teapot-coding-agent 0.37.185 → 0.37.187

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.
@@ -1125,23 +1125,56 @@ export class Agent {
1125
1125
  this.bgExits = [];
1126
1126
  }
1127
1127
  }
1128
+ /**
1129
+ * goal.md is `body` + a metadata block, and the block is found by POSITION-INDEPENDENT parsing.
1130
+ *
1131
+ * The reader used to strip `status:`/`updated:` only when they sat at the very END. The writer appends `verify:` AFTER
1132
+ * them, so on any goal with a verification contract the strip loop could not remove anything and the metadata lines
1133
+ * became part of the goal TEXT — reported as "セッションリストアするとgoalが汚染される" with a goal that ended in
1134
+ * `status: paused / updated: …` (#253).
1135
+ *
1136
+ * The block is the LAST run of metadata lines in the file: find the last `status:` line and keep it only if everything
1137
+ * after it is metadata (`updated:`, `verify:` and the contract's own lines). A goal whose prose mentions "status:" stays
1138
+ * intact, because prose is not followed by an `updated:` line.
1139
+ */
1128
1140
  parseGoalFile(text) {
1129
- const all = [...text.matchAll(/status:\s*(\w+)/gi)];
1130
- const last = all[all.length - 1]?.[1];
1131
- const status = last === "done" ? "done" : last === "paused" ? "paused" : "active";
1132
- let body = text.trim();
1133
- for (let i = 0; i < 4; i++) {
1134
- const stripped = body.replace(/\n+(?:status|updated):[^\n]*$/i, "").trimEnd();
1135
- if (stripped === body)
1141
+ const lines = text.replace(/\r\n/g, "\n").split("\n");
1142
+ let blockAt = -1;
1143
+ for (let i = lines.length - 1; i >= 0; i--) {
1144
+ if (/^status:\s*\w+\s*$/i.test(lines[i])) {
1145
+ // everything below must be metadata, or this "status:" belongs to the prose
1146
+ let ok = true;
1147
+ let inVerify = false;
1148
+ for (let j = i + 1; j < lines.length; j++) {
1149
+ const l = lines[j];
1150
+ if (!l.trim())
1151
+ continue;
1152
+ if (/^verify:/i.test(l)) {
1153
+ inVerify = true; // the contract's own lines are arbitrary text
1154
+ continue;
1155
+ }
1156
+ if (inVerify)
1157
+ continue;
1158
+ if (/^(updated|status):/i.test(l))
1159
+ continue;
1160
+ ok = false;
1161
+ break;
1162
+ }
1163
+ if (ok)
1164
+ blockAt = i;
1136
1165
  break;
1137
- body = stripped;
1166
+ }
1138
1167
  }
1168
+ const bodyLines = blockAt === -1 ? lines : lines.slice(0, blockAt);
1169
+ const meta = blockAt === -1 ? [] : lines.slice(blockAt);
1170
+ const metaText = meta.join("\n");
1171
+ const statusWord = /status:\s*(\w+)/i.exec(metaText)?.[1]?.toLowerCase();
1172
+ const status = statusWord === "done" ? "done" : statusWord === "paused" ? "paused" : "active";
1173
+ let body = bodyLines.join("\n").trimEnd();
1139
1174
  let verify = "";
1140
- const vm = body.match(/\nverify:\s*\n([\s\S]+)$/i) ?? body.match(/^verify:\s*(.+)$/im);
1141
- if (vm) {
1175
+ const vm = /\nverify:\s*\n([\s\S]+)$/i.exec(metaText) ?? /^verify:\s*(.+)$/im.exec(metaText);
1176
+ if (vm)
1142
1177
  verify = vm[1].trim();
1143
- body = (body.slice(0, vm.index).trimEnd());
1144
- }
1145
1178
  const out = { text: body, status, updatedAt: new Date().toISOString() };
1146
1179
  if (verify)
1147
1180
  out.verify = verify;
@@ -116,27 +116,32 @@ export function claimsContextLimit(text) {
116
116
  export function usageLine(used, budget, window) {
117
117
  const fmt = (n) => (n >= 1_000 ? `${Math.round(n / 1_000)}k` : String(Math.max(0, Math.round(n))));
118
118
  /**
119
- * The percentage is of the WINDOW, and the compaction threshold is stated separately.
