agent-sh 0.1.0 → 0.3.0

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.
@@ -16,8 +16,11 @@ export declare class AcpClient {
16
16
  private terminalCounter;
17
17
  private fileWatcher;
18
18
  private pendingToolCalls;
19
+ private autoCancelled;
19
20
  private pendingToolCounter;
20
21
  private agentInfo;
22
+ private modes;
23
+ private currentModeId;
21
24
  constructor(opts: {
22
25
  bus: EventBus;
23
26
  contextManager: ContextManager;
@@ -28,6 +31,12 @@ export declare class AcpClient {
28
31
  * Send a user query to the agent.
29
32
  */
30
33
  sendPrompt(query: string): Promise<void>;
34
+ /**
35
+ * Silently cancel the prompt after a shell tool completes.
36
+ * Unlike user-initiated cancel(), this doesn't show "(cancelled)" —
37
+ * the tool already ran, we just skip the unnecessary LLM follow-up.
38
+ */
39
+ private autoCancel;
31
40
  /**
32
41
  * Cancel the current prompt and force-recover shell mode.
33
42
  */
@@ -48,10 +57,25 @@ export declare class AcpClient {
48
57
  version: string;
49
58
  } | null;
50
59
  getModel(): string | undefined;
60
+ /**
61
+ * Get the current mode (e.g. thinking level).
62
+ */
63
+ getCurrentMode(): {
64
+ id: string;
65
+ name: string;
66
+ } | null;
51
67
  /**
52
68
  * Check if agent is connected.
53
69
  */
54
70
  isConnected(): boolean;
71
+ /**
72
+ * Parse modes from a session response and notify listeners.
73
+ */
74
+ private updateModes;
75
+ /**
76
+ * Cycle to the next session mode.
77
+ */
78
+ private cycleMode;
55
79
  private log;
56
80
  /**
57
81
  * Create the Client handler that responds to agent requests.
@@ -21,9 +21,12 @@ export class AcpClient {
21
21
  terminalDonePromises = new Map();
22
22
  terminalCounter = 0;
23
23
  fileWatcher;
24
- pendingToolCalls = new Map();
24
+ pendingToolCalls = new Map(); // toolCallId → title
25
+ autoCancelled = false;
25
26
  pendingToolCounter = 0;
26
27
  agentInfo = null;
28
+ modes = [];
29
+ currentModeId = null;
27
30
  constructor(opts) {
28
31
  this.bus = opts.bus;
29
32
  this.contextManager = opts.contextManager;
@@ -32,10 +35,20 @@ export class AcpClient {
32
35
  }
33
36
  async start() {
34
37
  this.log(`Starting agent: ${this.config.agentCommand} ${this.config.agentArgs.join(" ")}`);
35
- // Spawn the agent subprocess
36
- // Spawn the agent wait briefly to catch ENOENT and other spawn errors
38
+ // Spawn the agent subprocess with the user's full shell environment
39
+ // (includes vars from .zshrc/.bashrc that process.env may not have).
40
+ // Merge in any runtime env vars set by extensions (e.g. AGENT_SH_SOCKET)
41
+ // that weren't present when shellEnv was captured at startup.
