@hover-dev/core 0.17.0 → 0.19.0
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/dist/engine.d.ts +16 -39
- package/dist/engine.d.ts.map +1 -1
- package/dist/engine.js +18 -67
- package/dist/specs/pageObjectManifest.d.ts.map +1 -1
- package/dist/specs/pageObjectManifest.js +11 -10
- package/dist/specs/replayGrounded.d.ts.map +1 -1
- package/dist/specs/writeApiSpec.d.ts +36 -0
- package/dist/specs/writeApiSpec.d.ts.map +1 -0
- package/dist/specs/writeApiSpec.js +94 -0
- package/package.json +5 -22
- package/dist/agents/argv.d.ts +0 -11
- package/dist/agents/argv.d.ts.map +0 -1
- package/dist/agents/argv.js +0 -23
- package/dist/agents/claude.d.ts +0 -3
- package/dist/agents/claude.d.ts.map +0 -1
- package/dist/agents/claude.js +0 -220
- package/dist/agents/codex.d.ts +0 -19
- package/dist/agents/codex.d.ts.map +0 -1
- package/dist/agents/codex.js +0 -231
- package/dist/agents/detect.d.ts +0 -46
- package/dist/agents/detect.d.ts.map +0 -1
- package/dist/agents/detect.js +0 -80
- package/dist/agents/gemini.d.ts +0 -17
- package/dist/agents/gemini.d.ts.map +0 -1
- package/dist/agents/gemini.js +0 -186
- package/dist/agents/index.d.ts +0 -6
- package/dist/agents/index.d.ts.map +0 -1
- package/dist/agents/index.js +0 -5
- package/dist/agents/invoke.d.ts +0 -12
- package/dist/agents/invoke.d.ts.map +0 -1
- package/dist/agents/invoke.js +0 -93
- package/dist/agents/qwen.d.ts +0 -17
- package/dist/agents/qwen.d.ts.map +0 -1
- package/dist/agents/qwen.js +0 -172
- package/dist/agents/registry.d.ts +0 -19
- package/dist/agents/registry.d.ts.map +0 -1
- package/dist/agents/registry.js +0 -30
- package/dist/agents/shared.d.ts +0 -28
- package/dist/agents/shared.d.ts.map +0 -1
- package/dist/agents/shared.js +0 -35
- package/dist/agents/types.d.ts +0 -194
- package/dist/agents/types.d.ts.map +0 -1
- package/dist/agents/types.js +0 -23
- package/dist/index.d.ts +0 -3
- package/dist/index.d.ts.map +0 -1
- package/dist/index.js +0 -2
- package/dist/mcp/actuateServer.d.ts +0 -3
- package/dist/mcp/actuateServer.d.ts.map +0 -1
- package/dist/mcp/actuateServer.js +0 -594
- package/dist/mcp/sourceFence.d.ts +0 -23
- package/dist/mcp/sourceFence.d.ts.map +0 -1
- package/dist/mcp/sourceFence.js +0 -79
- package/dist/mcp/sourceServer.d.ts +0 -3
- package/dist/mcp/sourceServer.d.ts.map +0 -1
- package/dist/mcp/sourceServer.js +0 -191
- package/dist/modes.d.ts +0 -39
- package/dist/modes.d.ts.map +0 -1
- package/dist/modes.js +0 -34
- package/dist/playwright/cdpStatus.d.ts +0 -14
- package/dist/playwright/cdpStatus.d.ts.map +0 -1
- package/dist/playwright/cdpStatus.js +0 -52
- package/dist/playwright/preflight.d.ts +0 -31
- package/dist/playwright/preflight.d.ts.map +0 -1
- package/dist/playwright/preflight.js +0 -82
- package/dist/playwright/preflightCache.d.ts +0 -27
- package/dist/playwright/preflightCache.d.ts.map +0 -1
- package/dist/playwright/preflightCache.js +0 -21
- package/dist/playwright/resolveMcpConfig.d.ts +0 -61
- package/dist/playwright/resolveMcpConfig.d.ts.map +0 -1
- package/dist/playwright/resolveMcpConfig.js +0 -84
- package/dist/plugin-api.d.ts +0 -237
- package/dist/plugin-api.d.ts.map +0 -1
- package/dist/plugin-api.js +0 -52
- package/dist/qa/classify.d.ts +0 -38
- package/dist/qa/classify.d.ts.map +0 -1
- package/dist/qa/classify.js +0 -138
- package/dist/runSession.d.ts +0 -53
- package/dist/runSession.d.ts.map +0 -1
- package/dist/runSession.js +0 -96
- package/dist/service/cdpHandlers.d.ts +0 -24
- package/dist/service/cdpHandlers.d.ts.map +0 -1
- package/dist/service/cdpHandlers.js +0 -50
- package/dist/service/cdpHint.d.ts +0 -41
- package/dist/service/cdpHint.d.ts.map +0 -1
- package/dist/service/cdpHint.js +0 -158
- package/dist/service/conventions.d.ts +0 -8
- package/dist/service/conventions.d.ts.map +0 -1
- package/dist/service/conventions.js +0 -42
- package/dist/service/relayHandlers.d.ts +0 -28
- package/dist/service/relayHandlers.d.ts.map +0 -1
- package/dist/service/relayHandlers.js +0 -105
- package/dist/service/saveHandlers.d.ts +0 -50
- package/dist/service/saveHandlers.d.ts.map +0 -1
- package/dist/service/saveHandlers.js +0 -77
- package/dist/service/types.d.ts +0 -158
- package/dist/service/types.d.ts.map +0 -1
- package/dist/service/types.js +0 -26
- package/dist/service.d.ts +0 -54
- package/dist/service.d.ts.map +0 -1
- package/dist/service.js +0 -1772
- package/dist/specs/businessMap.d.ts +0 -29
- package/dist/specs/businessMap.d.ts.map +0 -1
- package/dist/specs/businessMap.js +0 -95
- package/dist/specs/extractPageObjects.d.ts +0 -18
- package/dist/specs/extractPageObjects.d.ts.map +0 -1
- package/dist/specs/extractPageObjects.js +0 -98
- package/dist/specs/optimizeSpecWithAgent.d.ts +0 -9
- package/dist/specs/optimizeSpecWithAgent.d.ts.map +0 -1
- package/dist/specs/optimizeSpecWithAgent.js +0 -39
package/dist/agents/claude.js
DELETED
|
@@ -1,220 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Running-cost accumulator across one parser lifecycle. Claude Code's
|
|
3
|
-
* stream-json sometimes carries `total_cost_usd` on intermediate events; when
|
|
4
|
-
* it does we just forward it. When it doesn't, we estimate from token usage
|
|
5
|
-
* using public per-million pricing. Hardcoded rates because they change
|
|
6
|
-
* rarely and we'd rather not network-fetch a price table from a sandbox.
