Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Configure RTK (command output), Caveman (prose), and stacked compression modes. Manage language packs, custom rules, and test prompt compression reducing tokens by 60–90%.
.claude/skills/diegosouzapw-omni-compression/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 141% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 40% | 0% |
| case-03 | ✗→✓ | ▲ Improved | -9% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 97% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 24% | 0% |
<!-- generated by src/lib/agentSkills/generator.ts; manual edits will be overwritten -->
Configure RTK (command output), Caveman (prose), and stacked compression modes. Manage language packs, custom rules, and test prompt compression reducing tokens by 60–90%.
All requests require a valid Bearer token or session cookie. Obtain a token via POST /api/auth/login or configure REQUIRE_API_KEY=false for local development.
Preview compression for a message payload
bashcurl -X POST https://localhost:20128/api/compression/preview \ -H "Authorization: Bearer $OMNIROUTE_TOKEN" \ -H "Content-Type: application/json" \ -d '{}'
List Caveman compression language packs
bashcurl https://localhost:20128/api/compression/language-packs \ -H "Authorization: Bearer $OMNIROUTE_TOKEN"
List Caveman compression rule metadata
bashcurl https://localhost:20128/api/compression/rules \ -H "Authorization: Bearer $OMNIROUTE_TOKEN"
POST compression › compare
bashcurl -X POST https://localhost:20128/api/compression/compare \ -H "Authorization: Bearer $OMNIROUTE_TOKEN" \ -H "Content-Type: application/json" \ -d '{}'
POST compression › compare › verify
bashcurl -X POST https://localhost:20128/api/compression/compare/verify \ -H "Authorization: Bearer $OMNIROUTE_TOKEN" \ -H "Content-Type: application/json" \ -d '{}'
GET compression › engines
bashcurl https://localhost:20128/api/compression/engines \ -H "Authorization: Bearer $OMNIROUTE_TOKEN"
POST compression › retrieve
bashcurl -X POST https://localhost:20128/api/compression/retrieve \ -H "Authorization: Bearer $OMNIROUTE_TOKEN" \ -H "Content-Type: application/json" \ -d '{}'
See the full OpenAPI specification at GET /api/openapi/spec or docs/openapi.yaml for detailed request/response schemas.
<!-- skill:custom-start --> <!-- Migrated from skills/omniroute-compression/SKILL.md (preserved curated content) -->
Requires OMNIROUTE_URL and OMNIROUTE_KEY. See entry-point SKILL for setup.
OmniRoute compresses token payloads before forwarding to providers. No code changes required — set it once, it applies to all requests transparently.
| Engine | Best for | Typical savings | | ------------------------- | ------------------------------------ | --------------- | | RTK | Terminal / build / test / git output | 60–90% | | Caveman | Human prose, chat history | 46% input | | Stacked (rtk → caveman) | Mixed coding sessions | 78–95% | | MCP accessibility filter | Browser/accessibility tool results | 60–80% |
bashcurl $OMNIROUTE_URL/api/settings/compression \ -H "Authorization: Bearer $OMNIROUTE_KEY"
bashcurl -X PUT $OMNIROUTE_URL/api/settings/compression \ -H "Authorization: Bearer $OMNIROUTE_KEY" \ -H "Content-Type: application/json" \ -d '{ "mode": "rtk", "enabled": true }'
bashcurl -X PUT $OMNIROUTE_URL/api/settings/compression \ -H "Authorization: Bearer $OMNIROUTE_KEY" \ -H "Content-Type: application/json" \ -d '{ "mode": "stacked", "enabled": true, "stackedPipeline": ["rtk", "caveman"] }'
bashcurl -X PUT $OMNIROUTE_URL/api/settings/compression \ -H "Authorization: Bearer $OMNIROUTE_KEY" \ -H "Content-Type: application/json" \ -d '{ "mode": "standard", "enabled": true }'
Caveman intensities: lite (safe), standard (balanced), aggressive (long sessions), ultra (context recovery).
bashcurl -X POST $OMNIROUTE_URL/api/compression/preview \ -H "Authorization: Bearer $OMNIROUTE_KEY" \ -H "Content-Type: application/json" \ -d '{ "mode": "rtk", "text": "$ npm test\n> jest\n\nPASS src/a.test.ts (2.1s)\nPASS src/b.test.ts (1.8s)\n..." }'
Response includes compressed, original_length, compressed_length, savings_pct.
When OmniRoute is used with browser/Playwright MCP tools, it automatically compresses verbose accessibility-tree tool results. Enabled by default; configure thresholds:
bashcurl -X PUT $OMNIROUTE_URL/api/settings/compression \ -H "Authorization: Bearer $OMNIROUTE_KEY" \ -H "Content-Type: application/json" \ -d '{ "mcpAccessibility": { "enabled": true, "collapseThreshold": 30, "maxTextChars": 50000 } }'
collapseThreshold: collapse sibling lines when ≥ N repeats (default 30). maxTextChars: hard truncate after N chars with navigation hint (default 50000).
