chainlesschain 0.81.0 → 0.132.0

package/src/commands/quantization.js

@@ -20,8 +20,51 @@ import {
   cancelJob,
   deleteJob,
   getQuantizationStats,
+  /* V2 (Phase 20) */
+  MODEL_MATURITY_V2,
+  JOB_TICKET_V2,
+  getDefaultMaxActiveModelsPerOwnerV2,
+  getMaxActiveModelsPerOwnerV2,
+  setMaxActiveModelsPerOwnerV2,
+  getDefaultMaxRunningJobsPerOwnerV2,
+  getMaxRunningJobsPerOwnerV2,
+  setMaxRunningJobsPerOwnerV2,
+  getDefaultModelIdleMsV2,
+  getModelIdleMsV2,
+  setModelIdleMsV2,
+  getDefaultJobStuckMsV2,
+  getJobStuckMsV2,
+  setJobStuckMsV2,
+  registerModelV2,
+  getModelV2,
+  setModelMaturityV2,
+  activateModel,
+  deprecateModel,
+  retireModel,
+  touchModelUsage,
+  enqueueJobTicketV2,
+  getJobTicketV2,
+  setJobTicketStatusV2,
+  startJobTicket,
+  completeJobTicket,
+  failJobTicket,
+  cancelJobTicket,
+  getActiveModelCount,
+  getRunningJobCount,
+  autoRetireIdleModels,
+  autoFailStuckJobTickets,
+  getQuantizationStatsV2,
 } from "../lib/quantization.js";
 
+function _parseMetaV2(s) {
+  if (!s) return undefined;
+  try {
+    return JSON.parse(s);
+  } catch {
+    throw new Error(`--metadata must be valid JSON`);
+  }
+}
+
 function _dbFromCtx(cmd) {
   const root = cmd?.parent?.parent ?? cmd?.parent;
   return root?._db;
@@ -276,5 +319,313 @@ export function registerQuantizationCommand(program) {
         console.log(`Avg duration: ${s.avgDurationMs}ms`);
     });
 
+  /* ═════════════════════════════════════════════════════ *
+   *  Phase 20 V2 — Model Maturity + Job Ticket Lifecycle
+   * ═════════════════════════════════════════════════════ */
+
+  quant
+    .command("model-maturities-v2")
+    .description("List Phase 20 V2 model maturity states")
+    .option("--json", "JSON output")
+    .action((opts) => {
+      const v = Object.values(MODEL_MATURITY_V2);
+      if (opts.json) return console.log(JSON.stringify(v, null, 2));
+      for (const s of v) console.log(s);
+    });
+
+  quant
+    .command("job-ticket-lifecycles-v2")
+    .description("List Phase 20 V2 job-ticket lifecycle states")
+    .option("--json", "JSON output")
+    .action((opts) => {
+      const v = Object.values(JOB_TICKET_V2);
+      if (opts.json) return console.log(JSON.stringify(v, null, 2));
+      for (const s of v) console.log(s);
+    });
+
+  quant
+    .command("default-max-active-models-per-owner")
+    .description("Show default V2 per-owner active-model cap")
+    .action(() => console.log(getDefaultMaxActiveModelsPerOwnerV2()));
+
+  quant
+    .command("max-active-models-per-owner")
+    .description("Show current V2 per-owner active-model cap")
+    .action(() => console.log(getMaxActiveModelsPerOwnerV2()));
+
+  quant
+    .command("set-max-active-models-per-owner <n>")
+    .description("Set V2 per-owner active-model cap")
+    .action((n) => console.log(setMaxActiveModelsPerOwnerV2(n)));
+
+  quant
+    .command("default-max-running-jobs-per-owner")
+    .description("Show default V2 per-owner running-job cap")
+    .action(() => console.log(getDefaultMaxRunningJobsPerOwnerV2()));
+
+  quant
+    .command("max-running-jobs-per-owner")
+    .description("Show current V2 per-owner running-job cap")
+    .action(() => console.log(getMaxRunningJobsPerOwnerV2()));
+
+  quant
+    .command("set-max-running-jobs-per-owner <n>")
+    .description("Set V2 per-owner running-job cap")
+    .action((n) => console.log(setMaxRunningJobsPerOwnerV2(n)));
+
+  quant
