@valescoagency/runway 0.9.0 → 0.10.1

package/dist/dashboard/storage.js

@@ -21,9 +21,13 @@ const SCHEMA = `
     parent_span_id TEXT,
     issue_identifier TEXT NOT NULL,
     issue_id TEXT,
+    issue_title TEXT,
+    issue_labels TEXT,
     branch TEXT,
     outcome_kind TEXT,
     outcome_detail TEXT,
+    pr_url TEXT,
+    hitl_reason TEXT,
     start_time_unix_nano TEXT NOT NULL,
     end_time_unix_nano TEXT NOT NULL,
     status_code INTEGER,
@@ -38,6 +42,9 @@ const SCHEMA = `
   CREATE INDEX IF NOT EXISTS idx_issue_processes_trace_id
     ON issue_processes(trace_id);
 
+  CREATE INDEX IF NOT EXISTS idx_issue_processes_span_id
+    ON issue_processes(span_id);
+
   CREATE TABLE IF NOT EXISTS raw_spans (
     trace_id TEXT NOT NULL,
     span_id TEXT NOT NULL,
@@ -63,6 +70,128 @@ const SCHEMA = `
   CREATE INDEX IF NOT EXISTS idx_agent_iterations_issue_process
     ON agent_iterations(trace_id, issue_process_id, iteration_index);
 `;
+const DEFAULT_AGGREGATE_WINDOW = 30;
+/**
+ * VA-399: SQL VIEW that computes the evaluator-facing aggregates over
+ * the last N drains. N is interpolated at view-creation time because
+ * SQLite views can't take parameters — when the dashboard process
+ * starts with a different `DASHBOARD_AGGREGATE_WINDOW`, the view is
+ * dropped and recreated with the new LIMIT.
+ *
+ * Median uses the "average of the two middle values when N is even,
+ * the middle value when N is odd" convention; p95 uses the
+ * nearest-rank method (smallest observed value whose rank meets-or-
+ * exceeds 95%). See `read-model.md` for the field-by-field contract.
+ *
+ * `reviewer_rejection_rate` keys on the detail prefix emitted by
+ * `src/review.ts` ("Sub-agent review rejected: ..."). It's a subset
+ * of `hitl_escape_rate` — a review rejection routes to HITL, so both
+ * rates count the same row.
+ */
+function aggregatesViewDdl(windowDrains) {
+    // windowDrains is the only spot we interpolate rather than
+    // parameter-bind (CREATE VIEW can't take params). Coerce to a
+    // positive integer so a hostile env var can't smuggle SQL through.
+    const n = Math.max(1, Math.floor(windowDrains));
+    return `
+    DROP VIEW IF EXISTS evaluator_aggregates_v1;
+    CREATE VIEW evaluator_aggregates_v1 AS
+    WITH recent_drains AS (
+      SELECT trace_id
+      FROM drains
+      ORDER BY CAST(start_time_unix_nano AS INTEGER) DESC
+      LIMIT ${n}
+    ),
+    process_rows AS (
+      SELECT
+        ip.trace_id,
+        ip.span_id,
+        ip.outcome_kind,
+        COALESCE(ip.outcome_detail, '') AS outcome_detail,
+        CASE
+          WHEN instr(ip.issue_identifier, '-') > 0
+            THEN substr(ip.issue_identifier, 1, instr(ip.issue_identifier, '-') - 1)
+          ELSE ip.issue_identifier
+        END AS category,
+        (CAST(ip.end_time_unix_nano AS INTEGER) - CAST(ip.start_time_unix_nano AS INTEGER)) / 1000000 AS wall_time_ms,
+        (
+          SELECT COUNT(*) FROM agent_iterations a
+          WHERE a.trace_id = ip.trace_id AND a.issue_process_id = ip.span_id
+        ) AS iteration_count
+      FROM issue_processes ip
+      WHERE ip.trace_id IN (SELECT trace_id FROM recent_drains)
+    ),
+    wt_ranked AS (
+      SELECT
+        category,
+        wall_time_ms,
+        ROW_NUMBER() OVER (PARTITION BY category ORDER BY wall_time_ms) AS rn,
