@sliday/tamp 0.2.12 → 0.2.14

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 CHANGED
@@ -1,6 +1,6 @@
1
1
  # Tamp
2
2
 
3
- **Token compression proxy for coding agents.** 33.9% fewer input tokens, zero code changes. Works with Claude Code, Codex, Aider, Cursor, Cline, Windsurf, and any OpenAI-compatible agent.
3
+ **Token compression proxy for coding agents.** 33.9% 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
@@ -19,7 +19,6 @@ Tamp auto-detects your agent's API format and compresses tool result blocks befo
19
19
  ```
20
20
  Claude Code ──► Tamp (localhost:7778) ──► Anthropic API
21
21
  Aider/Cursor ──► │ ──► OpenAI API
22
- Codex CLI ─────► │ ──► OpenAI API
23
22
  Gemini CLI ────► │ ──► Google AI API
24
23
  │
25
24
  ├─ JSON → minify whitespace
@@ -35,18 +34,21 @@ Gemini CLI ────► │ ──► Google AI API
35
34
  |--------|----------|--------|
36
35
  | Anthropic Messages | `POST /v1/messages` | Claude Code |
37
36
  | OpenAI Chat Completions | `POST /v1/chat/completions` | Aider, Cursor, Cline, Windsurf, OpenCode |
38
- | OpenAI Responses | `POST /v1/responses` | Codex CLI |
39
37
  | Google Gemini | `POST .../generateContent` | Gemini CLI |
40
38
 
39
+ > **Why not Codex?** Codex CLI uses OpenAI's Responses API (`POST /v1/responses`), which has a different request shape than Chat Completions — `input[]` with `function_call_output` items instead of `messages[]` with `role: tool`. We had early support but pulled it because the Responses API is still evolving and Codex sends zstd-compressed bodies that add another layer of complexity. We'll revisit once the format stabilizes.
40
+
41
41
  ### Compression Stages
42
42
 
43
43
  | Stage | What it does | When it applies |
44
44
  |-------|-------------|-----------------|
45
45
  | `minify` | Strips JSON whitespace | Pretty-printed JSON objects/arrays |
46
46
  | `toon` | Columnar [TOON encoding](https://github.com/nicholasgasior/toon-format) | Homogeneous arrays (file listings, routes, deps) |
47
- | `llmlingua` | Neural text compression via [LLMLingua](https://github.com/microsoft/LLMLingua) sidecar | Natural language text (requires sidecar) |
47
+ | `strip-lines` | Removes line-number prefixes | Read tool output (` 1→...`) |
48
+ | `whitespace` | Collapses blank lines, trims trailing spaces | CLI output, source code |
49
+ | `llmlingua` | Neural text compression via [LLMLingua-2](https://github.com/microsoft/LLMLingua) | Natural language text (auto-starts sidecar) |
48
50
 
49
- `minify` and `toon` are enabled by default. Add LLMLingua with `TAMP_STAGES=minify,toon,llmlingua`.
51
+ All 5 stages enabled by default. On first launch, an interactive prompt lets you toggle methods. Use `-y` to skip the prompt.
50
52
 
51
53
  ## Quick Start
52
54
 
@@ -83,12 +85,6 @@ export OPENAI_API_BASE=http://localhost:7778
83
85
  aider
84
86
  ```
85
87
 
86
- **Codex CLI:**
87
- Add to `~/.codex/config.toml`:
88
- ```toml
89
- openai_base_url = "http://localhost:7778"
90
- ```
91
-
92
88
  **Cursor / Cline / Windsurf:**
93
89
  Set the API base URL to `http://localhost:7778` in your editor's settings.
94
90
 
