Less context to read, more room to work.

Remove formatting overhead from JSON. Uniform arrays can share their column names in TOON or a tabular representation instead of repeating keys on every row. Requests for JSON keep JSON.

Collapse repeated lines with repetition counts and reduce noisy build or progress output. Reversible repetition encoding can reconstruct the original; condensed progress views retain a recovery reference.

Condense test reports and stack traces around failures, assertions and relevant frames. Build summaries keep diagnostics in view instead of repeating successful compilation steps. Retrieve the full output when the summary is not enough.

Diff windowing shortens surrounding context. Source skeletons expose declarations and signatures before full function bodies. These are reading aids, not replacement files or patches; omitted code stays behind a recovery reference.

Strip bulky page or vector markup into a smaller structural or text view. The original remains available for exact attributes, coordinates or details that the compact view leaves out.

Specialist views condense notebook content and dependency lockfiles. Large generated assets and embedded base64 data can become short notices with recovery handles, instead of filling the agent context. These views are intentionally incomplete.

Remove ANSI decoration, repeated blank space and table padding. Specialist operations can alias repeated long identifiers or keep only the head and tail of oversized output, with the original available on demand.

When the task only asks for paths, a working directory, an HTTP status or a supported pass/fail verdict, deterministic extraction returns that signal instead of the whole command transcript.

An available encoder ranks chunks against the question and selects relevant lines from large outputs. Selected lines keep their original order. This is extractive selection, not a promise that every omitted line is irrelevant.

A handle references stored content. Retrieve the original, request a line range, grep within it or use supported structured queries. Token estimates and previews help the agent decide how much to read. Store searches only cover content actually available there.

The local specialist can receive a bounded pack of question-relevant excerpts, answer over a handle, select candidates or compare multiple supplied outputs. Source spans let the main agent check the answer without starting with every full document.

The Librarian exposes a compact set of common operation hints and a paginated catalog for the rest. Bots include an index of their owned skills; the agent requests a skill body when it needs it. The skill instructions themselves are not summarized.

When Token Saver and Graphify are enabled and the Graphify CLI is available, the host maintains a repository graph and gives the agent query-first instructions. Targeted graph queries can reduce broad file exploration. This is a real host integration; the similarly named Librarian catalog operation does not itself run the CLI.

Model-assisted operations condense command output, long text and watch updates. Compact wire representations and short language summaries serve their specific modes. Local inference uses the installed model; an explicitly configured compatible API uses that endpoint instead.

Extract or group test failures, CI jobs, compiler and linter diagnostics, coverage, snapshots, build and install logs. Specialist prompts also cover container errors, infrastructure plans, Kubernetes output, dependency conflicts, code signing, notarization, accessibility and vulnerability reports. They summarize evidence; they do not certify a passing build.

Ask for file roles, imports, symbols, JSON shapes, schema fields, tables, relevant API endpoints or database structures. Summarize HTTP/HAR and GraphQL output, query plans, benchmarks, crashes, profiler stacks, processes, debugger output, sanitizers, memory reports, system-call traces and binary inspection. The main agent can retrieve details before acting.

Prepare smaller views of git history, blame, conflicts, PR/MR discussions, tickets and acceptance criteria. Supplied chat, email and meeting transcripts can become decisions, questions or action lists. Other operations summarize subagent output, classify its outcome, merge log timelines, compare metrics, identify contradictions and select release-note ranges.

Candidate operations can suggest plans, review notes, test plans, searches, failure hypotheses, refactors, issue triage, SQL or release copy. They may draft commit messages, PR descriptions and translations where supported. The main agent must verify them. Suggested commands are not automatically executed, and the local model does not silently replace your chosen model.

Automatic selection uses the 1.7B MLX model on Macs with up to 16 GiB of RAM and E4B above 16 GiB. Settings also offer explicit model selection or an OpenAI-compatible endpoint. Deterministic routes do not require generation; encoder and model routes depend on available runtimes. External generation sends eligible content to the configured endpoint.