|
|
7
|
-
*
|
|
8
|
-
* The estimate is empirically about 1.5–2× the authoritative
|
|
9
|
-
* `total_cost_usd` that arrives on the final `result` event (overhead from
|
|
10
|
-
* tool-definition tokens we don't see + Claude's actual cache hit ratios
|
|
11
|
-
* vs. our pessimistic accounting). That's fine for the widget's purpose:
|
|
12
|
-
* give the user a signal of cost direction and order-of-magnitude so they
|
|
13
|
-
* know when to hit Stop. The done card displays the final authoritative
|
|
14
|
-
* number from `total_cost_usd`, so the user always sees the ground truth.
|
|
15
|
-
*/
|
|
16
|
-
const PRICE_PER_M_USD = {
|
|
17
|
-
// claude-sonnet-4 / 4.5 / 4.6 / 4.7 — all priced the same as of 2026
|
|
18
|
-
sonnet: { in: 3, out: 15, cacheCreate: 3.75, cacheRead: 0.3 },
|
|
19
|
-
opus: { in: 15, out: 75, cacheCreate: 18.75, cacheRead: 1.5 },
|
|
20
|
-
haiku: { in: 1, out: 5, cacheCreate: 1.25, cacheRead: 0.1 },
|
|
21
|
-
};
|
|
22
|
-
function estimateCostUsd(modelHint, usage) {
|
|
23
|
-
// Match the most specific tier by substring. e.g. 'claude-sonnet-4-6' → sonnet.
|
|
24
|
-
const m = (modelHint ?? 'sonnet').toLowerCase();
|
|
25
|
-
const tier = m.includes('opus') ? PRICE_PER_M_USD.opus :
|
|
26
|
-
m.includes('haiku') ? PRICE_PER_M_USD.haiku :
|
|
27
|
-
PRICE_PER_M_USD.sonnet;
|
|
28
|
-
return ((usage.input_tokens ?? 0) * tier.in +
|
|
29
|
-
(usage.output_tokens ?? 0) * tier.out +
|
|
30
|
-
(usage.cache_creation_input_tokens ?? 0) * tier.cacheCreate +
|
|
31
|
-
(usage.cache_read_input_tokens ?? 0) * tier.cacheRead) / 1_000_000;
|
|
32
|
-
}
|
|
33
|
-
function claudeState(state) {
|
|
34
|
-
// First touch on this state object — seed the keys we read below.
|
|
35
|
-
if (typeof state.runningCost !== 'number') {
|
|
36
|
-
state.runningCost = 0;
|
|
37
|
-
state.runningTurns = 0;
|
|
38
|
-
state.runningTokens = 0;
|
|
39
|
-
state.runningModel = undefined;
|
|
40
|
-
}
|
|
41
|
-
return state;
|
|
42
|
-
}
|
|
43
|
-
/** Every built-in claude-code tool that has nothing to do with driving a
|
|
44
|
-
* browser. Combined with `--strict-mcp-config` + an allow-list of mcp__*
|
|
45
|
-
* ids, this leaves Claude with only the Playwright MCP (plus any
|
|
46
|
-
* plugin-contributed MCPs) as a usable tool surface. */
|
|
47
|
-
const CLAUDE_DEFAULT_DISALLOWED_TOOLS = [
|
|
48
|
-
// file / shell / data access — never appropriate for browser driving
|
|
49
|
-
'Bash', 'BashOutput', 'KillBash',
|
|
50
|
-
'Edit', 'MultiEdit', 'Write', 'Read', 'NotebookEdit',
|
|
51
|
-
// `Task` was renamed `Agent` in newer Claude Code — it spawns a subagent with
|
|
52
|
-
// the full tool surface (Bash/Write/…), a real sandbox escape, so deny BOTH
|
|
53
|
-
// the current and legacy name. (Only this one targeted name was added; do NOT
|
|
54
|
-
// add MCP-machinery tools like WaitForMcpServers here — denying those breaks
|
|
55
|
-
// the MCP tool path and the agent loses its real browser tools.)
|
|
56
|
-
'Grep', 'Glob', 'Task', 'Agent', 'TodoWrite',
|
|
57
|
-
'WebFetch', 'WebSearch',
|
|
58
|
-
// plan / worktree / cron / notification — irrelevant in -p mode
|
|
59
|
-
'EnterPlanMode', 'ExitPlanMode',
|
|
60
|
-
'EnterWorktree', 'ExitWorktree',
|
|
61
|
-
'CronCreate', 'CronDelete', 'CronList',
|
|
62
|
-
'PushNotification', 'RemoteTrigger',
|
|
63
|
-
// task & tool introspection added in claude 2.1.x — let through and
|
|
64
|
-
// the agent will burn turns exploring instead of executing
|
|
65
|
-
'ToolSearch',
|
|
66
|
-
'Monitor', 'TaskOutput', 'TaskStop',
|
|
67
|
-
'AskUserQuestion',
|
|
68
|
-
'ShareOnboardingGuide',
|
|
69
|
-
// Skills are loaded independently of the --allowedTools allow-list, so an
|
|
70
|
-
// allow-list of `mcp__playwright` does NOT block the `Skill` tool. Left
|
|
71
|
-
// through, the agent burns a turn "checking for a project skill first" and
|
|
72
|
-
// pollutes the crystallized spec with a junk `When · Skill` step. Deny it.
