agent-coord-mcp 0.26.19 → 0.26.21
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/README.md +82 -0
- package/dist/capabilities.js +57 -1
- package/dist/capabilities.js.map +1 -1
- package/dist/gated-head.js +130 -0
- package/dist/gated-head.js.map +1 -0
- package/dist/server.js +23 -0
- package/dist/server.js.map +1 -1
- package/dist/tools/queue-write.js +431 -0
- package/dist/tools/queue-write.js.map +1 -0
- package/dist/tools/records.js +406 -70
- package/dist/tools/records.js.map +1 -1
- package/dist/tools/registry.js +34 -5
- package/dist/tools/registry.js.map +1 -1
- package/dist/tools/shared.js.map +1 -1
- package/dist/tools/stall.js +2 -1
- package/dist/tools/stall.js.map +1 -1
- package/dist/tools/transport.js +82 -42
- package/dist/tools/transport.js.map +1 -1
- package/dist/tools/tree-provenance.js +107 -0
- package/dist/tools/tree-provenance.js.map +1 -0
- package/dist/tools/work.js +95 -3
- package/dist/tools/work.js.map +1 -1
- package/dist/transports/config.js +82 -0
- package/dist/transports/config.js.map +1 -0
- package/dist/transports/index.js +113 -0
- package/dist/transports/index.js.map +1 -0
- package/dist/transports/tmux.js +140 -0
- package/dist/transports/tmux.js.map +1 -0
- package/dist/transports/types.js +86 -0
- package/dist/transports/types.js.map +1 -0
- package/hooks/peek-coord.mjs +0 -0
- package/hooks/tmux-pusher.mjs +33 -3
- package/package.json +14 -11
- package/scripts/coord-attention-clock.mjs +0 -0
- package/scripts/coord-node.sh +0 -0
- package/scripts/coord-stall-clock.mjs +0 -0
- package/scripts/coord-token.mjs +0 -0
- package/scripts/probe-tmux-liveness.sh +0 -0
- package/scripts/spawn-agent.sh +0 -0
- package/scripts/stop-agent.sh +0 -0
- package/scripts/typed-record-stats.mjs +0 -0
- package/src/capabilities.ts +104 -1
- package/src/gated-head.ts +134 -0
- package/src/server.ts +29 -0
- package/src/tools/queue-write.ts +485 -0
- package/src/tools/records.ts +409 -40
- package/src/tools/registry.ts +36 -5
- package/src/tools/shared.ts +12 -36
- package/src/tools/stall.ts +2 -1
- package/src/tools/transport.ts +96 -43
- package/src/tools/tree-provenance.ts +136 -0
- package/src/tools/work.ts +95 -3
- package/src/transports/config.ts +110 -0
- package/src/transports/index.ts +126 -0
- package/src/transports/tmux.ts +177 -0
- package/src/transports/types.ts +201 -0
package/src/tools/registry.ts
CHANGED
|
@@ -1,4 +1,5 @@
|
|
|
1
1
|
import { detachAgentTool } from "./transport.js";
|
|
2
|
+
import { isLocallyProbeable, isRemoteTmuxKind, targetOf } from "../transports/index.js";
|
|
2
3
|
import { readAway, secondCoordinatorRefusal } from "./away.js";
|
|
3
4
|
import { randomUUID } from "node:crypto";
|
|
4
5
|
import { existsSync, openSync, watch } from "node:fs";
|
|
@@ -419,6 +420,36 @@ export async function listAgentsTool() {
|
|
|
419
420
|
return { agents, evicted, proseOnly };
|
|
420
421
|
}
|
|
421
422
|
|
|
423
|
+
/**
|
|
424
|
+
* EVERY marker on disk, READ-ONLY.
|
|
425
|
+
*
|
|
426
|
+
* `loadLiveTransports` below deletes any marker it judges not-live, which is
|
|
427
|
+
* right for the delivery paths that own that garbage collection and WRONG for a
|
|
428
|
+
* diagnostic: `capabilities` must be able to describe the fleet without changing
|
|
429
|
+
* it. Using the reaping loader from a read-only verb made a mixed-fleet check
|
|
430
|
+
* delete the very marker it was reporting (caught by its own test — the second
|
|
431
|
+
* disagreeing agent vanished between write and read).
|
|
432
|
+
*
|
|
433
|
+
* Liveness is reported per marker rather than filtered on, because a STALE marker
|
|
434
|
+
* naming a different transport is more alarming than a live one, not less: it is
|
|
435
|
+
* a seat that may come back on the wrong transport. Filtering it out would make
|
|
436
|
+
* the remote kind — whose liveness is heartbeat-based and therefore absent
|
|
437
|
+
* without a registry entry — systematically invisible to this check.