Identical compression requests can reuse a cached result. Cached embeddings avoid repeating local ranking work. This improves local efficiency; a cache hit does not by itself mean the paid provider charged fewer tokens.

A verified duplicate can point back to its stored original. Search and similarity notices help locate previous evidence without treating similar text as identical. Baseline comparisons focus on changed failures; scoped correction examples can help later specialist calls.

Compact response envelopes, on-demand catalog entries and removal of model thinking traces keep specialist replies focused. Stable local prompt construction and cache-drift checks help avoid needless local overhead; they do not rewrite a provider’s safety instructions or guarantee its prompt-cache pricing.

Admitted runs receive instructions for compact wire messages and repeated-span references. RCW1 is requested only when its full token count is at most 45% of equivalent terse text; otherwise the text stays as text. Literal values, negation, order and approvals must survive decoding. This eligibility rule is not a measured 55% reduction for every conversation.

A smaller tool result can also reduce context repeated in later turns. The app estimates that continuing benefit from permanent per-thread saved totals and run counts, with a floor derived from reported cache reads when larger. This is a counterfactual estimate, separate from compression-cache hits and the provider’s actual cache discount.

Managed wrappers intercept eligible noisy command output, and explicit distill pipes compress supplied text. Provider hooks apply only where output replacement is supported. Exact-output tools, skill bodies, permission surfaces and verbatim-required content are protected; support for one provider does not imply identical coverage for every provider. In the current implementation, Claude has a managed PostToolUse replacement hook; Codex and Grok rely on eligible shell wrappers and explicit pipes, not universal rewriting of native file reads. Commands such as cat, rg, sed and jq pass through unchanged unless the agent explicitly asks distill for an extract.

Large CLI input is split into bounded compression requests instead of exceeding a single request budget. MCP retrieval lets the agent recover stored output or ask for a smaller view. Chunking enables compression at larger sizes; it is not an extra savings multiplier.

The implementation also has gated paths for unchanged rereads, cross-turn deduplication, old result or argument masking, queued-intent coalescing and compact rollover/handoff context. They require the matching host integration, activation and recovery checks. They are not universal live-session rewriting. The macOS host activates the rollout gates for admitted Token Saver runs; each operation still checks its own eligibility and recovery support. Separate history-compaction and session-learning preferences are off by default: the former can compact persisted output with recovery, while the latter can write learned correction rules into managed instructions. Their presence in code does not mean every user has enabled them.

Catalog entries for image OCR/diffs, audio transcription, video frames, document or spreadsheet extraction, graph/LSP/AST queries, file-change notices, watches and codemods need host-side tools and input. An entry is not proof of an active integration: unsupported operations return unavailable. Text already extracted can use the compression techniques above.

Auto is a separate routing mode for agent work. When active, the savings calculation adds a token-equivalent estimate equal to 25% of measured tool-output savings. This accounting allowance is not a measured token reduction from every routed call and does not guarantee a cheaper invoice or identical quality.

Before serving a view that drops details, the supported recovery path stores and verifies the original. Handles are tied to their recovery scope. If recovery cannot be established, that lossy delivery is refused or the original is used. Recoverability is different from lossless compression.

Depending on the route, checks protect literal values and markers, verify evidence spans, reject unsupported summaries and use semantic/NLI checks when available. Gain thresholds reject reductions that are not worthwhile. Timeouts and circuit breakers let the workflow continue with a safer path. These checks reduce risk; they cannot guarantee identical reasoning quality.

Local mode processes compression on your Mac; a configured external API sends eligible content to that endpoint. Your coding provider still receives its own requests. Secret and injection checks constrain specialist operations.

Delivered replacements record the difference between token estimates for the original and compact output. The headline Estimated save additionally includes shorter-answer estimates, Graphify activity, multi-turn reuse and Auto when applicable. It compares against a hypothetical bare agent, not a provider invoice. Rejected candidates do not add measured savings.