niahere 0.5.6 → 0.5.8

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/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "niahere",
3
- "version": "0.5.6",
3
+ "version": "0.5.8",
4
4
  "description": "A personal AI assistant daemon — chat, scheduled jobs, persona system, extensible via skills.",
5
5
  "type": "module",
6
6
  "scripts": {
package/src/agent/auth.ts CHANGED
@@ -2,6 +2,8 @@ import { existsSync, readFileSync } from "fs";
2
2
  import { homedir } from "os";
3
3
  import { join } from "path";
4
4
  import type { ProviderName } from "./models";
5
+ import { resolveClaudeCredential } from "./credentials";
6
+ import { getConfig } from "../utils/config";
5
7
 
6
8
  /**
7
9
  * Provider sign-in state, read from whatever each CLI stores on disk.
@@ -43,6 +45,17 @@ const defaultReader: AuthReader = {
43
45
  env: (k) => process.env[k],
44
46
  };
45
47
 
48
+ /** Config is optional here: a health check must never fail because config is
49
+ * unreadable, it must say so. */
50
+ function credentialConfig(): { anthropic_oauth_token: string | null; anthropic_api_key: string | null } {
51
+ try {
52
+ const c = getConfig();
53
+ return { anthropic_oauth_token: c.anthropic_oauth_token, anthropic_api_key: c.anthropic_api_key };
54
+ } catch {
55
+ return { anthropic_oauth_token: null, anthropic_api_key: null };
56
+ }
57
+ }
58
+
46
59
  export function claudeCredentialsPath(): string {
47
60
  return join(homedir(), ".claude", ".credentials.json");
48
61
  }
@@ -62,8 +75,16 @@ function ago(ms: number): string {
62
75
  export function claudeAuthStatus(now: number = Date.now(), reader: AuthReader = defaultReader): AuthStatus {
63
76
  const base = { provider: "claude" as const };
64
77
 
65
- if (reader.env("ANTHROPIC_API_KEY")) {
66
- return { ...base, state: "ok", detail: "API key (no expiry)" };
78
+ // Ask the same resolver the backend uses. Reading the environment here while
79
+ // the backend also reads config meant the check could report one credential
80
+ // while a different one served the turn — the exact ambiguity this exists to
81
+ // remove.
82
+ const credential = resolveClaudeCredential(credentialConfig(), reader.env);
83
+ if (credential.kind === "oauth_token") {
84
+ return { ...base, state: "ok", detail: "configured oauth token (subscription, no refresh needed)" };
85
+ }
86
+ if (credential.kind === "api_key") {
87
+ return { ...base, state: "ok", detail: "configured API key (metered — billed per token)" };
67
88
  }
68
89
 
69
90
  const path = claudeCredentialsPath();
@@ -97,7 +118,11 @@ export function claudeAuthStatus(now: number = Date.now(), reader: AuthReader =
97
118
  return { ...status, state: "stale", detail: `access token lapsed ${ago(now - access)} ago, renewable${plan}` };
98
119
  }
99
120
  if (access !== undefined) {
100
- return { ...status, state: "ok", detail: `valid for ${ago(access - now)}${plan}` };
121
+ return {
122
+ ...status,
123
+ state: "ok",
124
+ detail: `Claude Code login, valid for ${ago(access - now)}${plan} — renewed only when the CLI is used here`,
125
+ };
101
126
  }
102
127
  return { ...status, state: "unknown", detail: "credentials file carries no expiry" };
103
128
  }
@@ -90,6 +90,16 @@ function attribute(modelUsage: unknown): Record<string, unknown> | undefined {
90
90
  * stay backend-agnostic.
91
91
  */
92
92
  export class SdkNormalizer implements Normalizer {
93
+ /**
94
+ * The SDK's `apiKeySource` only describes where an *API key* came from: it
95
+ * reports "none" for every OAuth path, so it cannot tell a configured token
96
+ * from the CLI's inherited login. Nia resolved the credential itself, so it
97
+ * records that too — the SDK's value is kept because it is genuinely
98
+ * informative when an API key is in play.
99
+ */
100
+ constructor(private readonly credential?: string) {}
101
+
102
+ private apiKeySource: string | undefined;
93
103
  private accumulatedThinking = "";
94
104
  private lastThinkingLine = "";
95
105
 