|
|
438
|
+
*/
|
|
439
|
+
export async function readAllTransportMarkers(): Promise<
|
|
440
|
+
{ marker: TransportMarker; live: boolean }[]
|
|
441
|
+
> {
|
|
442
|
+
const out: { marker: TransportMarker; live: boolean }[] = [];
|
|
443
|
+
const reg = await readJson<AgentRegistry>(AGENTS_FILE, {});
|
|
444
|
+
const now = Date.now();
|
|
445
|
+
for (const fname of await listTransportFiles()) {
|
|
446
|
+
const marker = await readJson<TransportMarker | null>(path.join(TRANSPORT_DIR, fname), null);
|
|
447
|
+
if (!marker) continue; // unparseable: nothing to attribute, and not ours to delete
|
|
448
|
+
out.push({ marker, live: isMarkerLive(marker, reg, now) });
|
|
449
|
+
}
|
|
450
|
+
return out;
|
|
451
|
+
}
|
|
452
|
+
|
|
422
453
|
export async function loadLiveTransports(): Promise<Map<string, TransportMarker>> {
|
|
423
454
|
const out = new Map<string, TransportMarker>();
|
|
424
455
|
const reg = await readJson<AgentRegistry>(AGENTS_FILE, {});
|
|
@@ -439,7 +470,7 @@ export async function loadLiveTransports(): Promise<Map<string, TransportMarker>
|
|
|
439
470
|
// remote markers (tmux-push-remote, pid 0 on a foreign host) can't be — so we
|
|
440
471
|
// trust the registry heartbeat the remote pusher refreshes (within STALE_MS).
|
|
441
472
|
export function isMarkerLive(marker: TransportMarker, reg: AgentRegistry, now: number): boolean {
|
|
442
|
-
if (marker.transport
|
|
473
|
+
if (isRemoteTmuxKind(marker.transport)) {
|
|
443
474
|
const entry = reg[marker.agentId];
|
|
444
475
|
return !!entry && now - entry.lastHeartbeat < STALE_MS;
|
|
445
476
|
}
|
|
@@ -523,13 +554,13 @@ export async function liveClaimEvidence(agentId: string, now: number): Promise<C
|
|
|
523
554
|
if (marker && isMarkerLive(marker, reg, now)) {
|
|
524
555
|
markerLive = true;
|
|
525
556
|
reasons.push(
|
|
526
|
-
`live ${marker.transport} transport (pid ${marker.pid}${marker
|
|
557
|
+
`live ${marker.transport} transport (pid ${marker.pid}${targetOf(marker) ? `, pane ${targetOf(marker)}` : ""})`,
|
|
527
558
|
);
|
|
528
559
|
if (
|
|
529
|
-
marker.transport
|
|
530
|
-
marker
|
|
560
|
+
isLocallyProbeable(marker.transport) &&
|
|
561
|
+
targetOf(marker) &&
|
|
531
562
|
process.env.TMUX_PANE &&
|
|
532
|
-
marker
|
|
563
|
+
targetOf(marker) === process.env.TMUX_PANE
|
|
533
564
|
) {
|
|
534
565
|
samePane = true;
|
|
535
566
|
}
|
package/src/tools/shared.ts
CHANGED
|
@@ -88,42 +88,18 @@ export type AgentEntry = {
|
|
|
88
88
|
proseOnly?: { since: number; reason?: string };
|
|
89
89
|
};
|
|
90
90
|
|
|
91
|
-
|
|
92
|
-
|
|
93
|
-
|
|
94
|
-
|
|
95
|
-
|
|
96
|
-
|
|
97
|
-
|
|
98
|
-
|
|
99
|
-
|
|
100
|
-
|
|
101
|
-
|
|
102
|
-
|
|
103
|
-
// pusher — the class of bug that silently dropped /clear /compact in v0.8.1.
|
|
104
|
-
// Absent on markers written by older versions (treated as "unknown, skip").
|
|
105
|
-
scriptMtime?: number;
|
|
106
|
-
// Build identity (newest dist mtime, sampled at that server's module load)
|
|
107
|
-
// of the MCP server whose attach_agent stamped this marker. A marker
|
|
108
|
-
// stamped by a server predating the current on-disk build was written by
|
|
109
|
-
// attach/stamp logic the rebuild replaced — doctor's provenance check flags
|
|
110
|
-
// it for a session restart + re-attach. Absent on markers written by older
|
|
111
|
-
// versions (treated as "unknown, skip", deliberately mirroring scriptMtime).
|
|
112
|
-
serverBuildMtime?: number;
|
|
113
|
-
// Does this transport carry ROOM traffic, or DMs only?
|
|
114
|
-
//
|
|
115
|
-
// A pusher started with `--no-room` delivers inbox messages and nothing
|
|
116
|
-
// else, and until this field existed the marker looked identical to a full
|
|
117
|
-
// one — so `status` said `attached: true` and an agent sat with its room
|
|
118
|
-
// feed off while every reading said healthy. That is how worker-2 missed
|
|
119
|
-
// its channel traffic.
|
|
120
|
-
//
|
|
121
|
-
// ABSENT MEANS UNKNOWN, NEVER "ON" — deliberately mirroring scriptMtime and
|
|
122
|
-
// serverBuildMtime above. A marker from an older pusher cannot tell us, and
|
|
123
|
-
// reporting an unasked question as full capability is the defect this field
|
|
124
|
-
// exists to remove.
|
|
125
|
-
rooms?: boolean;
|
|
126
|
-
};
|
|
91
|
+
/**
|
|
92
|
+
* THE MARKER TYPE NOW LIVES IN THE SEAM, and is re-exported here so the dozens
|
|
93
|
+
* of `from "./shared.js"` imports keep working.
|
|
94
|
+
*
|
|
95
|
+
* It was defined here, and the transport extraction gave it a second definition
|
|
96
|
+
* in `transports/types.ts` — two structurally-similar types that a consumer
|
|
97
|
+
* could satisfy while missing a field, which is the duplication this task exists
|
|
98
|
+
* to remove rather than double. One definition, one home: the transport layer,
|
|
99
|
+
* which is what the field describes. The nine fields and their comments moved
|
|
100
|
+
* with it verbatim, including `target`'s dual-write rule.
