feat: add GitHub step evidence presets

runners/agent-runner/src/host-evidence.ts

@@ -12,6 +12,8 @@ export const HOST_EVIDENCE_ACTIONS = [
   'github_pr_status',
   'github_pr_diff_stat',
   'github_run_status',
+  'github_run_jobs',
+  'github_workflow_file',
 ] as const;
 
 export const DB_EVIDENCE_ACTIONS = [
@@ -29,6 +31,8 @@ export const GITHUB_EVIDENCE_ACTIONS = [
   'github_pr_status',
   'github_pr_diff_stat',
   'github_run_status',
+  'github_run_jobs',
+  'github_workflow_file',
 ] as const;
 
 export const ARTIFACT_EVIDENCE_KINDS = [

runners/agent-runner/src/ipc-host-evidence-tool.ts

@@ -114,6 +114,16 @@ export function registerHostEvidenceTools(
         .describe('Only for GitHub evidence, in owner/repo form.'),
       pr_number: z.number().int().positive().optional(),
       run_id: z.number().int().positive().optional(),
+      workflow_path: z
+        .string()
+        .optional()
+        .describe(
+          'Only for github_workflow_file, e.g. .github/workflows/ci.yml.',
+        ),
+      ref: z
+        .string()
+        .optional()
+        .describe('Only for github_workflow_file; branch, tag, or commit SHA.'),
       artifact_kind: z.enum(ARTIFACT_EVIDENCE_KINDS).optional(),
     },
     async (args) =>
@@ -129,6 +139,8 @@ export function registerHostEvidenceTools(
         repo: args.repo,
         pr_number: args.pr_number,
         run_id: args.run_id,
+        workflow_path: args.workflow_path,
+        ref: args.ref,
         artifact_kind: args.artifact_kind,
       }),
   );
@@ -176,6 +188,16 @@ export function registerHostEvidenceTools(
       repo: z.string().describe('GitHub repository in owner/repo form.'),
       pr_number: z.number().int().positive().optional(),
       run_id: z.number().int().positive().optional(),
+      workflow_path: z
+        .string()
+        .optional()
+        .describe(
+          'Only for github_workflow_file, e.g. .github/workflows/ci.yml.',
+        ),
+      ref: z
+        .string()
+        .optional()
+        .describe('Only for github_workflow_file; branch, tag, or commit SHA.'),
     },
     async (args) =>
       requestHostEvidence(options, {
@@ -183,6 +205,8 @@ export function registerHostEvidenceTools(
         repo: args.repo,
         pr_number: args.pr_number,
         run_id: args.run_id,
+        workflow_path: args.workflow_path,
+        ref: args.ref,
       }),
   );
 }

src/github-evidence.test.ts

@@ -2,7 +2,10 @@ import { describe, expect, it } from 'vitest';
 
 import {
   buildGitHubEvidenceCommand,
+  decodeGitHubContentBase64,
+  normalizeGitHubRef,
   normalizeGitHubRepo,
+  normalizeGitHubWorkflowPath,
   normalizePositiveInteger,
 } from './github-evidence.js';
 
@@ -18,6 +21,35 @@ describe('GitHub evidence helpers', () => {
     );
   });
 