@@ -97,6 +107,10 @@ export class SdkNormalizer implements Normalizer {
97
107
  const msg = message as any;
98
108
 
99
109
  if (msg.type === "system" && msg.subtype === "init") {
110
+ // apiKeySource names which credential served the session ('oauth' is
111
+ // Claude Code's own login). Recording it is what makes a silent switch
112
+ // of credential visible after the fact.
113
+ if (typeof msg.apiKeySource === "string") this.apiKeySource = msg.apiKeySource;
100
114
  return [{ type: "session", backendSessionId: msg.session_id }];
101
115
  }
102
116
 
@@ -211,6 +225,8 @@ export class SdkNormalizer implements Normalizer {
211
225
  terminal_reason: msg.terminal_reason,
212
226
  session_id: msg.session_id,
213
227
  subtype: msg.subtype,
228
+ api_key_source: this.apiKeySource,
229
+ credential: this.credential,
214
230
  usage: msg.usage,
215
231
  model_usage: attribute(msg.modelUsage),
216
232
  },
@@ -11,6 +11,7 @@ import { MessageStream } from "../message-stream";
11
11
  import { getSdkSkillsSetting } from "../../core/skills";
12
12
  import { getSdkHooks } from "../../core/sdk-hooks";
13
13
  import { getConfig } from "../../utils/config";
14
+ import { resolveClaudeCredential, credentialEnv } from "../credentials";
14
15
  import { sleep } from "../../utils/retry";
15
16
 
16
17
  /** The shape of the SDK `query()` handle the session consumes. Injected so the
@@ -106,6 +107,14 @@ class ClaudeSession implements AgentSession {
106
107
  // same cwd; jobs always run with a unique id and never auto-continued.
107
108
  if (this.ctx.interactive) options.continue = false;
108
109
  }
110
+ // Hand the CLI a credential Nia owns when one is configured. Without this
111
+ // it inherits ~/.claude/.credentials.json, which only refreshes when a
112
+ // human runs `claude` on this machine — the coupling that had Nia
113
+ // answering as codex for sixteen days.
114
+ const credential = resolveClaudeCredential(getConfig());
115
+ if (credential.envVar) {
116
+ options.env = credentialEnv(credential, process.env as Record<string, string>);
117
+ }
109
118
  if (this.ctx.outputSchema) options.outputFormat = { type: "json_schema", schema: this.ctx.outputSchema };
110
119
  if (this.ctx.mcpServers) options.mcpServers = this.ctx.mcpServers;
111
120
  if (this.ctx.subagents && Object.keys(this.ctx.subagents).length > 0) options.agents = this.ctx.subagents;
@@ -119,7 +128,7 @@ class ClaudeSession implements AgentSession {
119
128
  while (true) {
120
129
  if (!this.iterator || !this.stream) this.startQuery();
121
130
  this.stream!.push(text, attachments);
122
- const normalizer = new SdkNormalizer();
131
+ const normalizer = new SdkNormalizer(resolveClaudeCredential(getConfig()).kind);
123
132
  let retry = false;
124
133
 
