@phren/agent 0.3.20 → 0.3.22

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.
Files changed (48) hide show
  1. package/dist/agent-loop/index.js +113 -19
  2. package/dist/agent-loop/stream.js +3 -1
  3. package/dist/attach-images.js +57 -0
  4. package/dist/commands/info.js +6 -2
  5. package/dist/commands/model.js +55 -13
  6. package/dist/commands/session.js +79 -18
  7. package/dist/commands.js +9 -2
  8. package/dist/config.js +45 -0
  9. package/dist/context/clear-tool-results.js +43 -0
  10. package/dist/context/compactor.js +8 -2
  11. package/dist/context/usage.js +26 -0
  12. package/dist/cost.js +10 -3
  13. package/dist/headless.js +41 -0
  14. package/dist/index.js +180 -56
  15. package/dist/mcp-client.js +38 -1
  16. package/dist/mcp-prompts.js +94 -0
  17. package/dist/multi/child-entry.js +3 -1
  18. package/dist/multi/spawner.js +27 -15
  19. package/dist/permissions/allowlist.js +19 -4
  20. package/dist/permissions/checker.js +37 -20
  21. package/dist/permissions/kernel-sandbox.js +32 -10
  22. package/dist/permissions/rules.js +173 -0
  23. package/dist/permissions/shell-classify.js +173 -0
  24. package/dist/plan.js +6 -2
  25. package/dist/repl.js +16 -1
  26. package/dist/session/persist.js +9 -4
  27. package/dist/settings.js +14 -0
  28. package/dist/structured-output.js +85 -0
  29. package/dist/system-prompt.js +117 -14
  30. package/dist/tools/apply-patch.js +15 -1
  31. package/dist/tools/edit-file.js +8 -1
  32. package/dist/tools/file-state.js +88 -0
  33. package/dist/tools/glob.js +53 -35
  34. package/dist/tools/grep.js +139 -53
  35. package/dist/tools/lint-test.js +24 -0
  36. package/dist/tools/read-file.js +2 -0
  37. package/dist/tools/registry.js +8 -1
  38. package/dist/tools/search-support.js +141 -0
  39. package/dist/tools/shell.js +3 -1
  40. package/dist/tools/spawn-agent.js +15 -3
  41. package/dist/tools/syntax-check.js +116 -0
  42. package/dist/tools/write-file.js +11 -3
  43. package/dist/tui/components/App.js +5 -4
  44. package/dist/tui/components/Banner.js +1 -1
  45. package/dist/tui/components/ListPicker.js +32 -0
  46. package/dist/tui/ink-entry.js +57 -16
  47. package/dist/user-hooks.js +26 -4
  48. package/package.json +3 -2
package/dist/commands.js CHANGED
@@ -4,7 +4,7 @@ import { sessionCommand, historyCommand, compactCommand, diffCommand, gitCommand
4
4
  import { memCommand, askCommand } from "./commands/memory.js";
5
5
  import { reviewCommand } from "./commands/review.js";
6
6
  import { findSkill, getScopedSkills } from "@phren/cli/skill/registry";
7
- import { modelCommand, providerCommand, presetCommand } from "./commands/model.js";
7
+ import { modelCommand, providerCommand, presetCommand, reasoningCommand } from "./commands/model.js";
8
8
  import { configCommand } from "./commands/config.js";
9
9
  import { loadInputMode, saveInputMode, savePermissionMode } from "./settings.js";
10
10
  import { addAllow } from "./permissions/allowlist.js";
@@ -20,7 +20,7 @@ export function createCommandContext(session, contextLimit) {
20
20
  }
21
21
  const BUILTIN_COMMAND_NAMES = [
22
22
  "/help", "/turns", "/clear", "/cwd", "/files", "/cost", "/plan", "/undo",
23
- "/context", "/model", "/provider", "/preset", "/session", "/history",
23
+ "/context", "/model", "/provider", "/reasoning", "/preset", "/session", "/history",
24
24
