@miraland-labs/conduit-bridge 0.16.113 → 0.16.115

package/dist/cli.js

@@ -655,13 +655,6 @@ async function runner() {
                 await heartbeat(client, config, workspace, currentBrief, preflight, intentProof, bootstrapResult);
             }
             await renewLeases(client, config);
-            if (workspace && onlineDriverIds(config).length) {
-                // Observation runs before authoring: an owner waiting on an answer should not queue behind
-                // a long delivery, and the control plane already counted this slot as busy.
-                progressed = await executeNextInvestigation(client, config, workspace, brief, timeoutMs, {
-                    managedRoot: hb.on_shift?.managed_workspace_root ?? null,
-                }) || progressed;
-            }
             if (workspace && onlineDriverIds(config).length) {
                 // Read the workspace again, then publish it. The last brief stands if the read fails.
                 const publishWorkspaceState = async (options = {}) => {
@@ -671,10 +664,18 @@ async function runner() {
                     const hb = await heartbeat(client, config, workspace, currentBrief, preflight, undefined, bootstrapResult);
                     return { preflight, hb };
                 };
+                const publishHeartbeat = async () => {
+                    await publishWorkspaceState();
+                };
+                // Observation runs before authoring: an owner waiting on an answer should not queue behind
+                // a long delivery, and the control plane already counted this slot as busy.
+                progressed = await executeNextInvestigation(client, config, workspace, brief, timeoutMs, {
+                    managedRoot: hb.on_shift?.managed_workspace_root ?? null,
+                    heartbeat: publishHeartbeat,
+                    heartbeatIntervalMs: intervalMs,
+                }) || progressed;
                 progressed = await pumpExecutionSlots(client, config, workspace, brief, timeoutMs, {
-                    heartbeat: async () => {
-                        await publishWorkspaceState();
-                    },
+                    heartbeat: publishHeartbeat,
                     beforeClaim: async (driverId) => {
                         // Never let the five-minute heartbeat cache span a CLI upgrade into a certified claim.
                         // Publish the fresh probe first; the Control Plane then remains the authority for the

package/dist/investigation.js

@@ -17,9 +17,8 @@ import { pickDriverForClaim, resolveDriverFuel, supportsReadOnlyDiagnosis } from
 import { diffStatSince } from "./git-witness.js";
 import { switchManagedWorkspace } from "./on-shift-apply.js";
 import { execFile, spawn } from "node:child_process";
-import { mkdtemp, rm, writeFile } from "node:fs/promises";
+import { rm, writeFile } from "node:fs/promises";
 import { createRequire } from "node:module";
-import { tmpdir } from "node:os";
 import { join } from "node:path";
 import { promisify } from "node:util";
 import { isRunnableVerificationCommand } from "./execution-class.js";
@@ -97,9 +96,11 @@ const investigationBriefSchema = z.object({
 /** T122: a delivery verification is an investigation whose question carries this header. */
 export const DELIVERY_VERIFICATION_HEADER = "DELIVERY VERIFICATION";
 /** Owner-verbatim: the lane writes probe source; Bridge executes. Shown only when BRIEF KEYS are present. */
-export const DELIVERY_VERIFICATION_PROBE_INSTRUCTION = 'For every BRIEF KEY that states behaviour, add one entry to probes: either {"key":"[sN]","test":"<registered command that asserts it>","checks":"<one line: what it proves>"} or {"key":"[sN]","lang":"ts|py|sh","source":"<a short script that exits 0 only when the key holds at this commit and non-zero otherwise; it may import repository modules by absolute path from the working directory>","checks":"<one line>"}. You do not run probes; Bridge runs them after you finish. Keys that state no behaviour (wording, scope, docs) get no probe.';
+export const DELIVERY_VERIFICATION_PROBE_INSTRUCTION = 'For every BRIEF KEY that states behaviour, add one entry to probes: either {"key":"[sN]","test":"<registered command that asserts it>","checks":"<one line: what it proves>"} or {"key":"[sN]","lang":"ts|py|sh","source":"<a short script run from the repository root that prints PROBE HOLDS as its last line and exits 0 when the key holds at this commit and prints PROBE FAILS as its last line and exits 1 when it does not; any other ending is recorded as broken; it may import repository modules by paths relative to the repository root>","checks":"<one line>"}. You do not run probes; Bridge runs them after you finish. Keys that state no behaviour (wording, scope, docs) get no probe.';