42
+ const baseEnv = this.config.shellEnv ?? process.env;
43
+ const agentEnv = { ...baseEnv };
44
+ for (const [k, v] of Object.entries(process.env)) {
45
+ if (v !== undefined && !(k in agentEnv)) {
46
+ agentEnv[k] = v;
47
+ }
48
+ }
37
49
  this.agentProcess = spawn(this.config.agentCommand, this.config.agentArgs, {
38
50
  stdio: ["pipe", "pipe", process.env.DEBUG ? "inherit" : "ignore"],
51
+ env: agentEnv,
39
52
  });
40
53
  // Catch spawn errors (ENOENT, EACCES, etc.) before proceeding
41
54
  await new Promise((resolve, reject) => {
@@ -65,19 +78,23 @@ export class AcpClient {
65
78
  this.log("Creating ACP connection");
66
79
  // Create the client-side connection, providing our Client handler
67
80
  this.connection = new acp.ClientSideConnection((_agent) => this.createClientHandler(), stream);
68
- // Initialize the connection
81
+ // Initialize the connection with timeout
69
82
  this.log("Sending initialize request");
70
- const initResponse = await this.connection.initialize({
71
- protocolVersion: acp.PROTOCOL_VERSION,
72
- clientInfo: { name: "agent-sh", version: "0.1.0" },
73
- clientCapabilities: {
74
- terminal: true,
75
- fs: {
76
- readTextFile: true,
77
- writeTextFile: true,
83
+ const initTimeoutMs = 30000; // 30 seconds
84
+ const initResponse = await Promise.race([
85
+ this.connection.initialize({
86
+ protocolVersion: acp.PROTOCOL_VERSION,
87
+ clientInfo: { name: "agent-sh", version: "0.1.0" },
88
+ clientCapabilities: {
89
+ terminal: true,
90
+ fs: {
91
+ readTextFile: true,
92
+ writeTextFile: true,
93
+ },
78
94
  },
79
- },
80
- });
95
+ }),
96
+ new Promise((_, reject) => setTimeout(() => reject(new Error(`Initialize timeout after ${initTimeoutMs}ms`)), initTimeoutMs)),
97
+ ]);
81
98
  this.log("Initialize successful");
82
99
  // Store agent info for display
83
100
  if (initResponse.agentInfo) {
@@ -87,15 +104,27 @@ export class AcpClient {
87
104
  };
88
105
  this.log(`Agent info: ${this.agentInfo.name} v${this.agentInfo.version}`);
89
106
  }
90
- // Create a session
107
+ // Create a session — let extensions add MCP servers via pipe
91
108
  const cwd = this.contextManager.getCwd();
92
109
  this.log(`Creating new session with cwd: ${cwd}`);
93
- const sessionResponse = await this.connection.newSession({
110
+ const sessionConfig = this.bus.emitPipe("session:configure", {
94
111
  cwd,
95
112
  mcpServers: [],
96
113
  });
114
+ const sessionTimeoutMs = 30000; // 30 seconds
115
+ const sessionResponse = await Promise.race([
116
+ this.connection.newSession({
117
+ cwd: sessionConfig.cwd,
118
+ mcpServers: sessionConfig.mcpServers,
119
+ }),
120
+ new Promise((_, reject) => setTimeout(() => reject(new Error(`newSession timeout after ${sessionTimeoutMs}ms`)), sessionTimeoutMs)),
121
+ ]);
97
122
  this.sessionId = sessionResponse.sessionId;
98
123
  this.log(`Session created: ${this.sessionId}`);
124
+ // Parse session modes (thinking level, etc.)
125
+ this.updateModes(sessionResponse);
126
+ // Listen for mode cycle requests from input handler
127
+ this.bus.on("config:cycle", () => this.cycleMode());
99
128
  }
100
129
  /**
101
130
  * Send a user query to the agent.
@@ -109,6 +138,7 @@ export class AcpClient {
109
138
  this.bus.emit("agent:processing-start", {});
110
139
  await this.fileWatcher.snapshot();
111
140
  this.currentResponseText = "";
141
+ this.autoCancelled = false;
112
142
  let cancelled = false;
113
143
  // Emit agent query event (TUI renders echo+spinner, ContextManager records it)
114
144
  this.bus.emit("agent:query", { query });
@@ -116,19 +146,25 @@ export class AcpClient {
116
146
  const contextBlock = this.contextManager.getContext();
117
147
  try {
118
148
  this.log("sending prompt...");
119
- const response = await this.connection.prompt({
120
- sessionId: this.sessionId,
121
- prompt: [
122
- {
123
- type: "text",
124
- text: contextBlock + "\n" + query,
125
- },
126
- ],
127
- });
149
+ const promptTimeoutMs = 300000; // 5 minutes timeout for LLM response
150
+ const response = await Promise.race([
151
+ this.connection.prompt({
152
+ sessionId: this.sessionId,
153
+ prompt: [
154
+ {
155
+ type: "text",
156
+ text: contextBlock + "\n" + query,
157
+ },
158
+ ],
159
+ }),
160
+ new Promise((_, reject) => setTimeout(() => reject(new Error(`Prompt timeout after ${promptTimeoutMs}ms`)), promptTimeoutMs)),
161
+ ]);
128
162
  this.log(`prompt resolved: stopReason=${response.stopReason}`);
129
163
  if (response.stopReason === "cancelled") {
130
164
  cancelled = true;
131
- this.bus.emit("agent:cancelled", {});
165
+ if (!this.autoCancelled) {
166
+ this.bus.emit("agent:cancelled", {});
167
+ }
132
168
  }
133
169
  }
134
170
  catch (err) {
@@ -154,6 +190,18 @@ export class AcpClient {
154
190
  this.promptInProgress = false;
155
191
  }
156
192
  }
193
+ /**
194
+ * Silently cancel the prompt after a shell tool completes.