@@ -104,7 +100,7 @@ All configuration via environment variables:
104
100
  | `TAMP_UPSTREAM` | `https://api.anthropic.com` | Default upstream API URL |
105
101
  | `TAMP_UPSTREAM_OPENAI` | `https://api.openai.com` | Upstream for OpenAI-format requests |
106
102
  | `TAMP_UPSTREAM_GEMINI` | `https://generativelanguage.googleapis.com` | Upstream for Gemini-format requests |
107
- | `TAMP_STAGES` | `minify,toon` | Comma-separated compression stages |
103
+ | `TAMP_STAGES` | `minify,toon,strip-lines,whitespace,llmlingua` | Comma-separated compression stages |
108
104
  | `TAMP_MIN_SIZE` | `200` | Minimum content size (chars) to attempt compression |
109
105
  | `TAMP_LOG` | `true` | Enable request logging to stderr |
110
106
  | `TAMP_LOG_FILE` | _(none)_ | Write logs to file |
@@ -114,8 +110,14 @@ All configuration via environment variables:
114
110
  ### Recommended setup
115
111
 
116
112
  ```bash
117
- # Maximum compression
118
- TAMP_STAGES=minify,toon,llmlingua TAMP_LLMLINGUA_URL=http://localhost:8788 npx @sliday/tamp
113
+ # All stages enabled by default — just run:
114
+ npx @sliday/tamp
115
+
116
+ # Skip interactive prompt (CI/scripts):
117
+ npx @sliday/tamp -y
118
+
119
+ # Without LLMLingua (no Python needed):
120
+ TAMP_STAGES=minify,toon,strip-lines,whitespace npx @sliday/tamp -y
119
121
  ```
120
122
 
121
123
  ## Installation Methods
@@ -151,7 +153,7 @@ The installer clones to `~/.tamp`, adds `ANTHROPIC_BASE_URL` to your shell profi
151
153
 
152
154
  ## What Gets Compressed
153
155
 
154
- Tamp only compresses the **last user message** in each request (the most recent `tool_result` blocks). Historical messages are left untouched to avoid redundant recompression.
156
+ Tamp compresses `tool_result` blocks in **all messages** — not just the latest. Since each API call re-sends the full conversation history uncompressed, this compounds savings with every turn. An in-memory cache ensures identical content is only compressed once per session.
155
157
 
156
158
  | Content Type | Action | Example |
157
159
  |-------------|--------|---------|
@@ -159,7 +161,7 @@ Tamp only compresses the **last user message** in each request (the most recent
159
161
  | JSON with line numbers | Strip prefixes + minify | Read tool output (` 1→{...}`) |
160
162
  | Homogeneous JSON arrays | TOON encode | File listings, route tables, dependencies |
161
163
  | Already-minified JSON | Skip | Single-line JSON |
162
- | Source code (text) | Passthrough | `.ts`, `.py`, `.rs` files |
164
+ | Source code (text) | Strip line numbers + normalize whitespace | `.ts`, `.py`, `.rs` files |
163
165
  | `is_error: true` results | Skip entirely | Error tool results |
164
166
  | TOON-encoded content | Skip | Already compressed |
165
167
 
package/bin/tamp.js CHANGED
@@ -1,15 +1,16 @@
1
1
  #!/usr/bin/env node
2
2
  import { createProxy } from '../index.js'
3
3
  import { existsSync, readFileSync } from 'node:fs'
4
- import { spawn } from 'node:child_process'
4
+ import { spawn, execFileSync } from 'node:child_process'
5
5
  import { fileURLToPath } from 'node:url'
6
6
  import { dirname, join } from 'node:path'
7
+ import { checkbox } from '@inquirer/prompts'
8
+ import http from 'node:http'
7
9
 
8
10
  const __dirname = dirname(fileURLToPath(import.meta.url))
9
11
  const root = join(__dirname, '..')
10
12
  const pkg = JSON.parse(readFileSync(join(root, 'package.json'), 'utf-8'))
11
13
 
12
- // ANSI colors
13
14
  const c = {
14
15
  reset: '\x1b[0m',
15
16
  bold: '\x1b[1m',
@@ -20,12 +21,146 @@ const c = {
20
21
  magenta: '\x1b[35m',
21
22
  cyan: '\x1b[36m',
22
23
  bgGreen: '\x1b[42m',
23
- bgYellow: '\x1b[43m',
24
24
  }
25
25
 
