feat(results): publish selected local runs

Bead: av-hbv.5

Author: christso

apps/cli/src/commands/results/remote.ts

@@ -6,9 +6,11 @@ import {
   type EvaluationResult,
   type GitListedRun,
   type ResultsConfig,
+  ResultsRepoRunExistsError,
   type ResultsRepoStatus,
   directPushResults,
   directorySizeBytes,
+  findResultsRepoRun,
   getProject,
   getProjectForPath,
   getResultsRepoSyncStatus,
@@ -27,6 +29,7 @@ import {
   listResultFiles,
   listResultFilesFromRunsDir,
 } from '../inspect/utils.js';
+import { loadManifestResults } from './manifest.js';
 import {
   type RemoteRunTagState,
   assertWritableResultsRepo,
@@ -97,6 +100,33 @@ export interface RemoteResultsStatus extends ResultsRepoStatus {
   readonly run_count: number;
 }
 
+export interface LocalRunPublishPreview {
+  readonly sourceRunId: string;
+  readonly targetRepo: string;
+  readonly targetPath: string;
+  readonly targetRunId: string;
+  readonly remoteExists: boolean;
+  readonly replaceRequired: boolean;
+  readonly canPublish: boolean;
+  readonly blockReason?: string;
+  readonly remoteStatus: RemoteResultsStatus;
+}
+
+export interface LocalRunPublishResult extends LocalRunPublishPreview {
+  readonly published: boolean;
+  readonly replaced: boolean;
+}
+
+export class LocalRunPublishError extends Error {
+  readonly status: 400 | 409;
+
+  constructor(message: string, status: 400 | 409 = 400) {
+    super(message);
+    this.name = 'LocalRunPublishError';
+    this.status = status;
+  }
+}
+
 const REMOTE_RUN_PREFIX = 'remote::';
 const SIZE_WARNING_BYTES = 10 * 1024 * 1024;
 
@@ -122,6 +152,136 @@ function getRelativeRunPath(cwd: string, runDir: string): string {
   return experiment && experiment !== runName ? path.join(experiment, runName) : runName;
 }
 