125
134
  while (true) {
@@ -0,0 +1,77 @@
1
+ import type { Config } from "../types/config";
2
+
3
+ /**
4
+ * Which Claude credential the daemon should use, and where it came from.
5
+ *
6
+ * Nia used to have no answer to this: it inherited whatever `claude` had last
7
+ * written to `~/.claude/.credentials.json`, refreshed by a human opening a
8
+ * terminal on the same machine. When that stopped, Nia answered as codex for
9
+ * sixteen days and nothing said why. A credential the daemon is handed
10
+ * explicitly is one it can report on, and one that does not lapse because
11
+ * nobody logged in today.
12
+ */
13
+
14
+ export type CredentialKind = "oauth_token" | "api_key" | "claude_code_login";
15
+
16
+ export interface ClaudeCredential {
17
+ kind: CredentialKind;
18
+ /** The variable the CLI reads it from. Absent for the inherited login. */
19
+ envVar?: "CLAUDE_CODE_OAUTH_TOKEN" | "ANTHROPIC_API_KEY";
20
+ value?: string;
21
+ /**
22
+ * `subscription` rides the plan. `metered` bills per token — worth saying out
23
+ * loud, because Nia has run $552 in a week and that is an invoice on the API.
24
+ */
25
+ billing: "subscription" | "metered";
26
+ }
27
+
28
+ /** Every variable a credential could occupy, so switching never leaves a stale one behind. */
29
+ const TOKEN_VARS = ["CLAUDE_CODE_OAUTH_TOKEN", "ANTHROPIC_API_KEY"] as const;
30
+
31
+ const clean = (v: unknown): string | undefined => {
32
+ const s = typeof v === "string" ? v.trim() : "";
33
+ return s.length > 0 ? s : undefined;
34
+ };
35
+
36
+ export type EnvLookup = (key: string) => string | undefined;
37
+
38
+ /**
39
+ * Config first, then the ambient environment, then Claude Code's own login.
40
+ * Config wins so the daemon is not at the mercy of whatever shell launched it.
41
+ */
42
+ export function resolveClaudeCredential(
43
+ config: Pick<Config, "anthropic_oauth_token" | "anthropic_api_key">,
44
+ env: EnvLookup = (k) => process.env[k],
45
+ ): ClaudeCredential {
46
+ const oauth = clean(config.anthropic_oauth_token) ?? clean(env("CLAUDE_CODE_OAUTH_TOKEN"));
47
+ if (oauth) {
48
+ return { kind: "oauth_token", envVar: "CLAUDE_CODE_OAUTH_TOKEN", value: oauth, billing: "subscription" };
49
+ }
50
+ const key = clean(config.anthropic_api_key) ?? clean(env("ANTHROPIC_API_KEY"));
51
+ if (key) {
52
+ return { kind: "api_key", envVar: "ANTHROPIC_API_KEY", value: key, billing: "metered" };
53
+ }
54
+ return { kind: "claude_code_login", billing: "subscription" };
55
+ }
56
+
57
+ /**
58
+ * The environment for the spawned CLI. The base is passed through — it still
59
+ * needs PATH and HOME — with exactly one credential variable set, and any other
60
+ * cleared so a removed credential stops working immediately rather than at the
61
+ * next restart.
62
+ */
63
+ export function credentialEnv(credential: ClaudeCredential, base: Record<string, string>): Record<string, string> {
64
+ const env: Record<string, string> = {};
65
+ for (const [k, v] of Object.entries(base)) {
66
+ if ((TOKEN_VARS as readonly string[]).includes(k)) continue;
67
+ env[k] = v;
68
+ }
69
+ if (credential.envVar && credential.value) env[credential.envVar] = credential.value;
70
+ return env;
71
+ }
72
+
73
+ export function describeCredential(credential: ClaudeCredential): string {
74
+ if (credential.kind === "oauth_token") return "configured oauth token (subscription, long-lived)";
75
+ if (credential.kind === "api_key") return "configured API key (metered, billed per token)";
76
+ return "Claude Code's own login (refreshed by whoever last used the CLI here)";
77
+ }
@@ -0,0 +1,79 @@
1
+ /**
2