26
26
  function log(msg = '') { console.error(msg) }
27
27
 
28
- function printBanner(config) {
28
+ const STAGE_INFO = {
29
+ minify: 'Strip JSON whitespace (lossless)',
30
+ toon: 'Columnar array encoding (lossless)',
31
+ 'strip-lines':'Remove line-number prefixes from Read output',
32
+ whitespace: 'Collapse blank lines, trim trailing spaces',
33
+ llmlingua: 'Neural text compression via LLMLingua-2',
34
+ }
35
+
36
+ // --- Determine stages ---
37
+ const skipPrompt = process.argv.includes('-y') || process.argv.includes('--no-interactive') || !process.stdin.isTTY
38
+ let selectedStages
39
+
40
+ if (process.env.TAMP_STAGES) {
41
+ selectedStages = process.env.TAMP_STAGES.split(',').map(s => s.trim()).filter(Boolean)
42
+ } else if (skipPrompt) {
43
+ selectedStages = Object.keys(STAGE_INFO)
44
+ } else {
45
+ log('')
46
+ log(` ${c.bold}${c.cyan}Tamp${c.reset} ${c.dim}v${pkg.version}${c.reset} — Token compression proxy`)
47
+ log('')
48
+ selectedStages = await checkbox({
49
+ message: 'Compression methods (space to toggle, enter to confirm):',
50
+ choices: Object.entries(STAGE_INFO).map(([value, desc]) => ({
51
+ name: `${value.padEnd(12)} ${c.dim}— ${desc}${c.reset}`,
52
+ value,
53
+ checked: true,
54
+ })),
55
+ })
56
+ if (selectedStages.length === 0) {
57
+ log(` ${c.yellow}No methods selected — running as passthrough proxy.${c.reset}`)
58
+ }
59
+ }
60
+
61
+ process.env.TAMP_STAGES = selectedStages.join(',')
62
+
63
+ // --- Sidecar startup ---
64
+ let sidecarProc = null
65
+
66
+ async function checkPort(port) {
67
+ return new Promise(resolve => {
68
+ const req = http.get(`http://127.0.0.1:${port}/health`, (res) => {
69
+ res.resume()
70
+ resolve(res.statusCode === 200)
71
+ })
72
+ req.on('error', () => resolve(false))
73
+ req.setTimeout(1000, () => { req.destroy(); resolve(false) })
74
+ })
75
+ }
76
+
77
+ function hasCommand(cmd) {
78
+ try { execFileSync('which', [cmd], { stdio: 'ignore' }); return true } catch { return false }
79
+ }
80
+
81
+ async function startSidecar() {
82
+ const sidecarPort = 8788
83
+ const sidecarDir = join(root, 'sidecar')
84
+ const serverPy = join(sidecarDir, 'server.py')
85
+
86
+ // 1. Already running?
87
+ if (await checkPort(sidecarPort)) {
88
+ log(` ${c.green}✓${c.reset} LLMLingua-2 sidecar already running on port ${sidecarPort}`)
89
+ return `http://localhost:${sidecarPort}`
90
+ }
91
+
92
+ if (!existsSync(serverPy)) {
93
+ return null
94
+ }
95
+
96
+ log(` ${c.yellow}→${c.reset} Starting LLMLingua-2 sidecar ...`)
97
+
98
+ // 2. Try uv run (no venv needed)
99
+ if (hasCommand('uv')) {
100
+ try {
101
+ const proc = spawn('uv', [
102
+ 'run', '--with', 'fastapi,uvicorn,llmlingua',
103
+ 'uvicorn', 'server:app', '--host', '127.0.0.1', '--port', String(sidecarPort),
104
+ ], { cwd: sidecarDir, stdio: ['ignore', 'pipe', 'pipe'] })
105
+
106
+ const url = await waitForSidecar(proc, sidecarPort)
107
+ if (url) { sidecarProc = proc; return url }
108
+ proc.kill()
109
+ } catch { /* try next */ }
110
+ }
111
+
112
+ // 3. Try existing venv
113
+ const venvPython = join(sidecarDir, '.venv', 'bin', 'python')
114
+ if (existsSync(venvPython)) {
115
+ try {
116
+ const proc = spawn(venvPython, [
117
+ '-m', 'uvicorn', 'server:app', '--host', '127.0.0.1', '--port', String(sidecarPort),
118
+ ], { cwd: sidecarDir, stdio: ['ignore', 'pipe', 'pipe'] })
119
+
120
+ const url = await waitForSidecar(proc, sidecarPort)
121
+ if (url) { sidecarProc = proc; return url }
122
+ proc.kill()
123
+ } catch { /* fall through */ }
124
+ }
125
+
126
+ return null
127
+ }
128
+
129
+ function waitForSidecar(proc, port, timeout = 30000) {
130
+ return new Promise(resolve => {
131
+ const timer = setTimeout(() => resolve(null), timeout)
132
+ proc.stderr.on('data', (d) => {
133
+ if (d.toString().includes('Uvicorn running')) {
134
+ clearTimeout(timer)
135
+ log(` ${c.green}✓${c.reset} LLMLingua-2 sidecar ready on port ${c.bold}${port}${c.reset}`)
136
+ resolve(`http://localhost:${port}`)
137
+ }
138
+ })
139
+ proc.on('exit', () => { clearTimeout(timer); resolve(null) })
140
+ })
141
+ }
142
+
143
+ // Start sidecar if llmlingua is in selected stages
144
+ if (selectedStages.includes('llmlingua')) {
145
+ const sidecarUrl = await startSidecar()
146
+ if (sidecarUrl) {
147
+ process.env.TAMP_LLMLINGUA_URL = sidecarUrl
148
+ } else {
149
+ log(` ${c.yellow}!${c.reset} LLMLingua-2 sidecar not available`)
150
+ if (!hasCommand('uv')) {
151
+ log(` Install uv: ${c.cyan}curl -LsSf https://astral.sh/uv/install.sh | sh${c.reset}`)
152
+ }
153
+ log(` Continuing without neural compression.`)
154
+ log('')
155
+ selectedStages = selectedStages.filter(s => s !== 'llmlingua')
156
+ process.env.TAMP_STAGES = selectedStages.join(',')
157
+ }
158
+ }
159
+
160
+ // --- Start proxy ---
161
+ const { config, server } = createProxy()
162
+
163
+ function printBanner() {
29
164
  const url = `http://localhost:${config.port}`
30
165
 
