myrmo-mcp 0.28.4 → 0.29.0
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/doctor.js +1 -1
- package/dist/hooks/lookup.mjs +21 -3
- package/dist/hooks/on-failure.mjs +79 -17
- package/dist/hooks/redact.mjs +1 -1
- package/dist/server.d.ts +1 -1
- package/dist/server.js +5 -2
- package/dist/trail.v1.schema.json +5 -0
- package/package.json +1 -1
package/dist/doctor.js
CHANGED
|
@@ -51,7 +51,7 @@ export async function diagnose(input) {
|
|
|
51
51
|
name: "Reaching the colony",
|
|
52
52
|
detail: refused ? "the machine answered but nothing listens on that port" : `no answer (${ready.error})`,
|
|
53
53
|
fix: refused
|
|
54
|
-
? "Ask whoever runs the colony whether it is up
|
|
54
|
+
? "Ask whoever runs the colony whether it is up."
|
|
55
55
|
: "Check that you are on the team's network and that the address and port are right; from this machine try: curl " + input.url + "/readyz",
|
|
56
56
|
});
|
|
57
57
|
return checks;
|
package/dist/hooks/lookup.mjs
CHANGED
|
@@ -260,16 +260,32 @@ export async function lookupTask(task, { env = process.env, settings = {}, fetch
|
|
|
260
260
|
|
|
261
261
|
const KINDS = { procedure: "procedure", fact: "fact", fix: "known error" };
|
|
262
262
|
|
|
263
|
+
/**
|
|
264
|
+
* The hits worth showing unasked: a suggestion the agent did not ask for has to be clearly about the task, or it teaches
|
|
265
|
+
* the agent to ignore the next one. A hit below `min` is left out, and so is one far below the best (more than `spread`).
|
|
266
|
+
* Measured on an organisation's colony: the right entry scored 0.63 to 0.79, unrelated ones 0.55 to 0.59, so no single
|
|
267
|
+
* cut separates them, but the two together do. A hit without a score (an older colony) is kept.
|
|
268
|
+
*/
|
|
269
|
+
export function relevantHits(results, { min = 0.6, spread = 0.15 } = {}) {
|
|
270
|
+
const score = (hit) => Number(hit?.match?.score);
|
|
271
|
+
const scored = results.filter((hit) => Number.isFinite(score(hit)));
|
|
272
|
+
if (!scored.length) return results;
|
|
273
|
+
const best = Math.max(...scored.map(score));
|
|
274
|
+
return results.filter((hit) => !Number.isFinite(score(hit)) || (score(hit) >= min && score(hit) >= best - spread));
|
|
275
|
+
}
|
|
276
|
+
|
|
263
277
|
/**
|
|
264
278
|
* The short list for the agent, and the state that lets the Stop hook ask for reports on what was shown. No note when
|
|
265
279
|
* nothing matched.
|
|
266
280
|
*/
|
|
267
|
-
export function briefNote(task, body, state = {}, now = Date.now(), how = "") {
|
|
281
|
+
export function briefNote(task, body, state = {}, now = Date.now(), how = "", { min = 0.6, spread = 0.15 } = {}) {
|
|
268
282
|
// An entry already shown in this session is not shown again: the agent has it, and repeating an unrelated one on every
|
|
269
283
|
// prompt is noise.
