[Serhan-Asad]

Follows #3419

Lines changed: +1,413 / -125 across 24 files.

What is implemented

• PR body Runtime: parsing and automatic synchronization (src/autopilot-core.mjs, prompts/cli/autopilot-core.prompt.md):
  • Tightens runtimeVerdicts so prose lines that start with Runtime without a colon (for example, Runtime errors are caught...) are never parsed as malformed Runtime: verdict lines.
  • Mechanically formats and synchronizes verified Runtime: verdict lines (and Evidence: runtime= state) from Autopilot's saved runtime verification state when updating PR descriptions, while still requiring the underlying saved runtime checks to pass.
• Sandbox real-binding-unavailable handling (src/autopilot-core.mjs, prompts/cli/autopilot-core.prompt.md):
  • Treats saved rows with real-binding-unavailable (zero provider calls when no live binding exists in the local sandbox) as local skips rather than author refusals or runtime coverage losses, while verifying every test of a multi-test Story has a passing or skipped/service deciding check so unrun sibling tests are never excused.
  • Stops feeding generic operator-reconciliation client_advice into coding-agent repair prompts so worker runs focus on real code, test, and lint failures instead of stopping with false needs-person aborts.
• Process-tree manifest bounds, stopped-job cleanup recovery, PID-recycling safety, and suite concurrency caps (src/agent-job.mjs, src/autopilot-core.mjs, owning prompts):
  • Raises the wrapper process-tree ps polling interval from 250 ms to 2000 ms and deduplicates and bounds captured process manifests to 512 entries.
  • Recovers stopped or timed-out jobs cleanly in cleanupAgentJob instead of latching .processes.error forever (including when prior retryable snapshot attempts recorded wrapper: null, failing closed on resolution write errors, and skipping -job.pid group collection, liveness probes, SIGTERM, and SIGKILL whenever job.pid has been recycled to an unrelated process without job.token).
  • Bounds Autopilot's full-suite node --test concurrency (--test-concurrency) to Math.max(1, Math.min(6, shareCpus)) when shareCpus is configured via ports.machine or ports.env.PDD_CHECK_CPUS (or directly on a node --test command).
• Offline test network guard and credit-account skip guards (package.json, test/helpers/test-net-guard.mjs, src/login-pool.mjs, src/agent-adapters.mjs, src/autopilot-launcher.mjs, .github/workflows/ci.yml, scripts/vm-suite.sh, owning prompts):
  • Adds test/helpers/test-net-guard.mjs and loads it via --import in package.json scripts.test, CI, and scripts/vm-suite.sh to allow 127.0.0.1, ::1, localhost, 0.0.0.0, and Unix domain sockets while blocking external network calls with ERR_PDD_OFFLINE_TEST_NET_BLOCKED unless PDD_ALLOW_NET=1 or PDD_LIVE_CANARY=1.
  • Enforces credit and reset-eligible account skip guards across selectCloudPoolLogin, selectPoolLogin, loadLoginPool fallback normalization, runStepAgentProcess, and runAutopilotLauncher startup checks unless allowCredits: true is explicitly passed by a human CLI command.
• Prompts and ownership headers:
  • Verified current Generated from ... prompt headers across all touched owned files via pdd owner.

What is tested

• Focused local tests (Mac):
  • test/autopilot-hardening.test.mjs, test/test-home-isolation.test.mjs, test/agent-adapters.test.mjs, test/autopilot-launcher.test.mjs, test/autopilot-core.test.mjs, test/autopilot-core-evidence.test.mjs, test/autopilot-core-guard.test.mjs, test/autopilot-core-runtime-wait.test.mjs, test/agent-job.test.mjs, test/login-pool.test.mjs, test/cloud-accounts.test.mjs, test/file-owner.test.mjs, test/plain-prompts-and-frontmatter.test.mjs (0 failures).
• Full test suite on runner VM (pdd-vm):
  • Passed 6,242 / 6,243 (0 failures, 1 skipped).