31
166
  log('')
@@ -48,66 +183,19 @@ function printBanner(config) {
48
183
  log('')
49
184
 
50
185
  log(` ${c.bold}Compression:${c.reset}`)
51
- for (const stage of config.stages) {
52
- const icon = stage === 'llmlingua' ? `${c.green}▸${c.reset}` : `${c.green}▸${c.reset}`
53
- const label = stage === 'minify' ? 'JSON whitespace removal'
54
- : stage === 'toon' ? 'TOON columnar encoding'
55
- : stage === 'llmlingua' ? `LLMLingua-2 neural compression ${c.dim}(${config.llmLinguaUrl})${c.reset}`
56
- : stage
57
- log(` ${icon} ${c.cyan}${stage}${c.reset} — ${label}`)
186
+ for (const [stage, desc] of Object.entries(STAGE_INFO)) {
187
+ const active = config.stages.includes(stage)
188
+ const icon = active ? `${c.green}✓${c.reset}` : `${c.dim}✗${c.reset}`
189
+ const extra = stage === 'llmlingua' && active && config.llmLinguaUrl ? ` ${c.dim}(${config.llmLinguaUrl})${c.reset}` : ''
190
+ log(` ${icon} ${active ? c.cyan : c.dim}${stage}${c.reset} — ${active ? desc : c.dim + desc + c.reset}${extra}`)
58
191
  }
59
192
  log('')
60
193
  }
61
194
 