+    .command("default-model-idle-ms")
+    .description("Show default V2 model-idle threshold")
+    .action(() => console.log(getDefaultModelIdleMsV2()));
+
+  quant
+    .command("model-idle-ms")
+    .description("Show current V2 model-idle threshold")
+    .action(() => console.log(getModelIdleMsV2()));
+
+  quant
+    .command("set-model-idle-ms <ms>")
+    .description("Set V2 model-idle threshold (ms)")
+    .action((ms) => console.log(setModelIdleMsV2(ms)));
+
+  quant
+    .command("default-job-stuck-ms")
+    .description("Show default V2 job-stuck threshold")
+    .action(() => console.log(getDefaultJobStuckMsV2()));
+
+  quant
+    .command("job-stuck-ms")
+    .description("Show current V2 job-stuck threshold")
+    .action(() => console.log(getJobStuckMsV2()));
+
+  quant
+    .command("set-job-stuck-ms <ms>")
+    .description("Set V2 job-stuck threshold (ms)")
+    .action((ms) => console.log(setJobStuckMsV2(ms)));
+
+  quant
+    .command("active-model-count")
+    .description("Count active V2 models (optionally scoped by owner)")
+    .option("-o, --owner <id>", "Owner ID")
+    .action((opts) => console.log(getActiveModelCount(opts.owner)));
+
+  quant
+    .command("running-job-count")
+    .description("Count running V2 job tickets (optionally scoped by owner)")
+    .option("-o, --owner <id>", "Owner ID")
+    .action((opts) => console.log(getRunningJobCount(opts.owner)));
+
+  quant
+    .command("register-model-v2 <model-id>")
+    .description("Register a V2 model")
+    .requiredOption("-o, --owner <id>", "Owner ID")
+    .option("-f, --family <name>", "Model family (e.g. llama)")
+    .option("-i, --initial-status <status>", "Initial status")
+    .option("-m, --metadata <json>", "Metadata JSON")
+    .option("--json", "JSON output")
+    .action((modelId, opts) => {
+      const db = _dbFromCtx(quant);
+      const rec = registerModelV2(db, {
+        modelId,
+        ownerId: opts.owner,
+        family: opts.family,
+        initialStatus: opts.initialStatus,
+        metadata: _parseMetaV2(opts.metadata),
+      });
+      if (opts.json) return console.log(JSON.stringify(rec, null, 2));
+      console.log(`Registered model ${modelId} (${rec.status})`);
+    });
+
+  quant
+    .command("model-v2 <model-id>")
+    .description("Show a V2 model")
+    .option("--json", "JSON output")
+    .action((id, opts) => {
+      const rec = getModelV2(id);
+      if (!rec) {
+        console.error(`Unknown model: ${id}`);
+        process.exitCode = 1;
+        return;
+      }
+      if (opts.json) return console.log(JSON.stringify(rec, null, 2));
+      console.log(`${rec.modelId} [${rec.status}] owner=${rec.ownerId}`);
+    });
+
+  quant
+    .command("set-model-maturity-v2 <model-id> <status>")
+    .description("Transition V2 model maturity")
+    .option("-r, --reason <text>", "Reason")
+    .option("-m, --metadata <json>", "Metadata patch (JSON)")
+    .action((id, status, opts) => {
+      const db = _dbFromCtx(quant);
+      const rec = setModelMaturityV2(db, id, status, {
+        reason: opts.reason,
+        metadata: _parseMetaV2(opts.metadata),
+      });
+      console.log(`${id} → ${rec.status}`);
+    });
+
+  quant
+    .command("activate-model <model-id>")
+    .description("Transition a V2 model to ACTIVE")
+    .option("-r, --reason <text>", "Reason")
+    .action((id, opts) => {
+      const rec = activateModel(_dbFromCtx(quant), id, opts.reason);
