soltrinox.compressor is a Cursor-only extension that supplies bounded forward
context to Cursor Agent Chat from locally maintained compressed session state.
Local state is projected into a bounded pack ordered as HOT_SET, typed graph
lines, and query-ranked chunks, then injected into Agent Chat.
Status: v0.1.0 ready for human Open VSX publish. See docs/PUBLISHING.md.
84% fewer forwarded · 139k vs 862k · ≤1,024 / turn, median 783
On a 199-prompt corpus (14–15 Aug 2026), packed inject forwarded about
one sixth the estimated tokens of full-corpus replay
(84% fewer). Memory-inject path (chars/4),
not a Cursor billing export.
On this inject-path comparison, packed volume is about
18¢ on the replay dollar
(~6× as far for the same inject budget).
Estimated forwarded tokens vs full-corpus replay — not a Cursor billing export.
How this was measured
- In Cursor, install comPREssOR (
soltrinox.compressor) from Open VSX or a release VSIX. - Reload Cursor. On first allowed activation, comPREssOR provisions its runtime, writes the env file, installs the hook shim, merges hook entries, and deploys its user rule and skill.
Development sideload:
cd extension && npm ci && npm run package
# Install the produced VSIX in Cursor.Before each turn, some systems re-send prior prompt text into the model context. comPREssOR instead maintains local state and injects a budgeted pack. On a 199-prompt corpus, that pack was about one sixth the estimated tokens of replaying the ingested corpus (~18¢ on the replay dollar). Separately, a vocabulary-bag baseline was smaller than the pack but hit zero fixture recall — so “most compressed” is the wrong leaderboard.
| Approach | What it puts in the prompt | Measured here? | Size / utility note |
|---|---|---|---|
Paste / replay prior prompts into additional_context |
Growing corpus every turn | Yes — inject corpus | $1.00 replay dollar; packed ≈ 18¢ (~6× as far) |
| Last-N truncation | Tail of the thread only | No | Cheap; can drop early decisions |
| One-shot summary | Prose digest | No | Can drop paths/open items |
| Vocabulary / bag compression | Tiny word list | Yes — SDK probe | 27 tokens, recall 0.00 — size wins, utility fails |
| Codebase RAG | Repo chunks | No / N/A | Different problem; comPREssOR is not RAG |
| comPREssOR pack | HOT_SET → typed → ranked ≤1,024 |
Yes — both probes | Inject: ~1/6 replay volume; SDK: 184 tokens, recall 0.33, ~34% fewer billed on that probe |
18¢ / 6× is a ratio illustration of inject-path volume vs replaying this
corpus as additional_context, not a Cursor invoice. Net spend drops only if
the gist replaces a paste or post-compact replay. Native history is still sent.
Full card walk + two-probe table → docs/PERFORMANCE.md.
Long Agent Chat sessions often lose earlier decisions or require expensive raw
transcript replay. comPREssOR keeps local compressed state and injects a bounded
text payload ordered as HOT_SET, typed graph lines, and query-ranked chunks.
The observable outcome is continuity for paths, open items, decisions, and selected spans without forwarding the full transcript on every turn.
Scope boundary: comPREssOR is a context-assembly mechanism. It is not codebase RAG, not a model, not hidden-state transport, and not a guarantee of better answers. Hooks add a gist; they do not strip Cursor's native chat history. Net spend drops only if the gist replaces a paste or post-compact replay. For the inject-corpus walk-through, read docs/PERFORMANCE.md. For theory and the separate SDK probe, read docs/WHY.md.
- Cursor desktop only. Unsupported hosts may install a VSIX, but activation refuses side effects and writes nothing under the Cursor data directory.
- Python 3.11+ is required. On first allowed activation, the extension provisions a private venv and installs the bundled engine wheel into it.
- Hooks are fail-open. If the shim, venv, state directory, or engine is unavailable, the Agent Chat turn proceeds without injected context.
- The hook path does not need a Cursor API key and the extension does not write one into managed configuration.
- Local state defaults to
$HOME/.cursor/context-graphs/. - Settings are projected into
$HOME/.cursor/chat-compressor.env, preserving unmanaged lines. - User hooks install
$HOME/.cursor/hooks/chat-compressor.shand merge four events into$HOME/.cursor/hooks.json.
Cursor-only behavior and the sideload boundary are documented in docs/COMPATIBILITY.md. System behavior and settings are documented in docs/SYSTEM.md.
Keep related work in one Agent conversation when continuity matters. Write durable facts as explicit paths, open items, decisions, and headings. Mark completed work directly so old open items can be superseded.
Tune chatCompressor.forwardBudget when recall needs change. Prefer packed
memory over pasting a full transcript for continuity, and attach or cite source
files directly when exact wording is required. Use project hooks only when a
repository should explicitly carry hook configuration.
The hero figures above are the 199-prompt inject corpus (chars/4), not billed
cost. A separate lab/live SDK probe compared raw replay, a legacy vocabulary-bag
replica, and the current comPREssOR pack. The inbound baseline was about
78,876 chars / 19,719 estimated tokens.
| Arm | Final-turn estimated forward tokens | entity_recall |
Live billed total |
|---|---|---|---|
| Raw replay | 19938 | 1.00 | 31971 |
| Legacy vocabulary bag | 27 | 0.00 | 22352 |
| comPREssOR pack | 184 | 0.33 | 21050 |
entity_recall is a fixture term-hit proxy, not answer correctness. Live billed
usage includes the SDK envelope, so billed input is not the same as gist-only
payload size. On this probe, the comPREssOR pack used about 34% fewer billed
tokens than raw replay. Do not merge that 34% billed result with the 84%
inject-path figure. Details: docs/WHY.md.
- Cursor desktop. See docs/COMPATIBILITY.md.
- Python 3.11 or newer, discoverable on the machine or configured with
chatCompressor.pythonPath. - Network access on first activation when Python wheels need to be installed.
- docs/PERFORMANCE.md: 199-prompt inject corpus, card sequence, locked numerals, and the two-probe table.
- docs/WHY.md: why bounded forward context can beat raw replay, plus the lab/live SDK probe.
- docs/SYSTEM.md: architecture, lifecycle, settings, and use patterns.
- docs/ARCHITECTURE.md: documentation index.
- docs/HOOK_CONTRACT.md: hook events and fail-open defaults.
- docs/PUBLISHING.md: Open VSX publishing checklist.
Apache-2.0. See LICENSE.
author :: Rosario
roles :: developer · architect · mathematician
ENI6MA.com— authentication product surface. Mechanism: mint a one-shot proof bound to a single request (apps, agents, paper). Outcome: the verifier observes allow/deny for that action; a spent proof does not replay as standing authority. Scope: Public / Cloud / Sovereign stacks under the published reference architecture and claim model — not a reusable password/token vault.RosarioCyber.com— Rosario Cybernetics research lab. Focus: cybersecurity, cryptography, and AI-safety research that feeds ENI6MA (password-free auth demos, audit-trail posture, seminar/colloquium material). Contact path for partnerships/licensing sits with the lab.
© 2026 Rosario. All rights reserved.
Source copyright: Rosario (developer, architect, mathematician).
Licensed under Apache-2.0 for redistribution terms — see LICENSE.
Pack local state. Forward a bounded gist. Keep the rest on disk.