Changelog

All revisions published to the LM Studio Hub as rootmonster/context-compressor.

v0.7.0 — first published at Hub revision 21

Universal thinking-model support: reasoning markers from all major families are stripped (<think>, [THINK], ◁think▷, <reasoning>, gpt-oss Harmony channels), including unterminated blocks; the engine's parsed non-reasoning content is preferred when available, so a summarizer reply that truncated mid-reasoning can no longer be cached as a chunk summary (previously it was cached permanently). /no_think now targets qwen3/QwQ/SmolLM3 (/no_think) and GLM (/nothink) instead of any identifier containing "qwen" — qwen2.5 no longer receives a stray suffix. Assistant reasoning in history is stripped before summarization.
No-system-role models (Gemma-style): a one-time per-model probe detects templates that reject system messages; the compressed view and all internal summarizer prompts then fold the system text into the first user message instead of crashing. Template render failures now surface as a warning instead of silently disabling exact prompt-fit verification.
Small context windows: keep-recent/chunk/reserved-output/ summarizer-max are auto-scaled to the model's context length (with a one-time notice), so compaction actually fires on 4k-8k models instead of never triggering.
Silent failures surfaced: unknown context length now warns (auto-compaction can't run without it); models not trained for tool use get tools disabled WITH a visible notice (previously silent, and session tools could crash strict templates); hitting the 24-pass safety cap now says so.
Summarizer prompts: agentic-workspace instructions are included only when the chunk actually contains agentic-workspace/v1 tool results. A summarizerPromptMode setting (auto/full/compact) lets the shorter compact variant be forced by hand; bench evidence (2026-08-24, two models) showed the full 10-section prompt beating compact on recall at both 3B and 18B, so auto now always uses full — compact remains available as a manual override and as the automatic rescue prompt when a model's reasoning runs away mid-summary.
Honest token accounting: estimates now count message text rather than internal JSON serialization (so /usage numbers and savings percentages are accurate), and CJK/base64-heavy tool results are truncated by a character-class-aware estimate instead of a flat 4-chars-per-token rule (previously up to ~4x under-truncation).
Fixes: canceled-compaction message reported one fewer saved chunk than reality; reasoning-only replies were rendered twice when promoted to visible text; malformed tool calls no longer return undefined to the engine; per-model state is now bounded and no longer leaks between chats (tool-schema token counts, one-time warnings).
Cache schema 2 → 3: previously cached summaries regenerate on the next compaction (one-time cost).
The source suite grew from 150 to 272 test cases.
Live-validated: bench matrix run 2026-08-24 against 8 live model families in LM Studio, zero plugin failures. This run drove three live fixes: the truncation-headroom retry, the structural-digest fallback for compaction that must never hard-fail, and the compact-prompt rescue for a model whose reasoning runs away mid-summary.

Revision 21 (v0.6.0)

Revision 19 (v0.5.0)

Attachment memory: before compressing past an attachment, its content is preserved — documents become fact-sheets via parseDocument (exact values verbatim), images get described by vision models. Extracted once per file identifier, cached at ~/.lm-context-compressor/attachments.json (LRU 500, atomic writes); failures degrade to an explicit "content not retained" marker and are never persisted. Toggle: "Attachment memory" (default on). /usage reports preserved files.
Benchmark suite: bench/ with deterministic fixtures (coding agent, research, contradictory updates, tool spam), seeded-fact recall after compaction and after forced consolidation, reduction, latency, structural and cache-determinism checks. npm run bench:static (model-free) and (against the local server); results committed in bench/results/.

Revision 18 (v0.4.0)

Bounded summary growth: when accumulated chunk summaries exceed the new summary budget (default 12000 tokens), the oldest are consolidated into one denser summary (rolling L1) with provenance and verbatim identifiers preserved. Summaries can no longer swallow the context — the field failure where a 528k chat plateaued at 63% prompt usage is fixed; /compress heals an already-stuck chat immediately.
Floor reachability: the compact-down-to floor is clamped at runtime to what the summary budget + kept-verbatim + system prompt actually allow, with a one-time warning when the configured floor is unreachable.
The exact-fit safety loop tries consolidation before chunk-compaction (compacting more chunks ADDS summary tokens when summaries dominate).
Transient engine failures ("terminated", connection resets) get one automatic retry, gated to passes with provably zero tool executions and zero visible output — exactly-once side effects preserved. Chunk summarization and consolidation retry unconditionally (no side effects).