195
+ * Unlike user-initiated cancel(), this doesn't show "(cancelled)" —
196
+ * the tool already ran, we just skip the unnecessary LLM follow-up.
197
+ */
198
+ autoCancel() {
199
+ if (!this.connection || !this.sessionId || !this.promptInProgress)
200
+ return;
201
+ this.log("auto-cancel: shell tool completed, skipping LLM follow-up");
202
+ this.autoCancelled = true;
203
+ this.connection.cancel({ sessionId: this.sessionId }).catch(() => { });
204
+ }
157
205
  /**
158
206
  * Cancel the current prompt and force-recover shell mode.
159
207
  */
@@ -185,13 +233,18 @@ export class AcpClient {
185
233
  async resetSession() {
186
234
  if (!this.connection)
187
235
  return;
188
- const sessionResponse = await this.connection.newSession({
236
+ const sessionConfig = this.bus.emitPipe("session:configure", {
189
237
  cwd: this.contextManager.getCwd(),
190
238
  mcpServers: [],
191
239
  });
240
+ const sessionResponse = await this.connection.newSession({
241
+ cwd: sessionConfig.cwd,
242
+ mcpServers: sessionConfig.mcpServers,
243
+ });
192
244
  this.sessionId = sessionResponse.sessionId;
193
245
  this.lastResponseText = "";
194
246
  this.currentResponseText = "";
247
+ this.updateModes(sessionResponse);
195
248
  }
196
249
  /**
197
250
  * Get the text of the last agent response (for /copy).
@@ -208,6 +261,14 @@ export class AcpClient {
208
261
  getModel() {
209
262
  return this.config.model;
210
263
  }
264
+ /**
265
+ * Get the current mode (e.g. thinking level).
266
+ */
267
+ getCurrentMode() {
268
+ if (!this.currentModeId)
269
+ return null;
270
+ return this.modes.find((m) => m.id === this.currentModeId) ?? null;
271
+ }
211
272
  /**
212
273
  * Check if agent is connected.
213
274
  */
@@ -216,6 +277,45 @@ export class AcpClient {
216
277
  // Session ID may not be set yet if we're still initializing
217
278
  return this.connection !== null && this.agentInfo !== null;
218
279
  }
280
+ /**
281
+ * Parse modes from a session response and notify listeners.
282
+ */
283
+ updateModes(response) {
284
+ const modes = response.modes;
285
+ if (!modes)
286
+ return;
287
+ if (modes.availableModes) {
288
+ this.modes = modes.availableModes.map((m) => ({
289
+ id: m.id,
290
+ name: m.name || m.id,
291
+ }));
292
+ }
293
+ if (modes.currentModeId) {
294
+ this.currentModeId = modes.currentModeId;
295
+ }
296
+ this.bus.emit("config:changed", {});
297
+ }
298
+ /**
299
+ * Cycle to the next session mode.