  "/compact", "/diff", "/git", "/mem", "/ask", "/resume", "/review", "/config", "/spawn", "/agents",
25
25
  "/allow",
26
26
  "/mode", "/permissions", "/verbose", "/theme", "/agent", "/rewind", "/fork", "/promote",
@@ -38,6 +38,12 @@ export function setCustomCommands(commands) {
38
38
  COMMAND_NAMES.length = 0;
39
39
  COMMAND_NAMES.push(...BUILTIN_COMMAND_NAMES, ...customCommands.map((command) => `/${command.name}`));
40
40
  }
41
+ /** Extra command names for completion (MCP prompts, found after servers connect). */
42
+ export function addCommandNames(names) {
43
+ for (const name of names)
44
+ if (!COMMAND_NAMES.includes(name))
45
+ COMMAND_NAMES.push(name);
46
+ }
41
47
  export function loadAndRegisterCustomCommands(cwd = process.cwd()) {
42
48
  const commands = loadCustomCommands(cwd);
43
49
  setCustomCommands(commands);
@@ -122,6 +128,7 @@ export function handleCommand(input, ctx) {
122
128
  case "/context": return contextCommand(parts, ctx);
123
129
  case "/model": return modelCommand(parts, ctx);
124
130
  case "/provider": return providerCommand(parts, ctx);
131
+ case "/reasoning": return reasoningCommand(parts, ctx);
125
132
  case "/preset": return presetCommand(parts, ctx);
126
133
  case "/session": return sessionCommand(parts, ctx);
127
134
  case "/history": return historyCommand(parts, ctx);
package/dist/config.js CHANGED
@@ -1,5 +1,6 @@
1
1
  import * as os from "node:os";
2
2
  import * as path from "node:path";
3
+ import { parseRuleList } from "./permissions/rules.js";
3
4
  import { normalizeReasoningEffort } from "./models.js";
4
5
  import { loadPermissionMode } from "./settings.js";
5
6
  export const AGENT_CONFIG_DIR = ".phren-agent";
@@ -34,8 +35,11 @@ Options:
34
35
  --no-subagents Disable spawn_agent/send_message/list_agents in one-shot mode
35
36
  --no-llm-compact Use regex prune summaries instead of LLM compaction
36
37
  --sandbox <mode> Kernel write-fence for shell (bwrap): off, auto (default), require
38
+ --no-network Run shell commands without network (bwrap or Seatbelt; fails closed)
37
39
  --permissions <mode> Permission mode: suggest (default), auto-confirm, full-auto
38
40
  --yolo Full-auto permissions — no confirmations (alias for --permissions full-auto)
41
+ --allowedTools <rules> Allow these without asking, comma-separated: read_file, shell(npm test), edit_file(src/**)
42
+ --disallowedTools <rules> Deny these in every mode, same syntax
39
43
  --interactive, -i Interactive REPL mode (multi-turn conversation)
40
44
  --resume, -c Resume the newest session's conversation (task optional)
41
45
  --session <id> Resume a specific session by id or unique id prefix
@@ -43,10 +47,18 @@ Options:
43
47
  -p, --print Headless: no prompts (tool approvals are denied), clean stdout
44
48
  --output-format <f> With -p: text (final message), json (one result object),
45
49
  stream-json (NDJSON events + result). Implies -p
50
+ --json-schema <s> With -p: end with a JSON value matching this schema (inline JSON or a
51
+ file), in the result's structured_output; implies -p
52
+ --input-format <f> With -p: text (the task, default) or stream-json (user
53
+ messages as JSON lines on stdin, one turn each, same session;
54
+ needs --output-format stream-json)
46
55
  --lint-cmd <cmd> Override auto-detected lint command
47
56
  --test-cmd <cmd> Override auto-detected test command
57