|
|
101
|
+
*/
|
|
102
|
+
export type { TransportMarker } from "../transports/types.js";
|
|
127
103
|
|
|
128
104
|
/**
|
|
129
105
|
* How to REPORT a transport's room capability.
|
package/src/tools/stall.ts
CHANGED
|
@@ -15,6 +15,7 @@
|
|
|
15
15
|
* from "nothing ran". A check that only speaks when it fires cannot be told from
|
|
16
16
|
* a broken one.
|
|
17
17
|
*/
|
|
18
|
+
import { isLocallyProbeable } from "../transports/index.js";
|
|
18
19
|
import { existsSync, readFileSync, writeFileSync, mkdirSync } from "node:fs";
|
|
19
20
|
import { execFileSync } from "node:child_process";
|
|
20
21
|
import path from "node:path";
|
|
@@ -311,7 +312,7 @@ export async function stallCheckTool(args: { repo?: string; stallMinutes?: numbe
|
|
|
311
312
|
// agent with no transport, or a REMOTE one where the heartbeat genuinely
|
|
312
313
|
// is the liveness mechanism, still HITs on a dead heartbeat.
|
|
313
314
|
const marker = liveTransports.get(agentId);
|
|
314
|
-
if (age > limit && marker && marker.transport
|
|
315
|
+
if (age > limit && marker && isLocallyProbeable(marker.transport)) {
|
|
315
316
|
unmeasurable.push({
|
|
316
317
|
agentId,
|
|
317
318
|
value: marker.transport,
|
package/src/tools/transport.ts
CHANGED
|
@@ -1,4 +1,16 @@
|
|
|
1
1
|
import { loadLiveTransports, isMarkerLive, isPidAlive } from "./registry.js";
|
|
2
|
+
import {
|
|
3
|
+
TMUX_PUSH,
|
|
4
|
+
registerTmuxHost,
|
|
5
|
+
type TmuxHost,
|
|
6
|
+
isLocallyProbeable,
|
|
7
|
+
isTmuxKind,
|
|
8
|
+
paneExists,
|
|
9
|
+
probePane,
|
|
10
|
+
tmuxVersion,
|
|
11
|
+
targetOf,
|
|
12
|
+
tmuxAvailable,
|
|
13
|
+
} from "../transports/index.js";
|
|
2
14
|
import { newestMtimeUnder, onDiskBuildMtime, onDiskSourceMtime, SERVER_BUILD_MTIME, SERVER_BUILD_SHA, BUILD_DIR } from "../build.js";
|
|
3
15
|
import { prefixOf, prefixVerdict } from "../prefix.js";
|
|
4
16
|
import { execFileSync } from "node:child_process";
|
|
@@ -119,17 +131,11 @@ export async function pingTool(args: { from: string; to: string; echo?: boolean
|
|
|
119
131
|
let paneAlive: boolean | undefined;
|
|
120
132
|
if (marker) {
|
|
121
133
|
transportLive = isMarkerLive(marker, reg, now);
|
|
122
|
-
if (transportLive && marker.transport
|
|
134
|
+
if (transportLive && isLocallyProbeable(marker.transport) && targetOf(marker)) {
|
|
123
135
|
// The pusher can outlive its pane (agent window closed) — probe the pane.
|
|
124
|
-
// `has-session`
|
|
125
|
-
//
|
|
126
|
-
|
|
127
|
-
// pane alive. Pinned to the BEHAVIOUR, not a version: measured identical on
|
|
128
|
-
// tmux 3.6b and 3.7b, and a version-pinned claim rots on the next upgrade.
|
|
129
|
-
// Positive control, both directions: bogus target -> has-session exit 1,
|
|
130
|
-
// display-message exit 0; live pane -> both exit 0.
|
|
131
|
-
const probe = spawnSync("tmux", ["has-session", "-t", marker.tmuxTarget]);
|
|
132
|
-
paneAlive = probe.status === 0;
|
|
136
|
+
// Why `has-session` and not `display-message`, with the positive control
|
|
137
|
+
// both ways, is documented once on `paneExists`.
|
|
138
|
+
paneAlive = paneExists(targetOf(marker)!);
|
|
133
139
|
}
|
|
134
140
|
}
|
|
135
141
|
|
|
@@ -146,7 +152,7 @@ export async function pingTool(args: { from: string; to: string; echo?: boolean
|
|
|
146
152
|
// same shape as `stall_clock_status` before Task 13.1. A REMOTE pusher, or
|
|
147
153
|
// an agent with no probeable marker at all, has no pid to fall back on —
|
|
148
154
|
// there heartbeat genuinely IS the liveness mechanism, unchanged.
|
|
149
|
-
const heartbeatIsValidSignal = !marker || marker.transport
|
|
155
|
+
const heartbeatIsValidSignal = !marker || !isLocallyProbeable(marker.transport);
|
|
150
156
|
const alive = reachable || (heartbeatIsValidSignal && heartbeatFresh);
|
|
151
157
|
|
|
152
158
|
let echoSent = false;
|
|
@@ -196,7 +202,7 @@ export const CONTROL_COMMANDS = ["clear", "compact", "reload-skills"] as const;
|
|
|
196
202
|
// Transports whose pusher can actually TYPE a slash command into a live CLI.