|
|
270
284
|
const seen = Array.isArray(state.briefSeen) ? state.briefSeen : [];
|
|
271
285
|
const idOf = (hit) => String(hit?.trail_id ?? "").replace(/[^0-9a-fA-F-]/g, "").slice(0, 36);
|
|
272
|
-
const hits = (Array.isArray(body?.results) ? body.results : []
|
|
286
|
+
const hits = relevantHits(Array.isArray(body?.results) ? body.results : [], { min, spread })
|
|
287
|
+
.filter((hit) => !seen.includes(idOf(hit)))
|
|
288
|
+
.slice(0, 3);
|
|
273
289
|
if (!hits.length) return { note: null, state };
|
|
274
290
|
const ids = [];
|
|
275
291
|
const lines = hits.map((hit, i) => {
|
|
@@ -281,7 +297,9 @@ export function briefNote(task, body, state = {}, now = Date.now(), how = "") {
|
|
|
281
297
|
const about = clip(t.problem?.summary || t.problem?.error_message || t.problem?.error_type, 160);
|
|
282
298
|
const o = hit.outcomes ?? {};
|
|
283
299
|
const strength = Number(hit.strength ?? 0);
|
|
284
|
-
|
|
300
|
+
const due = /^\d{4}-\d{2}-\d{2}$/.exec(String(hit.review_due ?? ""))?.[0];
|
|
301
|
+
const review = due ? ` [review due since ${due}: may be out of date]` : "";
|
|
302
|
+
return `${i + 1}. [${kind}] ${about} (strength ${Number.isFinite(strength) ? strength.toFixed(2) : "?"}, worked ${num(o.worked, "0")}, failed ${num(o.failed, "0")})${review} id ${id}`;
|
|
285
303
|
});
|
|
286
304
|
const note = [
|
|
287
305
|
"Myrmo (your organisation's private colony) already has knowledge that may help with this task:",
|
|
@@ -21,7 +21,7 @@
|
|
|
21
21
|
// 5. After a failed command, the hook searches the colony itself and shows a strong match (lookup.mjs).
|
|
22
22
|
// 6. When the user sends a request, the hook asks the colony what the organisation already knows about that task and
|
|
23
23
|
// shows a short list (UserPromptSubmit; lookup.mjs `briefNote`). MYRMO_HOOK_BRIEF=off turns only this off.
|
|
24
|
-
// 7. When the agent stops after a turn in which commands failed or knowledge was shown, it is asked once to report what
|
|
24
|
+
// 7. (In the public colony too, only for trails shown and not reported on.) When the agent stops after a turn in which commands failed or knowledge was shown, it is asked once to report what
|
|
25
25
|
// it relied on and to share what the next person's agent would need (Stop hook): what is learned in a session
|
|
26
26
|
// otherwise leaves with it.
|
|
27
27
|
//
|
|
@@ -51,6 +51,7 @@ const SHELLS = new Set(["Bash", "PowerShell"]);
|
|
|
51
51
|
const SEARCH_TOOL = /^mcp__.*myrmo_search$/;
|
|
52
52
|
const REPORT_TOOL = /^mcp__.*myrmo_report$/;
|
|
53
53
|
const PUBLISH_TOOL = /^mcp__.*myrmo_publish$/;
|
|
54
|
+
const BRIEF_TOOL = /^mcp__.*myrmo_brief$/;
|
|
54
55
|
// The id each trail is listed under in a search reply (`## Trail 1 of 2 · id <uuid>`).
|
|
55
56
|
const TRAIL_ID = /^## Trail \d+ of \d+ · id ([0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12})$/i;
|
|
56
57
|
|
|
@@ -183,6 +184,29 @@ function segmentsWithPipes(text) {
|
|
|
183
184
|
return out.map((w) => w.filter((x) => !/^\d*[<>]&?\d*$/.test(x) && !/^\d*>&\d+$/.test(x)));
|
|
184
185
|
}
|
|
185
186
|
|
|
187
|
+
// A tool saying it was called wrongly (a mistyped flag): the agent's own slip, fixed by reading the line, not knowledge to
|
|
188
|
+
// look up or share.
|
|
189
|
+
const USAGE_LINE = /\b(?:unknown (?:shorthand )?flag|unknown option|unrecognized option|invalid option|illegal option)\b|^\s*usage:/im;
|
|
190
|
+
|
|
191
|
+
/** Whether the command ends in a test whose "false" is an answer: `... ; [ -f x ] && echo found`, also as the last
|
|
192
|
+
* statement of a loop body. Exit 1 then says the condition did not hold, not that anything failed. */
|
|
193
|
+
export function endsInTest(command) {
|
|
194
|
+
const body = String(command).trim().replace(/(?:\s*;?\s*\b(?:done|fi)\b)+\s*$/, "");
|
|
195
|
+
const last = body.split(/;|\n|\bdo\b|\bthen\b/).at(-1)?.trim() ?? "";
|
|
196
|
+
// After `||` or a pipe the exit code is another command's: `test -d x || mkdir x` failing is mkdir failing.