+ --typecheck-cmd <cmd> Override the auto-detected type check run after edits
48
58
  --mcp <command> Connect to an MCP server via stdio (repeatable)
49
59
  --mcp-config <path> Load MCP server config from JSON file
60
+ --trust-project-mcp Load this project's .mcp.json and .phren-agent/mcp.json (remembered)
61
+ --strict-mcp-config Use only --mcp-config and --mcp servers, no default config files
50
62
  --team <name> Start in team mode with named team coordination
51
63
  --multi Start in multi-agent TUI mode
52
64
  --dry-run Show system prompt and exit
@@ -92,6 +104,8 @@ export function parseArgs(argv) {
92
104
  task: "",
93
105
  permissions: "suggest",
94
106
  permissionsExplicit: false,
107
+ allowedTools: [],
108
+ disallowedTools: [],
95
109
  maxTurns: 50,
96
110
  budget: null,
97
111
  plan: false,
@@ -102,6 +116,7 @@ export function parseArgs(argv) {
102
116
  listSessions: false,
103
117
  print: false,
104
118
  outputFormat: "text",
119
+ inputFormat: "text",
105
120
  noSubagents: false,
106
121
  noLlmCompact: false,
107
122
  sandbox: "auto",
@@ -136,6 +151,9 @@ export function parseArgs(argv) {
136
151
  else if (arg === "--no-llm-compact") {
137
152
  args.noLlmCompact = true;
138
153
  }
154
+ else if (arg === "--no-network") {
155
+ args.noNetwork = true;
156
+ }
139
157
  else if (arg === "--sandbox" && argv[i + 1]) {
140
158
  const mode = argv[++i];
141
159
  if (mode === "off" || mode === "auto" || mode === "require") {
@@ -164,6 +182,18 @@ export function parseArgs(argv) {
164
182
  else if (arg === "--print" || arg === "-p") {
165
183
  args.print = true;
166
184
  }
185
+ else if (arg === "--json-schema" && argv[i + 1]) {
186
+ args.jsonSchema = argv[++i];
187
+ args.print = true;
188
+ }
189
+ else if (arg === "--input-format" && argv[i + 1]) {
190
+ const format = argv[++i];
191
+ if (format !== "text" && format !== "stream-json")
192
+ throw new Error(`Unknown --input-format "${format}". Use text or stream-json.`);
193
+ args.inputFormat = format;
194
+ if (format === "stream-json")
195
+ args.print = true;
196
+ }
167
197
  else if (arg === "--output-format" && argv[i + 1]) {
168
198
  const format = argv[++i];
169
199
  if (format === "text" || format === "json" || format === "stream-json") {
@@ -183,12 +213,21 @@ export function parseArgs(argv) {
183
213
  else if (arg === "--test-cmd" && argv[i + 1]) {
184
214
  args.testCmd = argv[++i];
185
215
  }
216
+ else if (arg === "--typecheck-cmd" && argv[i + 1]) {
217
+ args.typecheckCmd = argv[++i];
218
+ }
186
219
  else if (arg === "--mcp" && argv[i + 1]) {
187
220
  args.mcp.push(argv[++i]);
188
221
  }
189
222
  else if (arg === "--mcp-config" && argv[i + 1]) {
190
223
  args.mcpConfig = argv[++i];
191
224
  }
225
+ else if (arg === "--trust-project-mcp") {
226
+ args.trustProjectMcp = true;
227
+ }
228
+ else if (arg === "--strict-mcp-config") {
229
+ args.strictMcpConfig = true;
230
+ }
192
231
  else if (arg === "--team" && argv[i + 1]) {
193
232
  args.team = argv[++i];
194
233
  }
@@ -228,6 +267,12 @@ export function parseArgs(argv) {
228
267
  else if (arg === "--budget" && argv[i + 1]) {
229
268
  args.budget = parseFloat(argv[++i]) || null;
230
269
  }
270
+ else if ((arg === "--allowedTools" || arg === "--allowed-tools") && argv[i + 1]) {
271
+ args.allowedTools.push(...parseRuleList(argv[++i]));
272
+ }
273