+      console.log(`${id} → ${rec.status}`);
+    });
+
+  quant
+    .command("deprecate-model <model-id>")
+    .description("Transition a V2 model to DEPRECATED")
+    .option("-r, --reason <text>", "Reason")
+    .action((id, opts) => {
+      const rec = deprecateModel(_dbFromCtx(quant), id, opts.reason);
+      console.log(`${id} → ${rec.status}`);
+    });
+
+  quant
+    .command("retire-model <model-id>")
+    .description("Transition a V2 model to RETIRED")
+    .option("-r, --reason <text>", "Reason")
+    .action((id, opts) => {
+      const rec = retireModel(_dbFromCtx(quant), id, opts.reason);
+      console.log(`${id} → ${rec.status}`);
+    });
+
+  quant
+    .command("touch-model-usage <model-id>")
+    .description("Bump lastUsedAt for a V2 model")
+    .action((id) => {
+      const rec = touchModelUsage(id);
+      console.log(`${id} lastUsedAt=${rec.lastUsedAt}`);
+    });
+
+  quant
+    .command("enqueue-job-ticket-v2 <ticket-id>")
+    .description("Enqueue a V2 quantization job ticket")
+    .requiredOption("-o, --owner <id>", "Owner ID")
+    .requiredOption("-M, --model <id>", "Model ID")
+    .requiredOption("-t, --quant-type <type>", "Quant type (gguf/gptq)")
+    .option("-l, --level <lvl>", "Level (e.g. Q4_K_M)")
+    .option("-m, --metadata <json>", "Metadata JSON")
+    .option("--json", "JSON output")
+    .action((ticketId, opts) => {
+      const db = _dbFromCtx(quant);
+      const rec = enqueueJobTicketV2(db, {
+        ticketId,
+        ownerId: opts.owner,
+        modelId: opts.model,
+        quantType: opts.quantType,
+        level: opts.level,
+        metadata: _parseMetaV2(opts.metadata),
+      });
+      if (opts.json) return console.log(JSON.stringify(rec, null, 2));
+      console.log(`Enqueued ticket ${ticketId} (${rec.status})`);
+    });
+
+  quant
+    .command("job-ticket-v2 <ticket-id>")
+    .description("Show a V2 job ticket")
+    .option("--json", "JSON output")
+    .action((id, opts) => {
+      const rec = getJobTicketV2(id);
+      if (!rec) {
+        console.error(`Unknown ticket: ${id}`);
+        process.exitCode = 1;
+        return;
+      }
+      if (opts.json) return console.log(JSON.stringify(rec, null, 2));
+      console.log(`${rec.ticketId} [${rec.status}] owner=${rec.ownerId}`);
+    });
+
+  quant
+    .command("set-job-ticket-status-v2 <ticket-id> <status>")
+    .description("Transition V2 job ticket status")
+    .option("-r, --reason <text>", "Reason")
+    .option("-m, --metadata <json>", "Metadata patch (JSON)")
+    .action((id, status, opts) => {
+      const db = _dbFromCtx(quant);
+      const rec = setJobTicketStatusV2(db, id, status, {
+        reason: opts.reason,
+        metadata: _parseMetaV2(opts.metadata),
+      });
+      console.log(`${id} → ${rec.status}`);
+    });
+
+  quant
+    .command("start-job-ticket <ticket-id>")
+    .description("Transition a V2 job ticket to RUNNING")
+    .option("-r, --reason <text>", "Reason")
+    .action((id, opts) => {
+      const rec = startJobTicket(_dbFromCtx(quant), id, opts.reason);
+      console.log(`${id} → ${rec.status}`);
+    });
+
+  quant
+    .command("complete-job-ticket <ticket-id>")
+    .description("Transition a V2 job ticket to COMPLETED")
+    .option("-r, --reason <text>", "Reason")
+    .action((id, opts) => {
+      const rec = completeJobTicket(_dbFromCtx(quant), id, opts.reason);
+      console.log(`${id} → ${rec.status}`);