|
|
73
|
-
'Skill',
|
|
74
|
-
// Playwright MCP's arbitrary-JS tools. browser_run_code_unsafe /
|
|
75
|
-
// browser_evaluate run any JS in the page — a real prompt-injection exfil
|
|
76
|
-
// path (fetch a token out, read localStorage) that punches through the
|
|
77
|
-
// "Playwright MCP only" sandbox, and their output can't be translated into
|
|
78
|
-
// a deterministic Playwright spec anyway (it lands as a `// TODO`). Agents
|
|
79
|
-
// drive via click/fill/select and read state via snapshot instead.
|
|
80
|
-
'mcp__playwright__browser_run_code_unsafe',
|
|
81
|
-
'mcp__playwright__browser_evaluate',
|
|
82
|
-
];
|
|
83
|
-
export const claudeAgent = {
|
|
84
|
-
id: 'claude',
|
|
85
|
-
binName: 'claude',
|
|
86
|
-
protocol: 'argv',
|
|
87
|
-
streamFormat: 'stream-json',
|
|
88
|
-
sandboxStrength: 'hard',
|
|
89
|
-
defaultDisallowedTools: CLAUDE_DEFAULT_DISALLOWED_TOOLS,
|
|
90
|
-
// F7 optimize is text-in/text-out refinement — haiku is plenty and ~cheap.
|
|
91
|
-
cheapModel: 'haiku',
|
|
92
|
-
display: {
|
|
93
|
-
label: 'Claude Code',
|
|
94
|
-
tagline: 'Anthropic — best-in-class browser driving, hard tool sandbox',
|
|
95
|
-
homepage: 'https://docs.claude.com/claude-code',
|
|
96
|
-
installHint: 'npm install -g @anthropic-ai/claude-code',
|
|
97
|
-
},
|
|
98
|
-
buildArgs(opts) {
|
|
99
|
-
const args = ['-p', opts.prompt];
|
|
100
|
-
args.push('--output-format', 'stream-json', '--verbose');
|
|
101
|
-
args.push('--permission-mode', 'dontAsk');
|
|
102
|
-
if (opts.mcpConfig) {
|
|
103
|
-
args.push('--mcp-config', opts.mcpConfig, '--strict-mcp-config');
|
|
104
|
-
}
|
|
105
|
-
if (opts.allowedTools?.length) {
|
|
106
|
-
args.push('--allowedTools', ...opts.allowedTools);
|
|
107
|
-
}
|
|
108
|
-
if (opts.disallowedTools?.length) {
|
|
109
|
-
args.push('--disallowedTools', ...opts.disallowedTools);
|
|
110
|
-
}
|
|
111
|
-
if (opts.maxBudgetUsd != null) {
|
|
112
|
-
args.push('--max-budget-usd', String(opts.maxBudgetUsd));
|
|
113
|
-
}
|
|
114
|
-
if (opts.maxTurns != null) {
|
|
115
|
-
args.push('--max-turns', String(opts.maxTurns));
|
|
116
|
-
}
|
|
117
|
-
if (opts.model) {
|
|
118
|
-
args.push('--model', opts.model);
|
|
119
|
-
}
|
|
120
|
-
if (opts.effort) {
|
|
121
|
-
args.push('--effort', opts.effort);
|
|
122
|
-
}
|
|
123
|
-
if (opts.sessionId) {
|
|
124
|
-
args.push('--resume', opts.sessionId);
|
|
125
|
-
}
|
|
126
|
-
if (opts.appendSystemPrompt && opts.appendSystemPrompt.trim().length > 0) {
|
|
127
|
-
args.push('--append-system-prompt', opts.appendSystemPrompt);
|
|
128
|
-
}
|
|
129
|
-
return args;
|
|
130
|
-
},
|
|
131
|
-
parseEvent(line, state = {}) {
|
|
132
|
-
if (!line.trim())
|
|
133
|
-
return [];
|
|
134
|
-
let ev;
|
|
135
|
-
try {
|
|
136
|
-
ev = JSON.parse(line);
|
|
137
|
-
}
|
|
138
|
-
catch {
|
|
139
|
-
return [{ kind: 'raw', line }];
|
|
140
|
-
}
|
|
141
|
-
const s = claudeState(state);
|
|
142
|
-
const out = [];
|
|
143
|
-
if (ev.type === 'system' && ev.subtype === 'init') {
|
|
144
|
-
// Fresh session — reset the cost/turn accumulator.
|
|
145
|
-
s.runningCost = 0;
|
|
146
|
-
s.runningTurns = 0;
|
|
147
|
-
s.runningTokens = 0;
|
|
148
|
-
s.runningModel = ev.model;
|
|
149
|
-
if (ev.session_id) {
|
|
150
|
-
out.push({ kind: 'session_start', sessionId: ev.session_id, model: ev.model });
|
|
151
|
-
}
|
|
152
|
-
for (const server of ev.mcp_servers ?? []) {
|
|
153
|
-
out.push({ kind: 'mcp_status', server: server.name, status: server.status });
|
|
154
|
-
}
|
|
155
|
-
return out;
|
|
156
|
-
}
|
|
157
|
-
if (ev.type === 'assistant') {
|
|
158
|
-
s.runningTurns += 1;
|
|
159
|
-
// Claude Code sometimes carries `total_cost_usd` on intermediate events;
|
|
160
|
-
// when present it's authoritative (server-computed, includes anything
|
|
161
|
-
// we'd miss). When absent, estimate from this turn's usage so the widget
|
|
162
|
-
// still shows a growing $ counter on long runs.
|
|
163
|
-
if (typeof ev.total_cost_usd === 'number') {
|
|
164
|
-
s.runningCost = ev.total_cost_usd;
|
|
165
|
-
}
|
|
166
|
-
else if (ev.message?.usage) {
|
|
167
|
-
s.runningCost += estimateCostUsd(s.runningModel ?? ev.message.model, ev.message.usage);
|
|
168
|
-
}
|
|
169
|
-
// Token consumption = fresh input + output only. We deliberately EXCLUDE
|
|
170
|
-
// cache_read (and cache_creation): Claude re-reads ~the whole context from
|
|
171
|
-
// cache every turn, so summing cache_read across turns inflates the total
|
|
172
|
-
// ~5-10× and diverges from what Claude Code reports. input+output tracks
|
|
173
|
-
// the new tokens processed, matching the user's mental model + CC's number.