|
|
197
|
+
if (/\|/.test(last)) return false;
|
|
198
|
+
const isTest = (part) => /^(?:\[\[?|test\b)/.test(part);
|
|
199
|
+
const parts = last.split(/&&/).map((part) => part.trim());
|
|
200
|
+
// `[ ... ]` alone, `[ ... ] && echo yes`, or a chain whose test comes right before the final echo.
|
|
201
|
+
return isTest(parts[0]) && parts.length <= 2
|
|
202
|
+
|| (parts.length >= 2 && isTest(parts.at(-2)) && /^(?:echo|printf|true|:)(?:\s|$)/.test(parts.at(-1)));
|
|
203
|
+
}
|
|
204
|
+
|
|
205
|
+
/** A line in the shape a program gives its messages, `program: what happened`, and not data such as `key=value`. */
|
|
206
|
+
function toolMessage(line) {
|
|
207
|
+
return /^[\w.\/-]+:\s+\S/.test(line) && !/\w=\S/.test(line) ? line : "";
|
|
208
|
+
}
|
|
209
|
+
|
|
186
210
|
const HELP_LINE = /\[Help \d+\]|re-run maven|for more information about the errors|to see the full stack trace|^\[(?:ERROR|FATAL)\]\s*$/i;
|
|
187
211
|
|
|
188
212
|
/** The user's settings file (the same one the SDKs and the MCP server read). Missing or damaged means defaults. */
|
|
@@ -323,6 +347,11 @@ export function decide(input, { now = Date.now(), state = {}, env = process.env,
|
|
|
323
347
|
}
|
|
324
348
|
return quiet(state);
|
|
325
349
|
}
|
|
350
|
+
// The agent read the knowledge the hook suggested (myrmo_brief) or searched itself: only then may it be asked to report.
|
|
351
|
+
if (BRIEF_TOOL.test(String(tool)) || (SEARCH_TOOL.test(String(tool)) && state.brief && !state.brief.used)) {
|
|
352
|
+
if (state.brief && !state.brief.reported) state = { ...state, brief: { ...state.brief, used: true } };
|
|
353
|
+
if (BRIEF_TOOL.test(String(tool))) return quiet(state);
|
|
354
|
+
}
|
|
326
355
|
// 4b. The agent reported: nothing left to remind. A trail that worked for the open problem: what it cost with help.
|
|
327
356
|
if (REPORT_TOOL.test(String(tool))) {
|
|
328
357
|
const brief = state.brief ? { ...state.brief, reported: true } : state.brief;
|
|
@@ -336,7 +365,7 @@ export function decide(input, { now = Date.now(), state = {}, env = process.env,
|
|
|
336
365
|
}
|
|
337
366
|
// The agent shared what it learned: the failures so far need no further asking when it stops.
|
|
338
367
|
if (PUBLISH_TOOL.test(String(tool))) {
|
|
339
|
-
return quiet({ ...state, harvestedAt: Math.max(Number(state.harvestedAt ?? 0), Number(state.failures ?? 0)), publishedTs: now });
|
|
368
|
+
return quiet({ ...state, harvestedAt: Math.max(Number(state.harvestedAt ?? 0), Number(state.failures ?? 0)), harvestedResolved: Number(state.resolved ?? 0), publishedTs: now });
|
|
340
369
|
}
|
|
341
370
|
if (!SHELLS.has(tool)) return quiet(state);
|
|
342
371
|
|
|
@@ -357,10 +386,14 @@ export function decide(input, { now = Date.now(), state = {}, env = process.env,
|
|
|
357
386
|
// 1. A command failed.