+ * What counts as a reply, and what is just the model talking to itself.
3
+ *
4
+ * Two related jobs live here on purpose, because splitting them is how the
5
+ * codebase ended up with two sentinel parsers that disagreed:
6
+ *
7
+ * - `shouldSuppressReply` is the cross-channel guard. Control artifacts have
8
+ * escaped as real messages before — `nia send --help` once DM'd the flag
9
+ * itself — so every channel checks the same list.
10
+ * - `decideWatchReply` is the richer judgement a watch turn needs: it prefers
11
+ * a schema's answer, and can tell a bare sentinel from one tangled up with
12
+ * content.
13
+ */
14
+
15
+ export const WATCH_JUDGEMENT_SCHEMA: Record<string, unknown> = {
16
+ type: "object",
17
+ properties: {
18
+ reply: {
19
+ type: ["string", "null"],
20
+ description: "The message to post in the channel, or null to stay silent. Most turns are null.",
21
+ },
22
+ },
23
+ required: ["reply"],
24
+ additionalProperties: false,
25
+ };
26
+
27
+ /** Strip markdown fencing so a sentinel matches even when the model wraps it.
28
+ * Underscores are left alone — the sentinel contains one. */
29
+ export function cleanControlReply(text: string): string {
30
+ return text.replace(/[`*]/g, "").trim();
31
+ }
32
+
33
+ /** Outputs that are control artifacts rather than anything a person should see. */
34
+ const CONTROL_ARTIFACTS = new Set(["[NO_REPLY]", "--help", "-h"]);
35
+
36
+ /**
37
+ * Exact matches only. "Use `--help` to see the flags" is a real answer and must
38
+ * survive; a reply that is nothing but `--help` is the CLI leaking.
39
+ */
40
+ export function shouldSuppressReply(text: string): boolean {
41
+ const cleaned = cleanControlReply(text);
42
+ return !cleaned || CONTROL_ARTIFACTS.has(cleaned);
43
+ }
44
+
45
+ export interface WatchDecision {
46
+ /** Whether to post at all. */
47
+ send: boolean;
48
+ /** What to post. Empty when staying quiet. */
49
+ text: string;
50
+ /** Which path decided, so a drop in schema coverage is visible in the log. */
51
+ source: "structured" | "sentinel";
52
+ /** Set only on the sentinel path, when the model emitted both a sentinel and content. */
53
+ ambiguous?: boolean;
54
+ }
55
+
56
+ /**
57
+ * `structured` wins when it carries the `reply` key the schema requires —
58
+ * including an explicit null, which is a decision, not an absence.
59
+ *
60
+ * The sentinel fallback matches `[NO_REPLY]` anywhere, not just exactly: a watch
61
+ * turn that says both is confused, and staying quiet is the safer reading. That
62
+ * is deliberately looser than `shouldSuppressReply`, which guards ordinary
63
+ * replies where a substring match would swallow real answers.
64
+ */
65
+ export function decideWatchReply(structured: unknown, raw: string): WatchDecision {
66
+ if (structured && typeof structured === "object" && "reply" in structured) {
67
+ const reply = (structured as { reply: unknown }).reply;
68
+ const text = typeof reply === "string" ? reply.trim() : "";
69
+ return { send: text.length > 0, text, source: "structured" };
70
+ }
71
+
72
+ const trimmed = raw.trim();
73
+ const cleaned = cleanControlReply(trimmed);
74
+ if (!trimmed || cleaned.includes("[NO_REPLY]") || CONTROL_ARTIFACTS.has(cleaned)) {
75
+ const exact = !trimmed || CONTROL_ARTIFACTS.has(cleaned);
76
+ return { send: false, text: "", source: "sentinel", ambiguous: !exact };
77
+ }
78
+ return { send: true, text: trimmed, source: "sentinel" };
79
+ }
@@ -12,7 +12,7 @@ import { ChatSessions, chainLock } from "./common/chat-session";
12
12
  import { SlackAttachmentCache } from "./slack/attachments";
13
13
  import { SlackWatchReloader } from "./slack/watch";
14
14
 