|
|
174
|
-
if (ev.message?.usage) {
|
|
175
|
-
const u = ev.message.usage;
|
|
176
|
-
s.runningTokens += (u.input_tokens ?? 0) + (u.output_tokens ?? 0);
|
|
177
|
-
}
|
|
178
|
-
out.push({ kind: 'usage', costUsd: s.runningCost, turns: s.runningTurns, tokens: s.runningTokens });
|
|
179
|
-
for (const block of ev.message?.content ?? []) {
|
|
180
|
-
if (block.type === 'tool_use') {
|
|
181
|
-
const name = block.name ?? '';
|
|
182
|
-
const tool = name.replace(/^mcp__playwright__/, '');
|
|
183
|
-
out.push({
|
|
184
|
-
kind: 'tool_use',
|
|
185
|
-
tool,
|
|
186
|
-
input: block.input,
|
|
187
|
-
costUsdSnapshot: s.runningCost,
|
|
188
|
-
tokensSnapshot: s.runningTokens,
|
|
189
|
-
});
|
|
190
|
-
}
|
|
191
|
-
else if (block.type === 'text') {
|
|
192
|
-
const text = block.text?.trim();
|
|
193
|
-
if (text)
|
|
194
|
-
out.push({ kind: 'text', text });
|
|
195
|
-
}
|
|
196
|
-
}
|
|
197
|
-
return out;
|
|
198
|
-
}
|
|
199
|
-
if (ev.type === 'user') {
|
|
200
|
-
for (const block of ev.message?.content ?? []) {
|
|
201
|
-
if (block.type === 'tool_result') {
|
|
202
|
-
out.push({ kind: 'tool_result', isError: block.is_error });
|
|
203
|
-
}
|
|
204
|
-
}
|
|
205
|
-
return out;
|
|
206
|
-
}
|
|
207
|
-
if (ev.type === 'result') {
|
|
208
|
-
out.push({
|
|
209
|
-
kind: 'session_end',
|
|
210
|
-
turns: ev.num_turns,
|
|
211
|
-
costUsd: ev.total_cost_usd,
|
|
212
|
-
tokens: s.runningTokens,
|
|
213
|
-
isError: ev.is_error,
|
|
214
|
-
summary: ev.result,
|
|
215
|
-
});
|
|
216
|
-
return out;
|
|
217
|
-
}
|
|
218
|
-
return [];
|
|
219
|
-
},
|
|
220
|
-
};
|
package/dist/agents/codex.d.ts
DELETED
|
@@ -1,19 +0,0 @@
|
|
|
1
|
-
import type { AgentDescriptor, ParserState } from './types.js';
|
|
2
|
-
export declare const codexAgent: AgentDescriptor;
|
|
3
|
-
/**
|
|
4
|
-
* Test-only escape hatches. Tests pass a state object in and get the
|
|
5
|
-
* accumulated counters back — same shape as the parser sees during a real
|
|
6
|
-
* invocation, just driven by the test instead of by invokeAgent.
|
|
7
|
-
*/
|
|
8
|
-
export declare const __testing: {
|
|
9
|
-
freshState: () => ParserState;
|
|
10
|
-
resetCounters: (state: ParserState) => void;
|
|
11
|
-
getState: (state: ParserState) => {
|
|
12
|
-
runningCost: number;
|
|
13
|
-
runningTurns: number;
|
|
14
|
-
runningSessionId: string | undefined;
|
|
15
|
-
lastAgentMessage: string | undefined;
|
|
16
|
-
sawErrorEvent: boolean;
|
|
17
|
-
};
|
|
18
|
-
};
|
|
19
|
-
//# sourceMappingURL=codex.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"codex.d.ts","sourceRoot":"","sources":["../../src/agents/codex.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,eAAe,EAA8B,WAAW,EAAE,MAAM,YAAY,CAAC;AA6K3F,eAAO,MAAM,UAAU,EAAE,eAoKxB,CAAC;AAEF;;;;GAIG;AACH,eAAO,MAAM,SAAS;sBACJ,WAAW;2BACJ,WAAW;sBAChB,WAAW;;;;;;;CAU9B,CAAC"}
|
package/dist/agents/codex.js
DELETED
|
@@ -1,231 +0,0 @@
|
|
|
1
|
-
import { stripMcpPrefix } from './shared.js';
|
|
2
|
-
/**
|
|
3
|
-
* Pricing per million tokens. Keep in lockstep with claude.ts's table —
|
|
4
|
-
* approximate published OpenAI rates as of 2026. We are deliberately
|
|
5
|
-
* conservative (no cache-hit discount); cost shown to the user is therefore
|
|
6
|
-
* a high-water estimate, which is the right error direction for a "should
|
|
7
|
-
* I hit Stop now" UI signal.
|
|
8
|
-
*/
|
|
9
|
-
const PRICE_PER_M_USD = {
|
|
10
|
-
// gpt-5.5 / gpt-5.4 / gpt-5 — public per-million pricing is similar to
|
|
11
|
-
// claude opus; tune empirically when OpenAI publishes a stable price table
|
|
12
|
-
// for the Codex tier specifically.
|
|
13
|
-
'gpt-5.5': { in: 5, out: 25 },
|
|
14
|
-
'gpt-5.4': { in: 5, out: 25 },
|
|
15
|
-
'gpt-5': { in: 5, out: 25 },
|
|
16
|
-
// gpt-4.x kept for users on legacy --model
|
|
17
|
-
'gpt-4o': { in: 2.5, out: 10 },
|
|
18
|
-
'gpt-4': { in: 30, out: 60 },
|
|
19
|
-
};
|
|
20
|
-
// `modelHint` is currently always passed as undefined — the parser can't see
|
|
21
|
-
// the invocation's --model — so the default tier below is what gets used. The
|
|
22
|
-
// parameter is kept so a future caller that does have the model id can pass it.