|
|
197
203
|
// A control command is meaningless to a plain MCP poller, so send_command is
|
|
198
204
|
// gated to agents currently attached over one of these.
|
|
199
|
-
|
|
205
|
+
|
|
200
206
|
|
|
201
207
|
// Normalize "clear" / "/clear" / " /Clear " → "clear"; null if not allowlisted.
|
|
202
208
|
function normalizeControlCommand(raw: string): string | null {
|
|
@@ -208,7 +214,7 @@ function normalizeControlCommand(raw: string): string | null {
|
|
|
208
214
|
async function liveTmuxTargets(): Promise<Map<string, TransportMarker>> {
|
|
209
215
|
const all = await loadLiveTransports();
|
|
210
216
|
const out = new Map<string, TransportMarker>();
|
|
211
|
-
for (const [id, m] of all) if (
|
|
217
|
+
for (const [id, m] of all) if (isTmuxKind(m.transport)) out.set(id, m);
|
|
212
218
|
return out;
|
|
213
219
|
}
|
|
214
220
|
|
|
@@ -645,19 +651,13 @@ export async function attachAgentTool(args: {
|
|
|
645
651
|
};
|
|
646
652
|
}
|
|
647
653
|
|
|
648
|
-
// Validate target exists.
|
|
649
|
-
//
|
|
650
|
-
|
|
651
|
-
|
|
652
|
-
// pane alive. Pinned to the BEHAVIOUR, not a version: measured identical on
|
|
653
|
-
// tmux 3.6b and 3.7b, and a version-pinned claim rots on the next upgrade.
|
|
654
|
-
// Positive control, both directions: bogus target -> has-session exit 1,
|
|
655
|
-
// display-message exit 0; live pane -> both exit 0.
|
|
656
|
-
const probe = spawnSync("tmux", ["has-session", "-t", target]);
|
|
657
|
-
if (probe.status !== 0) {
|
|
654
|
+
// Validate target exists. The probe's discriminating power, and the control
|
|
655
|
+
// that proves it, are documented once on `paneExists`.
|
|
656
|
+
const targetProbe = probePane(target);
|
|
657
|
+
if (!targetProbe.exists) {
|
|
658
658
|
return {
|
|
659
659
|
ok: false,
|
|
660
|
-
error: `tmux target '${target}' not found: ${
|
|
660
|
+
error: `tmux target '${target}' not found: ${targetProbe.stderr}`,
|
|
661
661
|
};
|
|
662
662
|
}
|
|
663
663
|
|
|
@@ -723,8 +723,14 @@ export async function attachAgentTool(args: {
|
|
|
723
723
|
const scriptMtime = newestPusherSourceMtime();
|
|
724
724
|
const marker: TransportMarker = {
|
|
725
725
|
agentId: args.agentId,
|
|
726
|
-
transport:
|
|
726
|
+
transport: TMUX_PUSH,
|
|
727
727
|
pid,
|
|
728
|
+
// DUAL-WRITTEN, and the duplication is the point. `target` is what every
|
|
729
|
+
// consumer now reads (`targetOf`); `tmuxTarget` is what the code a merge
|
|
730
|
+
// revert restores reads. Writing only the new field would leave markers the
|
|
731
|
+
// old server cannot parse, and that failure does not degrade gracefully —
|
|
732
|
+
// it silences every attached lane at once.
|
|
733
|
+
target,
|
|
728
734
|
tmuxTarget: target,
|
|
729
735
|
since: Date.now(),
|
|
730
736
|
scriptMtime,
|
|
@@ -749,7 +755,7 @@ export async function attachAgentTool(args: {
|
|
|
749
755
|
return {
|
|
750
756
|
ok: true,
|
|
751
757
|
agentId: args.agentId,
|
|
752
|
-
transport:
|
|
758
|
+
transport: TMUX_PUSH,
|
|
753
759
|
tmuxTarget: target,
|
|
754
760
|
pid,
|
|
755
761
|
log,
|
|
@@ -1171,6 +1177,53 @@ async function scanStaleLocks(olderThanMs: number, now: number): Promise<{ path:
|
|
|
1171
1177
|
return out;
|
|
1172
1178
|
}
|
|
1173
1179
|
|
|
1180
|
+
/* ── wiring the seam's TMUX HOST (Phase 5.4 Task 3) ─────────────────────────── */
|
|
1181
|
+
|
|
1182
|
+
/**
|
|
1183
|
+
* The process-layer half of `TmuxTransport`, supplied by the module that owns
|
|
1184
|
+
* pusher spawn, receipts and marker files.
|
|
1185
|
+
*
|
|
1186
|
+
* ⚠ SCOPE, STATED RATHER THAN IMPLIED: Task 3 puts the seam in the IDENTITY and
|
|
1187
|
+
* DIAGNOSIS path — `capabilities` asks the live transport what it is, and the
|
|
1188
|
+
* mixed-fleet check reads markers through it. DELIVERY STILL RUNS THROUGH THE
|
|
1189
|
+
* EXISTING CODE PATHS. Task 2's gate was that nothing changes, and rerouting
|
|
1190
|
+
* every push through a new object would change the thing most likely to break
|
|
1191
|
+
* quietly. So the four delivery methods below THROW rather than no-op: a host
|
|
1192
|
+
* that silently accepted a push and dropped it would be the one failure this
|
|
1193
|
+
* fleet cannot observe, and an explicit throw is reachable only from code that
|
|
1194
|
+
* has not been written yet.