|
|
358
387
|
if (failed) {
|
|
359
388
|
if (input.is_interrupt === true || STOPPED.has(code) || isProbe(command) || answered(command, code)) return quiet(state);
|
|
389
|
+
if (code === 1 && endsInTest(command)) return quiet(state);
|
|
360
390
|
const output = errorText.replace(/^Exit code \d+\r?\n?/, "");
|
|
361
|
-
// A company's own tools print errors in no
|
|
362
|
-
//
|
|
363
|
-
|
|
391
|
+
// A company's own tools print errors in shapes no pattern knows (`migrate: lock 'migrate' is held by ci-runner-17`):
|
|
392
|
+
// the loose patterns take a line that says something failed, or else a last line in the shape tools give their
|
|
393
|
+
// messages (`program: message`). Nothing else is quoted or looked up: the last line of a loop that printed data
|
|
394
|
+
// (`bucket-a objects=95`) is not an error, and searching for it only teaches the agent to ignore the note.
|
|
395
|
+
const line = lastErrorLine(output, { loose: true }) || toolMessage(lastLine(output));
|
|
396
|
+
if (line && USAGE_LINE.test(line)) return quiet(state);
|
|
364
397
|
const problem = openProblem(state, now) ?? (line ? { at: now, line } : null);
|
|
365
398
|
const next = { ...state, failed: { key: commandKey(command), at: now, line }, failures: (state.failures ?? 0) + 1, problem };
|
|
366
399
|
const sig = signature(line);
|
|
@@ -377,6 +410,11 @@ export function decide(input, { now = Date.now(), state = {}, env = process.env,
|
|
|
377
410
|
|
|
378
411
|
if (mode === "failures" || event === undefined || isProbe(command)) return quiet(state);
|
|
379
412
|
|
|
413
|
+
// A command that failed now works: something was solved, which is what the Stop hook may ask to share.
|
|
414
|
+
if (state.failed && !state.failed.resolved && commandKey(command) === state.failed.key && now - state.failed.at < PROBLEM_MS) {
|
|
415
|
+
state = { ...state, failed: { ...state.failed, resolved: true }, resolved: Number(state.resolved ?? 0) + 1 };
|
|
416
|
+
}
|
|
417
|
+
|
|
380
418
|
// 3b. A command that failed earlier now works, and Myrmo had nothing: the agent may have solved something new.
|
|
381
419
|
const nomatch = state.nomatch;
|
|
382
420
|
if (nomatch && !nomatch.nudged && now - nomatch.at < 45 * 60_000 && state.failed && commandKey(command) === state.failed.key && !spent) {
|
|
@@ -450,24 +488,46 @@ export function decide(input, { now = Date.now(), state = {}, env = process.env,
|
|
|
450
488
|
}
|
|
451
489
|
|
|
452
490
|
/**
|
|
453
|
-
*
|
|
454
|
-
*
|
|
455
|
-
*
|
|
456
|
-
*
|
|
491
|
+
* 6b. The agent is about to stop in the public colony, with trails shown to it and not reported on: ask once for the
|
|
492
|
+
* report, and nothing else. Measured on the public colony, 8 reports followed 235 searches that returned trails: the
|
|
493
|
+
* reminder after a failed command succeeds again rarely applies (a fix is often checked by another command), and
|
|
494
|
+
* without the outcome a trail can neither rise nor fade. Not after the reminder of case 4 (it already asked), not for
|
|
495
|
+
* trails shown more than 12 hours ago, and once per set of trails.