|
|
23
|
-
function estimateCostUsd(modelHint, usage) {
|
|
24
|
-
const m = (modelHint ?? 'gpt-5.5').toLowerCase();
|
|
25
|
-
// Match by longest-prefix so 'gpt-5.5-mini' picks up the 'gpt-5.5' tier.
|
|
26
|
-
const tier = Object.entries(PRICE_PER_M_USD).find(([key]) => m.startsWith(key))?.[1] ??
|
|
27
|
-
PRICE_PER_M_USD['gpt-5.5'];
|
|
28
|
-
return ((usage.input_tokens ?? 0) * tier.in +
|
|
29
|
-
(usage.output_tokens ?? 0) * tier.out) / 1_000_000;
|
|
30
|
-
}
|
|
31
|
-
function codexState(state) {
|
|
32
|
-
if (typeof state.runningCost !== 'number') {
|
|
33
|
-
state.runningCost = 0;
|
|
34
|
-
state.runningTurns = 0;
|
|
35
|
-
state.runningTokens = 0;
|
|
36
|
-
state.runningSessionId = undefined;
|
|
37
|
-
state.lastAgentMessage = undefined;
|
|
38
|
-
state.sawErrorEvent = false;
|
|
39
|
-
state.itemTypeById = new Map();
|
|
40
|
-
}
|
|
41
|
-
return state;
|
|
42
|
-
}
|
|
43
|
-
function resetCodexCounters(s) {
|
|
44
|
-
s.runningCost = 0;
|
|
45
|
-
s.runningTurns = 0;
|
|
46
|
-
s.runningTokens = 0;
|
|
47
|
-
s.runningSessionId = undefined;
|
|
48
|
-
s.lastAgentMessage = undefined;
|
|
49
|
-
s.sawErrorEvent = false;
|
|
50
|
-
s.itemTypeById.clear();
|
|
51
|
-
}
|
|
52
|
-
/** The tool-restriction is codex's only sandbox (no --allowedTools flag), so it
|
|
53
|
-
* must mirror the hard-sandbox allow-list the service computes per mode — NOT a
|
|
54
|
-
* hardcoded "playwright only", which made codex refuse the active mode's plugin
|
|
55
|
-
* tools (api_request, replay_flow, hover-control, …). Built from opts.allowedTools. */
|
|
56
|
-
function codexDeveloperInstructions(allowedTools) {
|
|
57
|
-
const prefixes = (allowedTools && allowedTools.length ? allowedTools : ['mcp__playwright']).map((p) => `${p}__*`);
|
|
58
|
-
return [
|
|
59
|
-
'You are operating in Hover, a browser- and API-testing tool.',
|
|
60
|
-
`Use ONLY MCP tools whose name starts with one of these prefixes: ${prefixes.join(', ')}. They cover driving the browser, the Hover control + API-request tools, and reading source — everything the task needs.`,
|
|
61
|
-
'Do NOT call shell, file-edit, web-search, or any other built-in tool.',
|
|
62
|
-
'Do NOT navigate to a URL the user is already on; check the page state via `browser_snapshot` first.',
|
|
63
|
-
'When the task is complete, emit a short agent_message summary and stop.',
|
|
64
|
-
].join(' ');
|
|
65
|
-
}
|
|
66
|
-
export const codexAgent = {
|
|
67
|
-
id: 'codex',
|
|
68
|
-
binName: 'codex',
|
|
69
|
-
protocol: 'argv',
|
|
70
|
-
streamFormat: 'json-lines',
|
|
71
|
-
sandboxStrength: 'soft',
|
|
72
|
-
display: {
|
|
73
|
-
label: 'OpenAI Codex',
|
|
74
|
-
tagline: 'OpenAI — soft sandbox (no built-in tool deny-list)',
|
|
75
|
-
homepage: 'https://developers.openai.com/codex',
|
|
76
|
-
installHint: 'npm install -g @openai/codex',
|
|
77
|
-
},
|
|
78
|
-
buildArgs(opts) {
|
|
79
|
-
const args = ['exec'];
|
|
80
|
-
// Resume must come BEFORE the prompt positional. `codex exec resume <id>
|
|
81
|
-
// [prompt]` is the documented shape.
|
|
82
|
-
if (opts.sessionId) {
|
|
83
|
-
args.push('resume', opts.sessionId);
|
|
84
|
-
}
|
|
85
|
-
args.push(opts.prompt);
|
|
86
|
-
// JSONL streaming output.
|
|
87
|
-
args.push('--json');
|
|
88
|
-
// Never prompt for approval in headless mode.
|
|
89
|
-
args.push('--ask-for-approval', 'never');
|
|
90
|
-
// Soft sandbox: prevent shell side-effects on disk / network even when
|
|
91
|
-
// the agent tries to call its built-in shell. read-only is the strictest
|
|
92
|
-
// documented level.
|
|
93
|
-
args.push('--sandbox', 'read-only');
|
|
94
|
-
if (opts.model) {
|
|
95
|
-
args.push('--model', opts.model);
|
|
96
|
-
}
|
|
97
|
-
// Reasoning effort → the API's reasoning_effort, set via the `-c` TOML
|
|
98
|
-
// override (no stable long flag exists). Sits alongside the other -c keys.
|
|
99
|
-
if (opts.effort) {
|
|
100
|
-
args.push('-c', `model_reasoning_effort=${opts.effort}`);
|
|
101
|
-
}
|
|
102
|
-
// System-prompt injection. Codex has no --append-system-prompt; we route
|
|
103
|
-
// through `-c developer_instructions='...'`. Concatenate the standing
|
|
104
|
-
// Hover-mode instructions with whatever the caller passes (e.g. "user is
|
|
105
|
-
// already on http://localhost:5173/").
|
|
106
|
-
const base = codexDeveloperInstructions(opts.allowedTools);
|
|
107
|
-
const sysPrompt = opts.appendSystemPrompt && opts.appendSystemPrompt.trim().length > 0
|
|
108
|
-
? `${base} ${opts.appendSystemPrompt}`
|
|
109
|
-
: base;
|
|
110
|
-
args.push('-c', `developer_instructions=${JSON.stringify(sysPrompt)}`);
|
|
111
|
-
// MCP servers are configured in ~/.codex/config.toml at install time,
|
|
112
|
-
// not per-invocation. If the user passed an mcpConfig path, we don't
|
|
113
|
-
// have a way to forward it to codex — log a warning to stderr from the
|
|
114
|
-
// invoker so the user knows. (See invoke.ts wiring.)