+        COUNT(*) OVER (PARTITION BY category) AS cnt
+      FROM process_rows
+    ),
+    it_ranked AS (
+      SELECT
+        category,
+        iteration_count,
+        ROW_NUMBER() OVER (PARTITION BY category ORDER BY iteration_count) AS rn,
+        COUNT(*) OVER (PARTITION BY category) AS cnt
+      FROM process_rows
+    ),
+    wt_median AS (
+      SELECT category, AVG(wall_time_ms * 1.0) AS value
+      FROM wt_ranked
+      WHERE rn IN ((cnt + 1) / 2, (cnt / 2) + 1)
+      GROUP BY category
+    ),
+    wt_p95 AS (
+      SELECT category, MIN(wall_time_ms) AS value
+      FROM wt_ranked
+      WHERE rn >= (cnt * 95 + 99) / 100
+      GROUP BY category
+    ),
+    it_median AS (
+      SELECT category, AVG(iteration_count * 1.0) AS value
+      FROM it_ranked
+      WHERE rn IN ((cnt + 1) / 2, (cnt / 2) + 1)
+      GROUP BY category
+    ),
+    it_p95 AS (
+      SELECT category, MIN(iteration_count) AS value
+      FROM it_ranked
+      WHERE rn >= (cnt * 95 + 99) / 100
+      GROUP BY category
+    ),
+    rates AS (
+      SELECT
+        category,
+        COUNT(*) AS sample_size,
+        AVG(CASE WHEN outcome_kind = 'hitl' AND outcome_detail LIKE 'Sub-agent review rejected%'
+                 THEN 1.0 ELSE 0.0 END) AS reviewer_rejection_rate,
+        AVG(CASE WHEN outcome_kind = 'reverted' THEN 1.0 ELSE 0.0 END) AS revert_rate,
+        AVG(CASE WHEN outcome_kind = 'hitl' THEN 1.0 ELSE 0.0 END) AS hitl_escape_rate,
+        AVG(CASE WHEN outcome_kind = 'errored' THEN 1.0 ELSE 0.0 END) AS infra_error_rate
+      FROM process_rows
+      GROUP BY category
+    )
+    SELECT
+      r.category                AS category,
+      r.sample_size             AS sample_size,
+      itm.value                 AS median_iteration_count,
+      itp.value                 AS p95_iteration_count,
+      wtm.value                 AS median_wall_time_ms,
+      wtp.value                 AS p95_wall_time_ms,
+      r.reviewer_rejection_rate AS reviewer_rejection_rate,
+      r.revert_rate             AS revert_rate,
+      r.hitl_escape_rate        AS hitl_escape_rate,
+      r.infra_error_rate        AS infra_error_rate
+    FROM rates r
+    LEFT JOIN wt_median wtm ON wtm.category = r.category
+    LEFT JOIN wt_p95    wtp ON wtp.category = r.category
+    LEFT JOIN it_median itm ON itm.category = r.category
+    LEFT JOIN it_p95    itp ON itp.category = r.category
+    ORDER BY r.category;
+  `;
+}
 /**
  * Open (or create) a SQLite database at `path` and return a typed
  * `Storage` handle. Pass `:memory:` for tests — the in-memory db
@@ -72,9 +201,35 @@ const SCHEMA = `
  * OTel SDK retrying a flush) don't blow up the receiver — last writer
  * wins on (trace_id, span_id).
  */
-export function createStorage(path) {
+export function createStorage(path, opts = {}) {
     const db = new DatabaseSync(path);
     db.exec(SCHEMA);
+    // VA-387: idempotent column adds for DBs created against an older
+    // schema. `CREATE TABLE IF NOT EXISTS` won't migrate an existing
+    // table; SQLite has no `ADD COLUMN IF NOT EXISTS`, so we swallow
+    // the duplicate-column error individually. Runs BEFORE VA-399's
+    // view install — `evaluator_aggregates_v1` reads from
+    // `issue_processes`, so the columns it may query must exist first.