300
+ */
301
+ async cycleMode() {
302
+ if (!this.connection || !this.sessionId || this.modes.length === 0)
303
+ return;
304
+ const currentIdx = this.modes.findIndex((m) => m.id === this.currentModeId);
305
+ const nextIdx = (currentIdx + 1) % this.modes.length;
306
+ const nextMode = this.modes[nextIdx];
307
+ try {
308
+ await this.connection.setSessionMode({
309
+ sessionId: this.sessionId,
310
+ modeId: nextMode.id,
311
+ });
312
+ this.currentModeId = nextMode.id;
313
+ this.bus.emit("config:changed", {});
314
+ }
315
+ catch (err) {
316
+ this.log(`Failed to set mode: ${err}`);
317
+ }
318
+ }
219
319
  log(msg) {
220
320
  if (process.env.DEBUG) {
221
321
  process.stderr.write(`[agent-sh] ${msg}\n`);
@@ -282,29 +382,61 @@ export class AcpClient {
282
382
  break;
283
383
  }
284
384
  case "tool_call": {
285
- // Use toolCallId if available, otherwise generate a simple ID
286
385
  const toolId = update.toolCallId || `tool-${this.pendingToolCounter++}`;
287
- this.pendingToolCalls.set(toolId, true);
288
- this.bus.emit("agent:tool-started", { title: update.title, toolCallId: toolId });
386
+ this.pendingToolCalls.set(toolId, update.title ?? "");
387
+ this.bus.emit("agent:tool-started", {
388
+ title: update.title,
389
+ toolCallId: toolId,
390
+ kind: update.kind ?? undefined,
391
+ locations: update.locations?.map((l) => ({ path: l.path, line: l.line })),
392
+ rawInput: update.rawInput,
393
+ });
289
394
  break;
290
395
  }
291
396
  case "tool_call_update": {
292
- // Only show result when the tool completes, don't show tool call again
397
+ const toolId = update.toolCallId;
398
+ const toolTitle = toolId ? this.pendingToolCalls.get(toolId) : undefined;
293
399
  if (update.status === "completed" || update.status === "failed") {
294
- const toolId = update.toolCallId;
400
+ // Show content only on final status. Skip tools whose output the
401
+ // user already sees (user_shell → PTY) or is agent-only (shell_recall).
402
+ const skipOutput = toolTitle === "user_shell" || toolTitle === "shell_recall";
403
+ if (!skipOutput && update.content && Array.isArray(update.content)) {
404
+ for (const block of update.content) {
405
+ if (block.type === "content" && block.content?.type === "text" && block.content.text) {
406
+ this.bus.emit("agent:tool-output-chunk", { chunk: block.content.text });
407
+ }
408
+ }
409
+ }
295
410
  const exitCode = update.status === "completed" ? 0 : 1;
296
411
  if (toolId && this.pendingToolCalls.has(toolId)) {
297
412
  this.pendingToolCalls.delete(toolId);
298
- this.bus.emit("agent:tool-completed", { toolCallId: toolId, exitCode });
413
+ this.bus.emit("agent:tool-completed", {
414
+ toolCallId: toolId,
415
+ exitCode,
416
+ rawOutput: update.rawOutput,
417
+ });
299
418
  }
300
419
  else if (!toolId) {
301
- this.bus.emit("agent:tool-completed", { exitCode });
420
+ this.bus.emit("agent:tool-completed", { exitCode, rawOutput: update.rawOutput });
421
+ }
422
+ // Auto-cancel after shell tools complete — the command already
423
+ // ran in the user's PTY, no need for a second LLM round trip.
424
+ // The result is captured in shell context / shell_recall.
425
+ if (toolTitle === "user_shell" && update.status === "completed") {
426
+ this.autoCancel();
302
427
  }
303
428
  }
304
429
  break;
305
430
  }
431
+ case "current_mode_update": {
432
+ const modeId = update.currentModeId;
433
+ if (modeId) {
434
+ this.currentModeId = modeId;
435
+ this.bus.emit("config:changed", {});
436
+ }
437
+ break;
438
+ }
306
439
  default:
307
- // Ignore other update types for now
308
440
  break;
309
441
  }
310
442
  }
@@ -363,6 +495,7 @@ export class AcpClient {
363
495
  const { session, done } = executeCommand({
364
496
  command: fullCommand,
365
497
  cwd,
498
+ env: this.config.shellEnv,
366
499
  timeout: 60_000,
367
500
  maxOutputBytes: 256 * 1024,
368
501
  onOutput: (chunk) => {
@@ -5,6 +5,8 @@ export declare class ContextManager {
5
5
  private currentCwd;
6
6
  private sessionStart;
7
7
  private pendingToolCalls;
8
+ private firstPrompt;
9
+ private agentShellActive;
8
10
  constructor(bus: EventBus);
9
11
  getCwd(): string;
10
12
  /**
@@ -17,15 +19,16 @@ export declare class ContextManager {
17
19
  */
18
20
  search(query: string): string;
19
21
  /**
20
- * Return full untruncated content for specific exchange IDs.