|
|
115
|
-
// No equivalent for --max-budget-usd or --allowedTools.
|
|
116
|
-
return args;
|
|
117
|
-
},
|
|
118
|
-
parseEvent(line, state = {}) {
|
|
119
|
-
if (!line.trim())
|
|
120
|
-
return [];
|
|
121
|
-
let ev;
|
|
122
|
-
try {
|
|
123
|
-
ev = JSON.parse(line);
|
|
124
|
-
}
|
|
125
|
-
catch {
|
|
126
|
-
return [{ kind: 'raw', line }];
|
|
127
|
-
}
|
|
128
|
-
const s = codexState(state);
|
|
129
|
-
const out = [];
|
|
130
|
-
if (ev.type === 'thread.started') {
|
|
131
|
-
resetCodexCounters(s);
|
|
132
|
-
if (ev.thread_id) {
|
|
133
|
-
s.runningSessionId = ev.thread_id;
|
|
134
|
-
out.push({ kind: 'session_start', sessionId: ev.thread_id, model: ev.model });
|
|
135
|
-
}
|
|
136
|
-
return out;
|
|
137
|
-
}
|
|
138
|
-
if (ev.type === 'item.started' && ev.item) {
|
|
139
|
-
const it = ev.item;
|
|
140
|
-
if (it.id && it.type)
|
|
141
|
-
s.itemTypeById.set(it.id, it.type);
|
|
142
|
-
if (it.type === 'mcp_tool_call') {
|
|
143
|
-
// The exact field names aren't published. Read defensively: prefer
|
|
144
|
-
// `name`, fall back to `tool`. Same for input.
|
|
145
|
-
const rawName = it.name ?? it.tool ?? '';
|
|
146
|
-
const tool = stripMcpPrefix(rawName);
|
|
147
|
-
out.push({ kind: 'tool_use', tool, input: it.input ?? it.arguments, costUsdSnapshot: s.runningCost, tokensSnapshot: s.runningTokens });
|
|
148
|
-
}
|
|
149
|
-
else if (it.type === 'command_execution') {
|
|
150
|
-
// We DISCOURAGED this in developer_instructions but the agent can
|
|
151
|
-
// still try. Surface it so the user sees it happen.
|
|
152
|
-
out.push({ kind: 'tool_use', tool: 'shell', input: { command: it.command }, costUsdSnapshot: s.runningCost, tokensSnapshot: s.runningTokens });
|
|
153
|
-
}
|
|
154
|
-
return out;
|
|
155
|
-
}
|
|
156
|
-
if (ev.type === 'item.completed' && ev.item) {
|
|
157
|
-
const it = ev.item;
|
|
158
|
-
const recordedType = (it.id && s.itemTypeById.get(it.id)) || it.type;
|
|
159
|
-
if (recordedType === 'agent_message') {
|
|
160
|
-
const text = it.text?.trim();
|
|
161
|
-
if (text) {
|
|
162
|
-
s.lastAgentMessage = text;
|
|
163
|
-
out.push({ kind: 'text', text });
|
|
164
|
-
}
|
|
165
|
-
}
|
|
166
|
-
else if (recordedType === 'mcp_tool_call' || recordedType === 'command_execution') {
|
|
167
|
-
const isError = it.is_error === true ||
|
|
168
|
-
(typeof it.status === 'string' && /error|fail/i.test(it.status));
|
|
169
|
-
out.push({ kind: 'tool_result', isError });
|
|
170
|
-
}
|
|
171
|
-
return out;
|
|
172
|
-
}
|
|
173
|
-
if (ev.type === 'turn.completed') {
|
|
174
|
-
s.runningTurns += 1;
|
|
175
|
-
if (ev.usage) {
|
|
176
|
-
// The parser has no access to the invocation's --model, so we let
|
|
177
|
-
// estimateCostUsd fall back to its fixed default tier. Cost is a
|
|
178
|
-
// high-water "should I hit Stop now" signal, not an invoice.
|
|
179
|
-
s.runningCost += estimateCostUsd(undefined, ev.usage);
|
|
180
|
-
s.runningTokens += (ev.usage.input_tokens ?? 0) + (ev.usage.output_tokens ?? 0);
|
|
181
|
-
}
|
|
182
|
-
out.push({ kind: 'usage', costUsd: s.runningCost, turns: s.runningTurns, tokens: s.runningTokens });
|
|
183
|
-
return out;
|
|
184
|
-
}
|
|
185
|
-
// Codex emits various error envelopes; we conservatively match anything
|
|
186
|
-
// whose `type` contains 'error' or carries a top-level message string.
|
|
187
|
-
if (ev.type && /error/i.test(ev.type)) {
|
|
188
|
-
s.sawErrorEvent = true;
|
|
189
|
-
if (ev.message) {
|
|
190
|
-
out.push({ kind: 'text', text: `[codex] ${ev.message}` });
|
|
191
|
-
}
|
|
192
|
-
return out;
|
|
193
|
-
}
|
|
194
|
-
return [];
|
|
195
|
-
},
|
|
196
|
-
/**
|
|
197
|
-
* Codex doesn't emit a `session_end` line — the child process simply
|
|
198
|
-
* exits after the final `turn.completed`. We synthesize the terminator
|
|
199
|
-
* here so the widget sees the same shape it sees from claude.
|
|
200
|
-
*/
|
|
201
|
-
onStreamEnd(exitCode, state = {}) {
|
|
202
|
-
const s = codexState(state);
|
|
203
|
-
return {
|
|
204
|
-
kind: 'session_end',
|
|
205
|
-
turns: s.runningTurns,
|
|
206
|
-
costUsd: s.runningCost,
|
|
207
|
-
tokens: s.runningTokens,
|
|
208
|
-
isError: s.sawErrorEvent || (exitCode != null && exitCode !== 0),
|
|
209
|
-
summary: s.lastAgentMessage,
|
|
210
|
-
};
|
|
211
|
-
},
|
|
212
|
-
};
|
|
213
|
-
/**
|
|
214
|
-
* Test-only escape hatches. Tests pass a state object in and get the
|
|
215
|
-
* accumulated counters back — same shape as the parser sees during a real
|
|
216
|
-
* invocation, just driven by the test instead of by invokeAgent.