|
|
496
|
+
*/
|
|
497
|
+
function askToReport(state, now) {
|
|
498
|
+
const found = state.found;
|
|
499
|
+
const none = { reason: null, state };
|
|
500
|
+
if (!found || found.nudged || found.asked || !Array.isArray(found.ids) || !found.ids.length) return none;
|
|
501
|
+
if (now - found.at > 12 * 3_600_000) return none;
|
|
502
|
+
const reason =
|
|
503
|
+
`Myrmo: before you finish, call myrmo_report once for each trail it showed you (${found.ids.join(", ")}) that you tried: ` +
|
|
504
|
+
"worked, partially_worked, or failed if it was wrong or out of date, saying what changed; not_applicable if it was about something else. " +
|
|
505
|
+
"It is how good trails rise and stale ones fade. If you tried none, finish.";
|
|
506
|
+
return { reason, state: { ...state, found: { ...found, asked: true } } };
|
|
507
|
+
}
|
|
508
|
+
|
|
509
|
+
/**
|
|
510
|
+
* 6. The agent is about to stop. In a private colony, after a turn in which a failed command was made to work, ask it once
|
|
511
|
+
* to share what it learned before it finishes: Claude Code then continues with this as its next instruction. At most once
|
|
512
|
+
* every MYRMO_HARVEST_MINUTES (default 30) and three times a session, never while it is already continuing because of this
|
|
513
|
+
* hook, and only for what was solved since the last time it was asked. Knowledge the hook suggested is asked about only
|
|
514
|
+
* if the agent read it (myrmo_brief or myrmo_search): a suggestion it never opened was not relied on.
|
|
457
515
|
*/
|
|
458
516
|
export function decideStop(input, { now = Date.now(), state = {}, env = process.env, settings = {} } = {}) {
|
|
459
517
|
const none = { reason: null, state };
|
|
460
|
-
if (hookMode(env, settings) === "off" ||
|
|
518
|
+
if (hookMode(env, settings) === "off" || env.MYRMO_HARVEST === "off") return none;
|
|
461
519
|
if (input?.stop_hook_active === true) return none;
|
|
462
|
-
|
|
463
|
-
|
|
464
|
-
const
|
|
465
|
-
|
|
520
|
+
if (!privateColony(env, settings)) return askToReport(state, now);
|
|
521
|
+
const resolved = Number(state.resolved ?? 0);
|
|
522
|
+
const solvedSince = resolved > Number(state.harvestedResolved ?? 0);
|
|
523
|
+
// Knowledge shown at the start of the task, read by the agent and not reported on yet: worth asking once.
|
|
524
|
+
const shown = state.brief && state.brief.used && !state.brief.reported && !state.brief.asked ? state.brief : null;
|
|
525
|
+
if (!solvedSince && !shown) return none;
|
|
466
526
|
const minMs = Number(env.MYRMO_HARVEST_MINUTES ?? 30) * 60_000;
|
|
467
527
|
if ((state.harvests ?? 0) >= 3 || (state.harvestedTs && now - state.harvestedTs < minMs)) return none;
|
|
468
528
|
// Nothing new failed since this session last published: what is left is only reporting on what was shown, so ask that alone
|
|
469
529
|
// instead of asking again to share what was learned.
|
|
470
|
-
if (
|
|
530
|
+
if (!solvedSince && shown) {
|
|
471
531
|
const reason =
|
|
472
532
|
`Myrmo: before you finish, call myrmo_report for each entry it showed you for this task (${shown.ids.join(", ")}) that you actually relied on ` +
|
|
473
533
|
"(worked, partially_worked, or failed if it was wrong or out of date, saying what changed); skip the ones you did not use. If you used none, finish.";
|
|
@@ -486,7 +546,7 @@ export function decideStop(input, { now = Date.now(), state = {}, env = process.
