@sliday/tamp 0.2.14 → 0.3.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +20 -3
- package/bin/tamp.js +4 -1
- package/compress.js +110 -1
- package/config.js +1 -1
- package/package.json +5 -3
package/README.md
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
# Tamp
|
|
2
2
|
|
|
3
|
-
**Token compression proxy for coding agents.**
|
|
3
|
+
**Token compression proxy for coding agents.** 50.5% fewer input tokens, zero code changes. Works with Claude Code, Aider, Cursor, Cline, Windsurf, and any OpenAI-compatible agent.
|
|
4
4
|
|
|
5
5
|
```
|
|
6
6
|
npx @sliday/tamp
|
|
@@ -232,9 +232,26 @@ test/fixtures/ Sample API payloads
|
|
|
232
232
|
|
|
233
233
|
Claude Code sends the full conversation history on every API call. As a session progresses, tool results accumulate — file contents, directory listings, command outputs — all re-sent as input tokens on each request.
|
|
234
234
|
|
|
235
|
-
At $3/million input tokens (Sonnet 4), a 200-request session consuming 3M input tokens costs $9.
|
|
235
|
+
At $3/million input tokens (Sonnet 4.6), a 200-request session consuming 3M input tokens costs $9. With 32% compression on 60% compressible traffic, that's ~63K tokens and $0.19 saved per session. For Opus 4.6 ($5/MTok), savings are $0.32/session.
|
|
236
236
|
|
|
237
|
-
|
|
237
|
+
### Claude Max subscribers
|
|
238
|
+
|
|
239
|
+
Max plans have fixed token budgets. With 32% fewer input tokens per request, you get **47% more requests** from the same budget:
|
|
240
|
+
|
|
241
|
+
| Plan | Without Tamp | With Tamp |
|
|
242
|
+
|------|-------------|-----------|
|
|
243
|
+
| Max 5× ($100/mo) | 5× Pro | **7.4×** Pro |
|
|
244
|
+
| Max 20× ($200/mo) | 20× Pro | **29.4×** Pro |
|
|
245
|
+
|
|
246
|
+
### API cost savings
|
|
247
|
+
|
|
248
|
+
| Model | Per dev/month | 10-person team/year |
|
|
249
|
+
|-------|-------------|-------------------|
|
|
250
|
+
| Sonnet 4.6 ($3/MTok) | $28 | $3,400 |
|
|
251
|
+
| Opus 4.6 ($5/MTok) | $48 | $5,760 |
|
|
252
|
+
| Opus 4.6 extended ($10/MTok) | $96 | $11,520 |
|
|
253
|
+
|
|
254
|
+
Based on 5 sessions/day, 200 requests/session, 60% compressible traffic.
|
|
238
255
|
|
|
239
256
|
## License
|
|
240
257
|
|
package/bin/tamp.js
CHANGED
|
@@ -31,6 +31,9 @@ const STAGE_INFO = {
|
|
|
31
31
|
'strip-lines':'Remove line-number prefixes from Read output',
|
|
32
32
|
whitespace: 'Collapse blank lines, trim trailing spaces',
|
|
33
33
|
llmlingua: 'Neural text compression via LLMLingua-2',
|
|
34
|
+
dedup: 'Replace duplicate tool_results with references',
|
|
35
|
+
diff: 'Replace similar re-reads with unified diffs',
|
|
36
|
+
prune: 'Strip lockfile hashes, registry URLs, npm metadata',
|
|
34
37
|
}
|
|
35
38
|
|
|
36
39
|
// --- Determine stages ---
|
|
@@ -99,7 +102,7 @@ async function startSidecar() {
|
|
|
99
102
|
if (hasCommand('uv')) {
|
|
100
103
|
try {
|
|
101
104
|
const proc = spawn('uv', [
|
|
102
|
-
'run', '--with', 'fastapi,uvicorn,llmlingua',
|
|
105
|
+
'run', '--with', 'fastapi', '--with', 'uvicorn', '--with', 'llmlingua', '--with', 'mlx',
|
|
103
106
|
'uvicorn', 'server:app', '--host', '127.0.0.1', '--port', String(sidecarPort),
|
|
104
107
|
], { cwd: sidecarDir, stdio: ['ignore', 'pipe', 'pipe'] })
|
|
105
108
|
|
package/compress.js
CHANGED
|
@@ -1,5 +1,6 @@
|
|
|
1
1
|
import { encode } from '@toon-format/toon'
|
|
2
2
|
import { countTokens } from '@anthropic-ai/tokenizer'
|
|
3
|
+
import { createPatch } from 'diff'
|
|
3
4
|
import { tryParseJSON, classifyContent, stripLineNumbers } from './detect.js'
|
|
4
5
|
import { anthropic } from './providers.js'
|
|
5
6
|
|
|
@@ -19,6 +20,89 @@ function normalizeWhitespace(text) {
|
|
|
19
20
|
.replace(/\n{3,}/g, '\n\n')
|
|
20
21
|
}
|
|
21
22
|
|
|
23
|
+