+ else if ((arg === "--disallowedTools" || arg === "--disallowed-tools") && argv[i + 1]) {
274
+ args.disallowedTools.push(...parseRuleList(argv[++i]));
275
+ }
231
276
  else if (arg === "--yolo") {
232
277
  args.permissions = "full-auto";
233
278
  args.permissionsExplicit = true;
@@ -0,0 +1,43 @@
1
+ import { toolResultText } from "../providers/types.js";
2
+ export const DEFAULT_CLEAR = { keepRecent: 8, minChars: 2_000 };
3
+ const CLEARED_PREFIX = "[Earlier output cleared to save context";
4
+ /** Plan which old tool results to clear. Returns [] when there is nothing worth clearing. */
5
+ export function planToolResultClearing(messages, config = {}) {
6
+ const { keepRecent, minChars } = { ...DEFAULT_CLEAR, ...config };
7
+ const names = new Map();
8
+ const results = [];
9
+ messages.forEach((msg, index) => {
10
+ if (typeof msg.content === "string")
11
+ return;
12
+ for (const block of msg.content) {
13
+ if (block.type === "tool_use")
14
+ names.set(block.id, block.name);
15
+ else if (block.type === "tool_result")
16
+ results.push({ index, block });
17
+ }
18
+ });
19
+ const clear = new Map();
20
+ for (const { index, block } of results.slice(0, Math.max(0, results.length - keepRecent))) {
21
+ const text = toolResultText(block);
22
+ const hasImage = Array.isArray(block.content) && block.content.some((part) => part.type === "image");
23
+ if (text.startsWith(CLEARED_PREFIX) || (text.length < minChars && !hasImage))
24
+ continue;
25
+ if (!clear.has(index))
26
+ clear.set(index, new Set());
27
+ clear.get(index).add(block);
28
+ }
29
+ return [...clear].map(([index, blocks]) => {
30
+ const content = messages[index].content.map((block) => {
31
+ if (block.type !== "tool_result" || !blocks.has(block))
32
+ return block;
33
+ const text = toolResultText(block);
34
+ const tool = names.get(block.tool_use_id) ?? "the tool";
35
+ const firstLine = text.split("\n", 1)[0].slice(0, 120);
36
+ return {
37
+ ...block,
38
+ content: `${CLEARED_PREFIX}: ${tool} returned ${text.length} chars, starting "${firstLine}". Run it again if you need it.]`,
39
+ };
40
+ });
41
+ return { index, message: { ...messages[index], content } };
42
+ });
43
+ }
@@ -32,6 +32,12 @@ The original task; work completed so far; files created or modified; the current
32
32
  ## Knowledge
33
33
  A fenced json block: {"items":[{"text":"...","confidence":0.9,"kind":"finding|gotcha|decision"}]}
34
34
  Only durable, non-obvious, project-level knowledge worth remembering across sessions (root causes, architecture decisions with rationale, gotchas, workarounds). Confidence is YOUR certainty the item is true and durable, 0 to 1. Use {"items":[]} if nothing qualifies. Never include secrets or credentials.`;
35
+ function checkpointInstruction(focus) {
36
+ const trimmed = focus?.trim();
37
+ if (!trimmed)
38
+ return CHECKPOINT_INSTRUCTION;
39
+ return `${CHECKPOINT_INSTRUCTION}\n\nThe user asked the summary to focus on: ${trimmed}. Keep everything about it; shorten the rest.`;
40
+ }
35
41
  // ── Response parsing ─────────────────────────────────────────────────────────
36
42
  const KNOWLEDGE_HEADING_RE = /^##\s*Knowledge\b.*$/im;
37
43
  const SUMMARY_HEADING_RE = /^##\s*Checkpoint Summary\b.*$/im;
@@ -184,12 +190,12 @@ export async function compactWithLlm(provider, systemPrompt, messages, opts = {}
184
190
  // covers the prefix, so this costs roughly one summary generation.