|
|
|
486
546
|
"what, where something is configured), publish it with kind \"fact\". Keep internal names and addresses as printed; never include credentials or personal data. " +
|
|
487
547
|
"If nothing qualifies, say so in one line and finish.";
|
|
488
548
|
const brief = shown ? { ...shown, asked: true } : state.brief;
|
|
489
|
-
return { reason, state: { ...state, brief, harvestedAt: failures, harvestedTs: now, harvests: (state.harvests ?? 0) + 1 } };
|
|
549
|
+
return { reason, state: { ...state, brief, harvestedAt: Number(state.failures ?? 0), harvestedResolved: resolved, harvestedTs: now, harvests: (state.harvests ?? 0) + 1 } };
|
|
490
550
|
}
|
|
491
551
|
|
|
492
552
|
/** Without the file (a folder that cannot be trusted, a full disk) the notes are simply not throttled. */
|
|
@@ -515,7 +575,9 @@ async function main() {
|
|
|
515
575
|
const task = taskOf(input.prompt);
|
|
516
576
|
if (!task || hookMode(process.env, settings) === "off" || !briefEnabled(process.env, settings)) return;
|
|
517
577
|
const body = await lookupTask(task, { settings, agent: "myrmo-claude-code-hook/1" });
|
|
518
|
-
|
|
578
|
+
// MYRMO_HOOK_BRIEF_MIN: the lowest score a suggestion may have (default 0.6).
|
|
579
|
+
const min = Number(process.env.MYRMO_HOOK_BRIEF_MIN);
|
|
580
|
+
const out = briefNote(task, body, state, Date.now(), "", Number.isFinite(min) && min > 0 ? { min } : {});
|
|
519
581
|
save(file, state, out.state);
|
|
520
582
|
if (out.note) process.stdout.write(JSON.stringify({ hookSpecificOutput: { hookEventName: "UserPromptSubmit", additionalContext: out.note } }));
|
|
521
583
|
return;
|
package/dist/hooks/redact.mjs
CHANGED
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
// Generated from clients/typescript/packages/myrmo/src/redact.ts by clients/typescript/scripts/sync-plugin-redact.mjs.
|
|
2
2
|
// Do not edit: change the SDK, build it and run that script. The SDK's tests check that this copy is current.
|
|
3
3
|
// Client-side redaction: the same detectors the colony runs, applied before anything leaves the
|
|
4
|
-
// machine. Keep in sync with server
|
|
4
|
+
// machine. Keep in sync with the colony server's redaction and clients/python/src/myrmo/redact.py: all
|
|
5
5
|
// three must reproduce protocol/redact.v1.vectors.json exactly.
|
|
6
6
|
//
|
|
7
7
|
// Rules run in order, most specific first. A rule's replacement never matches a later rule (it
|
package/dist/server.d.ts
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
|
|
2
2
|
import { Colony, type PublishMode, type RejectionDetail } from "myrmo";
|
|
3
|
-
export declare const VERSION = "0.
|
|
3
|
+
export declare const VERSION = "0.29.0";
|
|
4
4
|
export interface ServerOptions {
|
|
5
5
|
colony: Colony;
|
|
6
6
|
/** Who this server acts as, for the agent to read: a prefix of the key's hash, or of the agent id. Never the key. */
|
package/dist/server.js
CHANGED
|
@@ -6,7 +6,7 @@ import { MyrmoError, detectEnvironment, formatResult, writeConfig } from "myrmo"
|
|
|
6
6
|
import { z } from "zod";
|
|
7
7
|
import { checkTrail } from "./check.js";
|
|
8
8
|
import { buildInstructions } from "./instructions.js";
|
|
9
|
-
export const VERSION = "0.