+    for (const sql of [
+        `ALTER TABLE issue_processes ADD COLUMN issue_title TEXT`,
+        `ALTER TABLE issue_processes ADD COLUMN issue_labels TEXT`,
+        `ALTER TABLE issue_processes ADD COLUMN pr_url TEXT`,
+        `ALTER TABLE issue_processes ADD COLUMN hitl_reason TEXT`,
+    ]) {
+        try {
+            db.exec(sql);
+        }
+        catch {
+            // Column already present — fresh CREATE TABLE path, or a prior
+            // dashboard boot ran the same migration.
+        }
+    }
+    // VA-399: install the evaluator-facing read-model view after the
+    // base tables exist (and after VA-387's column migrations above),
+    // but before any prepared statement is created — a
+    // `SELECT FROM evaluator_aggregates_v1` would otherwise race the
+    // DDL on first use.
+    db.exec(aggregatesViewDdl(opts.aggregateWindowDrains ?? DEFAULT_AGGREGATE_WINDOW));
     const insertDrain = db.prepare(`
     INSERT INTO drains (
       trace_id, span_id, start_time_unix_nano, end_time_unix_nano,
@@ -93,16 +248,21 @@ export function createStorage(path) {
     const insertIssueProcess = db.prepare(`
     INSERT INTO issue_processes (
       trace_id, span_id, parent_span_id, issue_identifier, issue_id,
-      branch, outcome_kind, outcome_detail,
+      issue_title, issue_labels, branch, outcome_kind, outcome_detail,
+      pr_url, hitl_reason,
       start_time_unix_nano, end_time_unix_nano, status_code, status_message
-    ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+    ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
     ON CONFLICT (trace_id, span_id) DO UPDATE SET
       parent_span_id = excluded.parent_span_id,
       issue_identifier = excluded.issue_identifier,
       issue_id = excluded.issue_id,
+      issue_title = excluded.issue_title,
+      issue_labels = excluded.issue_labels,
       branch = excluded.branch,
       outcome_kind = excluded.outcome_kind,
       outcome_detail = excluded.outcome_detail,
+      pr_url = excluded.pr_url,
+      hitl_reason = excluded.hitl_reason,
       start_time_unix_nano = excluded.start_time_unix_nano,
       end_time_unix_nano = excluded.end_time_unix_nano,
       status_code = excluded.status_code,
@@ -130,35 +290,40 @@ export function createStorage(path) {
   `);
     // Two list variants instead of one with conditional SQL — keeps
     // each prepared statement static.
-    const listAll = db.prepare(`
-    SELECT
+    const ISSUE_PROCESS_COLUMNS = `
       trace_id, span_id, parent_span_id, issue_identifier, issue_id,
-      branch, outcome_kind, outcome_detail,
+      issue_title, issue_labels, branch, outcome_kind, outcome_detail,
+      pr_url, hitl_reason,
       start_time_unix_nano, end_time_unix_nano, status_code, status_message,
       inserted_at
+  `;
+    const listAll = db.prepare(`
+    SELECT ${ISSUE_PROCESS_COLUMNS}
     FROM issue_processes
     ORDER BY inserted_at DESC, span_id DESC
     LIMIT ?
   `);
     const listByTrace = db.prepare(`
-    SELECT
-      trace_id, span_id, parent_span_id, issue_identifier, issue_id,
-      branch, outcome_kind, outcome_detail,
-      start_time_unix_nano, end_time_unix_nano, status_code, status_message,
-      inserted_at
+    SELECT ${ISSUE_PROCESS_COLUMNS}
     FROM issue_processes
     WHERE trace_id = ?
     ORDER BY inserted_at DESC, span_id DESC
     LIMIT ?
   `);
     const getProcessStmt = db.prepare(`
-    SELECT
-      trace_id, span_id, parent_span_id, issue_identifier, issue_id,
-      branch, outcome_kind, outcome_detail,
-      start_time_unix_nano, end_time_unix_nano, status_code, status_message,
-      inserted_at
+    SELECT ${ISSUE_PROCESS_COLUMNS}
     FROM issue_processes
     WHERE trace_id = ? AND span_id = ?