62
- // --- Auto-start LLMLingua-2 sidecar if needed ---
63
- let { config: finalConfig, server: finalServer } = createProxy()
64
-
65
- const needsSidecar = finalConfig.stages.includes('llmlingua') && !finalConfig.llmLinguaUrl
66
- const venvPython = join(root, 'sidecar', '.venv', 'bin', 'python')
67
- const serverPy = join(root, 'sidecar', 'server.py')
68
- const hasSidecar = existsSync(venvPython) && existsSync(serverPy)
69
-
70
- if (needsSidecar && hasSidecar) {
71
- const sidecarPort = 8788
72
- process.env.TAMP_LLMLINGUA_URL = `http://localhost:${sidecarPort}`
73
- const refreshed = createProxy()
74
- finalConfig = refreshed.config
75
- finalServer = refreshed.server
76
-
77
- log('')
78
- log(` ${c.yellow}→${c.reset} Starting LLMLingua-2 sidecar ...`)
79
-
80
- const sidecar = spawn(venvPython, ['-m', 'uvicorn', 'server:app', '--host', '127.0.0.1', '--port', String(sidecarPort)], {
81
- cwd: join(root, 'sidecar'),
82
- stdio: ['ignore', 'pipe', 'pipe'],
83
- })
84
-
85
- let ready = false
86
- sidecar.stderr.on('data', (d) => {
87
- const line = d.toString()
88
- if (!ready && line.includes('Uvicorn running')) {
89
- ready = true
90
- log(` ${c.green}✓${c.reset} LLMLingua-2 sidecar ready on ${c.bold}port ${sidecarPort}${c.reset}`)
91
- }
92
- })
93
-
94
- sidecar.on('exit', (code) => {
95
- if (code !== null && code !== 0) {
96
- log(` ${c.yellow}✗${c.reset} LLMLingua-2 sidecar exited (code ${code})`)
97
- }
98
- })
99
-
100
- process.on('exit', () => { sidecar?.kill() })
101
- process.on('SIGINT', () => { sidecar?.kill(); process.exit() })
102
- process.on('SIGTERM', () => { sidecar?.kill(); process.exit() })
103
- } else if (needsSidecar && !hasSidecar) {
104
- log('')
105
- log(` ${c.yellow}✗${c.reset} LLMLingua-2 sidecar not installed`)
106
- log(` Run: ${c.cyan}curl -fsSL tamp.dev/setup.sh | bash${c.reset}`)
107
- }
108
-
109
- const { config: cfg, server: srv } = { config: finalConfig, server: finalServer }
110
-
111
- srv.listen(cfg.port, () => {
112
- printBanner(cfg)
195
+ server.listen(config.port, () => {
196
+ printBanner()
113
197
  })
198
+
199
+ process.on('exit', () => sidecarProc?.kill())
200
+ process.on('SIGINT', () => { sidecarProc?.kill(); process.exit() })
201
+ process.on('SIGTERM', () => { sidecarProc?.kill(); process.exit() })
package/compress.js CHANGED
@@ -3,16 +3,45 @@ import { countTokens } from '@anthropic-ai/tokenizer'
3
3
  import { tryParseJSON, classifyContent, stripLineNumbers } from './detect.js'
4
4
  import { anthropic } from './providers.js'
5
5
 
6
+ const cache = new Map()
7
+ const MAX_CACHE = 500
8
+
9
+ function cacheKey(text) {
10
+ if (text.length < 128) return text
11
+ return `${text.length}:${text.slice(0, 64)}:${text.slice(-64)}`
12
+ }
13
+
14
+ export function clearCache() { cache.clear() }
15
+
16
+ function normalizeWhitespace(text) {
17
+ return text
18
+ .replace(/[ \t]+$/gm, '')
19
+ .replace(/\n{3,}/g, '\n\n')
20
+ }
21
+
6
22
  export function compressText(text, config) {
7
23
  if (text.length < config.minSize) return null
8
24
  const cls = classifyContent(text)
9
25
  if (cls === 'toon') return null
26
+
10
27
  if (cls === 'text') {
28
+ let processed = text
29
+ if (config.stages.includes('strip-lines')) {
30
+ const stripped = stripLineNumbers(text)
31
+ if (stripped !== text) processed = stripped
32
+ }
33
+ if (config.stages.includes('whitespace')) {
34
+ processed = normalizeWhitespace(processed)
35
+ }
36
+ if (processed.length < text.length * 0.9) {
37
+ return { text: processed, method: 'normalize', originalLen: text.length, compressedLen: processed.length, originalTokens: countTokens(text), compressedTokens: countTokens(processed) }
38
+ }
11
39
  if (config.stages.includes('llmlingua') && config.llmLinguaUrl) {
12
- return { async: true, text, cls }
40
+ return { async: true, text: processed, cls }
13
41
  }
14
42
  return null
15
43
  }
44
+
16
45
  if (cls !== 'json' && cls !== 'json-lined') return null
17
46
 