22
+ * Return content for specific exchange IDs.
23
+ * Optional start/end restrict to a line range (1-indexed).
21
24
  */
22
- expand(ids: number[]): string;
25
+ expand(ids: number[], start?: number, end?: number): string;
23
26
  /**
24
27
  * One-line summaries of last N exchanges.
25
28
  */
26
29
  getRecentSummary(n?: number): string;
27
30
  /**
28
- * Parse and handle __shell_recall commands.
31
+ * Parse and handle shell_recall commands.
29
32
  */
30
33
  handleRecallCommand(command: string): string;
31
34
  /**
@@ -37,7 +40,6 @@ export declare class ContextManager {
37
40
  private formatContext;
38
41
  private addExchange;
39
42
  private formatExchangeTruncated;
40
- private truncateForRecall;
41
43
  private formatExchangeFull;
42
44
  private exchangeOneLiner;
43
45
  private exchangeSearchText;
@@ -1,21 +1,18 @@
1
- const DEFAULT_WINDOW_SIZE = 20;
2
- const DEFAULT_BUDGET = 16384; // ~4K tokens at ~4 chars/token
3
- // Truncation thresholds (in lines)
4
- const SHELL_TRUNCATE_THRESHOLD = 30;
5
- const SHELL_HEAD_LINES = 10;
6
- const SHELL_TAIL_LINES = 10;
1
+ import { getSettings } from "./settings.js";
2
+ // Non-configurable thresholds (agent response and tool output follow shell settings)
7
3
  const AGENT_RESPONSE_TRUNCATE_THRESHOLD = 20;
8
4
  const AGENT_RESPONSE_HEAD_LINES = 15;
9
5
  const TOOL_TRUNCATE_THRESHOLD = 20;
10
6
  const TOOL_HEAD_LINES = 5;
11
7
  const TOOL_TAIL_LINES = 5;
12
- const RECALL_EXPAND_MAX_LINES = 500;
13
8
  export class ContextManager {
14
9
  exchanges = [];
15
10
  nextId = 1;
16
11
  currentCwd;
17
12
  sessionStart;
18
13
  pendingToolCalls = [];
14
+ firstPrompt = true;
15
+ agentShellActive = false; // true while user_shell command is executing
19
16
  constructor(bus) {
20
17
  this.currentCwd = process.cwd();
21
18
  this.sessionStart = Date.now();
@@ -30,11 +27,15 @@ export class ContextManager {
30
27
  exitCode: e.exitCode,
31
28
  outputLines: lines.length,
32
29
  outputBytes: e.output.length,
30
+ source: this.agentShellActive ? "agent" : "user",
33
31
  });
34
32
  });
35
33
  bus.on("shell:cwd-change", (e) => {
36
34
  this.currentCwd = e.cwd;
37
35
  });
36
+ // Track agent-initiated shell commands (user_shell tool)
37
+ bus.on("shell:agent-exec-start", () => { this.agentShellActive = true; });
38
+ bus.on("shell:agent-exec-done", () => { this.agentShellActive = false; });
38
39
  // ── Subscribe to agent events ──
39
40
  bus.on("agent:query", (e) => {
40
41
  this.pendingToolCalls = [];
@@ -85,7 +86,8 @@ export class ContextManager {
85
86
  * Build the <shell_context> block for the agent prompt.