Caveman supports language-aware rules for pt-BR, es, de, fr, ja:
bashcurl -X PUT $OMNIROUTE_URL/api/settings/compression \ -H "Authorization: Bearer $OMNIROUTE_KEY" \ -H "Content-Type: application/json" \ -d '{ "mode": "standard", "cavemanConfig": { "language": "pt-BR", "autoDetectLanguage": true } }'
omniroute_compression_status → current settings + savings analytics
omniroute_compression_configure → update mode/threshold/language
omniroute_set_compression_engine → switch engine at runtimebashcurl -X PUT $OMNIROUTE_URL/api/settings/compression \ -H "Authorization: Bearer $OMNIROUTE_KEY" \ -d '{ "enabled": false }'
400 invalid mode → use off, lite, standard, aggressive, ultra, rtk, or stacked400 invalid stackedPipeline → array must contain valid engine ids (rtk, caveman)<!-- skill:custom-end -->
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 11,626 | 10,185 | -12% | 1 | 1 | 0% | 1,095 | 2,636 | +141% | 0 | 0 | — |
case-02 | fail→pass | 16,453 | 11,397 | -31% | 1 | 1 | 0% | 2,160 | 3,034 | +40% | 0 | 0 | — |
case-03 | fail→pass | 21,540 | 9,912 | -54% | 1 | 1 | 0% | 2,878 | 2,621 | -9% | 0 | 0 | — |
case-04 | pass→pass | 18,271 | 15,061 | -18% | 1 | 1 | 0% | 2,503 | 3,800 | +52% | 0 | 0 | — |
case-05 | pass→pass | 21,177 | 13,411 | -37% | 1 | 1 | 0% | 2,478 | 3,181 | +28% | 0 | 0 | — |
case-06 | pass→pass | 15,104 | 3,007 | -80% | 1 | 1 | 0% | 1,661 | 2,219 | +34% | 0 | 0 | — |
case-07 | fail→pass | 13,871 | 9,749 | -30% | 1 | 1 | 0% | 1,287 | 2,539 | +97% | 0 | 0 | — |
case-08 | fail→pass | 18,561 | 9,791 | -47% | 1 | 1 | 0% | 1,995 | 2,471 | +24% | 0 | 0 | — |
case-09 | fail→pass | 11,960 | 7,524 | -37% | 1 | 1 | 0% | 1,106 | 2,266 | +105% | 0 | 0 | — |
case-10 | fail→pass | 14,239 | 10,235 | -28% | 1 | 1 | 0% | 2,187 | 2,807 | +28% | 0 | 0 | — |
case-11 | fail→pass | 16,764 | 6,858 | -59% | 1 | 1 | 0% | 1,890 | 2,079 | +10% | 0 | 0 | — |
case-16 | fail→pass | 16,462 | 7,285 | -56% | 1 | 1 | 0% | 1,706 | 2,234 | +31% | 0 | 0 | — |
case-12 | fail→pass | 14,572 | 2,521 | -83% | 1 | 1 | 0% | 1,371 | 2,268 | +65% | 0 | 0 | — |
case-13 | fail→pass | 11,121 | 7,483 | -33% | 1 | 1 | 0% | 974 | 2,289 | +135% | 0 | 0 | — |
case-14 | fail→pass | 16,162 | 7,577 | -53% | 1 | 1 | 0% | 1,551 | 2,270 | +46% | 0 | 0 | — |
case-15 | fail→pass | 11,931 | 7,212 | -40% | 1 | 1 | 0% | 1,744 | 2,161 | +24% | 0 | 0 | — |
case-17 | fail→pass | 18,896 | 7,601 | -60% | 1 | 1 | 0% | 2,128 | 2,263 | +6% | 0 | 0 | — |
case-18 | fail→pass | 14,338 | 7,537 | -47% | 1 | 1 | 0% | 1,723 | 2,253 | +31% | 0 | 0 | — |
case-19 | fail→pass | 15,309 | 6,979 | -54% | 1 | 1 | 0% | 1,422 | 2,183 | +54% | 0 | 0 | — |
case-20 | fail→pass | 20,656 | 5,270 | -74% | 1 | 1 | 0% | 2,679 | 2,816 | +5% | 0 | 0 | — |
case-21 | fail→pass | 16,932 | 8,231 | -51% | 1 | 1 | 0% | 1,783 | 2,320 | +30% | 0 | 0 | — |
case-22 | fail→pass | 13,542 | 7,060 | -48% | 1 | 1 | 0% | 1,213 | 2,172 | +79% | 0 | 0 | — |
case-23 | fail→pass | 9,571 | 9,297 | -3% | 1 | 1 | 0% | 1,325 | 2,247 | +70% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted. The headline lift of +87 percentage points is the difference between those two pass rates over the 23 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
| Model | Method | Date | Lift |
|---|---|---|---|
| gemini-3.6-flash | verified | 8/24/2026 | +73% |
| gemini-3.6-flash | verified | 8/9/2026 | +73% |
Other measured skills in the registry, with their headline benchmark lift.