15
- import { decideWatchReply } from "./common/watch-judgement";
15
+ import { decideWatchReply, shouldSuppressReply } from "./common/reply";
16
16
  import { createTurnPump } from "./common/coalesce";
17
17
 
18
18
  const logActivity = (status: string) => log.debug({ status }, "slack engine activity");
@@ -304,7 +304,8 @@ class SlackChannel implements Channel {
304
304
  const { result } = await state.engine.send(subcommand, {
305
305
  onActivity: logActivity,
306
306
  });
307
- await respond(result.trim() || "(no response)");
307
+ if (shouldSuppressReply(result)) return;
308
+ await respond(result.trim());
308
309
  } catch (err) {
309
310
  const errText = errMsg(err);
310
311
  await respond(`[error] ${errText}`);
@@ -23,6 +23,8 @@ import { errMsg, ignore } from "../utils/errors";
23
23
  import { sendMessage as twilioSendMessage } from "./twilio/rest";
24
24
  import { getTwilioServer } from "./twilio/server";
25
25
  import { ChatSessions, chainLock } from "./common/chat-session";
26
+ import { Message } from "../db/models";
27
+ import { shouldSuppressReply } from "./common/reply";
26
28
  import { ackTwiml, deliveryStatusAck, isAllowedSender } from "./twilio/shared";
27
29
 
28
30
  class SmsChannel implements Channel {
@@ -105,8 +107,13 @@ class SmsChannel implements Channel {
105
107
  // Twilio's ~15s webhook timeout when the engine takes longer.
106
108
  chainLock(state, async () => {
107
109
  try {
108
- const { result } = await state.engine.send(body);
109
- const reply = result.trim() || "(no response)";
110
+ const { result, messageId } = await state.engine.send(body);
111
+ const reply = result.trim();
112
+ if (shouldSuppressReply(reply)) {
113
+ if (messageId) await ignore(Message.updateDeliveryStatus(messageId, "sent"), "record suppressed delivery status");
114
+ log.info({ from }, "sms: agent chose not to reply");
115
+ return;
116
+ }
110
117
  await this.sendTo(from, reply);
111
118
  } catch (err) {
112
119
  log.error({ err, from }, "sms: engine error");
@@ -12,6 +12,7 @@ import { getMcpServers } from "../mcp";
12
12
  import { classifyMime, validateAttachment, prepareImage } from "../utils/attachment";
13
13
  import { getNiaHome } from "../utils/paths";
14
14
  import { ChatSessions, chainLock } from "./common/chat-session";
15
+ import { shouldSuppressReply } from "./common/reply";
15
16
 
16
17
  function safeExtension(filename?: string): string {
17
18
  const ext = filename?.split(".").pop();
@@ -187,7 +188,12 @@ class TelegramChannel implements Channel {
187
188
 
188
189
  try {
189
190
  const { result, messageId } = await state.engine.send(text, {}, attachments);
190
- const reply = result.trim() || "(no response)";
191
+ const reply = result.trim();
192
+ if (shouldSuppressReply(reply)) {
193
+ if (messageId) await ignore(Message.updateDeliveryStatus(messageId, "sent"), "record suppressed delivery status");
194
+ log.info({ chatId }, "telegram: agent chose not to reply");
195
+ return;
196
+ }
191
197
  try {
192
198
  await bot.api.sendMessage(chatId, reply);
193
199
  if (messageId) await ignore(Message.updateDeliveryStatus(messageId, "sent"), "record sent delivery status");
@@ -21,6 +21,7 @@ import { getMcpServers } from "../mcp";
21
21
  import { Message } from "../db/models";
22
22
  import { runMigrations } from "../db/migrate";
23
23
  import { ChatSessions, chainLock } from "./common/chat-session";
24
+ import { shouldSuppressReply } from "./common/reply";
24
25
  import { ackTwiml, deliveryStatusAck, isAllowedSender } from "./twilio/shared";
25
26
  import type { Attachment, Channel, ChatState, Outbound, TwilioConfig, WhatsappConfig, PhoneConfig } from "../types";
26
27
  import { getConfig } from "../utils/config";
@@ -201,7 +202,12 @@ class WhatsAppChannel implements Channel {
201
202
 