+  it('validates workflow evidence paths and refs', () => {
+    expect(
+      normalizeGitHubWorkflowPath('.github/workflows/display-analyzer.yml'),
+    ).toBe('.github/workflows/display-analyzer.yml');
+    expect(normalizeGitHubWorkflowPath('.github/workflows/ci.yaml')).toBe(
+      '.github/workflows/ci.yaml',
+    );
+    expect(() => normalizeGitHubWorkflowPath('README.md')).toThrow(
+      'Unsupported GitHub workflow path',
+    );
+    expect(() =>
+      normalizeGitHubWorkflowPath('.github/workflows/../ci.yml'),
+    ).toThrow('Unsupported GitHub workflow path');
+
+    expect(normalizeGitHubRef('prod')).toBe('prod');
+    expect(normalizeGitHubRef('feature/display-analyzer')).toBe(
+      'feature/display-analyzer',
+    );
+    expect(normalizeGitHubRef('00b971972e4b2e7ecf0c6d789f405fef7edeb258')).toBe(
+      '00b971972e4b2e7ecf0c6d789f405fef7edeb258',
+    );
+    expect(() => normalizeGitHubRef('../prod')).toThrow(
+      'Missing or invalid ref',
+    );
+    expect(() => normalizeGitHubRef('feature branch')).toThrow(
+      'Missing or invalid ref',
+    );
+  });
+
   it('builds fixed gh commands', () => {
     expect(
       buildGitHubEvidenceCommand({
@@ -42,5 +74,42 @@ describe('GitHub evidence helpers', () => {
         runId: 123,
       }).args,
     ).toContain('123');
+
+    expect(
+      buildGitHubEvidenceCommand({
+        action: 'github_run_jobs',
+        repo: 'owner/repo',
+        runId: 123,
+      }).args,
+    ).toEqual([
+      'run',
+      'view',
+      '123',
+      '--repo',
+      'owner/repo',
+      '--json',
+      'databaseId,name,status,conclusion,url,headBranch,headSha,jobs',
+    ]);
+
+    expect(
+      buildGitHubEvidenceCommand({
+        action: 'github_workflow_file',
+        repo: 'owner/repo',
+        workflowPath: '.github/workflows/display-analyzer-check.yml',
+        ref: 'feature/display-analyzer',
+      }),
+    ).toMatchObject({
+      args: [
+        'api',
+        'repos/owner/repo/contents/.github/workflows/display-analyzer-check.yml?ref=feature%2Fdisplay-analyzer',
+        '--jq',
+        '.content',
+      ],
+      decodeBase64Stdout: true,
+    });
+  });
+
+  it('decodes GitHub contents API workflow bodies', () => {
+    expect(decodeGitHubContentBase64('bmFtZTogQ0kK')).toBe('name: CI\n');
   });
 });

src/github-evidence.ts

@@ -4,6 +4,8 @@ export const GITHUB_EVIDENCE_ACTIONS = [
   'github_pr_status',
   'github_pr_diff_stat',
   'github_run_status',
+  'github_run_jobs',
+  'github_workflow_file',
 ] as const;
 
 export type GitHubEvidenceAction = (typeof GITHUB_EVIDENCE_ACTIONS)[number];
@@ -13,15 +15,20 @@ export interface GitHubEvidenceRequest {
   repo?: string;
   prNumber?: number;
   runId?: number;
+  workflowPath?: string;
+  ref?: string;
 }
 
 interface GitHubCommandSpec {
   file: string;
   args: string[];
   commandText: string;
+  decodeBase64Stdout?: boolean;
 }
 
 const REPO_PATTERN = /^[A-Za-z0-9_.-]+\/[A-Za-z0-9_.-]+$/;
+const WORKFLOW_PATH_PATTERN = /^\.github\/workflows\/[A-Za-z0-9_.-]+\.ya?ml$/;
+const REF_PATTERN = /^[A-Za-z0-9][A-Za-z0-9._/-]{0,255}$/;
 const COMMAND_TIMEOUT_MS = 10_000;
 const COMMAND_MAX_BUFFER = 2 * 1024 * 1024;
 const MAX_OUTPUT_CHARS = 24_000;
@@ -53,6 +60,32 @@ export function normalizePositiveInteger(
   return value;
 }
 