|
|
1195
|
+
*/
|
|
1196
|
+
const TMUX_HOST: TmuxHost = {
|
|
1197
|
+
attach: async () => {
|
|
1198
|
+
throw new Error("TmuxTransport.attach is not the delivery path yet — call attachAgentTool (Phase 5.4 Task 3 wires identity only)");
|
|
1199
|
+
},
|
|
1200
|
+
detach: async () => {
|
|
1201
|
+
throw new Error("TmuxTransport.detach is not the delivery path yet — call detachAgentTool");
|
|
1202
|
+
},
|
|
1203
|
+
push: async () => {
|
|
1204
|
+
throw new Error("TmuxTransport.push is not the delivery path yet — delivery runs through the pusher process");
|
|
1205
|
+
},
|
|
1206
|
+
sendControl: async () => {
|
|
1207
|
+
throw new Error("TmuxTransport.sendControl is not the delivery path yet — call sendCommandTool");
|
|
1208
|
+
},
|
|
1209
|
+
/**
|
|
1210
|
+
* Is the pusher behind this marker still running? Reuses `isPusherProcess`,
|
|
1211
|
+
* which checks the COMMAND of the pid rather than merely that a pid exists —
|
|
1212
|
+
* a recycled pid belonging to something else is not a live pusher.
|
|
1213
|
+
*/
|
|
1214
|
+
pusherAlive: (marker) => isPusherProcess(marker.pid),
|
|
1215
|
+
killPusher: (marker) => {
|
|
1216
|
+
try {
|
|
1217
|
+
process.kill(marker.pid, "SIGTERM");
|
|
1218
|
+
return true;
|
|
1219
|
+
} catch {
|
|
1220
|
+
return false;
|
|
1221
|
+
}
|
|
1222
|
+
},
|
|
1223
|
+
};
|
|
1224
|
+
|
|
1225
|
+
registerTmuxHost(TMUX_HOST);
|
|
1226
|
+
|
|
1174
1227
|
export const doctorSchema = {
|
|
1175
1228
|
fix: z.boolean().optional(),
|
|
1176
1229
|
maxFileBytes: z.number().int().positive().optional(),
|
|
@@ -1231,7 +1284,7 @@ export async function doctorTool(args: { fix?: boolean; maxFileBytes?: number })
|
|
|
1231
1284
|
const file = path.join(TRANSPORT_DIR, fname);
|
|
1232
1285
|
const marker = await readJson<TransportMarker | null>(file, null);
|
|
1233
1286
|
if (!marker || !isMarkerLive(marker, reg, now)) continue;
|
|
1234
|
-
if (marker.transport
|
|
1287
|
+
if (!isLocallyProbeable(marker.transport)) continue; // remote = can't verify (documented limit: can't stat another host)
|
|
1235
1288
|
if (marker.scriptMtime === undefined) {
|
|
1236
1289
|
// ABSENCE IS NOT EXEMPTION. The field's own writer once dropped it,
|
|
1237
1290
|
// and the silent skip here meant the check was disabled by the very
|
|
@@ -1339,7 +1392,7 @@ export async function doctorTool(args: { fix?: boolean; maxFileBytes?: number })
|
|
|
1339
1392
|
const file = path.join(TRANSPORT_DIR, fname);
|
|
1340
1393
|
const marker = await readJson<TransportMarker | null>(file, null);
|
|
1341
1394
|
if (!marker || !isMarkerLive(marker, reg, now)) continue;
|
|
1342
|
-
if (marker.transport
|
|
1395
|
+
if (!isLocallyProbeable(marker.transport)) continue; // remote = can't verify (documented limit: can't stat another host)
|
|
1343
1396
|
if (marker.serverBuildMtime === undefined) {
|
|
1344
1397
|
// ABSENCE IS NOT EXEMPTION — flipped in the same commit as the
|
|
1345
1398
|
// scriptMtime absence above, so the two checks can never disagree
|
|
@@ -1434,7 +1487,7 @@ export async function doctorTool(args: { fix?: boolean; maxFileBytes?: number })
|
|
|
1434
1487
|
const file = path.join(TRANSPORT_DIR, fname);
|
|
1435
1488
|
const marker = await readJson<TransportMarker | null>(file, null);
|
|
1436
1489
|
if (!marker || !isMarkerLive(marker, reg, now)) continue;
|
|
1437
|
-
if (marker.transport
|
|
1490
|
+
if (!isLocallyProbeable(marker.transport)) continue; // remote: the script lives on another host
|
|
1438
1491
|
const { scriptMtime, serverBuildMtime, agentId } = marker;
|
|
1439
1492
|
if (scriptMtime === undefined || serverBuildMtime === undefined || onDiskScript === undefined || onDiskServer === undefined) {
|
|
1440
1493
|
// A pane missing either stamp cannot be classified. Saying so is the
|
|
@@ -1475,20 +1528,20 @@ export async function doctorTool(args: { fix?: boolean; maxFileBytes?: number })
|
|
|
1475
1528
|
{
|
|
1476
1529
|
// Without a tmux binary we can't tell "wedged" from "can't probe" — skip
|
|
1477
1530
|
// rather than flag every local marker as dead.