// --- Stage: dedup ---
|
|
24
|
+
function deduplicateTargets(targets) {
|
|
25
|
+
const seen = new Map()
|
|
26
|
+
for (const target of targets) {
|
|
27
|
+
if (target.skip) continue
|
|
28
|
+
const key = cacheKey(target.text)
|
|
29
|
+
const prev = seen.get(key)
|
|
30
|
+
if (prev && prev.text === target.text) {
|
|
31
|
+
const ref = `[see tool_result in message ${prev.path[1]}, block ${prev.index} — identical content]`
|
|
32
|
+
target.compressed = ref
|
|
33
|
+
target.dedup = true
|
|
34
|
+
} else {
|
|
35
|
+
seen.set(key, target)
|
|
36
|
+
}
|
|
37
|
+
}
|
|
38
|
+
}
|
|
39
|
+
|
|
40
|
+
// --- Stage: diff ---
|
|
41
|
+
function diffTargets(targets) {
|
|
42
|
+
const seen = []
|
|
43
|
+
for (const target of targets) {
|
|
44
|
+
if (target.skip || target.dedup || target.compressed) continue
|
|
45
|
+
if (target.text.length < 200) { seen.push(target); continue }
|
|
46
|
+
|
|
47
|
+
for (const prev of seen) {
|
|
48
|
+
const sim = quickSimilarity(prev.text, target.text)
|
|
49
|
+
if (sim > 0.5 && sim < 1.0) {
|
|
50
|
+
const patch = createPatch('file', prev.text, target.text, '', '', { context: 1 })
|
|
51
|
+
// Strip the patch header (first 4 lines: ---, +++, index, etc.)
|
|
52
|
+
const lines = patch.split('\n')
|
|
53
|
+
const bodyStart = lines.findIndex(l => l.startsWith('@@'))
|
|
54
|
+
if (bodyStart === -1) continue
|
|
55
|
+
const diffBody = lines.slice(bodyStart).join('\n')
|
|
56
|
+
if (diffBody.length < target.text.length * 0.5) {
|
|
57
|
+
target.compressed = `[diff from tool_result in message ${prev.path[1]}, block ${prev.index}]:\n${diffBody}`
|
|
58
|
+
target.diffed = true
|
|
59
|
+
break
|
|
60
|
+
}
|
|
61
|
+
}
|
|
62
|
+
}
|
|
63
|
+
seen.push(target)
|
|
64
|
+
}
|
|
65
|
+
}
|
|
66
|
+
|
|
67
|
+
function quickSimilarity(a, b) {
|
|
68
|
+
if (Math.abs(a.length - b.length) > Math.max(a.length, b.length) * 0.5) return 0
|
|
69
|
+
const setA = new Set(a.split('\n'))
|
|
70
|
+
const setB = new Set(b.split('\n'))
|
|
71
|
+
const intersection = [...setA].filter(x => setB.has(x)).length
|
|
72
|
+
const union = new Set([...setA, ...setB]).size
|
|
73
|
+
return union > 0 ? intersection / union : 0
|
|
74
|
+
}
|
|
75
|
+
|
|
76
|
+
// --- Stage: prune ---
|
|
77
|
+
const PRUNE_KEYS = new Set(['integrity', 'shasum', '_id', '_from', '_resolved', '_integrity', '_nodeVersion', '_npmVersion', '_phantomChildren', '_requiredBy'])
|
|
78
|
+
|
|
79
|
+
function shouldPrune(key, val) {
|
|
80
|
+
if (PRUNE_KEYS.has(key)) return true
|
|
81
|
+
if (key === 'resolved' && typeof val === 'string' && val.startsWith('https://registry.')) return true
|
|
82
|
+
return false
|
|
83
|
+
}
|
|
84
|
+
|
|
85
|
+
function deepPrune(obj) {
|
|
86
|
+
if (Array.isArray(obj)) return obj.map(deepPrune)
|
|
87
|
+
if (obj === null || typeof obj !== 'object') return obj
|
|
88
|
+
const result = {}
|
|
89
|
+
for (const [key, val] of Object.entries(obj)) {
|
|
90
|
+
if (shouldPrune(key, val)) continue
|
|
91
|
+
result[key] = deepPrune(val)
|
|
92
|
+
}
|
|
93
|
+
return result
|
|
94
|
+
}
|
|
95
|
+
|
|
96
|
+
function pruneJSON(text) {
|
|
97
|
+
const { ok, value } = tryParseJSON(text)
|
|
98
|
+
if (!ok) return null
|
|
99
|
+
const pruned = deepPrune(value)
|
|
100
|
+
const result = JSON.stringify(pruned, null, 2)
|
|
101
|
+
if (result.length >= text.length * 0.95) return null
|
|
102
|
+
return result
|
|
103
|
+
}
|
|
104
|
+
|
|
105
|
+
// --- Core compression ---
|
|
22
106
|
export function compressText(text, config) {
|
|
23
107
|
if (text.length < config.minSize) return null
|
|
24
108
|
const cls = classifyContent(text)
|
|
@@ -44,7 +128,14 @@ export function compressText(text, config) {
|
|
|
44
128
|
|
|
45
129
|
if (cls !== 'json' && cls !== 'json-lined') return null
