@geraldmaron/construct 1.0.15 → 1.0.17

package/bin/construct-postinstall.mjs

@@ -34,6 +34,18 @@ const PKG_VERSION = (() => {
 
 const log = (msg) => process.stdout.write(`[construct-postinstall] ${msg}\n`);
 
+// Failures go to stderr as a prominent, actionable block so they are visible in
+// npm output instead of scrolling past as a benign-looking line. The script
+// stays exit-0 by design (see the staging catch): best-effort adapter staging
+// must not hard-fail a consumer's `npm install` — the package is installed and
+// usable, and `construct init` completes staging on demand.
+
+const fail = (msg, hint) => {
+  process.stderr.write(`\n[construct-postinstall] WARNING: ${msg}\n`);
+  if (hint) process.stderr.write(`[construct-postinstall]   -> ${hint}\n`);
+  process.stderr.write('\n');
+};
+
 if (process.env.CONSTRUCT_SKIP_POSTINSTALL === '1') {
   log('skipping (CONSTRUCT_SKIP_POSTINSTALL=1)');
   process.exit(0);
@@ -72,7 +84,7 @@ if (process.env.npm_config_global === 'true' || process.env.npm_config_global ==
       stdio: 'inherit',
     });
     if (result.status !== 0) {
-      log(`global sync failed (exit ${result.status}); run \`construct sync --global\` manually`);
+      fail(`Global front-door sync failed (exit ${result.status}).`, 'Run `construct sync --global` to complete it.');
     }
   }
   process.exit(0);
@@ -103,6 +115,9 @@ try {
     log,
   });
 } catch (err) {
-  log(`staging failed: ${err.message}; leaving project in a clean state`);
+  fail(`Adapter staging failed: ${err.message}`, 'The package is installed; run `npx construct init` in this project to complete setup.');
+  // Intentionally exit 0: staging is best-effort completion, and a non-zero exit
+  // here would fail the consumer's entire `npm install`. The failure is loud on
+  // stderr above and recoverable via `construct init`.
   process.exit(0);
 }

package/db/schema/010_cx_scores.sql

@@ -0,0 +1,51 @@
+-- db/schema/010_cx_scores.sql — Per-trace quality score log + skill quality correlation view.
+--
+-- Backs `construct skills correlate-quality`.  Producer: lib/mcp/tools/telemetry.mjs
+-- cxScore() writes a row here when CONSTRUCT_DB_URL (or DATABASE_URL) is set.
+-- Local JSONL/observation paths remain primary in solo mode; this table is
+-- primary in team/enterprise mode where multiple agents share a session
+-- timeline and the dashboard reads aggregate correlations.
+
+create table if not exists construct_cx_scores (
+  id           bigserial    primary key,
+  ts           timestamptz  not null,
+  trace_id     text         not null,
+  session_id   text,
+  agent_id     text,
+  name         text         not null default 'quality',
+  value        numeric      not null,
+  comment      text
+);
+
+create index if not exists construct_cx_scores_trace_idx
+  on construct_cx_scores (trace_id);
+
+create index if not exists construct_cx_scores_session_ts_idx
+  on construct_cx_scores (session_id, ts desc);
+
+create index if not exists construct_cx_scores_agent_ts_idx
+  on construct_cx_scores (agent_id, ts desc);
+
+-- Skill quality correlation view: joins skill invocations and quality
+-- scores by session_id so a reader can ask "for sessions where skill X was
+-- invoked, what's the median / mean / p10 quality score?" The view is a
+-- simple aggregate over the past 90 days; queries that need a longer
+-- window can pull straight from the underlying tables. Materialised view
+-- not used here because the correlation surface is small and the freshness
+-- expectation is "current session" not "yesterday's batch."
+
+create or replace view construct_skill_quality_correlation as
+select
+  inv.skill_id,
+  count(distinct inv.session_id)                            as sessions,
+  count(distinct sc.id)                                     as score_count,
+  round(avg(sc.value)::numeric, 3)                          as mean_score,
+  round((percentile_cont(0.5) within group (order by sc.value))::numeric, 3)  as median_score,
+  round((percentile_cont(0.10) within group (order by sc.value))::numeric, 3) as p10_score,
+  round((percentile_cont(0.90) within group (order by sc.value))::numeric, 3) as p90_score,
+  max(sc.ts)                                                as last_scored_at
+from construct_skill_invocations inv
+  inner join construct_cx_scores sc on sc.session_id = inv.session_id
+where inv.ts > now() - interval '90 days'
+  and sc.ts  > now() - interval '90 days'
+group by inv.skill_id;