+    });
+
+  quant
+    .command("fail-job-ticket <ticket-id>")
+    .description("Transition a V2 job ticket to FAILED")
+    .option("-r, --reason <text>", "Reason")
+    .action((id, opts) => {
+      const rec = failJobTicket(_dbFromCtx(quant), id, opts.reason);
+      console.log(`${id} → ${rec.status}`);
+    });
+
+  quant
+    .command("cancel-job-ticket <ticket-id>")
+    .description("Transition a V2 job ticket to CANCELED")
+    .option("-r, --reason <text>", "Reason")
+    .action((id, opts) => {
+      const rec = cancelJobTicket(_dbFromCtx(quant), id, opts.reason);
+      console.log(`${id} → ${rec.status}`);
+    });
+
+  quant
+    .command("auto-retire-idle-models")
+    .description("Flip idle V2 models (active+deprecated) → RETIRED")
+    .option("--json", "JSON output")
+    .action((opts) => {
+      const r = autoRetireIdleModels(_dbFromCtx(quant));
+      if (opts.json) return console.log(JSON.stringify(r, null, 2));
+      console.log(`Retired ${r.count} idle model(s)`);
+    });
+
+  quant
+    .command("auto-fail-stuck-job-tickets")
+    .description("Flip stuck RUNNING V2 tickets → FAILED")
+    .option("--json", "JSON output")
+    .action((opts) => {
+      const r = autoFailStuckJobTickets(_dbFromCtx(quant));
+      if (opts.json) return console.log(JSON.stringify(r, null, 2));
+      console.log(`Failed ${r.count} stuck ticket(s)`);
+    });
+
+  quant
+    .command("stats-v2")
+    .description("Phase 20 V2 statistics")
+    .option("--json", "JSON output")
+    .action((opts) => {
+      const s = getQuantizationStatsV2();
+      if (opts.json) return console.log(JSON.stringify(s, null, 2));
+      console.log(
+        `Models(V2)=${s.totalModelsV2} Tickets(V2)=${s.totalTicketsV2} ` +
+          `caps: active-models/owner=${s.maxActiveModelsPerOwner} running-jobs/owner=${s.maxRunningJobsPerOwner}`,
+      );
+      console.log("models-by-status:");
+      for (const [k, v] of Object.entries(s.modelsByStatus))
+        console.log(`  ${k.padEnd(12)} ${v}`);
+      console.log("tickets-by-status:");
+      for (const [k, v] of Object.entries(s.ticketsByStatus))
+        console.log(`  ${k.padEnd(12)} ${v}`);
+    });
+
   program.addCommand(quant);
 }

package/src/commands/rcache.js

@@ -0,0 +1,271 @@
+/**
+ * Response Cache V2 commands
+ * `cc rcache ...` — in-memory governance layer for cache profiles
+ * + refresh jobs, independent of SQLite llm_cache table.
+ */
+
+import { logger } from "../lib/logger.js";
+import {
+  PROFILE_MATURITY_V2,
+  REFRESH_JOB_LIFECYCLE_V2,
+  getMaxActiveProfilesPerOwnerV2,
+  setMaxActiveProfilesPerOwnerV2,
+  getMaxPendingRefreshJobsPerProfileV2,
+  setMaxPendingRefreshJobsPerProfileV2,
+  getProfileIdleMsV2,
+  setProfileIdleMsV2,
+  getRefreshStuckMsV2,
+  setRefreshStuckMsV2,
+  registerProfileV2,
+  getProfileV2,
+  listProfilesV2,
+  setProfileStatusV2,
+  activateProfileV2,
+  suspendProfileV2,
+  archiveProfileV2,
+  touchProfileV2,
+  getActiveProfileCountV2,
+  createRefreshJobV2,
+  getRefreshJobV2,
+  listRefreshJobsV2,
+  setRefreshJobStatusV2,
+  startRefreshJobV2,
+  completeRefreshJobV2,
+  failRefreshJobV2,
+  cancelRefreshJobV2,
+  getPendingRefreshJobCountV2,
+  autoSuspendIdleProfilesV2,
+  autoFailStuckRefreshJobsV2,
+  getResponseCacheStatsV2,
+} from "../lib/response-cache.js";
+
+const out = (obj) => console.log(JSON.stringify(obj, null, 2));
+const tryRun = (fn) => {
+  try {
+    fn();