|
|
46
130
|
|
|
47
|
-
|
|
131
|
+
let raw = cls === 'json-lined' ? stripLineNumbers(text) : text
|
|
132
|
+
|
|
133
|
+
// Prune low-value fields before minifying
|
|
134
|
+
if (config.stages.includes('prune')) {
|
|
135
|
+
const pruned = pruneJSON(raw)
|
|
136
|
+
if (pruned) raw = pruned
|
|
137
|
+
}
|
|
138
|
+
|
|
48
139
|
const { ok, value } = tryParseJSON(raw)
|
|
49
140
|
if (!ok) return null
|
|
50
141
|
|
|
@@ -106,9 +197,27 @@ async function compressBlock(text, config) {
|
|
|
106
197
|
|
|
107
198
|
export async function compressRequest(body, config, provider) {
|
|
108
199
|
const targets = provider.extract(body)
|
|
200
|
+
|
|
201
|
+
// Dedup: replace identical blocks with reference markers
|
|
202
|
+
if (config.stages.includes('dedup')) deduplicateTargets(targets)
|
|
203
|
+
|
|
204
|
+
// Diff: replace similar blocks with unified diffs
|
|
205
|
+
if (config.stages.includes('diff')) diffTargets(targets)
|
|
206
|
+
|
|
109
207
|
const stats = []
|
|
110
208
|
for (const target of targets) {
|
|
111
209
|
if (target.skip) { stats.push({ index: target.index, skipped: target.skip }); continue }
|
|
210
|
+
|
|
211
|
+
// Dedup/diff already set .compressed — record stats and skip compression
|
|
212
|
+
if (target.dedup) {
|
|
213
|
+
stats.push({ index: target.index, method: 'dedup', originalLen: target.text.length, compressedLen: target.compressed.length, originalTokens: countTokens(target.text), compressedTokens: countTokens(target.compressed) })
|
|
214
|
+
continue
|
|
215
|
+
}
|
|
216
|
+
if (target.diffed) {
|
|
217
|
+
stats.push({ index: target.index, method: 'diff', originalLen: target.text.length, compressedLen: target.compressed.length, originalTokens: countTokens(target.text), compressedTokens: countTokens(target.compressed) })
|
|
218
|
+
continue
|
|
219
|
+
}
|
|
220
|
+
|
|
112
221
|
const result = await compressBlock(target.text, config)
|
|
113
222
|
if (result) {
|
|
114
223
|
target.compressed = result.text
|
package/config.js
CHANGED
|
@@ -1,5 +1,5 @@
|
|
|
1
1
|
export function loadConfig(env = process.env) {
|
|
2
|
-
const stages = (env.TAMP_STAGES || 'minify,toon,strip-lines,whitespace,llmlingua').split(',').map(s => s.trim()).filter(Boolean)
|
|
2
|
+
const stages = (env.TAMP_STAGES || 'minify,toon,strip-lines,whitespace,llmlingua,dedup,diff,prune').split(',').map(s => s.trim()).filter(Boolean)
|
|
3
3
|
return Object.freeze({
|
|
4
4
|
port: parseInt(env.TAMP_PORT, 10) || 7778,
|
|
5
5
|
upstream: env.TAMP_UPSTREAM || 'https://api.anthropic.com',
|
package/package.json
CHANGED
|
@@ -9,8 +9,8 @@
|
|
|
9
9
|
"providers.js",
|
|
10
10
|
"stats.js"
|
|
11
11
|
],
|
|
12
|
-
"version": "0.
|
|
13
|
-
"description": "Token compression proxy for coding agents. Works with Claude Code, Aider, Cursor, Cline, Windsurf.
|
|
12
|
+
"version": "0.3.1",
|
|
13
|
+
"description": "Token compression proxy for coding agents. Works with Claude Code, Aider, Cursor, Cline, Windsurf. 50.5% fewer input tokens.",
|
|
14
14
|
"type": "module",
|
|
15
15
|
"main": "index.js",
|
|
16
16
|
"bin": {
|
|
@@ -20,7 +20,8 @@
|
|
|
20
20
|
"start": "node bin/tamp.js",
|
|
21
21
|
"test": "node --test test/*.test.js",
|
|
22
22
|
"test:sidecar": "node test/capture-golden.js --verify",
|
|
23
|
-
"bench:semantic": "node bench/semantic-eval.js"
|
|
23
|
+
"bench:semantic": "node bench/semantic-eval.js",
|
|
24
|
+
"test:quality": "node bench/quality-eval.js"
|
|
24
25
|
},
|
|
25
26
|
"keywords": [
|
|
26
27
|
"claude",
|
|
@@ -41,6 +42,7 @@
|
|
|
41
42
|
"@anthropic-ai/tokenizer": "^0.0.4",
|
|
42
43
|
"@inquirer/prompts": "^8.3.2",
|
|
43
44
|
"@toon-format/toon": "^2.1.0",
|
|
45
|
+
"diff": "^8.0.4",
|
|
44
46
|
"fzstd": "^0.1.1"
|
|
45
47
|
}
|
|
46
48
|
}
|