119
+ * EVERY number, named once.
120
120
  *
121
- * It used to be `used / budget`, where `budget` is the compaction threshold — so a 900k window that compacts at 750k
122
- * read as "82% used" at 615k and the denominator was a number the model could not see anywhere else. Reported as
123
- * "生のサイズじゃなくてcompactボーダーまでにしたのはなぜ? 消費/フルサイズ とcompactまでの残り値にすべきでは?".
121
+ * The line used to mix two denominators without naming either: the percentage was of the WINDOW while "until
122
+ * auto-compaction" was measured from the BUDGET (the compaction threshold). A model reading
123
+ * "context 1% used (15k/1000k tokens, 735k until compaction)" had three different numbers in front of it and concluded it
124
+ * could not tell what they meant — "compact後のコンテキストウィンドウ残量報告が間違っている?" (#249). So the line states the
125
+ * percentage with its window, the threshold as its own figure, and the room left before that threshold:
124
126
  *
125
- * context 68% used (615k/900k tokens, 135k until compaction)
126
- *
127
- * When the window is unknown the budget is the only denominator available, and the line says so by omitting the
128
- * window entirely rather than pretending they are the same thing.
127
+ * context 1% used (15k/1000k tokens) · auto-compacts at 750k · 735k until auto-compaction
128
+ *
129
+ * When the window is unknown the budget is the only denominator there is, and the line says so rather than pretending.
129
130
  */
130
131
  const full = Number.isFinite(window) && (window ?? 0) > 0 ? window : 0;
131
- const denom = full || (Number.isFinite(budget) && budget > 0 ? budget : 0);
132
+ const hasBudget = Number.isFinite(budget) && budget > 0;
133
+ const denom = full || (hasBudget ? budget : 0);
132
134
  if (!denom)
133
135
  return `context ${fmt(used)} tokens used`;
134
136
  const pct = Math.max(0, Math.min(100, Math.round((used / denom) * 100)));
135
- const until = Number.isFinite(budget) && budget > 0 ? Math.max(0, budget - used) : null;
136
- const size = full ? `${fmt(used)}/${fmt(full)} tokens` : `${fmt(used)}/${fmt(budget)} tokens (compaction threshold; full window unknown)`;
137
- return until === null
138
- ? `context ${pct}% used (${size})`
139
- : `context ${pct}% used (${size}, ${fmt(until)} until compaction)`;
137
+ const size = full
138
+ ? `${fmt(used)}/${fmt(full)} tokens`
139
+ : `${fmt(used)}/${fmt(budget)} tokens — the compaction threshold; full window unknown`;
140
+ if (!hasBudget)
141
+ return `context ${pct}% used (${size})`;
142
+ const until = Math.max(0, budget - used);
143
+ const threshold = full ? ` · auto-compacts at ${fmt(budget)}` : "";
144
+ return `context ${pct}% used (${size})${threshold} · ${fmt(until)} until auto-compaction`;
140
145
  }
141
146
  /**
142
147
  * …and the OTHER thing that gets reported as "context is exhausted": the model has run out of its own ideas.
@@ -19,10 +19,27 @@ export class McpRegistry {
19
19
  auths = new Map();
20
20
  store = null;
21
21
  flows = new Map();
22
+ /**
23
+ * A counter that changes whenever the tool set can have changed.
24
+ *
25
+ * The deferred catalogue (`search_tools`/`run_tool`) is cached per ToolContext, and a cache keyed only by that context
26
+ * cannot see the registry change underneath it: after re-authorising a server the tools were populated here but
27
+ * `search_tools` kept serving the catalogue built before the grant — "MCP再認証してもsearch_toolsで出てこない、叩けない"
28
+ * (#243). A version is the smallest thing the cache can compare.