202
203
  try {
203
204
  const { result, messageId } = await state.engine.send(userText || "(media only)", {}, attachments);
204
- const reply = result.trim() || "(no response)";
205
+ const reply = result.trim();
206
+ if (shouldSuppressReply(reply)) {
207
+ if (messageId) await ignore(Message.updateDeliveryStatus(messageId, "sent"), "record suppressed delivery status");
208
+ log.info({ from }, "whatsapp: agent chose not to reply");
209
+ return;
210
+ }
205
211
  try {
206
212
  await this.sendTextTo(from, reply);
207
213
  if (messageId) await ignore(Message.updateDeliveryStatus(messageId, "sent"), "record sent delivery status");
@@ -1,28 +1,38 @@
1
1
  import { getConfig, updateRawConfig } from "../utils/config";
2
2
  import { getPaths } from "../utils/paths";
3
3
  import { errMsg } from "../utils/errors";
4
- import { fail, ICON_PASS, ICON_FAIL } from "../utils/cli";
4
+ import { fail, ICON_PASS, ICON_FAIL, parseArgs } from "../utils/cli";
5
5
  import { log } from "../utils/log";
6
6
 
7
+ const SEND_USAGE = "Usage: nia send [-c channel] [--to <slack-channel-id>] [--thread <ts>] <message>";
8
+
9
+ export interface SendArgs {
10
+ channel?: string;
11
+ toChannelId?: string;
12
+ threadTs?: string;
13
+ message: string;
14
+ help: boolean;
15
+ }
16
+
17
+ /** Extracted so the parse is testable — `nia send --help` used to DM the flag. */
18
+ export function parseSendArgs(args: string[] = process.argv.slice(3)): SendArgs {
19
+ const parsed = parseArgs(args);
20
+ return {
21
+ channel: parsed.getString("channel") || parsed.getString("c"),
22
+ toChannelId: parsed.getString("to"),
23
+ threadTs: parsed.getString("thread"),
24
+ message: parsed.positional.join(" "),
25
+ help: parsed.help,
26
+ };
27
+ }
28
+
7
29
  export async function sendCommand(): Promise<void> {
8
- const args = process.argv.slice(3);
9
- let channel: string | undefined;
10
- let toChannelId: string | undefined;
11
- let threadTs: string | undefined;
12
- const msgParts: string[] = [];
13
- for (let i = 0; i < args.length; i++) {
14
- if ((args[i] === "--channel" || args[i] === "-c") && args[i + 1]) {
15
- channel = args[++i];
16
- } else if (args[i] === "--to" && args[i + 1]) {
17
- toChannelId = args[++i];
18
- } else if (args[i] === "--thread" && args[i + 1]) {
19
- threadTs = args[++i];
20
- } else {
21
- msgParts.push(args[i]);
22
- }
30
+ let { channel, toChannelId, threadTs, message, help } = parseSendArgs();
31
+ if (help) {
32
+ console.log(SEND_USAGE);
33
+ return;
23
34
  }
24
- const message = msgParts.join(" ");
25
- if (!message) fail("Usage: nia send [-c channel] [--to <slack-channel-id>] [--thread <ts>] <message>");
35
+ if (!message) fail(SEND_USAGE);
26
36
 
27
37
  // --to implies slack channel
28
38
  if (toChannelId) channel = channel || "slack";
@@ -112,6 +112,11 @@ export interface Config {
112
112
  activeHours: { start: string; end: string };
113
113
  database_url: string;
114
114
  log_level: string;
115
+ /** Long-lived token from `claude setup-token`. Preferred: it keeps the
116
+ * subscription and does not depend on anyone logging in on this machine. */
117
+ anthropic_oauth_token: string | null;
118
+ /** Metered per-token billing — a different bill from the subscription. */
119
+ anthropic_api_key: string | null;
115
120
  gemini_api_key: string | null;
116
121
  sessionFinalization: SessionFinalizationConfig;
117
122
  channels: ChannelsConfig;
@@ -15,6 +15,8 @@ const DEFAULTS: Config = {
15
15
  activeHours: { start: "00:00", end: "23:59" },
16
16
  database_url: DEFAULT_DATABASE_URL,
17
17
  log_level: "info",
18
+ anthropic_oauth_token: null,
19
+ anthropic_api_key: null,
18
20
  gemini_api_key: null,
19
21
  sessionFinalization: {
20
22
  enabled: true,
@@ -129,6 +131,11 @@ export function loadConfig(): Config {
129
131
  const log_level = process.env.LOG_LEVEL || (typeof raw.log_level === "string" ? raw.log_level : DEFAULTS.log_level);
130
132
 