+  } catch (err) {
+    logger.error(err.message);
+    process.exit(1);
+  }
+};
+
+export function registerRCacheCommand(program) {
+  const root = program
+    .command("rcache")
+    .description("Response cache V2 — profile maturity + refresh jobs");
+
+  root
+    .command("profile-maturities-v2")
+    .description("List V2 profile maturity states")
+    .action(() => out(Object.values(PROFILE_MATURITY_V2)));
+
+  root
+    .command("refresh-job-lifecycles-v2")
+    .description("List V2 refresh-job lifecycle states")
+    .action(() => out(Object.values(REFRESH_JOB_LIFECYCLE_V2)));
+
+  root
+    .command("stats-v2")
+    .description("V2 response cache stats")
+    .action(() => out(getResponseCacheStatsV2()));
+
+  root
+    .command("get-max-active-profiles-v2")
+    .description("Get max active profiles per owner (V2)")
+    .action(() =>
+      out({ maxActiveProfilesPerOwner: getMaxActiveProfilesPerOwnerV2() }),
+    );
+
+  root
+    .command("set-max-active-profiles-v2 <n>")
+    .description("Set max active profiles per owner (V2)")
+    .action((n) =>
+      tryRun(() => {
+        setMaxActiveProfilesPerOwnerV2(Number(n));
+        out({ maxActiveProfilesPerOwner: getMaxActiveProfilesPerOwnerV2() });
+      }),
+    );
+
+  root
+    .command("get-max-pending-refresh-jobs-v2")
+    .description("Get max pending refresh jobs per profile (V2)")
+    .action(() =>
+      out({
+        maxPendingRefreshJobsPerProfile: getMaxPendingRefreshJobsPerProfileV2(),
+      }),
+    );
+
+  root
+    .command("set-max-pending-refresh-jobs-v2 <n>")
+    .description("Set max pending refresh jobs per profile (V2)")
+    .action((n) =>
+      tryRun(() => {
+        setMaxPendingRefreshJobsPerProfileV2(Number(n));
+        out({
+          maxPendingRefreshJobsPerProfile:
+            getMaxPendingRefreshJobsPerProfileV2(),
+        });
+      }),
+    );
+
+  root
+    .command("get-profile-idle-ms-v2")
+    .description("Get profile idle threshold (V2)")
+    .action(() => out({ profileIdleMs: getProfileIdleMsV2() }));
+
+  root
+    .command("set-profile-idle-ms-v2 <ms>")
+    .description("Set profile idle threshold (V2)")
+    .action((ms) =>
+      tryRun(() => {
+        setProfileIdleMsV2(Number(ms));
+        out({ profileIdleMs: getProfileIdleMsV2() });
+      }),
+    );
+
+  root
+    .command("get-refresh-stuck-ms-v2")
+    .description("Get refresh-job stuck threshold (V2)")
+    .action(() => out({ refreshStuckMs: getRefreshStuckMsV2() }));
+
+  root
+    .command("set-refresh-stuck-ms-v2 <ms>")
+    .description("Set refresh-job stuck threshold (V2)")
+    .action((ms) =>
+      tryRun(() => {
+        setRefreshStuckMsV2(Number(ms));
+        out({ refreshStuckMs: getRefreshStuckMsV2() });
+      }),
+    );
+
+  root
+    .command("active-profile-count-v2 <ownerId>")
+    .description("Active profile count for owner (V2)")
+    .action((ownerId) =>
+      out({ ownerId, count: getActiveProfileCountV2(ownerId) }),
+    );
+
+  root
+    .command("pending-refresh-job-count-v2 <profileId>")
+    .description("Pending refresh-job count for profile (V2)")
+    .action((profileId) =>
+      out({ profileId, count: getPendingRefreshJobCountV2(profileId) }),
+    );
+
+  root
+    .command("register-profile-v2 <id>")
+    .description("Register a V2 profile")
+    .requiredOption("-o, --owner <id>", "owner id")
+    .requiredOption("-l, --label <label>", "profile label")