185
191
  const prefixMessages = [
186
192
  ...messages.slice(0, plan.endIndex + 1),
187
- { role: "user", content: CHECKPOINT_INSTRUCTION },
193
+ { role: "user", content: checkpointInstruction(opts.focus) },
188
194
  ];
189
195
  // A timeout or turn abort settles the wait; finally() then cancels the
190
196
  // request itself so it doesn't keep generating (and billing) detached.
191
197
  const cancel = new AbortController();
192
- const response = await withTimeout(provider.chat(systemPrompt, prefixMessages, [], cancel.signal), config.timeoutMs, opts.signal).finally(() => cancel.abort());
198
+ const response = await withTimeout(provider.chat(systemPrompt, prefixMessages, opts.tools ?? [], cancel.signal), config.timeoutMs, opts.signal).finally(() => cancel.abort());
193
199
  if (response.usage && opts.costTracker) {
194
200
  recordTokenUsage(opts.costTracker, response.usage);
195
201
  }
@@ -0,0 +1,26 @@
1
+ import { estimateMessageTokens, estimateTokens } from "./token-counter.js";
2
+ /** Record the size the provider reported, after its reply is in the log. */
3
+ export function reportedContext(usage, log) {
4
+ if (!usage)
5
+ return undefined;
6
+ const prompt = usage.input_tokens + (usage.cache_read_input_tokens ?? 0) + (usage.cache_creation_input_tokens ?? 0);
7
+ // A provider that reports nothing useful (0 prompt tokens) leaves the estimate in charge.
8
+ if (prompt <= 0)
9
+ return undefined;
10
+ return { tokens: prompt + usage.output_tokens, messageCount: log.getMessages().length, logLength: log.length };
11
+ }
12
+ /** Tokens the next request will carry: reported plus estimated growth, or the estimate alone. */
13
+ export function contextTokens(systemPrompt, messages, log, reported) {
14
+ if (reported && messages.length >= reported.messageCount && log.length >= reported.logLength && !replacedSince(log, reported.logLength)) {
15
+ return reported.tokens + estimateMessageTokens(messages.slice(reported.messageCount));
16
+ }
17
+ return estimateTokens(systemPrompt) + estimateMessageTokens(messages);
18
+ }
19
+ function replacedSince(log, from) {
20
+ const events = log.all;
21
+ for (let i = from; i < events.length; i++) {
22
+ if (events[i].type === "log/replace")
23
+ return true;
24
+ }
25
+ return false;
26
+ }
package/dist/cost.js CHANGED
@@ -36,12 +36,12 @@ export function resolvePricing(model, provider, baseUrl) {
36
36
  /** Prompt-cache writes cost this much more than ordinary input. */
37
37
  export const CACHE_WRITE_MULTIPLIER = 1.25;
38
38
  export function createCostTracker(model, budget = null, provider, baseUrl) {
39
- const { pricing, metered } = resolvePricing(model, provider, baseUrl);
39
+ let { pricing, metered } = resolvePricing(model, provider, baseUrl);
40
40
  // Without a cache price, hits bill as ordinary input (no discount assumed).
41
- const cacheReadPer1M = pricing.cacheReadPer1M ?? pricing.inputPer1M;
41
+ let cacheReadPer1M = pricing.cacheReadPer1M ?? pricing.inputPer1M;
42
42
  // Cache writes bill at 1.25x input (Anthropic's 5-minute cache, the one
43
43
  // the agent asks for); providers that cache implicitly report none.