|
|
1478
|
-
const
|
|
1531
|
+
const tmuxIsAvailable = tmuxAvailable();
|
|
1479
1532
|
const wedged: { agentId: string; pid: number; file: string; target: string; isPusher: boolean }[] = [];
|
|
1480
|
-
if (
|
|
1533
|
+
if (tmuxIsAvailable) {
|
|
1481
1534
|
for (const fname of await listTransportFiles()) {
|
|
1482
1535
|
const file = path.join(TRANSPORT_DIR, fname);
|
|
1483
1536
|
const marker = await readJson<TransportMarker | null>(file, null);
|
|
1484
1537
|
if (!marker || !isMarkerLive(marker, reg, now)) continue;
|
|
1485
|
-
if (marker.transport
|
|
1538
|
+
if (!isLocallyProbeable(marker.transport)) continue; // remote = no local pane to probe
|
|
1486
1539
|
if (!marker.tmuxTarget) continue; // no target recorded, can't probe
|
|
1487
1540
|
// has-session actually validates the target and fails on a dead
|
|
1488
1541
|
// pane/session; `display-message -p -t <target> <literal>` does NOT
|
|
1489
1542
|
// (tmux 3.6b exits 0 for any target, even a just-killed one, when
|
|
1490
1543
|
// the format string has no #{...} needing that target resolved).
|
|
1491
|
-
const probe =
|
|
1544
|
+
const probe = { status: paneExists(targetOf(marker)!) ? 0 : 1 };
|
|
1492
1545
|
if (probe.status === 0) continue; // pane alive
|
|
1493
1546
|
// The marker's pid being alive does not make it OUR pid — see
|
|
1494
1547
|
// isPusherProcess. Record the verdict now so `fix` only ever signals
|
|
@@ -1523,7 +1576,7 @@ export async function doctorTool(args: { fix?: boolean; maxFileBytes?: number })
|
|
|
1523
1576
|
level: wedged.length ? "warn" : "ok",
|
|
1524
1577
|
detail: wedged.length
|
|
1525
1578
|
? `${wedged.length} local pusher(s) alive (pid) but their tmux pane is gone — looks attached, delivers nothing. ${fix ? "Reaped (SIGTERM + marker cleared)." : "Run doctor with fix:true to SIGTERM and clear the marker."}`
|
|
1526
|
-
:
|
|
1579
|
+
: tmuxIsAvailable
|
|
1527
1580
|
? "no wedged local pushers (pid-alive, pane-dead)"
|
|
1528
1581
|
: "tmux not available — skipped wedged-pusher pane probe",
|
|
1529
1582
|
fixable: true,
|
|
@@ -1768,9 +1821,8 @@ export async function doctorTool(args: { fix?: boolean; maxFileBytes?: number })
|
|
|
1768
1821
|
// Without a tmux binary we cannot probe a pane at all — every stale
|
|
1769
1822
|
// agent is reported as an orphan candidate rather than silently split,
|
|
1770
1823
|
// same posture `wedged-local-pushers` takes.
|
|
1771
|
-
const
|
|
1772
|
-
const paneAlive = (target: string): boolean =>
|
|
1773
|
-
tmuxAvailable && spawnSync("tmux", ["has-session", "-t", target]).status === 0;
|
|
1824
|
+
const tmuxIsAvailable = tmuxAvailable();
|
|
1825
|
+
const paneAlive = (target: string): boolean => tmuxIsAvailable && paneExists(target);
|
|
1774
1826
|
|
|
1775
1827
|
const paneConfirmed: string[] = [];
|
|
1776
1828
|
const orphans: string[] = [];
|
|
@@ -1779,8 +1831,9 @@ export async function doctorTool(args: { fix?: boolean; maxFileBytes?: number })
|
|
|
1779
1831
|
if (now - a.lastHeartbeat <= EVICT_MS) continue;
|
|
1780
1832
|
const age = `${Math.floor((now - a.lastHeartbeat) / 3600000)}h`;
|
|
1781
1833
|
const marker = markerByAgent.get(id);
|
|
1782
|
-
|
|
1783
|
-
|
|
1834
|
+
const markerTarget = marker ? targetOf(marker) : undefined;
|
|
1835
|
+
if (isLocallyProbeable(marker?.transport) && markerTarget && paneAlive(markerTarget)) {
|
|
1836
|
+
paneConfirmed.push(`${id} (${age}, pane '${markerTarget}' still current)`);
|
|
1784
1837
|
} else {
|
|
1785
1838
|
orphans.push(`${id} (${age})`);
|
|
1786
1839
|
}
|
|
@@ -1808,7 +1861,7 @@ export async function doctorTool(args: { fix?: boolean; maxFileBytes?: number })
|
|
|
1808
1861
|
level: orphans.length ? "warn" : "ok",
|
|
1809
1862
|
detail: orphans.length
|
|
1810
1863
|
? `${orphans.length} stale agent(s) have no live tmux pane behind them — permanent, will not clear on their own (unregister or let eviction drop them)`
|
|
1811
|
-
:
|
|
1864
|
+
: tmuxIsAvailable
|
|
1812
1865
|
? `no orphans among ${stale.length} stale agent(s)${stale.length ? " — all pane-confirmed current" : ""}`
|
|
1813
1866
|
: "tmux not available — could not distinguish orphans from pane-confirmed stale agents",
|
|
1814
1867
|
fixable: false,
|
|
@@ -2026,8 +2079,8 @@ export async function doctorTool(args: { fix?: boolean; maxFileBytes?: number })
|
|
|
2026
2079
|
// identity is walk-up-to-.git + rev-parse (kit monorepo root); omitted
|
|
2027
2080
|
// entirely when there is no checkout — never invented.