+    .action((id, opts) =>
+      tryRun(() =>
+        out(registerProfileV2(id, { ownerId: opts.owner, label: opts.label })),
+      ),
+    );
+
+  root
+    .command("get-profile-v2 <id>")
+    .description("Get a V2 profile")
+    .action((id) => out(getProfileV2(id)));
+
+  root
+    .command("list-profiles-v2")
+    .description("List V2 profiles")
+    .option("-o, --owner <id>", "filter by owner")
+    .option("-s, --status <status>", "filter by status")
+    .action((opts) =>
+      out(listProfilesV2({ ownerId: opts.owner, status: opts.status })),
+    );
+
+  root
+    .command("set-profile-status-v2 <id> <next>")
+    .description("Set V2 profile status")
+    .action((id, next) => tryRun(() => out(setProfileStatusV2(id, next))));
+
+  root
+    .command("activate-profile-v2 <id>")
+    .description("Activate a V2 profile")
+    .action((id) => tryRun(() => out(activateProfileV2(id))));
+
+  root
+    .command("suspend-profile-v2 <id>")
+    .description("Suspend a V2 profile")
+    .action((id) => tryRun(() => out(suspendProfileV2(id))));
+
+  root
+    .command("archive-profile-v2 <id>")
+    .description("Archive a V2 profile")
+    .action((id) => tryRun(() => out(archiveProfileV2(id))));
+
+  root
+    .command("touch-profile-v2 <id>")
+    .description("Touch a V2 profile")
+    .action((id) => tryRun(() => out(touchProfileV2(id))));
+
+  root
+    .command("create-refresh-job-v2 <id>")
+    .description("Create a V2 refresh job")
+    .requiredOption("-p, --profile <id>", "profile id")
+    .option("-k, --kind <kind>", "job kind", "warm")
+    .action((id, opts) =>
+      tryRun(() =>
+        out(
+          createRefreshJobV2(id, { profileId: opts.profile, kind: opts.kind }),
+        ),
+      ),
+    );
+
+  root
+    .command("get-refresh-job-v2 <id>")
+    .description("Get a V2 refresh job")
+    .action((id) => out(getRefreshJobV2(id)));
+
+  root
+    .command("list-refresh-jobs-v2")
+    .description("List V2 refresh jobs")
+    .option("-p, --profile <id>", "filter by profile")
+    .option("-s, --status <status>", "filter by status")
+    .action((opts) =>
+      out(
+        listRefreshJobsV2({
+          profileId: opts.profile,
+          status: opts.status,
+        }),
+      ),
+    );
+
+  root
+    .command("set-refresh-job-status-v2 <id> <next>")
+    .description("Set V2 refresh-job status")
+    .action((id, next) => tryRun(() => out(setRefreshJobStatusV2(id, next))));
+
+  root
+    .command("start-refresh-job-v2 <id>")
+    .description("Start a V2 refresh job")
+    .action((id) => tryRun(() => out(startRefreshJobV2(id))));
+
+  root
+    .command("complete-refresh-job-v2 <id>")
+    .description("Complete a V2 refresh job")
+    .action((id) => tryRun(() => out(completeRefreshJobV2(id))));
+
+  root
+    .command("fail-refresh-job-v2 <id>")
+    .description("Fail a V2 refresh job")
+    .action((id) => tryRun(() => out(failRefreshJobV2(id))));
+
+  root
+    .command("cancel-refresh-job-v2 <id>")
+    .description("Cancel a V2 refresh job")
+    .action((id) => tryRun(() => out(cancelRefreshJobV2(id))));
+
+  root
+    .command("auto-suspend-idle-profiles-v2")
+    .description("Auto-suspend idle V2 profiles")
+    .action(() => out(autoSuspendIdleProfilesV2()));
+
+  root
+    .command("auto-fail-stuck-refresh-jobs-v2")
+    .description("Auto-fail stuck V2 refresh jobs")
+    .action(() => out(autoFailStuckRefreshJobsV2()));
+}