29
+ */
30
+ version = 0;
31
+ /** the current tool-set version; see `version` */
32
+ stateVersion() {
33
+ return this.version;
34
+ }
35
+ bump() {
36
+ this.version++;
37
+ }
22
38
  init(dataDir) {
23
39
  this.store = new McpAuthStore(dataDir);
24
40
  }
25
41
  apply(configs) {
42
+ this.bump();
26
43
  const wanted = new Set(configs.map((c) => c.name));
27
44
  for (const [name, st] of this.servers) {
28
45
  if (!wanted.has(name)) {
@@ -59,6 +76,7 @@ export class McpRegistry {
59
76
  return this.servers.get(name);
60
77
  }
61
78
  recordSuccess(name, tools) {
79
+ this.bump();
62
80
  const st = this.servers.get(name);
63
81
  if (!st)
64
82
  return;
@@ -226,6 +244,7 @@ export class McpRegistry {
226
244
  const conn = await connectMcp(st.config, { provider: auth });
227
245
  st.connection = conn;
228
246
  st.tools = await conn.tools();
247
+ this.bump();
229
248
  st.status = "connected";
230
249
  st.oauthRequired = false;
231
250
  await this.finishFlow(name, pending);
@@ -251,6 +270,7 @@ export class McpRegistry {
251
270
  await conn.close().catch(() => { });
252
271
  st.connection = live;
253
272
  st.tools = await live.tools();
273
+ this.bump();
254
274
  st.status = "connected";
255
275
  st.oauthRequired = false;
256
276
  st.lastError = undefined;
@@ -1502,10 +1502,34 @@ export const TOOLS = [
1502
1502
  async run(args, ctx) {
1503
1503
  if (args.background === true) {
1504
1504
  const id = startBackgroundShell(str(args.command), ctx);
1505
+ /**
1506
+ * The result has to say two things the model kept getting wrong (#231, #235).
1507
+ *
1508
+ * - "you are woken when it exits" was already there, but models still polled or kept busy-working while a long
1509
+ * build ran ("モデルがバックエンドbashを待つために止まるってことを知らない / ポーリングしたり、余計な作業をし続けたり").
1510
+ * So the WAKE is stated as an instruction: stopping the round is how you wait, and you do not need to poll.
1511
+ * - and it lists the other jobs still running, because a model that starts three builds forgets the first two
1512
+ * ("そうじゃないと起動したこと忘れるモデルがある").
1513
+ */
1514
+ const running = [...bgMapFor(ctx.cwd).values()].filter((j) => !j.exited);
1515
+ const others = running.filter((j) => j.id !== id);
1516
+ const list = running
1517
+ .map((j) => {
1518
+ const secs = Math.round((Date.now() - j.startedAt) / 1000);
1519
+ const when = secs >= 60 ? `${Math.round(secs / 60)}m` : `${secs}s`;
1520
+ const cmd = j.cmd.replace(/\s+/g, " ").slice(0, 90);
1521
+ return ` ${j.id}${j.id === id ? " (just started)" : ""} · ${when} · ${cmd}`;
1522
+ })
1523
+ .join("\n");
1505
1524
  return {
1506
1525
  ok: true,
1507
- result: `started in background (job ${id}) — you are notified when it exits (code + output tail); ` +
1508
- `do NOT poll it. Output: bash_output(job_id="${id}"). Stop: bash_output(job_id="${id}", action="kill").`,
1526
+ result: `started in background (job ${id}). ` +
1527
+ `You are WOKEN when it exits (exit code + output tail), so waiting is not your job: do not poll it, do not ` +
1528
+ `sleep, and do not start another round just to check — end your turn (or do other work) and the harness will ` +
1529
+ `call you when it finishes.\n` +
1530
+ `Output mid-run: bash_output(job_id="${id}"). Stop: bash_output(job_id="${id}", action="kill").\n` +
1531
+ `running background jobs for this workspace (${running.length}):\n${list}` +
1532