|
|
217
|
-
*/
|
|
218
|
-
export const __testing = {
|
|
219
|
-
freshState: () => ({}),
|
|
220
|
-
resetCounters: (state) => resetCodexCounters(codexState(state)),
|
|
221
|
-
getState: (state) => {
|
|
222
|
-
const s = codexState(state);
|
|
223
|
-
return {
|
|
224
|
-
runningCost: s.runningCost,
|
|
225
|
-
runningTurns: s.runningTurns,
|
|
226
|
-
runningSessionId: s.runningSessionId,
|
|
227
|
-
lastAgentMessage: s.lastAgentMessage,
|
|
228
|
-
sawErrorEvent: s.sawErrorEvent,
|
|
229
|
-
};
|
|
230
|
-
},
|
|
231
|
-
};
|
package/dist/agents/detect.d.ts
DELETED
|
@@ -1,46 +0,0 @@
|
|
|
1
|
-
import type { AgentDescriptor } from './types.js';
|
|
2
|
-
/**
|
|
3
|
-
* Find a binary on PATH. Returns absolute path or null.
|
|
4
|
-
* macOS/Linux uses `which`; Windows uses `where`.
|
|
5
|
-
*/
|
|
6
|
-
export declare function resolveOnPath(binName: string): Promise<string | null>;
|
|
7
|
-
export declare function resolveBinForAgent(descriptor: AgentDescriptor): Promise<string | null>;
|
|
8
|
-
export interface DetectedAgent {
|
|
9
|
-
descriptor: AgentDescriptor;
|
|
10
|
-
binPath: string;
|
|
11
|
-
}
|
|
12
|
-
/**
|
|
13
|
-
* Scan PATH for every agent in the registry. Returns only the ones found,
|
|
14
|
-
* in registry insertion order.
|
|
15
|
-
*
|
|
16
|
-
* Probes all agents in parallel — each `which`/`where` call is ~50-200ms
|
|
17
|
-
* and the registry's growth target is 7-10 agents, so serial would noticeably
|
|
18
|
-
* lag the first widget hello / agent-dropdown open.
|
|
19
|
-
*/
|
|
20
|
-
export declare function detectAgents(): Promise<DetectedAgent[]>;
|
|
21
|
-
export interface AgentAvailability {
|
|
22
|
-
id: string;
|
|
23
|
-
label: string;
|
|
24
|
-
tagline?: string;
|
|
25
|
-
sandboxStrength: 'hard' | 'soft';
|
|
26
|
-
installed: boolean;
|
|
27
|
-
binPath?: string;
|
|
28
|
-
homepage?: string;
|
|
29
|
-
installHint?: string;
|
|
30
|
-
}
|
|
31
|
-
/**
|
|
32
|
-
* Like `detectAgents`, but also includes registered-but-not-installed agents
|
|
33
|
-
* so the widget can render them dimmed with an install hint. Order matches
|
|
34
|
-
* the registry. Probes run in parallel — see `detectAgents`.
|
|
35
|
-
*/
|
|
36
|
-
export declare function listAgentAvailability(): Promise<AgentAvailability[]>;
|
|
37
|
-
/**
|
|
38
|
-
* Pick the agent we should default to when the user / Vite plugin didn't
|
|
39
|
-
* specify one. Prefer the explicit hint if it's installed; otherwise the
|
|
40
|
-
* first registered agent that's installed; finally null if nothing matches.
|
|
41
|
-
*
|
|
42
|
-
* `preferredId` is typically `process.env.HOVER_AGENT` or the value the user
|
|
43
|
-
* picked in the widget last session (persisted by the widget to localStorage).
|
|
44
|
-
*/
|
|
45
|
-
export declare function pickPrimaryAgent(preferredId?: string): Promise<DetectedAgent | null>;
|
|
46
|
-
//# sourceMappingURL=detect.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"detect.d.ts","sourceRoot":"","sources":["../../src/agents/detect.ts"],"names":[],"mappings":"AAGA,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,YAAY,CAAC;AAIlD;;;GAGG;AACH,wBAAsB,aAAa,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAAC,MAAM,GAAG,IAAI,CAAC,CAS3E;AAED,wBAAsB,kBAAkB,CAAC,UAAU,EAAE,eAAe,GAAG,OAAO,CAAC,MAAM,GAAG,IAAI,CAAC,CAE5F;AAED,MAAM,WAAW,aAAa;IAC5B,UAAU,EAAE,eAAe,CAAC;IAC5B,OAAO,EAAE,MAAM,CAAC;CACjB;AAED;;;;;;;GAOG;AACH,wBAAsB,YAAY,IAAI,OAAO,CAAC,aAAa,EAAE,CAAC,CAO7D;AAED,MAAM,WAAW,iBAAiB;IAChC,EAAE,EAAE,MAAM,CAAC;IACX,KAAK,EAAE,MAAM,CAAC;IACd,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,eAAe,EAAE,MAAM,GAAG,MAAM,CAAC;IACjC,SAAS,EAAE,OAAO,CAAC;IACnB,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,WAAW,CAAC,EAAE,MAAM,CAAC;CACtB;AAED;;;;GAIG;AACH,wBAAsB,qBAAqB,IAAI,OAAO,CAAC,iBAAiB,EAAE,CAAC,CAgB1E;AAED;;;;;;;GAOG;AACH,wBAAsB,gBAAgB,CAAC,WAAW,CAAC,EAAE,MAAM,GAAG,OAAO,CAAC,aAAa,GAAG,IAAI,CAAC,CAU1F"}
|
package/dist/agents/detect.js
DELETED
|
@@ -1,80 +0,0 @@
|
|
|
1
|
-
import { execFile } from 'node:child_process';
|
|
2
|
-
import { promisify } from 'node:util';
|
|
3
|
-
import { AGENTS, listAgents } from './registry.js';
|
|
4
|
-
const execFileAsync = promisify(execFile);
|
|
5
|
-
/**
|
|
6
|
-
* Find a binary on PATH. Returns absolute path or null.
