refactor(tests): modular mock-free test suite with cross-platform CI

[ofriw]

Note: This PR was generated by an AI agent. If you'd like to talk with other humans, drop by our Discord!

---

What

This PR replaces the single monolithic 3,840-line test file with per-module unit suites, adds a process-isolated E2E layer, extracts shared test infrastructure (singleton container, harness factory, module loader), and wires cross-platform CI that exercises the full suite on Linux, macOS, and Windows.

Why

The old agenticoding.test.ts was a single file that grew organically. It used ad-hoc mocks and stubs, imported production module-level singletons directly (leaking state between tests), and had no cross-platform coverage. The new architecture enforces no-mocks testing — every suite verifies external invariants, constraints, or user-facing contracts through only real code paths. The E2E layer runs real tool invocations against a minimal pi host in a child process, testing the actual tool contracts without mocking the SDK.

What changed

Test infrastructure:

• Extracted three module globals (write lock, frame scheduler, async context) into RuntimeSingletons — a single container the test harness swaps atomically via __setSingletons(), eliminating per-test patches
• Created createTestHarness() in tests/test-utils.ts — one call per test isolates singletons and captures console output; teardown restores originals
• Added scripts/run-node-test.mjs — runner that loads the custom module bootstrap via --import (replacing the deprecated --experimental-loader flag)
• Migrated test-loader.mjs to resolve SDK packages by walking node_modules instead of hardcoding an absolute path — works from any install location

Unit test reorganization:

• tests/unit/handoff.test.ts — handoff tool, command, and compaction contracts
• tests/unit/notebook.test.ts — write/read/index lifecycle, overwrites, rehydration, stale detection, concurrent write isolation
• tests/unit/render-snapshots.test.ts — golden-file snapshot tests for TUI render output (indicator levels, spawn call/result frames, nested session views)
• tests/unit/register-loader.test.ts — validates the module loader bootstrap resolves correctly
• tests/unit/runtime-singletons.test.ts — singleton container atomic swap, pending-write protection
• tests/unit/spawn-render.test.ts — TUI rendering of spawn lifecycle (expanded/collapsed, success/error/abort, truncation markers)
• tests/unit/spawn.test.ts (migrated from the original) — spawn execution logic, tool inheritance, session lifecycle, cleanup guarantees, cancellation
• tests/unit/state-invariants.test.ts — property-based tests using fast-check to verify runtime singletons never silently drop pending writes
• tests/unit/system-prompt.test.ts — system prompt injection at before_agent_start
• tests/unit/topic.test.ts — notebook topic lifecycle (set, override, human-authoritative)
• tests/unit/tui-indicators.test.ts — TUI status/widget indicators for context usage
• tests/unit/watchdog.test.ts — primacy-zone advisory injection at thresholds

E2E layer:

• tests/e2e/test-host.ts — minimal pi host REPL that loads the real extension, no TUI
• tests/e2e/pty-harness.ts — process-isolated child-process harness for stdin/stdout communication (no PTY dependency, works on Windows)
• tests/e2e/basic.test.ts — tool registration, notebook round-trips, topic lifecycle, handoff state, command registration, error handling

CI:

• Cross-platform matrix: Ubuntu (Node 22 + 24), macOS (Node 24), Windows (Node 24)
• Steps: npm ci → type check → security audit → unit tests → E2E tests → upload results
• Git LFS snapshots with .gitattributes enforcing LF line endings
• Snapshot update controlled via UPDATE_SNAPSHOTS=1 env var
• Test artifacts retained 30 days

SDK upgrade: Bumped peerDependencies to ^0.78.1 and adapted to new type surfaces (CustomEntry, TextContent, TypeBox Static, widened message content union). Added typebox, fast-check as dev dependencies.

Breaking changes: None to the extension's public API or behavior. The package.json description and keywords were updated to reflect the current primitive set (notebook replaces ledger). Node engine minimum set to >=22.

---

Attached is an agent optimized description of the changes in this PR - AGENT_REVIEW.md

[grzegorznowak]

great work, no major findings @ofriw 💪🏾