|
|
9
|
+
export const VERSION = "0.29.0"; // x-release-please-version
|
|
10
10
|
const PUBLIC_HOSTS = new Set(["myrmo.dev", "www.myrmo.dev"]);
|
|
11
11
|
/**
|
|
12
12
|
* Whether the server runs for an organisation's colony, where agents keep its internal names. Never for the public
|
|
@@ -55,7 +55,8 @@ const PRIVATE_REMOVE = "This colony is private to your organisation: keep its in
|
|
|
55
55
|
const KNOWLEDGE_FORMAT = `
|
|
56
56
|
This colony also keeps knowledge that does not come from an error. Set "kind":
|
|
57
57
|
- "procedure": a multi-step task your organisation repeats, done and checked. { kind: "procedure", problem: { summary: its goal (20+ chars), task_context: when it applies, in the words a person would use to ask for it, failed_approaches?: the pitfalls }, solution: { steps: [in order], shell_commands_executed?: [...] }, source?: { type: agent|handover|document|person, ref, revision }, tags? }
|
|
58
|
-
- "fact": something about your organisation's systems an agent needs and cannot guess (which machine does what, where something is configured, a rule not written in the code). { kind: "fact", problem: { summary: the fact, task_context: when it matters }, solution: { steps: [details] }, source?, tags
|
|
58
|
+
- "fact": something about your organisation's systems an agent needs and cannot guess (which machine does what, where something is configured, a rule not written in the code). { kind: "fact", problem: { summary: the fact, task_context: when it matters }, solution: { steps: [details] }, source?, tags?, review_after?: "YYYY-MM-DD" }
|
|
59
|
+
Set review_after ("YYYY-MM-DD") on a fact that describes a state that changes (a count, a size, which databases a deployment has, what is deployed where): from that day it is shown as due for review, so pick a day a few weeks to a few months ahead, as long as you would trust the fact. Leave it out for what does not age: a rule, a procedure, where something is configured, which machine does what.
|
|
59
60
|
Procedures and facts need no failed attempts, environment or verification.`;
|
|
60
61
|
const publishDescription = (minFailedAttempts, privateColony = false) => `${privateColony ? "Share what you learned with your organisation's colony: a fix, a procedure or a fact (see kind below)." : "Publish a fix to Myrmo so the next agent does not repeat your work."}
|
|
61
62
|
Use only when ALL are true: you solved an error after ${minFailedAttempts === 1 ? "at least one failed attempt" : `at least ${minFailedAttempts} failed attempts`}, you verified the fix, and either myrmo_search found no matching trail or the trails it found failed or only partly worked for you (report them with myrmo_report first, then publish your own fix as an alternative).
|
|
@@ -81,6 +82,8 @@ Use preview: true to check the payload without publishing.`;
|
|
|
81
82
|
export function noteLine(note, reason) {
|
|
82
83
|
if (note === "shown_with_trail")
|
|
83
84
|
return " Your note will be shown to the next agents with the trail.";
|
|
85
|
+
if (note === "shown_as_author_note")
|
|
86
|
+
return " Your report did not count (you wrote the trail), but your note will be shown with it as the author's note.";
|
|
84
87
|
if (note === "not_counted")
|
|
85
88
|
return " Your note was not kept either.";
|
|
86
89
|
if (note === "withheld") {
|
|
@@ -35,6 +35,11 @@
|
|
|
35
35
|
"enum": ["fix", "procedure", "fact"],
|
|
36
36
|
"default": "fix"
|
|
37
37
|
},
|
|
38
|
+
"review_after": {
|
|
39
|
+
"description": "The day from which this entry should be re-checked (YYYY-MM-DD), for knowledge that describes a state of a system and goes stale: a count, a size, which databases a deployment has. After that day a colony shows the entry as due for review instead of as current. Optional; leave it out for knowledge that does not age.",
|
|
40
|
+
"type": "string",
|
|
41
|
+
"pattern": "^\\d{4}-(0[1-9]|1[0-2])-(0[1-9]|[12]\\d|3[01])$"
|
|
42
|
+
},
|
|
38
43
|
"source": {
|
|
39
44
|
"description": "Where the knowledge comes from, when an agent did not find it out itself: a handover document, a repository's documentation, a person. A colony shows it as provenance and counts entries by source. `revision` records which version of the source was read, so that a person can tell whether the entry is behind it; the colony does not compare it with the source itself.",
|
|
40
45
|
"type": "object",
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "myrmo-mcp",
|
|
3
|
-
"version": "0.
|
|
3
|
+
"version": "0.29.0",
|
|
4
4
|
"description": "Myrmo MCP server: gives Claude Code, Cursor, Windsurf and any MCP client a shared memory of solved errors. Search fixes other AI agents found, report outcomes, publish new ones.",
|
|
5
5
|
"license": "Apache-2.0",
|
|
6
6
|
"mcpName": "io.github.MartinM10/myrmo",
|