 const BRIEF_JSON_EXAMPLE = '{"summary":"one paragraph: what the change does and whether it meets the criteria","findings":["grounded fact with file/symbol"],"verification":["command — result"],"limitations":["what you could not settle"],"criteria":[{"criterion":"c1","assessment":"supported|unsupported|unclear","reason":"what you saw"}],"command_failures":["command that exited non-zero — the decisive output line"],"files_read":0,"bytes_read":0,"commands_run":["command you actually ran"]}';
 const BRIEF_JSON_EXAMPLE_WITH_PROBES = '{"summary":"one paragraph: what the change does and whether it meets the criteria","findings":["grounded fact with file/symbol"],"verification":["command — result"],"limitations":["what you could not settle"],"criteria":[{"criterion":"c1","assessment":"supported|unsupported|unclear","reason":"what you saw"}],"command_failures":["command that exited non-zero — the decisive output line"],"probes":[],"files_read":0,"bytes_read":0,"commands_run":["command you actually ran"]}';
+/** Named in both prompts so the lane knows each list's maximum before it writes. */
+export const BRIEF_LIMITS_SENTENCE = "Limits: findings at most 12, verification at most 8, limitations at most 8, criteria at most 40, command_failures at most 10, commands_run at most 10, probes at most 20; each text entry at most 1000 characters, commands_run entries at most 200 characters. Keep the most decisive entries.";
 export function isDeliveryVerification(assignment) {
     return assignment.question.trimStart().startsWith(DELIVERY_VERIFICATION_HEADER);
 }
@@ -121,6 +122,7 @@ export function buildInvestigationPrompt(assignment, commit, context) {
         "",
         "Return only one fenced ```json object with exactly this shape:",
         '{"summary":"direct answer to the question","findings":["grounded fact with file/symbol"],"verification":["command you actually ran"],"limitations":["what you could not settle"],"files_read":0,"bytes_read":0,"commands_run":["command you actually ran"]}',
+        BRIEF_LIMITS_SENTENCE,
         "summary is a string; the other list fields are arrays of strings. Close the fence and add no prose after it.",
     ].join("\n");
 }
@@ -164,6 +166,7 @@ export function buildDeliveryVerificationPrompt(assignment, commit, context) {
         "",
         "Return only one fenced ```json object with exactly this shape:",
         keys.length ? BRIEF_JSON_EXAMPLE_WITH_PROBES : BRIEF_JSON_EXAMPLE,
+        BRIEF_LIMITS_SENTENCE,
         keys.length
             ? "criteria has one entry per ACCEPTANCE key. probes has at most 20 entries, one per BRIEF KEY that states behaviour. Close the fence and add no prose after it."
             : "criteria has one entry per ACCEPTANCE key. Close the fence and add no prose after it.",
@@ -208,6 +211,83 @@ function settleCommandsRun(gates, reported) {
 export const UNUSABLE_BRIEF = "Investigation returned no usable brief";
 /** T130: how much of an unusable reply travels with the failure, so the next run can be diagnosed. */
 const UNUSABLE_BRIEF_TAIL_CHARS = 800;
+const BRIEF_SUMMARY_MAX = 4_000;
+const BRIEF_ITEM_MAX = 1_000;
+const BRIEF_FILES_READ_MAX = 24;
+const BRIEF_BYTES_READ_MAX = 512 * 1024;
+const BRIEF_COMMAND_RUN_MAX = 200;
+const BRIEF_LIST_MAX = {
+    findings: 12,
+    verification: 8,
+    limitations: 8,
+    criteria: 40,
+    command_failures: 10,
+    commands_run: 10,
+    probes: 20,
+};
+function clipBriefText(value, max) {
+    return value.length > max ? value.slice(0, max) : value;
+}
+function clipBriefStringField(record, key, max) {
+    const value = record[key];
+    if (typeof value === "string")
+        record[key] = clipBriefText(value, max);
+}
+function fitBriefStrings(items, max) {
+    return items.map((item) => typeof item === "string" ? clipBriefText(item, max) : item);
+}
+function fitCriteriaItems(items) {
+    return items.map((item) => {
+        if (!item || typeof item !== "object" || Array.isArray(item))
+            return item;
+        const row = { ...item };
+        clipBriefStringField(row, "criterion", BRIEF_ITEM_MAX);
+        clipBriefStringField(row, "reason", BRIEF_ITEM_MAX);
+        return row;
+    });
+}
+/**
+ * Cut an over-full brief to the contract before schema parse. Only size is fitted; a wrong type,
+ * a missing summary, or an unknown assessment still fails as today. Probe entries are kept as
+ * written — clipping them would change what Bridge later executes.