86
87
  * Pipeline: window → truncate → format
87
88
  */
88
- getContext(budget = DEFAULT_BUDGET) {
89
+ getContext(budget) {
90
+ budget ??= getSettings().contextBudget;
89
91
  let exchanges = this.applyWindow(this.exchanges);
90
92
  exchanges = this.applyTruncation(exchanges, budget);
91
93
  return this.formatContext(exchanges);
@@ -141,9 +143,10 @@ export class ContextManager {
141
143
  return parts.join("\n");
142
144
  }
143
145
  /**
144
- * Return full untruncated content for specific exchange IDs.
146
+ * Return content for specific exchange IDs.
147
+ * Optional start/end restrict to a line range (1-indexed).
145
148
  */
146
- expand(ids) {
149
+ expand(ids, start, end) {
147
150
  const results = [];
148
151
  for (const id of ids) {
149
152
  const ex = this.exchanges.find((e) => e.id === id);
@@ -151,7 +154,25 @@ export class ContextManager {
151
154
  results.push(`#${id}: not found`);
152
155
  continue;
153
156
  }
154
- results.push(this.formatExchangeFull(ex));
157
+ const text = this.formatExchangeFull(ex);
158
+ const lines = text.split("\n");
159
+ const total = lines.length;
160
+ if (start != null || end != null) {
161
+ // Line range requested
162
+ const s = Math.max(0, (start ?? 1) - 1);
163
+ const e = end ?? total;
164
+ results.push(lines.slice(s, e).join("\n") +
165
+ `\n[showing lines ${s + 1}-${Math.min(e, total)} of ${total}]`);
166
+ }
167
+ else if (total > getSettings().recallExpandMaxLines) {
168
+ // Too large — tell the agent to narrow down
169
+ results.push(`#${ex.id}: output is ${total} lines, too large to expand fully. ` +
170
+ `Use start/end params to select a line range (e.g. start=1, end=50), ` +
171
+ `or use search with a regex to find specific content.`);
172
+ }
173
+ else {
174
+ results.push(text);
175
+ }
155
176
  }
156
177
  return results.join("\n\n");
157
178
  }
@@ -165,21 +186,21 @@ export class ContextManager {
165
186
  return recent.map((ex) => this.exchangeOneLiner(ex)).join("\n");
166
187
  }
167
188
  /**
168
- * Parse and handle __shell_recall commands.
189
+ * Parse and handle shell_recall commands.
169
190
  */
170
191
  handleRecallCommand(command) {
171
- const args = command.replace(/^__shell_recall\s*/, "").trim();
192
+ const args = command.replace(/^_*shell_recall\s*/, "").trim();
172
193
  if (!args || args === "--help") {
173
194
  return [
174
195
  "Usage:",
175
- " __shell_recall Browse recent exchanges",
176
- " __shell_recall --search <query> Search all exchanges",
177
- " __shell_recall --expand <id,...> Show full content of exchanges",
196
+ " shell_recall Browse recent exchanges",
197
+ " shell_recall --search <query> Search all exchanges",
198
+ " shell_recall --expand <id,...> Show full content of exchanges",
178
199
  "",
179
200
  "Examples:",
180
- ' __shell_recall --search "test fail"',
181
- " __shell_recall --expand 41",
182
- " __shell_recall --expand 41,42,43",
201
+ ' shell_recall --search "test fail"',
202
+ " shell_recall --expand 41",
203
+ " shell_recall --expand 41,42,43",
183
204
  ].join("\n");
184
205
  }
185
206
  const searchMatch = args.match(/^--search\s+(?:"([^"]+)"|(\S+))/);
@@ -205,10 +226,12 @@ export class ContextManager {
205
226
  clear() {
206
227
  this.exchanges = [];
207
228
  this.pendingToolCalls = [];
229
+ this.firstPrompt = true;
208
230
  // Don't reset nextId — IDs should be globally unique within a session
209
231
  }
210
232
  // ── Pipeline stages ───────────────────────────────────────────
211