|
|
7
|
-
* macOS/Linux uses `which`; Windows uses `where`.
|
|
8
|
-
*/
|
|
9
|
-
export async function resolveOnPath(binName) {
|
|
10
|
-
const tool = process.platform === 'win32' ? 'where' : 'which';
|
|
11
|
-
try {
|
|
12
|
-
const { stdout } = await execFileAsync(tool, [binName]);
|
|
13
|
-
const first = stdout.split('\n')[0]?.trim();
|
|
14
|
-
return first && first.length > 0 ? first : null;
|
|
15
|
-
}
|
|
16
|
-
catch {
|
|
17
|
-
return null;
|
|
18
|
-
}
|
|
19
|
-
}
|
|
20
|
-
export async function resolveBinForAgent(descriptor) {
|
|
21
|
-
return resolveOnPath(descriptor.binName);
|
|
22
|
-
}
|
|
23
|
-
/**
|
|
24
|
-
* Scan PATH for every agent in the registry. Returns only the ones found,
|
|
25
|
-
* in registry insertion order.
|
|
26
|
-
*
|
|
27
|
-
* Probes all agents in parallel — each `which`/`where` call is ~50-200ms
|
|
28
|
-
* and the registry's growth target is 7-10 agents, so serial would noticeably
|
|
29
|
-
* lag the first widget hello / agent-dropdown open.
|
|
30
|
-
*/
|
|
31
|
-
export async function detectAgents() {
|
|
32
|
-
const descriptors = listAgents();
|
|
33
|
-
const paths = await Promise.all(descriptors.map(d => resolveBinForAgent(d)));
|
|
34
|
-
return descriptors.flatMap((descriptor, i) => {
|
|
35
|
-
const binPath = paths[i];
|
|
36
|
-
return binPath ? [{ descriptor, binPath }] : [];
|
|
37
|
-
});
|
|
38
|
-
}
|
|
39
|
-
/**
|
|
40
|
-
* Like `detectAgents`, but also includes registered-but-not-installed agents
|
|
41
|
-
* so the widget can render them dimmed with an install hint. Order matches
|
|
42
|
-
* the registry. Probes run in parallel — see `detectAgents`.
|
|
43
|
-
*/
|
|
44
|
-
export async function listAgentAvailability() {
|
|
45
|
-
const descriptors = listAgents();
|
|
46
|
-
const paths = await Promise.all(descriptors.map(d => resolveBinForAgent(d)));
|
|
47
|
-
return descriptors.map((descriptor, i) => {
|
|
48
|
-
const binPath = paths[i];
|
|
49
|
-
return {
|
|
50
|
-
id: descriptor.id,
|
|
51
|
-
label: descriptor.display.label,
|
|
52
|
-
tagline: descriptor.display.tagline,
|
|
53
|
-
sandboxStrength: descriptor.sandboxStrength,
|
|
54
|
-
installed: binPath != null,
|
|
55
|
-
binPath: binPath ?? undefined,
|
|
56
|
-
homepage: descriptor.display.homepage,
|
|
57
|
-
installHint: descriptor.display.installHint,
|
|
58
|
-
};
|
|
59
|
-
});
|
|
60
|
-
}
|
|
61
|
-
/**
|
|
62
|
-
* Pick the agent we should default to when the user / Vite plugin didn't
|
|
63
|
-
* specify one. Prefer the explicit hint if it's installed; otherwise the
|
|
64
|
-
* first registered agent that's installed; finally null if nothing matches.
|
|
65
|
-
*
|
|
66
|
-
* `preferredId` is typically `process.env.HOVER_AGENT` or the value the user
|
|
67
|
-
* picked in the widget last session (persisted by the widget to localStorage).
|
|
68
|
-
*/
|
|
69
|
-
export async function pickPrimaryAgent(preferredId) {
|
|
70
|
-
if (preferredId) {
|
|
71
|
-
const descriptor = AGENTS[preferredId];
|
|
72
|
-
if (descriptor) {
|
|
73
|
-
const binPath = await resolveBinForAgent(descriptor);
|
|
74
|
-
if (binPath)
|
|
75
|
-
return { descriptor, binPath };
|
|
76
|
-
}
|
|
77
|
-
}
|
|
78
|
-
const detected = await detectAgents();
|
|
79
|
-
return detected[0] ?? null;
|
|
80
|
-
}
|
package/dist/agents/gemini.d.ts
DELETED
|
@@ -1,17 +0,0 @@
|
|
|
1
|
-
import type { AgentDescriptor, ParserState } from './types.js';
|
|
2
|
-
export declare const geminiAgent: AgentDescriptor;
|
|
3
|
-
/**
|
|
4
|
-
* Test-only escape hatches, same pattern as cursor.ts / codex.ts.
|
|
5
|
-
*/
|
|
6
|
-
export declare const __testing: {
|
|
7
|
-
freshState: () => ParserState;
|
|
8
|
-
resetCounters: (state: ParserState) => void;
|
|
9
|
-
getState: (state: ParserState) => {
|
|
10
|
-
runningTurns: number;
|
|
11
|
-
runningSessionId: string | undefined;
|
|
12
|
-
runningModel: string | undefined;
|
|
13
|
-
lastAssistantText: string | undefined;
|
|
14
|
-
sawErrorEvent: boolean;
|
|
15
|
-
};
|
|
16
|
-
};
|
|
17
|
-
//# sourceMappingURL=gemini.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"gemini.d.ts","sourceRoot":"","sources":["../../src/agents/gemini.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,eAAe,EAA8B,WAAW,EAAE,MAAM,YAAY,CAAC;AA+J3F,eAAO,MAAM,WAAW,EAAE,eAiJzB,CAAC;AAEF;;GAEG;AACH,eAAO,MAAM,SAAS;sBACJ,WAAW;2BACJ,WAAW;sBAChB,WAAW;;;;;;;CAU9B,CAAC"}
|