+  `);
+    // VA-387: span_id is unique in practice (random 64-bit ids); the
+    // detail route at `/issue-processes/:id` keys on span_id alone so
+    // operators don't have to type the trace_id in URLs.
+    const getProcessBySpanStmt = db.prepare(`
+    SELECT ${ISSUE_PROCESS_COLUMNS}
+    FROM issue_processes
+    WHERE span_id = ?
+    ORDER BY inserted_at DESC
+    LIMIT 1
   `);
     const listIterations = db.prepare(`
     SELECT
@@ -169,11 +334,16 @@ export function createStorage(path) {
     WHERE trace_id = ? AND issue_process_id = ?
     ORDER BY iteration_index ASC
   `);
+    const selectAggregates = db.prepare(`SELECT * FROM evaluator_aggregates_v1`);
     const saveDrain = (d) => {
         insertDrain.run(d.traceId, d.spanId, d.startTimeUnixNano, d.endTimeUnixNano, asInt(d.attempts), asInt(d.opened), asInt(d.hitl), asInt(d.errored), asInt(d.statusCode), d.statusMessage);
     };
     const saveIssueProcess = (p) => {
-        insertIssueProcess.run(p.traceId, p.spanId, p.parentSpanId, p.issueIdentifier, p.issueId, p.branch, p.outcomeKind, p.outcomeDetail, p.startTimeUnixNano, p.endTimeUnixNano, asInt(p.statusCode), p.statusMessage);
+        insertIssueProcess.run(p.traceId, p.spanId, p.parentSpanId, p.issueIdentifier, p.issueId, p.issueTitle, 
+        // VA-387: labels round-trip as a JSON array string. Keeping them
+        // in one column avoids a label-many-to-many table for a feature
+        // that's read-only on the dashboard side.
+        p.issueLabels.length === 0 ? null : JSON.stringify(p.issueLabels), p.branch, p.outcomeKind, p.outcomeDetail, p.prUrl, p.hitlReason, p.startTimeUnixNano, p.endTimeUnixNano, asInt(p.statusCode), p.statusMessage);
     };
     const saveAgentIteration = (a) => {
         insertAgentIteration.run(a.traceId, a.spanId, a.issueProcessSpanId, asInt(a.iterationIndex), a.startTimeUnixNano, a.endTimeUnixNano, a.sandcastleRunId, a.exitStatus);
@@ -192,6 +362,10 @@ export function createStorage(path) {
         const row = getProcessStmt.get(traceId, spanId);
         return row ? rowToIssueProcess(row) : undefined;
     };
+    const getIssueProcessBySpanId = (spanId) => {
+        const row = getProcessBySpanStmt.get(spanId);
+        return row ? rowToIssueProcess(row) : undefined;
+    };
     const listAgentIterations = (traceId, issueProcessSpanId) => {
         const rows = listIterations.all(traceId, issueProcessSpanId);
         return rows.map(rowToAgentIteration);
@@ -227,6 +401,7 @@ export function createStorage(path) {
             .all(traceId, issueProcessSpanId, ...names);
         return rows.map(rowToPhaseSpan);
     };
+    const listAggregates = () => selectAggregates.all().map(rowToAggregate);
     const close = () => {
         db.close();
     };
@@ -237,8 +412,10 @@ export function createStorage(path) {
         saveRawSpan,
         listIssueProcesses,
         getIssueProcess,
+        getIssueProcessBySpanId,
         listAgentIterations,
         listPhaseSpans,
+        listAggregates,
         close,
     };
 }
@@ -256,9 +433,13 @@ function rowToIssueProcess(row) {
         parentSpanId: nullableStr(r.parent_span_id),
         issueIdentifier: String(r.issue_identifier ?? ""),
         issueId: nullableStr(r.issue_id),
+        issueTitle: nullableStr(r.issue_title),
+        issueLabels: parseLabels(r.issue_labels),
         branch: nullableStr(r.branch),
         outcomeKind: nullableStr(r.outcome_kind),
         outcomeDetail: nullableStr(r.outcome_detail),
+        prUrl: nullableStr(r.pr_url),
+        hitlReason: nullableStr(r.hitl_reason),
         startTimeUnixNano: String(r.start_time_unix_nano ?? ""),
         endTimeUnixNano: String(r.end_time_unix_nano ?? ""),
         statusCode: nullableNum(r.status_code),
@@ -266,6 +447,26 @@ function rowToIssueProcess(row) {
         insertedAt: String(r.inserted_at ?? ""),
     };
 }
+/**
+ * VA-387: decode the JSON-encoded `issue_labels` column back into a
+ * string array. A row stored before the column existed (or one with
+ * NULL / malformed JSON) collapses to an empty list.