Revision 17 (v0.3.0)

Tool definitions are now measured and included in context budgeting; a "Tool definitions" line appears in /usage and the usage tool.
Hard stop with an actionable message when the prompt (incl. tool schemas) cannot fit the context at all — no more cryptic engine 400s.
Tool-session initialization failure is surfaced visibly instead of silently proceeding without external tools.
Aggregate tool-output budget per reply (default 20000 tokens) as a second line of defense after the per-result cap.
Side-effect boundary tests added to the testing protocol; version bumped to 0.3.0.

Revision 16

Docs overhaul: feature comparison, troubleshooting guide, changelog.
License change: MIT → Context Compressor Community License 1.0 (free for personal + internal professional use; commercial embedding requires a separate license). Revisions 1–15 remain MIT.

Revision 15

Mid-task compaction: agentic turns run in capped tool-round passes (default 6); between passes the grown context is measured and compacted, then the task continues. Marathon tool loops can no longer overflow the context window.
Reasoning promoted into replies is sanitized of internal LM Studio markers.

Revision 14

Aborted or crashed /compact runs can no longer render an empty reply: the report block is created before the work starts and upgraded in place.

Revision 13

Chunk summarization skips hidden reasoning on qwen-family models (/no_think), cutting compaction wall time substantially.

Revision 12

Chunk-granular resume made visible ((N cached chunks reused), cancel message notes saved progress); empty assistant messages from canceled replies are filtered from the model's view.

Revision 11

Compact-to-floor hysteresis (default 55% of context).
contextOverflowPolicy: stopAtLimit everywhere — the engine's own truncation could drop the only user message and crash strict templates.
Exact templated-prompt verification with reserved output tokens and up to three safety compaction rounds; meter and /usage report the rendered count.
SDK 1.4.0 → 1.5.0; tool results correlated by ToolCallContext.callId.
Summarizer hardening: excerpts treated as untrusted data with provenance; repeated file reads deduplicated to latest version + change note.

Revision 10

Cache persistence serialized and coalesced (concurrent-write race found in external review, reproduced with an injected-latency test, fixed).

Revisions 1–9

Initial release and hardening: prediction-loop-handler architecture, chunked byte-stable summaries with a content-addressed persistent cache, /compress·/compact·/usage commands, get_context_usage tool, task-aware summary schema, tool passthrough with result capping, live progress bar and prompt meter, template compatibility fixes (single system message, synthetic user message), tool-error feedback, empty-reply fallbacks, crash-proof top-level guard.

v0.7.0 — first published at Hub revision 21

Universal thinking-model support: reasoning markers from all major families are stripped (<think>, [THINK], ◁think▷, <reasoning>, gpt-oss Harmony channels), including unterminated blocks; the engine's parsed non-reasoning content is preferred when available, so a summarizer reply that truncated mid-reasoning can no longer be cached as a chunk summary (previously it was cached permanently). /no_think now targets qwen3/QwQ/SmolLM3 (/no_think) and GLM (/nothink) instead of any identifier containing "qwen" — qwen2.5 no longer receives a stray suffix. Assistant reasoning in history is stripped before summarization.

No-system-role models (Gemma-style): a one-time per-model probe detects templates that reject system messages; the compressed view and all internal summarizer prompts then fold the system text into the first user message instead of crashing. Template render failures now surface as a warning instead of silently disabling exact prompt-fit verification.

Small context windows: keep-recent/chunk/reserved-output/ summarizer-max are auto-scaled to the model's context length (with a one-time notice), so compaction actually fires on 4k-8k models instead of never triggering.

Silent failures surfaced: unknown context length now warns (auto-compaction can't run without it); models not trained for tool use get tools disabled WITH a visible notice (previously silent, and session tools could crash strict templates); hitting the 24-pass safety cap now says so.

Summarizer prompts: agentic-workspace instructions are included only when the chunk actually contains agentic-workspace/v1 tool results. A summarizerPromptMode setting (auto/full/compact) lets the shorter compact variant be forced by hand; bench evidence (2026-08-24, two models) showed the full 10-section prompt beating compact on recall at both 3B and 18B, so auto now always uses full — compact remains available as a manual override and as the automatic rescue prompt when a model's reasoning runs away mid-summary.