+ */
+export function fitBriefToContract(candidate) {
+    if (!candidate || typeof candidate !== "object" || Array.isArray(candidate)) {
+        return { fitted: candidate, cuts: [] };
+    }
+    const brief = { ...candidate };
+    const cuts = [];
+    const fitNamedList = (key, fitItems) => {
+        const value = brief[key];
+        if (!Array.isArray(value))
+            return;
+        const max = BRIEF_LIST_MAX[key];
+        if (value.length > max)
+            cuts.push({ list: key, kept: max, total: value.length });
+        brief[key] = fitItems(value.slice(0, max));
+    };
+    fitNamedList("findings", (items) => fitBriefStrings(items, BRIEF_ITEM_MAX));
+    fitNamedList("verification", (items) => fitBriefStrings(items, BRIEF_ITEM_MAX));
+    fitNamedList("limitations", (items) => fitBriefStrings(items, BRIEF_ITEM_MAX));
+    fitNamedList("criteria", fitCriteriaItems);
+    fitNamedList("command_failures", (items) => fitBriefStrings(items, BRIEF_ITEM_MAX));
+    fitNamedList("commands_run", (items) => fitBriefStrings(items, BRIEF_COMMAND_RUN_MAX));
+    fitNamedList("probes", (items) => items);
+    clipBriefStringField(brief, "summary", BRIEF_SUMMARY_MAX);
+    if (typeof brief.files_read === "number" && brief.files_read > BRIEF_FILES_READ_MAX) {
+        brief.files_read = BRIEF_FILES_READ_MAX;
+    }
+    if (typeof brief.bytes_read === "number" && brief.bytes_read > BRIEF_BYTES_READ_MAX) {
+        brief.bytes_read = BRIEF_BYTES_READ_MAX;
+    }
+    if (cuts.length) {
+        const notes = cuts.map((cut) => clipBriefText(`Bridge kept the first ${cut.kept} of ${cut.total} ${cut.list}`, BRIEF_ITEM_MAX));
+        const existing = Array.isArray(brief.limitations) ? [...brief.limitations] : [];
+        const room = Math.max(0, BRIEF_LIST_MAX.limitations - notes.length);
+        brief.limitations = [...existing.slice(0, room), ...notes];
+    }
+    return { fitted: brief, cuts };
+}
 /**
  * T130: an unusable brief used to fail with five words. The first production verification retry
  * (2026-09-09, acme-sim / forge R20) returned "no usable brief" twice and nobody could say whether the
@@ -222,7 +302,7 @@ export function parseInvestigationBrief(text) {
     catch {
         throw new Error(`${UNUSABLE_BRIEF}: no JSON object in the reply. Reply tail: ${redactSecrets(boundedTail(text, UNUSABLE_BRIEF_TAIL_CHARS))}`);
     }
-    const parsed = investigationBriefSchema.safeParse(candidate);
+    const parsed = investigationBriefSchema.safeParse(fitBriefToContract(candidate).fitted);
     if (!parsed.success) {
         const issues = parsed.error.issues.slice(0, 3).map((issue) => `${issue.path.join(".") || "(root)"}: ${issue.message}`).join("; ");
         throw new Error(`${UNUSABLE_BRIEF}: the JSON does not match the brief contract (${issues}). Reply tail: ${redactSecrets(boundedTail(text, UNUSABLE_BRIEF_TAIL_CHARS))}`);
@@ -354,11 +434,8 @@ function probeInterpreter(language) {
     return "bash";
 }
 function probeFilename(language) {
-    if (language === "ts")
-        return "probe.mts";
-    if (language === "py")
-        return "probe.py";
-    return "probe.sh";
+    const extension = language === "ts" ? "mts" : language === "py" ? "py" : "sh";
+    return `.conduit-probe-${crypto.randomUUID()}.${extension}`;
 }
 function probeArgv(language, file) {
     if (language === "ts")
@@ -452,13 +529,11 @@ export async function runDeliveryVerificationProbes(input) {
             continue;
         const language = probe.lang;
         const command = `${probeInterpreter(language)} ${probe.key}`;
-        let dir = null;
+        const file = join(input.workspace, probeFilename(language));
         let exitCode = -1;
         let stdout = "";
         let stderr = "";
         try {
-            dir = await mkdtemp(join(tmpdir(), "conduit-probe-"));
-            const file = join(dir, probeFilename(language));
             await writeFile(file, source, { encoding: "utf8" });
             const result = await execProbe(probeArgv(language, file), input.workspace);
             exitCode = result.code;
@@ -470,8 +545,7 @@ export async function runDeliveryVerificationProbes(input) {