+ (others.length ? `\nDo not start a duplicate of a job that is already here.` : ""),
1509
1533
  };
1510
1534
  }
1511
1535
  const explicit = args.timeout_ms !== undefined && args.timeout_ms !== null;
@@ -2212,20 +2236,21 @@ export function repairToolInput(name, rawArgs) {
2212
2236
  */
2213
2237
  const catalogs = new WeakMap();
2214
2238
  function catalogFor(ctx) {
2215
- let c = catalogs.get(ctx);
2216
- if (!c) {
2217
- c = createCatalog(mcp.states().map((st) => ({
2218
- id: st.config.name,
2219
- name: st.config.name,
2220
- description: st.config.description,
2221
- enabled: st.status !== "disabled",
2222
- status: st.status,
2223
- tools: st.tools ?? [],
2224
- })),
2225
- // the built-ins are already in the request, so the catalog must know how much room is left
2226
- { builtInCount: TOOLS.length });
2227
- catalogs.set(ctx, c);
2228
- }
2239
+ const version = mcp.stateVersion();
2240
+ const hit = catalogs.get(ctx);
2241
+ if (hit && hit.version === version)
2242
+ return hit.catalog;
2243
+ const c = createCatalog(mcp.states().map((st) => ({
2244
+ id: st.config.name,
2245
+ name: st.config.name,
2246
+ description: st.config.description,
2247
+ enabled: st.status !== "disabled",
2248
+ status: st.status,
2249
+ tools: st.tools ?? [],
2250
+ })),
2251
+ // the built-ins are already in the request, so the catalog must know how much room is left
2252
+ { builtInCount: TOOLS.length });
2253
+ catalogs.set(ctx, { version, catalog: c });
2229
2254
  return c;
2230
2255
  }
2231
2256
  /** called at a turn boundary so each turn starts with nothing discovered */
package/dist/master.js CHANGED
@@ -750,13 +750,19 @@ export class Master {
750
750
  if (!parent)
751
751
  return;
752
752
  this.wakeAncestorWaiters(ac.parent);
753
- if (this.reportedRunning.has(agentId) && SETTLED.has(status))
753
+ /**
754
+ * An `error` child is NOT announced twice.
755
+ *
756
+ * The child's own `error` event already reaches the parent as "Sub-agent X hit an error: …" (see `onChildEvent`),
757
+ * so the generic "settled without having called finish()" note only repeated it with weaker wording — reported as
758
+ * "サブエージェントがエラー落ちしたときの親への通知がびみょい" (#245). The generic note exists for statuses that have
759
+ * no event of their own (`idle`, `stopped`, `waiting`).
760
+ */
761
+ if (this.reportedRunning.has(agentId) && SETTLED.has(status) && status !== "error")
754
762
  parent.enqueuePrompt(`[harness] Sub-agent @${agentId} is now ${status} without having called finish(), so it has not reported a result.` +
755
- (status === "error"
756
- ? ` Check its last output before deciding whether to re-spawn it.`
757
- : status === "stopped"
758
- ? ` It may have been stopped mid-task; message it to resume, or re-spawn it.`
759
- : ` If you expected a result, message it or re-spawn it.`), "harness");
763
+ (status === "stopped"
764
+ ? ` It may have been stopped mid-task; message it to resume, or re-spawn it.`
765
+ : ` If you expected a result, message it or re-spawn it.`), "harness");
760
766
  if (status === "running")
761
767
  this.reportedRunning.add(agentId);
762
768
  else if (SETTLED.has(status))
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "teapot-coding-agent",
3
- "version": "0.37.185",
3
+ "version": "0.37.187",
4
4
  "description": "A lightweight, always-on multi-agent harness for AI coding agents",
5
5
  "type": "module",
6
6
  "license": "AGPL-3.0-or-later",