+function toPosixPath(p: string): string {
+  return p.split(path.sep).join('/');
+}
+
+function getPublishStatusBlockReason(status: RemoteResultsStatus): string | undefined {
+  if (!status.configured) {
+    return 'Remote results repo is not configured for this project.';
+  }
+
+  if (status.sync_status === 'syncing') {
+    return 'Project sync is already in progress. Wait for it to finish before publishing a run.';
+  }
+
+  if (status.sync_status && status.sync_status !== 'clean') {
+    return 'Sync Project before publishing a selected run so pending result metadata is handled first.';
+  }
+
+  if (status.blocked) {
+    return status.block_reason ?? 'Remote results repo sync is blocked.';
+  }
+
+  return undefined;
+}
+
+function getExperimentFromRunId(runId: string): string {
+  const separatorIndex = runId.lastIndexOf('::');
+  return separatorIndex === -1 ? 'default' : runId.slice(0, separatorIndex);
+}
+
+async function buildLocalRunPublishPreview(params: {
+  readonly cwd: string;
+  readonly config: Required<ResultsConfig>;
+  readonly meta: Pick<SourcedResultFileMeta, 'source' | 'path' | 'raw_filename'>;
+  readonly projectId?: string;
+}): Promise<LocalRunPublishPreview & { readonly relativeRunPath: string }> {
+  if (params.meta.source !== 'local') {
+    throw new LocalRunPublishError('Selected run publish is only available for local runs');
+  }
+
+  const runDir = path.dirname(params.meta.path);
+  const relativeRunPath = toPosixPath(getRelativeRunPath(params.cwd, runDir));
+  const targetPath = path.posix.join('.agentv/results/runs', relativeRunPath);
+  const existingRun = await findResultsRepoRun(params.config, params.meta.raw_filename);
+  const remoteStatus = await getRemoteResultsStatus(params.cwd, params.projectId);
+  const blockReason = getPublishStatusBlockReason(remoteStatus);
+  const remoteExists = existingRun !== undefined;
+
+  return {
+    sourceRunId: params.meta.raw_filename,
+    targetRepo: params.config.repo,
+    targetPath,
+    targetRunId: params.meta.raw_filename,
+    remoteExists,
+    replaceRequired: remoteExists,
+    canPublish: !blockReason && !remoteExists,
+    ...(blockReason && { blockReason }),
+    remoteStatus,
+    relativeRunPath,
+  };
+}
+
+export async function previewLocalRunPublish(
+  cwd: string,
+  meta: Pick<SourcedResultFileMeta, 'source' | 'path' | 'raw_filename'>,
+  projectId?: string,
+): Promise<LocalRunPublishPreview> {
+  const config = await loadNormalizedResultsConfig(cwd, projectId);
+  if (!config) {
+    throw new LocalRunPublishError('Remote results repo is not configured for this project', 409);
+  }
+
+  const { relativeRunPath: _relativeRunPath, ...preview } = await buildLocalRunPublishPreview({
+    cwd,
+    config,
+    meta,
+    projectId,
+  });
+  return preview;
+}
+
+export async function publishLocalRun(params: {
+  readonly cwd: string;
+  readonly meta: Pick<SourcedResultFileMeta, 'source' | 'path' | 'raw_filename'>;
+  readonly projectId?: string;
+  readonly replace?: boolean;
+}): Promise<LocalRunPublishResult> {
+  const config = await loadNormalizedResultsConfig(params.cwd, params.projectId);
+  if (!config) {
+    throw new LocalRunPublishError('Remote results repo is not configured for this project', 409);
+  }
+
+  const preview = await buildLocalRunPublishPreview({
+    cwd: params.cwd,
+    config,
+    meta: params.meta,
+    projectId: params.projectId,
+  });
+  if (preview.blockReason) {
+    throw new LocalRunPublishError(preview.blockReason, 409);
+  }
+  if (preview.remoteExists && params.replace !== true) {
+    throw new ResultsRepoRunExistsError(preview.targetRunId, preview.relativeRunPath);
+  }
+
+  await maybeWarnLargeArtifact(path.dirname(params.meta.path));
+  const results = loadManifestResults(params.meta.path);
+  const pushed = await directPushResults({
+    config,
+    sourceDir: path.dirname(params.meta.path),
+    destinationPath: preview.relativeRunPath,
+    commitMessage: buildCommitTitle({
+      cwd: params.cwd,
+      run_dir: path.dirname(params.meta.path),
+      test_files: [],
+      results,
+      eval_summaries: [],
+      experiment: getExperimentFromRunId(preview.targetRunId),
+    }),
+    replaceExisting: params.replace === true,
+  });
+
+  invalidateGitRunsCache(config.path);
+  const refreshed = await previewLocalRunPublish(params.cwd, params.meta, params.projectId);
+  return {
+    ...refreshed,
+    published: pushed,
+    replaced: preview.remoteExists && params.replace === true,
+  };
+}
+
 function buildCommitTitle(payload: RemoteExportPayload): string {
   const passed = payload.results.filter((result) => result.score >= DEFAULT_THRESHOLD).length;
   const avgScore =

apps/cli/src/commands/results/serve.ts

@@ -40,6 +40,7 @@ import { command, flag, number, option, optional, positional, string } from 'cmd
 import {
   DEFAULT_CATEGORY,
   type EvaluationResult,
+  ResultsRepoRunExistsError,
   addProject,
   getProject,
   loadConfig,
@@ -72,12 +73,16 @@ import {
   resolveResultSourcePath,
 } from './manifest.js';
 import {
+  LocalRunPublishError,
+  type LocalRunPublishPreview,
   type SourcedResultFileMeta,
   clearRemoteRunTags,
   ensureRemoteRunAvailable,
   findRunById,
   getRemoteResultsStatus,
   listMergedResultFiles,
+  previewLocalRunPublish,
+  publishLocalRun,
   readRemoteRunTagState,
   setRemoteRunTags,
   syncRemoteResults,
@@ -362,6 +367,34 @@ function remoteMetadataErrorStatus(error: unknown): 400 | 409 {
   return 400;
 }
 