18
47
  const raw = cls === 'json-lined' ? stripLineNumbers(text) : text
@@ -56,11 +85,23 @@ async function compressWithLLMLingua(text, config) {
56
85
  }
57
86
 
58
87
  async function compressBlock(text, config) {
88
+ const key = cacheKey(text)
89
+ if (cache.has(key)) return cache.get(key)
90
+
59
91
  const sync = compressText(text, config)
92
+ let result
60
93
  if (sync && sync.async) {
61
- return compressWithLLMLingua(text, config)
94
+ result = await compressWithLLMLingua(sync.text, config)
95
+ } else {
96
+ result = sync
97
+ }
98
+
99
+ if (cache.size >= MAX_CACHE) {
100
+ const firstKey = cache.keys().next().value
101
+ cache.delete(firstKey)
62
102
  }
63
- return sync
103
+ cache.set(key, result)
104
+ return result
64
105
  }
65
106
 
66
107
  export async function compressRequest(body, config, provider) {
package/config.js CHANGED
@@ -1,5 +1,5 @@
1
1
  export function loadConfig(env = process.env) {
2
- const stages = (env.TAMP_STAGES || 'minify,toon').split(',').map(s => s.trim()).filter(Boolean)
2
+ const stages = (env.TAMP_STAGES || 'minify,toon,strip-lines,whitespace,llmlingua').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
@@ -9,7 +9,7 @@
9
9
  "providers.js",
10
10
  "stats.js"
11
11
  ],
12
- "version": "0.2.12",
12
+ "version": "0.2.14",
13
13
  "description": "Token compression proxy for coding agents. Works with Claude Code, Aider, Cursor, Cline, Windsurf. 33.9% fewer input tokens.",
14
14
  "type": "module",
15
15
  "main": "index.js",
@@ -39,6 +39,7 @@
39
39
  "homepage": "https://github.com/sliday/tamp",
40
40
  "dependencies": {
41
41
  "@anthropic-ai/tokenizer": "^0.0.4",
42
+ "@inquirer/prompts": "^8.3.2",
42
43
  "@toon-format/toon": "^2.1.0",
43
44
  "fzstd": "^0.1.1"
44
45
  }
package/providers.js CHANGED
@@ -7,29 +7,26 @@ const anthropic = {
7
7
  const targets = []
8
8
  if (!body?.messages?.length) return targets
9
9
 
10
- let lastUserIdx = -1
11
- for (let i = body.messages.length - 1; i >= 0; i--) {
12
- if (body.messages[i].role === 'user') { lastUserIdx = i; break }
13
- }
14
- if (lastUserIdx === -1) return targets
15
-
16
- const msg = body.messages[lastUserIdx]
10
+ for (let mi = 0; mi < body.messages.length; mi++) {
11
+ const msg = body.messages[mi]
12
+ if (msg.role !== 'user') continue
17
13
 
18
- if (typeof msg.content === 'string') {
19
- targets.push({ path: ['messages', lastUserIdx, 'content'], text: msg.content })
20
- } else if (Array.isArray(msg.content)) {
21
- for (let i = 0; i < msg.content.length; i++) {
22
- const block = msg.content[i]
23
- if (block.type !== 'tool_result') continue
24
- if (block.is_error) { targets.push({ skip: 'error', index: i }); continue }
14
+ if (typeof msg.content === 'string') {
15
+ targets.push({ path: ['messages', mi, 'content'], text: msg.content })
16
+ } else if (Array.isArray(msg.content)) {
17
+ for (let i = 0; i < msg.content.length; i++) {
18
+ const block = msg.content[i]
19
+ if (block.type !== 'tool_result') continue
20
+ if (block.is_error) { targets.push({ skip: 'error', index: i }); continue }
25
21
 