44
- const cacheWritePer1M = pricing.inputPer1M * CACHE_WRITE_MULTIPLIER;
44
+ let cacheWritePer1M = pricing.inputPer1M * CACHE_WRITE_MULTIPLIER;
45
45
  const price = (input, output, cacheRead, cacheWrite) => (input / 1_000_000) * pricing.inputPer1M +
46
46
  (output / 1_000_000) * pricing.outputPer1M +
47
47
  (cacheRead / 1_000_000) * cacheReadPer1M +
@@ -61,6 +61,13 @@ export function createCostTracker(model, budget = null, provider, baseUrl) {
61
61
  tracker.totalCacheWriteTokens += cacheWriteTokens;
62
62
  tracker.totalCost += price(inputTokens, outputTokens, cacheReadTokens, cacheWriteTokens);
63
63
  },
64
+ reprice(nextModel, nextProvider, nextBaseUrl) {
65
+ ({ pricing, metered } = resolvePricing(nextModel, nextProvider, nextBaseUrl));
66
+ cacheReadPer1M = pricing.cacheReadPer1M ?? pricing.inputPer1M;
67
+ cacheWritePer1M = pricing.inputPer1M * CACHE_WRITE_MULTIPLIER;
68
+ // Money already spent on a metered model still counts against --budget.
69
+ tracker.metered = metered || tracker.totalCost > 0;
70
+ },
64
71
  isOverBudget() {
65
72
  return tracker.metered && budget !== null && tracker.totalCost >= budget;
66
73
  },
package/dist/headless.js CHANGED
@@ -4,6 +4,7 @@ const SUBTYPE = {
4
4
  budget: "error_budget",
5
5
  plan_rejected: "error_plan_rejected",
6
6
  aborted: "cancelled",
7
+ hook_blocked: "error_hook_blocked",
7
8
  };
8
9
  export function buildHeadlessResult(opts) {
9
10
  const subtype = opts.stopReason === "error" ? "error_during_execution" : SUBTYPE[opts.stopReason];
@@ -111,6 +112,46 @@ export function parseOutputFormat(raw) {
111
112
  return raw;
112
113
  return null;
113
114
  }
115
+ /**
116
+ * One `--input-format stream-json` line: Claude Code's shape,
117
+ * {"type":"user","message":{"role":"user","content":"…" | [{"type":"text","text":"…"}]}}.
118
+ * Returns the prompt text, or an error naming what was wrong.
119
+ */
120
+ export function parseStreamJsonInput(line) {
121
+ let event;
122
+ try {
123
+ event = JSON.parse(line);
124
+ }
125
+ catch {
126
+ return { error: "not JSON" };
127
+ }
128
+ const e = event;
129
+ if (e?.type !== "user" || !e.message || e.message.role !== "user")
130
+ return { error: "expected {\"type\":\"user\",\"message\":{\"role\":\"user\",…}}" };
131
+ const content = e.message.content;
132
+ const prompt = typeof content === "string"
133
+ ? content
134
+ : Array.isArray(content)
135
+ ? content.filter((b) => b?.type === "text" && typeof b.text === "string").map((b) => b.text).join("\n")
136
+ : "";
137
+ return prompt.trim() ? { prompt } : { error: "the message has no text" };
138
+ }
139
+ /** Non-empty lines from a stream, as they arrive. */
140
+ export async function* readLines(stream = process.stdin) {
141
+ let buffer = "";
142
+ for await (const chunk of stream) {
143
+ buffer += typeof chunk === "string" ? chunk : chunk.toString("utf-8");
144
+ let nl;
145
+ while ((nl = buffer.indexOf("\n")) >= 0) {
146
+ const line = buffer.slice(0, nl).trim();
147
+ buffer = buffer.slice(nl + 1);
148
+ if (line)
149
+ yield line;
150
+ }
151
+ }
152
+ if (buffer.trim())
153
+ yield buffer.trim();
154
+ }
114
155
  /** Read the whole of stdin (for `echo task | phren-agent -p`). */
115
156
  export async function readStdin(stream = process.stdin) {
116
157
  const chunks = [];