+function localRunPublishToWire(
+  preview: LocalRunPublishPreview & { published?: boolean; replaced?: boolean },
+) {
+  return {
+    source_run_id: preview.sourceRunId,
+    target_repo: preview.targetRepo,
+    target_path: preview.targetPath,
+    target_run_id: preview.targetRunId,
+    remote_exists: preview.remoteExists,
+    replace_required: preview.replaceRequired,
+    can_publish: preview.canPublish,
+    ...(preview.blockReason && { block_reason: preview.blockReason }),
+    remote_status: preview.remoteStatus,
+    ...(preview.published !== undefined && { published: preview.published }),
+    ...(preview.replaced !== undefined && { replaced: preview.replaced }),
+  };
+}
+
+function localRunPublishErrorStatus(error: unknown): 400 | 409 | 500 {
+  if (error instanceof LocalRunPublishError) {
+    return error.status;
+  }
+  if (error instanceof ResultsRepoRunExistsError) {
+    return 409;
+  }
+  return 500;
+}
+
 async function ensureRunReadable(
   searchDir: string,
   meta: SourcedResultFileMeta,
@@ -1193,9 +1226,10 @@ function getLocalRunsRoot(searchDir: string): string {
 function validateLocalCompletedRun(
   searchDir: string,
   meta: SourcedResultFileMeta,
+  actionName = 'Run combine',
 ): { ok: true } | { error: string; status: 400 | 409 } {
   if (meta.source === 'remote') {
-    return { error: 'Run combine is only available for local runs', status: 400 };
+    return { error: `${actionName} is only available for local runs`, status: 400 };
   }
   if (getActiveRunStatus(meta.path) === 'starting' || getActiveRunStatus(meta.path) === 'running') {
     return { error: 'Run is still active', status: 409 };
@@ -1214,6 +1248,52 @@ function validateLocalCompletedRun(
   return { error: 'Run workspace is outside the local results directory', status: 400 };
 }
 
+async function handleRunPublishPreview(c: C, { searchDir, projectId }: DataContext) {
+  const filename = c.req.param('filename') ?? '';
+  const meta = await findRunById(searchDir, filename, projectId);
+  if (!meta) return c.json({ error: 'Run not found' }, 404);
+  const safe = validateLocalCompletedRun(searchDir, meta, 'Selected run publish');
+  if ('error' in safe) return c.json({ error: safe.error }, safe.status);
+
+  try {
+    const preview = await previewLocalRunPublish(searchDir, meta, projectId);
+    return c.json(localRunPublishToWire(preview));
+  } catch (err) {
+    return c.json({ error: (err as Error).message }, localRunPublishErrorStatus(err));
+  }
+}
+
+async function handleRunPublish(c: C, { searchDir, projectId }: DataContext) {
+  const filename = c.req.param('filename') ?? '';
+  const meta = await findRunById(searchDir, filename, projectId);
+  if (!meta) return c.json({ error: 'Run not found' }, 404);
+  const safe = validateLocalCompletedRun(searchDir, meta, 'Selected run publish');
+  if ('error' in safe) return c.json({ error: safe.error }, safe.status);
+
+  let replace = false;
+  try {
+    const body = (await c.req.json()) as unknown;
+    if (body && typeof body === 'object' && !Array.isArray(body)) {
+      const value = (body as Record<string, unknown>).replace;
+      if (value !== undefined && typeof value !== 'boolean') {
+        return c.json({ error: 'replace must be a boolean' }, 400);
+      }
+      replace = value === true;
+    } else if (body !== undefined) {
+      return c.json({ error: 'Invalid payload' }, 400);
+    }
+  } catch {
+    return c.json({ error: 'Invalid JSON' }, 400);
+  }
+
+  try {
+    const result = await publishLocalRun({ cwd: searchDir, meta, projectId, replace });
+    return c.json(localRunPublishToWire(result));
+  } catch (err) {
+    return c.json({ error: (err as Error).message }, localRunPublishErrorStatus(err));
+  }
+}
+
 async function handleRunsCombine(c: C, { searchDir, projectId }: DataContext) {
   let body: unknown;
   try {
@@ -1538,6 +1618,13 @@ export function createApp(
     }
     return handleRunsCombine(c, defaultCtx);
   });
+  app.get('/api/runs/:filename/publish', (c) => handleRunPublishPreview(c, defaultCtx));
+  app.post('/api/runs/:filename/publish', (c) => {
+    if (readOnly) {
+      return c.json({ error: 'Dashboard is running in read-only mode' }, 403);
+    }
+    return handleRunPublish(c, defaultCtx);
+  });
   app.put('/api/runs/:filename/tags', (c) => {
     if (readOnly) {
       return c.json({ error: 'Dashboard is running in read-only mode' }, 403);
@@ -1679,6 +1766,15 @@ export function createApp(
     }
     return withProject(c, handleRunsCombine);
   });
+  app.get('/api/projects/:projectId/runs/:filename/publish', (c) =>
+    withProject(c, handleRunPublishPreview),
+  );
+  app.post('/api/projects/:projectId/runs/:filename/publish', (c) => {
+    if (readOnly) {
+      return c.json({ error: 'Dashboard is running in read-only mode' }, 403);
+    }
+    return withProject(c, handleRunPublish);
+  });
   app.put('/api/projects/:projectId/runs/:filename/tags', (c) => {
     if (readOnly) {
       return c.json({ error: 'Dashboard is running in read-only mode' }, 403);