26
- if (typeof block.content === 'string') {
27
- targets.push({ path: ['messages', lastUserIdx, 'content', i, 'content'], text: block.content, index: i })
28
- } else if (Array.isArray(block.content)) {
29
- for (let j = 0; j < block.content.length; j++) {
30
- const sub = block.content[j]
31
- if (sub.type === 'text') {
32
- targets.push({ path: ['messages', lastUserIdx, 'content', i, 'content', j, 'text'], text: sub.text, index: i })
22
+ if (typeof block.content === 'string') {
23
+ targets.push({ path: ['messages', mi, 'content', i, 'content'], text: block.content, index: i })
24
+ } else if (Array.isArray(block.content)) {
25
+ for (let j = 0; j < block.content.length; j++) {
26
+ const sub = block.content[j]
27
+ if (sub.type === 'text') {
28
+ targets.push({ path: ['messages', mi, 'content', i, 'content', j, 'text'], text: sub.text, index: i })
29
+ }
33
30
  }
34
31
  }
35
32
  }
@@ -54,19 +51,17 @@ const openai = {
54
51
  return method === 'POST' && (
55
52
  url.startsWith('/v1/chat/completions') ||
56
53
  url.startsWith('/chat/completions') ||
57
- url.startsWith('/v1/responses') ||
58
- url.startsWith('/responses')
54
+ url.startsWith('/v1/responses')
59
55
  )
60
56
  },
61
57
  normalizeUrl(url) {
62
58
  if (url.startsWith('/chat/completions')) return '/v1' + url
63
- if (url.startsWith('/responses')) return '/v1' + url
64
59
  return url
65
60
  },
66
61
  extract(body) {
67
62
  const targets = []
68
63
 
69
- // Responses API format: body.input array
64
+ // Responses API format: body.input array (Codex CLI)
70
65
  if (body?.input?.length) {
71
66
  for (let i = 0; i < body.input.length; i++) {
72
67
  const item = body.input[i]
@@ -89,20 +84,9 @@ const openai = {
89
84
  // Chat Completions format
90
85
  if (!body?.messages?.length) return targets
91
86
 
92
- // Find last assistant message with tool_calls
93
- let lastAssistantIdx = -1
94
- for (let i = body.messages.length - 1; i >= 0; i--) {
95
- if (body.messages[i].role === 'assistant' && body.messages[i].tool_calls?.length) {
96
- lastAssistantIdx = i
97
- break
98
- }
99
- }
100
- if (lastAssistantIdx === -1) return targets
101
-
102
- // Collect all subsequent role:tool messages
103
- for (let i = lastAssistantIdx + 1; i < body.messages.length; i++) {
87
+ for (let i = 0; i < body.messages.length; i++) {
104
88
  const msg = body.messages[i]
105
- if (msg.role !== 'tool') break
89
+ if (msg.role !== 'tool') continue
106
90
  if (typeof msg.content === 'string') {
107
91
  targets.push({ path: ['messages', i, 'content'], text: msg.content, index: i })
108
92
  }
@@ -129,8 +113,7 @@ const gemini = {
129
113
  const targets = []
130
114
  if (!body?.contents?.length) return targets
131
115
 
132
- // Find last content with functionResponse parts
133
- for (let ci = body.contents.length - 1; ci >= 0; ci--) {
116
+ for (let ci = 0; ci < body.contents.length; ci++) {
134
117
  const content = body.contents[ci]
135
118
  if (!content.parts?.length) continue
136
119
  for (let pi = 0; pi < content.parts.length; pi++) {
@@ -145,7 +128,6 @@ const gemini = {
145
128
  wasObject: typeof resp !== 'string',
146
129
  })
147
130
  }
148
- if (targets.length) break
149
131
  }
150
132
  return targets
151
133
  },
@@ -155,7 +137,6 @@ const gemini = {
155
137
  let obj = body
156
138
  const path = t.path
157
139
  for (let i = 0; i < path.length - 1; i++) obj = obj[path[i]]
158
- // If original was object, try to parse compressed back to object
159
140
  if (t.wasObject) {
160
141
  try {
161
142
  obj[path[path.length - 1]] = JSON.parse(t.compressed)