             exitCode = -1;
         }
         finally {
-            if (dir)
-                await rm(dir, { recursive: true, force: true }).catch(() => undefined);
+            await rm(file, { force: true }).catch(() => undefined);
         }
         witnesses.push(witnessRecord({
             key: probe.key, kind: "probe", command, exitCode, stdout, stderr, checks,
@@ -495,11 +569,12 @@ const deliveredWitnessSchema = z.object({
 function stampInstrument(witnesses, instrument) {
     return witnesses.map((witness) => ({ ...witness, instrument }));
 }
-function markSelfCheckBroken(witnesses) {
+function markSelfCheckBroken(witnesses, reason) {
+    const tail = boundedTail(`${DELIVERED_PROBE_SELF_CHECK_TAIL}: ${reason}`, PROBE_TAIL_CHARS);
     return stampInstrument(witnesses.map((witness) => ({
         ...witness,
         verdict: "broken",
-        tail: DELIVERED_PROBE_SELF_CHECK_TAIL,
+        tail,
     })), "running");
 }
 function toStepWitness(witness) {
@@ -569,16 +644,18 @@ export async function runDeliveryVerificationProbesForAttempt(input) {
         return stampInstrument(await runRunning(), "running");
     }
     const runDelivered = input.runDeliveredProbes ?? ((args) => runDeliveredProbeInstrument(args));
+    let selfCheckDetail = null;
     try {
         const check = await runDelivered({ workspace: input.workspace, probes: KNOWN_ANSWER_PROBES });
         if (knownAnswerProbesPassed(check)) {
             return stampInstrument(await runDelivered({ workspace: input.workspace, probes: input.probes }), "delivered");
         }
+        selfCheckDetail = `gave ${check.map((witness) => witness.verdict).join(", ")}`;
     }
-    catch {
-        // Known-answer failure and a delivered-instrument throw take the same path: running probes, then broken.
+    catch (error) {
+        selfCheckDetail = redactSecrets(error instanceof Error ? error.message : String(error));
     }
-    return markSelfCheckBroken(await runRunning());
+    return markSelfCheckBroken(await runRunning(), selfCheckDetail ?? "");
 }
 async function settle(client, investigationId, body) {
     await client.request(`/runner/v1/investigations/${investigationId}/settle`, {
@@ -675,6 +752,19 @@ export async function executeNextInvestigation(client, config, workspace, brief,
         }).catch((error) => console.error(`Investigation ${assignment.id} lease renewal failed: ${redactSecrets(error instanceof Error ? error.message : "unknown")}`));
     }, options.renewIntervalMs ?? INVESTIGATION_RENEW_INTERVAL_MS);
     renewTimer.unref?.();
+    let heartbeatRunning = false;
+    const heartbeat = options.heartbeat;
+    const heartbeatTimer = heartbeat
+        ? setInterval(() => {
+            if (heartbeatRunning)
+                return;
+            heartbeatRunning = true;
+            void heartbeat()
+                .catch((error) => console.error(`Heartbeat failed during investigation: ${redactSecrets(error instanceof Error ? error.message : "unknown")}`))
+                .finally(() => { heartbeatRunning = false; });
+        }, options.heartbeatIntervalMs ?? 15_000)
+        : null;
+    heartbeatTimer?.unref?.();
     try {
         attemptWorkspace = await createAttemptWorktree({
             sourceWorkspace: workspace,
@@ -795,6 +885,8 @@ export async function executeNextInvestigation(client, config, workspace, brief,
     }
     finally {
         clearInterval(renewTimer);
+        if (heartbeatTimer)
+            clearInterval(heartbeatTimer);
         if (attemptWorkspace) {
             await removeAttemptWorktree(workspace, attemptWorkspace)
                 .catch((error) => console.error(`Investigation worktree release failed: ${redactSecrets(error instanceof Error ? error.message : "unknown")}`));

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@miraland-labs/conduit-bridge",
-  "version": "0.16.113",
+  "version": "0.16.115",
   "description": "Conduit Bridge CLI \u2014 join, connect, disconnect, multi-driver lanes, and run Claude Code / Codex / Cursor / OpenCode / Pi / Kiro / Antigravity / Grok Build agents for a Conduit organization",
   "type": "module",
   "bin": {