Honest token accounting: estimates now count message text rather than internal JSON serialization (so /usage numbers and savings percentages are accurate), and CJK/base64-heavy tool results are truncated by a character-class-aware estimate instead of a flat 4-chars-per-token rule (previously up to ~4x under-truncation).

Fixes: canceled-compaction message reported one fewer saved chunk than reality; reasoning-only replies were rendered twice when promoted to visible text; malformed tool calls no longer return undefined to the engine; per-model state is now bounded and no longer leaks between chats (tool-schema token counts, one-time warnings).

Cache schema 2 → 3: previously cached summaries regenerate on the next compaction (one-time cost).

The source suite grew from 150 to 272 test cases.

Live-validated: bench matrix run 2026-08-24 against 8 live model families in LM Studio, zero plugin failures. This run drove three live fixes: the truncation-headroom retry, the structural-digest fallback for compaction that must never hard-fail, and the compact-prompt rescue for a model whose reasoning runs away mid-summary.

Revision 19 (v0.5.0)

Attachment memory: before compressing past an attachment, its content is preserved — documents become fact-sheets via parseDocument (exact values verbatim), images get described by vision models. Extracted once per file identifier, cached at ~/.lm-context-compressor/attachments.json (LRU 500, atomic writes); failures degrade to an explicit "content not retained" marker and are never persisted. Toggle: "Attachment memory" (default on). /usage reports preserved files.

Benchmark suite: bench/ with deterministic fixtures (coding agent, research, contradictory updates, tool spam), seeded-fact recall after compaction and after forced consolidation, reduction, latency, structural and cache-determinism checks. npm run bench:static (model-free) and (against the local server); results committed in bench/results/.

Revision 18 (v0.4.0)

Bounded summary growth: when accumulated chunk summaries exceed the new summary budget (default 12000 tokens), the oldest are consolidated into one denser summary (rolling L1) with provenance and verbatim identifiers preserved. Summaries can no longer swallow the context — the field failure where a 528k chat plateaued at 63% prompt usage is fixed; /compress heals an already-stuck chat immediately.

Floor reachability: the compact-down-to floor is clamped at runtime to what the summary budget + kept-verbatim + system prompt actually allow, with a one-time warning when the configured floor is unreachable.

The exact-fit safety loop tries consolidation before chunk-compaction (compacting more chunks ADDS summary tokens when summaries dominate).

Transient engine failures ("terminated", connection resets) get one automatic retry, gated to passes with provably zero tool executions and zero visible output — exactly-once side effects preserved. Chunk summarization and consolidation retry unconditionally (no side effects).

Revision 17 (v0.3.0)

Tool definitions are now measured and included in context budgeting; a "Tool definitions" line appears in /usage and the usage tool.

Hard stop with an actionable message when the prompt (incl. tool schemas) cannot fit the context at all — no more cryptic engine 400s.

Tool-session initialization failure is surfaced visibly instead of silently proceeding without external tools.

Aggregate tool-output budget per reply (default 20000 tokens) as a second line of defense after the per-result cap.

Side-effect boundary tests added to the testing protocol; version bumped to 0.3.0.

Revision 15

Mid-task compaction: agentic turns run in capped tool-round passes (default 6); between passes the grown context is measured and compacted, then the task continues. Marathon tool loops can no longer overflow the context window.

Reasoning promoted into replies is sanitized of internal LM Studio markers.

Revision 11

Compact-to-floor hysteresis (default 55% of context).

contextOverflowPolicy: stopAtLimit everywhere — the engine's own truncation could drop the only user message and crash strict templates.

Exact templated-prompt verification with reserved output tokens and up to three safety compaction rounds; meter and /usage report the rendered count.

SDK 1.4.0 → 1.5.0; tool results correlated by ToolCallContext.callId.

Summarizer hardening: excerpts treated as untrusted data with provenance; repeated file reads deduplicated to latest version + change note.

Revisions 1–9

Initial release and hardening: prediction-loop-handler architecture, chunked byte-stable summaries with a content-addressed persistent cache, /compress·/compact·/usage commands, get_context_usage tool, task-aware summary schema, tool passthrough with result capping, live progress bar and prompt meter, template compatibility fixes (single system message, synthetic user message), tool-error feedback, empty-reply fallbacks, crash-proof top-level guard.