|
|
2028
2081
|
{
|
|
2029
|
-
const
|
|
2030
|
-
const tmuxOk =
|
|
2082
|
+
const tmuxReported = tmuxVersion();
|
|
2083
|
+
const tmuxOk = tmuxReported !== undefined;
|
|
2031
2084
|
const ident = resolveServerIdentity();
|
|
2032
2085
|
const loc = ident.branch && ident.sha
|
|
2033
2086
|
? `path=${ident.path} version=${ident.version} ${ident.branch}@${ident.sha.slice(0, 12)}`
|
|
@@ -2045,7 +2098,7 @@ export async function doctorTool(args: { fix?: boolean; maxFileBytes?: number })
|
|
|
2045
2098
|
check: "environment",
|
|
2046
2099
|
level: tmuxOk ? "ok" : "warn",
|
|
2047
2100
|
detail: tmuxOk
|
|
2048
|
-
? `root=${ROOT}; node=${process.execPath}; ${loc}; tmux=${
|
|
2101
|
+
? `root=${ROOT}; node=${process.execPath}; ${loc}; tmux=${tmuxReported || "present"}`
|
|
2049
2102
|
: `root=${ROOT}; node=${process.execPath}; ${loc}; tmux NOT on PATH — the tmux-push transport will not work`,
|
|
2050
2103
|
fixable: false,
|
|
2051
2104
|
items,
|
|
@@ -0,0 +1,136 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* ⟨q-c1af2db3⟩ — WHICH TREE DID THIS VERB JUST READ, AND HOW OLD IS IT?
|
|
3
|
+
*
|
|
4
|
+
* ⛔ THE BUG THIS EXISTS FOR ROUTED THE FLEET. `next_unblocked` reads
|
|
5
|
+
* `docs/QUEUE.md` out of whatever working tree it is handed, and that tree
|
|
6
|
+
* belongs to nobody: its freshness is a property of whoever last worked in it.
|
|
7
|
+
* Measured 2026-09-12 16:36 — the primary checkout sat FOUR commits behind
|
|
8
|
+
* `origin/main`, its `QUEUE.md` had zero hits for a tag committed six minutes
|
|
9
|
+
* earlier, and the verb returned the PRE-EDIT text as `next`.
|
|
10
|
+
*
|
|
11
|
+
* ⛆ AND THE HARM LANDED IN BOTH DIRECTIONS INSIDE TEN MINUTES: a row was ruled
|
|
12
|
+
* un-claimable from an observation that was actually a stale file, and a worker
|
|
13
|
+
* was told to hold on a row that was already split. ⭐ ***A true conclusion
|
|
14
|
+
* reached through an untrue reading is the shape, and here the INSTRUMENT
|
|
15
|
+
* supplied the untrue reading — the answer was well-formed, internally
|
|
16
|
+
* consistent, `accounting.reconciles` was TRUE, and nothing in it said which
|
|
17
|
+
* tree it read.***
|
|
18
|
+
*
|
|
19
|
+
* ⛔⛆⛆ THE TRAP THAT BREAKS THE OBVIOUS FIX, AND IT IS WHY THIS FETCHES:
|
|
20
|
+
*
|
|
21
|
+
* a checkout that has not fetched CANNOT REPORT that it has not fetched.
|
|
22
|
+
*
|
|
23
|
+
* `git rev-list --left-right --count origin/main...HEAD` in an unfetched clone
|
|
24
|
+
* compares HEAD against a STALE `origin/main` ref and returns `0 0` — *"I am
|
|
25
|
+
* perfectly current"* — while the real remote has moved. Measured: HEAD
|
|
26
|
+
* `1cab5c2`, its own `origin/main` also `1cab5c2`, true `origin/main` `558365d`,
|
|
27
|
+
* three commits behind, reported as ZERO. ⭐⭐ ***A distance of `0` from an
|
|
28
|
+
* unfetched ref and a distance of `0` from a fetched one are BYTE-IDENTICAL and
|
|
29
|
+
* mean opposite things. That distinction is the whole point of this module: the
|
|
30
|
+
* ref is fetched IN THIS CALL, and when the fetch fails the answer is
|
|
31
|
+
* `"unknown"` and never `0`.***
|
|
32
|
+
*
|
|
33
|
+
* ⚠ AND FETCHING IS NOT THE FIX — that is the row's named anti-control.
|
|
34
|
+
* Fetching updates REFS; the verb reads the TREE, which is still stale
|
|
35
|
+
* afterwards. So a fetch only makes the staleness *legible*; the caller must
|
|
36
|
+
* still refuse or flag. This module reports; it never silently repairs, because
|
|
37
|
+
* making someone else's working tree current is a destructive write into a
|
|
38
|
+
* checkout a seat may be mid-work in.
|
|
39
|
+
*/
|
|
40
|
+
import { execFileSync } from "node:child_process";
|
|
41
|
+
import path from "node:path";
|
|
42
|
+
|
|
43
|
+
export type TreeProvenance = {
|
|
44
|
+
/** The tree actually read, resolved — not the string the caller passed. */
|
|
45
|
+
path: string;
|
|
46
|
+
head: string | null;
|
|
47
|
+
/** The ref the distance is measured against, e.g. `origin/main`. */
|
|
48
|
+
base: string;
|
|
49
|
+
baseSha: string | null;
|
|
50
|
+
/**
|
|
51
|
+
* Commits the tree is behind `base`.