+ */
+function parseLabels(v) {
+    if (v === null || v === undefined)
+        return [];
+    if (typeof v !== "string")
+        return [];
+    try {
+        const parsed = JSON.parse(v);
+        if (!Array.isArray(parsed))
+            return [];
+        return parsed.filter((x) => typeof x === "string");
+    }
+    catch {
+        return [];
+    }
+}
 function rowToAgentIteration(row) {
     const r = row;
     return {
@@ -302,3 +503,18 @@ function nullableNum(v) {
     const n = Number(v);
     return Number.isFinite(n) ? n : null;
 }
+function rowToAggregate(row) {
+    const r = row;
+    return {
+        category: String(r.category ?? ""),
+        sampleSize: Number(r.sample_size ?? 0),
+        medianIterationCount: nullableNum(r.median_iteration_count),
+        p95IterationCount: nullableNum(r.p95_iteration_count),
+        medianWallTimeMs: nullableNum(r.median_wall_time_ms),
+        p95WallTimeMs: nullableNum(r.p95_wall_time_ms),
+        reviewerRejectionRate: Number(r.reviewer_rejection_rate ?? 0),
+        revertRate: Number(r.revert_rate ?? 0),
+        hitlEscapeRate: Number(r.hitl_escape_rate ?? 0),
+        infraErrorRate: Number(r.infra_error_rate ?? 0),
+    };
+}

package/dist/dashboard/views.js

@@ -60,7 +60,8 @@ export function renderListView(rows) {
 function renderRow(r) {
     const kind = r.outcomeKind ?? "pending";
     const outcomeCls = `outcome outcome-${escapeHtml(kind)}`;
-    const href = `/issue/${encodeURIComponent(r.traceId)}/${encodeURIComponent(r.spanId)}`;
+    // VA-387: canonical detail link uses the span_id alone.
+    const href = `/issue-processes/${encodeURIComponent(r.spanId)}`;
     return `<tr>
     <td class="id"><a href="${escapeHtml(href)}">${escapeHtml(r.issueIdentifier)}</a></td>
     <td class="${outcomeCls}">${escapeHtml(kind)}</td>
@@ -83,6 +84,18 @@ export function renderDetailView(vm) {
     const ip = vm.issueProcess;
     const kind = ip.outcomeKind ?? "pending";
     const outcomeCls = `outcome outcome-${escapeHtml(kind)}`;
+    const titleLine = ip.issueTitle
+        ? `<div class="title">${escapeHtml(ip.issueTitle)}</div>`
+        : "";
+    const prLine = ip.prUrl
+        ? `<div><span class="label">PR:</span><a href="${escapeHtml(ip.prUrl)}" rel="noopener noreferrer" target="_blank">${escapeHtml(ip.prUrl)}</a></div>`
+        : "";
+    // HITL reason is only rendered when present (HITL or errored runs).
+    // For opened/reverted outcomes we omit the row entirely rather than
+    // showing an empty label.