+export function normalizeGitHubWorkflowPath(value?: string): string {
+  const workflowPath = value?.trim();
+  if (!workflowPath || !WORKFLOW_PATH_PATTERN.test(workflowPath)) {
+    throw new Error(`Unsupported GitHub workflow path for evidence: ${value}`);
+  }
+  return workflowPath;
+}
+
+export function normalizeGitHubRef(value?: string): string {
+  const ref = value?.trim();
+  if (
+    !ref ||
+    ref.includes('..') ||
+    ref.includes('//') ||
+    ref.endsWith('/') ||
+    !REF_PATTERN.test(ref)
+  ) {
+    throw new Error(`Missing or invalid ref for GitHub evidence`);
+  }
+  return ref;
+}
+
+export function decodeGitHubContentBase64(value: string): string {
+  return Buffer.from(value.replace(/\s+/g, ''), 'base64').toString('utf8');
+}
+
 export function buildGitHubEvidenceCommand(
   request: GitHubEvidenceRequest,
 ): GitHubCommandSpec {
@@ -102,6 +135,35 @@ export function buildGitHubEvidenceCommand(
         commandText: `gh ${args.join(' ')}`,
       };
     }
+    case 'github_run_jobs': {
+      const runId = normalizePositiveInteger(request.runId, 'run_id');
+      const args = [
+        'run',
+        'view',
+        String(runId),
+        '--repo',
+        repo,
+        '--json',
+        'databaseId,name,status,conclusion,url,headBranch,headSha,jobs',
+      ];
+      return {
+        file: 'gh',
+        args,
+        commandText: `gh ${args.join(' ')}`,
+      };
+    }
+    case 'github_workflow_file': {
+      const workflowPath = normalizeGitHubWorkflowPath(request.workflowPath);
+      const ref = normalizeGitHubRef(request.ref);
+      const endpoint = `repos/${repo}/contents/${workflowPath}?ref=${encodeURIComponent(ref)}`;
+      const args = ['api', endpoint, '--jq', '.content'];
+      return {
+        file: 'gh',
+        args,
+        commandText: `gh ${args.join(' ')}`,
+        decodeBase64Stdout: true,
+      };
+    }
   }
 }
 
@@ -142,9 +204,12 @@ export function runGitHubEvidenceCommand(
           );
           return;
         }
+        const normalizedStdout = command.decodeBase64Stdout
+          ? decodeGitHubContentBase64(stdout)
+          : stdout;
         resolve({
           command: command.commandText,
-          stdout: truncateGitHubEvidenceText(stdout),
+          stdout: truncateGitHubEvidenceText(normalizedStdout),
           stderr: truncateGitHubEvidenceText(stderr),
         });
       },

src/host-evidence-ipc.test.ts

@@ -95,6 +95,8 @@ describe('host evidence IPC', () => {
       repo: undefined,
       prNumber: undefined,
       runId: undefined,
+      workflowPath: undefined,
+      ref: undefined,
       artifactKind: undefined,
       sourceGroup: 'other-group',
       isMain: false,

src/host-evidence.test.ts

@@ -13,6 +13,8 @@ describe('host evidence helpers', () => {
     expect(isHostEvidenceAction('db_paired_task_status')).toBe(true);
     expect(isHostEvidenceAction('ejclaw_deploy_state')).toBe(true);
     expect(isHostEvidenceAction('github_pr_status')).toBe(true);
+    expect(isHostEvidenceAction('github_run_jobs')).toBe(true);
+    expect(isHostEvidenceAction('github_workflow_file')).toBe(true);
     expect(isHostEvidenceAction('rm -rf /')).toBe(false);
   });
 

src/host-evidence.ts

@@ -44,6 +44,8 @@ export interface HostEvidenceRequest {
   repo?: string;
   prNumber?: number;
   runId?: number;
+  workflowPath?: string;
+  ref?: string;
   artifactKind?: string;
   sourceGroup?: string;
   isMain?: boolean;
@@ -282,6 +284,8 @@ export async function runHostEvidenceRequest(
         repo: request.repo,
         prNumber: request.prNumber,
         runId: request.runId,
+        workflowPath: request.workflowPath,
+        ref: request.ref,
       });
       commandText = githubResult.command;
       return {

src/ipc-task-processor/host-evidence.ts

@@ -54,6 +54,9 @@ export async function handleHostEvidenceRequest(
     repo: typeof data.repo === 'string' ? data.repo : undefined,
     prNumber: typeof data.pr_number === 'number' ? data.pr_number : undefined,
     runId: typeof data.run_id === 'number' ? data.run_id : undefined,
+    workflowPath:
+      typeof data.workflow_path === 'string' ? data.workflow_path : undefined,
+    ref: typeof data.ref === 'string' ? data.ref : undefined,
     artifactKind:
       typeof data.artifact_kind === 'string' ? data.artifact_kind : undefined,
     sourceGroup,

src/ipc-types.ts

@@ -153,6 +153,8 @@ export interface TaskIpcPayload {
   repo?: string;
   pr_number?: number;
   run_id?: number;
+  workflow_path?: string;
+  ref?: string;
   artifact_kind?: string;
   profile?: string;
   expected_snapshot_id?: string;