- applyWindow(exchanges, windowSize = DEFAULT_WINDOW_SIZE) {
233
+ applyWindow(exchanges, windowSize) {
234
+ windowSize ??= getSettings().contextWindowSize;
212
235
  return exchanges.slice(-windowSize);
213
236
  }
214
237
  applyTruncation(exchanges, budget) {
@@ -217,7 +240,8 @@ export class ContextManager {
217
240
  // Pass 1: per-type truncation
218
241
  for (const ex of result) {
219
242
  if (ex.type === "shell_command") {
220
- ex.output = truncateOutput(ex.output, SHELL_TRUNCATE_THRESHOLD, SHELL_HEAD_LINES, SHELL_TAIL_LINES, ex.id);
243
+ const s = getSettings();
244
+ ex.output = truncateOutput(ex.output, s.shellTruncateThreshold, s.shellHeadLines, s.shellTailLines, ex.id);
221
245
  }
222
246
  else if (ex.type === "agent_response") {
223
247
  ex.response = truncateHead(ex.response, AGENT_RESPONSE_TRUNCATE_THRESHOLD, AGENT_RESPONSE_HEAD_LINES, ex.id);
@@ -232,13 +256,13 @@ export class ContextManager {
232
256
  const ex = result[i];
233
257
  const before = this.exchangeSize(ex);
234
258
  if (ex.type === "shell_command") {
235
- ex.output = `[output omitted, use __shell_recall --expand ${ex.id}]`;
259
+ ex.output = `[output omitted, use shell_recall tool to expand id ${ex.id}]`;
236
260
  }
237
261
  else if (ex.type === "tool_execution") {
238
- ex.output = `[output omitted, use __shell_recall --expand ${ex.id}]`;
262
+ ex.output = `[output omitted, use shell_recall tool to expand id ${ex.id}]`;
239
263
  }
240
264
  else if (ex.type === "agent_response") {
241
- ex.response = `[response omitted, use __shell_recall --expand ${ex.id}]`;
265
+ ex.response = `[response omitted, use shell_recall tool to expand id ${ex.id}]`;
242
266
  }
243
267
  totalSize -= before - this.exchangeSize(ex);
244
268
  }
@@ -248,9 +272,22 @@ export class ContextManager {
248
272
  const elapsed = Math.round((Date.now() - this.sessionStart) / 60000);
249
273
  const totalCount = this.exchanges.length;
250
274
  let out = "<shell_context>\n";
275
+ if (this.firstPrompt) {
276
+ out += `You are an AI assistant living inside agent-sh, a shell-first terminal.\n`;
277
+ out += `The user interacts with a real shell (PTY) and sends you queries inline. You are there to help them with their tasks.\n`;
278
+ out += `\n`;
279
+ out += `IMPORTANT tool usage rules:\n`;
280
+ out += `- user_shell runs commands in the user's live shell (PTY). The user sees output directly — no summary needed.\n`;
281
+ out += `- Your internal tools (bash, read, write, ls, etc.) run in an isolated subprocess. The user CANNOT see their output.\n`;
282
+ out += `- When the user asks to see, list, view, or display anything, ALWAYS use user_shell. NEVER use internal tools like ls/read/bash for display — the user won't see it.\n`;
283
+ out += `- Only use internal tools when YOU need to reason about content silently (e.g. reading a file to answer a question about it).\n`;
284
+ out += `- After a user_shell command, the user already saw the output. Do NOT repeat or summarize it.\n`;
285
+ out += `- You can browse or search session history with shell_recall.\n`;
286
+ out += `\n`;
287
+ this.firstPrompt = false;
288
+ }
251
289
  out += `cwd: ${this.currentCwd}\n`;
252
290
  out += `session: ${totalCount} exchanges, ${elapsed}m elapsed\n`;
253
- out += `[hint: run \`__shell_recall --search "query"\` or \`__shell_recall --expand ID\` to retrieve truncated content]\n`;
254
291
  for (const ex of exchanges) {
255
292
  out += "\n" + this.formatExchangeTruncated(ex);
256
293
  }