131
133
  // Gemini API key — env var overrides config
134
+ const anthropic_oauth_token =
135
+ process.env.CLAUDE_CODE_OAUTH_TOKEN ||
136
+ (typeof raw.anthropic_oauth_token === "string" ? raw.anthropic_oauth_token : null);
137
+ const anthropic_api_key =
138
+ process.env.ANTHROPIC_API_KEY || (typeof raw.anthropic_api_key === "string" ? raw.anthropic_api_key : null);
132
139
  const gemini_api_key =
133
140
  process.env.GEMINI_API_KEY || (typeof raw.gemini_api_key === "string" ? raw.gemini_api_key : null);
134
141
 
@@ -256,6 +263,8 @@ export function loadConfig(): Config {
256
263
  activeHours: { start, end },
257
264
  database_url,
258
265
  log_level,
266
+ anthropic_oauth_token,
267
+ anthropic_api_key,
259
268
  gemini_api_key,
260
269
  sessionFinalization,
261
270
  channels: {
@@ -1,61 +0,0 @@
1
- /**
2
- * A watch turn answers one question: say something, or stay quiet.
3
- *
4
- * That answer used to travel as a `[NO_REPLY]` sentinel inside prose, which
5
- * means parsing a decision out of an answer that was free to phrase it any way
6
- * it liked. It mostly worked — and 47 times it produced the sentinel *and*
7
- * content, where the code has to guess which the model meant.
8
- *
9
- * A schema removes the guess. The sentinel path stays as the fallback, because
10
- * a backend can decline to produce structured output and a watch that stops
11
- * deciding is worse than one that decides the old way.
12
- */
13
-
14
- export const WATCH_JUDGEMENT_SCHEMA: Record<string, unknown> = {
15
- type: "object",
16
- properties: {
17
- reply: {
18
- type: ["string", "null"],
19
- description: "The message to post in the channel, or null to stay silent. Most turns are null.",
20
- },
21
- },
22
- required: ["reply"],
23
- additionalProperties: false,
24
- };
25
-
26
- export interface WatchDecision {
27
- /** Whether to post at all. */
28
- send: boolean;
29
- /** What to post. Empty when staying quiet. */
30
- text: string;
31
- /** Which path decided, so a drop in schema coverage is visible in the log. */
32
- source: "structured" | "sentinel";
33
- /** Set only on the sentinel path, when the model emitted both a sentinel and content. */
34
- ambiguous?: boolean;
35
- }
36
-
37
- /** Strip markdown fencing so a sentinel matches even when the model wraps it.
38
- * Underscores are left alone — the sentinel contains one. */
39
- export function cleanSentinel(text: string): string {
40
- return text.replace(/[`*]/g, "").trim();
41
- }
42
-
43
- /**
44
- * `structured` wins when it carries the `reply` key the schema requires —
45
- * including an explicit null, which is a decision, not an absence.
46
- */
47
- export function decideWatchReply(structured: unknown, raw: string): WatchDecision {
48
- if (structured && typeof structured === "object" && "reply" in structured) {
49
- const reply = (structured as { reply: unknown }).reply;
50
- const text = typeof reply === "string" ? reply.trim() : "";
51
- return { send: text.length > 0, text, source: "structured" };
52
- }
53
-
54
- const trimmed = raw.trim();
55
- const cleaned = cleanSentinel(trimmed);
56
- if (!trimmed || cleaned.includes("[NO_REPLY]")) {
57
- const exact = !trimmed || cleaned === "[NO_REPLY]";
58
- return { send: false, text: "", source: "sentinel", ambiguous: !exact };
59
- }
60
- return { send: true, text: trimmed, source: "sentinel" };
61
- }