|
|
52
|
+
*
|
|
53
|
+
* ⛔ `"unknown"` WHEN THE REF COULD NOT BE FRESHENED IN THIS CALL. Never `0`
|
|
54
|
+
* in that case: `0` is a claim of currency and an unfetched ref cannot make it.
|
|
55
|
+
*/
|
|
56
|
+
behind: number | "unknown";
|
|
57
|
+
/** Whether the ref this distance rests on was freshened in THIS call. */
|
|
58
|
+
fetched: boolean;
|
|
59
|
+
fetchError?: string;
|
|
60
|
+
dirty: boolean | "unknown";
|
|
61
|
+
/** True only when the tree is MEASURABLY behind a ref fetched in this call. */
|
|
62
|
+
stale: boolean;
|
|
63
|
+
/** Present whenever the answer cannot be trusted as current. */
|
|
64
|
+
warning?: string;
|
|
65
|
+
};
|
|
66
|
+
|
|
67
|
+
const run = (repo: string, args: string[]): string | null => {
|
|
68
|
+
try {
|
|
69
|
+
return execFileSync("git", args, {
|
|
70
|
+
cwd: repo,
|
|
71
|
+
encoding: "utf8",
|
|
72
|
+
stdio: ["ignore", "pipe", "ignore"],
|
|
73
|
+
timeout: 20_000,
|
|
74
|
+
}).trim();
|
|
75
|
+
} catch {
|
|
76
|
+
return null;
|
|
77
|
+
}
|
|
78
|
+
};
|
|
79
|
+
|
|
80
|
+
/**
|
|
81
|
+
* What the caller needs to say out loud about the tree it just read.
|
|
82
|
+
*
|
|
83
|
+
* @param repo the working tree a verb was handed
|
|
84
|
+
* @param baseBranch the branch the tree should be current with
|
|
85
|
+
*/
|
|
86
|
+
export function treeProvenance(repo: string, baseBranch = "main"): TreeProvenance {
|
|
87
|
+
const resolved = run(repo, ["rev-parse", "--show-toplevel"]) ?? path.resolve(repo);
|
|
88
|
+
const base = `origin/${baseBranch}`;
|
|
89
|
+
const head = run(resolved, ["rev-parse", "HEAD"]);
|
|
90
|
+
const status = run(resolved, ["status", "--porcelain"]);
|
|
91
|
+
|
|
92
|
+
// ⛔ FETCH FIRST, AND THE ORDER IS THE POINT. Every number below is worthless
|
|
93
|
+
// if it rests on a ref the tree last updated at some unknown past moment.
|
|
94
|
+
const fetched = run(resolved, ["fetch", "--quiet", "origin", baseBranch]) !== null;
|
|
95
|
+
|
|
96
|
+
const out: TreeProvenance = {
|
|
97
|
+
path: resolved,
|
|
98
|
+
head,
|
|
99
|
+
base,
|
|
100
|
+
baseSha: null,
|
|
101
|
+
behind: "unknown",
|
|
102
|
+
fetched,
|
|
103
|
+
dirty: status === null ? "unknown" : status.length > 0,
|
|
104
|
+
stale: false,
|
|
105
|
+
};
|
|
106
|
+
|
|
107
|
+
if (!fetched) {
|
|
108
|
+
out.fetchError = `could not fetch origin/${baseBranch} — the distance cannot be computed`;
|
|
109
|
+
out.warning =
|
|
110
|
+
`TREE FRESHNESS UNKNOWN: ${resolved} could not reach origin/${baseBranch} in this call, so its ` +
|
|
111
|
+
`distance is UNKNOWN rather than 0. A checkout that has not fetched cannot report that it has not ` +
|
|
112
|
+
`fetched — an unfetched ref answers "0 behind" and means nothing.`;
|
|
113
|
+
return out;
|
|
114
|
+
}
|
|
115
|
+
|
|
116
|
+
out.baseSha = run(resolved, ["rev-parse", base]);
|
|
117
|
+
const counts = run(resolved, ["rev-list", "--left-right", "--count", `${base}...HEAD`]);
|
|
118
|
+
const behind = counts ? Number(counts.split(/\s+/)[0]) : NaN;
|
|
119
|
+
if (!Number.isFinite(behind)) {
|
|
120
|
+
out.warning = `TREE FRESHNESS UNKNOWN: could not count ${base}...HEAD in ${resolved}.`;
|
|
121
|
+
return out;
|
|
122
|
+
}
|
|
123
|
+
|
|
124
|
+
out.behind = behind;
|
|
125
|
+
out.stale = behind > 0;
|
|
126
|
+
if (out.stale) {
|
|
127
|
+
// ⚠ FETCHING DID NOT FIX THE TREE — it only made this sentence possible.
|
|
128
|
+
// The file the verb read is still the old one.
|
|
129
|
+
out.warning =
|
|
130
|
+
`STALE TREE: ${resolved} is ${behind} commit(s) behind ${base} (HEAD ${head?.slice(0, 7)}, ` +
|
|
131
|
+
`${base} ${out.baseSha?.slice(0, 7)}). The answer above was read from THAT tree's files, so a row ` +
|
|
132
|
+
`filed or amended on ${base} since is invisible here and a superseded row can be offered as current. ` +
|
|
133
|
+
`Fetching refreshed the REF, not the working tree.`;
|
|
134
|
+
}
|
|
135
|
+
return out;
|
|
136
|
+
}
|