+    const hitlLine = ip.hitlReason
+        ? `<div><span class="label">HITL reason:</span><span class="detail">${escapeHtml(ip.hitlReason)}</span></div>`
+        : "";
     return `<!doctype html>
 <html lang="en">
 <head>
@@ -90,6 +103,7 @@ export function renderDetailView(vm) {
   <title>${escapeHtml(ip.issueIdentifier)} · runway dashboard</title>
   <style>${SHARED_STYLE}
     .breadcrumb { color: #9ca3af; margin-bottom: 16px; font-size: 12px; }
+    .title { font-size: 14px; color: #d4d4d8; margin: -8px 0 12px; }
     .meta { margin: 4px 0 16px; }
     .meta .label { color: #9ca3af; margin-right: 4px; }
     .timeline { position: relative; height: 28px; background: #18181b;
@@ -113,8 +127,11 @@ export function renderDetailView(vm) {
 <body>
   <div class="breadcrumb"><a href="/">← all issue processes</a></div>
   <h1>${escapeHtml(ip.issueIdentifier)} · <span class="${outcomeCls}">${escapeHtml(kind)}</span></h1>
+  ${titleLine}
   <div class="meta">
     <div><span class="label">branch:</span><code>${escapeHtml(ip.branch ?? "—")}</code></div>
+    ${prLine}
+    ${hitlLine}
     <div><span class="label">detail:</span><span class="detail">${escapeHtml(ip.outcomeDetail ?? "")}</span></div>
     <div><span class="label">seen at:</span>${escapeHtml(ip.insertedAt)}</div>
   </div>

package/dist/finalize.js

@@ -1,10 +1,23 @@
 import { Effect } from "effect";
+import { rebaseOntoBase } from "./git.js";
 /**
- * Push the agent branch, open the PR, transition the Linear issue to
- * the in-review status, and link the PR back on the issue.
+ * VA-419: rebase the agent branch onto the latest `origin/<baseBranch>`,
+ * then push and open the PR. If the rebase hits a conflict, restore
+ * pre-rebase state and surface a `rebase-conflict` outcome — the
+ * orchestrator routes it to HITL so the operator can reconcile
+ * manually (or let VA-417 reset the branch on the next drain).
  */
 export const finalize = (issue, deps, branch) => Effect.gen(function* () {
     const { config, cwd, baseBranch, linear, github } = deps;
+    const rebase = yield* rebaseOntoBase(cwd, baseBranch, branch).pipe(Effect.withSpan("rebaseOntoBase"));
+    if (rebase.kind === "conflict") {
+        return {
+            kind: "rebase-conflict",
+            baseBranch,
+            conflictedFiles: rebase.conflictedFiles,
+            reviewVerdict: "REVIEW: APPROVED",
+        };
+    }
     yield* github.pushBranch(cwd, branch).pipe(Effect.withSpan("pushBranch"));
     const prBody = buildPrBody(issue);
     const prUrl = yield* github
@@ -20,6 +33,25 @@ export const finalize = (issue, deps, branch) => Effect.gen(function* () {
     yield* linear.comment(issue.id, `Runway opened a PR for review: ${prUrl}`);
     return { kind: "opened", detail: prUrl };
 });
+/**
+ * VA-419: format the HITL message body for a rebase-conflict outcome.
+ * Names the base branch and the conflicted file list, and quotes the
+ * reviewer's APPROVED verdict so the operator knows the agent's diff
+ * was good before the conflict surfaced.
+ */
+export function formatRebaseConflictReason(args) {
+    const fileLines = args.conflictedFiles.length
+        ? args.conflictedFiles.map((f) => `  - ${f}`).join("\n")
+        : "  (no conflicted files reported by git — inspect manually)";
+    return [
+        `Upstream base \`${args.baseBranch}\` advanced during this drain; the rebase`,
+        "onto the latest base produced conflicts in:",
+        fileLines,
+        `Review was APPROVED before the rebase (\`${args.reviewVerdict}\`); the agent's diff`,
+        "is good but needs operator reconciliation against the new base. Re-run runway",
+        "after rebasing manually, or let VA-417 handle it on the next drain.",
+    ].join("\n");
+}
 // VA-412: `Closes` (not `Refs`) is the Linear GitHub-integration magic
 // word that auto-transitions the issue to Done on PR merge. `Refs`
 // only attaches the PR to the issue and leaves it stuck In Progress.