@@ -269,7 +306,8 @@ export class ContextManager {
269
306
  formatExchangeTruncated(ex) {
270
307
  switch (ex.type) {
271
308
  case "shell_command": {
272
- let s = `#${ex.id} [shell] $ ${ex.command}\n`;
309
+ const label = ex.source === "agent" ? "agent → shell" : "shell";
310
+ let s = `#${ex.id} [${label} cwd:${ex.cwd}] $ ${ex.command}\n`;
273
311
  if (ex.output)
274
312
  s += indent(ex.output, " ") + "\n";
275
313
  if (ex.exitCode !== null)
@@ -299,20 +337,12 @@ export class ContextManager {
299
337
  }
300
338
  }
301
339
  }
302
- truncateForRecall(text) {
303
- const lines = text.split("\n");
304
- if (lines.length <= RECALL_EXPAND_MAX_LINES)
305
- return text;
306
- const half = RECALL_EXPAND_MAX_LINES / 2;
307
- return (lines.slice(0, half).join("\n") +
308
- `\n[... ${lines.length - RECALL_EXPAND_MAX_LINES} more lines ...]\n` +
309
- lines.slice(-half).join("\n"));
310
- }
311
340
  formatExchangeFull(ex) {
312
341
  switch (ex.type) {
313
342
  case "shell_command": {
314
- const output = this.truncateForRecall(ex.output);
315
- let s = `#${ex.id} [shell] $ ${ex.command} (${ex.outputLines} lines, ${ex.outputBytes} bytes)\n`;
343
+ const label = ex.source === "agent" ? "agent → shell" : "shell";
344
+ const output = ex.output;
345
+ let s = `#${ex.id} [${label}] $ ${ex.command} (${ex.outputLines} lines, ${ex.outputBytes} bytes)\n`;
316
346
  if (output)
317
347
  s += output + "\n";
318
348
  if (ex.exitCode !== null)
@@ -324,10 +354,9 @@ export class ContextManager {
324
354
  case "agent_response":
325
355
  return `#${ex.id} [agent]\n${ex.response}`;
326
356
  case "tool_execution": {
327
- const output = this.truncateForRecall(ex.output);
328
357
  let s = `#${ex.id} [tool] ${ex.tool} (${ex.outputLines} lines, ${ex.outputBytes} bytes)\n`;
329
- if (output)
330
- s += output + "\n";
358
+ if (ex.output)
359
+ s += ex.output + "\n";
331
360
  if (ex.exitCode !== null)
332
361
  s += `exit ${ex.exitCode}\n`;
333
362
  return s;
@@ -336,8 +365,10 @@ export class ContextManager {
336
365
  }
337
366
  exchangeOneLiner(ex) {
338
367
  switch (ex.type) {
339
- case "shell_command":
340
- return `#${ex.id} shell: ${ex.command} (${ex.outputLines} lines, exit ${ex.exitCode ?? "?"})`;
368
+ case "shell_command": {
369
+ const label = ex.source === "agent" ? "agent shell" : "shell";
370
+ return `#${ex.id} ${label} [cwd:${ex.cwd}]: ${ex.command} (${ex.outputLines} total lines, exit ${ex.exitCode ?? "?"})`;
371
+ }
341
372
  case "agent_query":
342
373
  return `#${ex.id} query: ${ex.query}`;
343
374
  case "agent_response": {
@@ -381,7 +412,7 @@ function truncateOutput(text, threshold, headLines, tailLines, id) {
381
412
  const omitted = lines.length - headLines - tailLines;
382
413
  return [
383
414
  ...lines.slice(0, headLines),
384
- `[... ${omitted} lines truncated, use __shell_recall --expand ${id} to see full output ...]`,
415
+ `[... ${omitted} lines truncated, use shell_recall tool with expand and id ${id} to see full output ...]`,
385
416
  ...lines.slice(-tailLines),
386
417
  ].join("\n");
387
418
  }
@@ -391,7 +422,7 @@ function truncateHead(text, threshold, headLines, id) {
391
422
  return text;
392
423
  return [
393
424
  ...lines.slice(0, headLines),
394
- `[... truncated, use __shell_recall --expand ${id} for full response ...]`,
425
+ `[... truncated, use shell_recall tool with expand and id ${id} for full response ...]`,
395
426
  ].join("\n");
396
427
  }
397
428
  function indent(text, prefix) {