter).

What has this session been working on? A heat-ranked table of the files the flight recorder knows about — from this server's own reads plus any transcript sweeps the hooks performed (last 24h) — with sizes, touch counts, and staleness vs current disk state. Use after a compaction to see what is recoverable, or anytime for orientation.