e2e-engineering 1.0.0 → 1.1.0

package/README.md

@@ -172,7 +172,7 @@ The risks
 
 ## Part 3 — Establish shared language before building
 
-### 3.1 Why “Grill Me” is useful
+### 3.1 Why “Grill Me” is useful (@mattpocock)
 
 The original **grill-me** skill asks the AI to interview the user relentlessly until both sides reach a shared understanding. It walks down the design tree and resolves dependencies between decisions one by one.
 
@@ -193,7 +193,7 @@ Use this when an idea is still vague.
 
 ---
 
-### 3.2 Prefer “Grill with Docs” when there is a codebase
+### 3.2 Prefer “Grill with Docs” when there is a codebase (@mattpocock)
 
 The newer workflow replaces pure grill-me with **grill-with-docs** when a codebase exists. The problem with grill-me alone is that good terminology may emerge during the conversation but not get documented. Then the user has to re-explain the same domain concepts again in future sessions.
 

package/dist/marketplace/plugins/e2e-engineering/skills/e2e-engineering/SKILL.md

@@ -64,7 +64,7 @@ Repeat until COMPLETE (all stories `status: done`):
 1. **Compute ready set** — stories whose `depends_on` are all `done` AND own `status: todo`.
 2. **Fan-out** — dispatch each ready story to its OWN git worktree + subagent (use EnterWorktree). Inject [constitution](./constitution.md) + the story + its testCases into each subagent. Subagent runs [tdd](./impl/tdd.md): gap-check → red-green-refactor → automate its FEATURE e2e → return SUMMARY ONLY.
    - **HARD GATE 2 — TDD red before green.** Each subagent must write a failing test before production code. Enforced inside tdd.md.
-   - **HARD GATE 3 — debug escalation.** Subagent fails 3 fix attempts → orchestrator re-dispatches ONCE with [systematic-debugging](./impl/systematic-debugging.md) (4-phase root-cause). Still red → mark story `blocked`, append `## Blocked` in progress.txt, keep draining the ready set. Escalate to human ONLY on stall (no ready work left, or every remaining story depends on a blocked one).
+   - **HARD GATE 3 — debug escalation.** Subagent fails 3 fix attempts → orchestrator re-dispatches ONCE with [systematic-debugging](./impl/systematic-debugging.md) (4-phase root-cause). Still red → mark story `blocked`, append `## Blocked` in progress.txt, keep draining the ready set. Escalate to human ONLY on stall (no ready work left, or every remaining story depends on a blocked one). Emit `<e2e-stall reason="all-stories-blocked" />` before escalating.
 3. **Fan-in (orchestrator, serial — sole writer):** for each returned summary:
    - Per-slice review: check summary against the story spec + constitution. Drift → bounce back to the subagent, do not merge.
    - Merge the worktree branch into baseBranch. Resolve conflicts (never discard work).
@@ -88,7 +88,7 @@ Entry: gate 5 passed.
 1. [review](./post-impl/review.md) — fresh-context, full-diff, cross-slice audit by a clean reviewer. Findings ranked by severity.
 2. [human-qa](./post-impl/human-qa.md) — walk the Manual test-case set (the disposition `Manual` cases). Single human-approval chokepoint: in ONE touch the human approves QA sign-off AND batched `## Pending Amendments` → promote to [constitution](./constitution.md) (bump version) or drop.
 
-Task close: extract durable learnings, ensure amendments resolved, progress.txt resets on the NEXT task.
+Task close: extract durable learnings, ensure amendments resolved, progress.txt resets on the NEXT task. Emit `<e2e-complete stories="N" />` (N = total story count from prd.json).
 
 ---
 

package/dist/marketplace/plugins/e2e-engineering/skills/e2e-engineering/cross/context-checkpoint.md

@@ -1,6 +1,17 @@
 # context-checkpoint — 65% snapshot
 
-When context reaches 65% (hook-injected %), save a checkpoint and end the session so a fresh one resumes cleanly. Provenance: ralph checkpoint/phase-transition. ADR 0002.
+When context reaches 65% (hook-injected %), **set a checkpoint flag** — do NOT interrupt immediately. Finish the current in-flight task to its next fan-in boundary, then save the checkpoint and end the session so a fresh one resumes cleanly. Provenance: ralph checkpoint/phase-transition. ADR 0002.
+
+## When to trigger
+
+| State at 65% | Action |
+|---|---|
+| Between slices / at fan-in / idle | Checkpoint immediately |
+| Mid-slice TDD loop | Finish current slice, checkpoint at fan-in |
+| Mid-subagent (spawned, not returned) | Wait for subagent result, checkpoint after fan-in |
+| Mid-user-message reply | Complete reply, then checkpoint |
+
+Never abort mid-task. The 65% signal means "next safe stop, not right now."
 
 ## What to write (three files)
 1. **prd.json** — already maintained live by the orchestrator. Ensure `status` of every story is current.
@@ -14,10 +25,32 @@ When context reaches 65% (hook-injected %), save a checkpoint and end the sessio
 ## Suggested skill   # which e2e-engineering sub-skill the fresh session should invoke
 ```
 
-## Then
-End the session. The fresh session runs [phase-transition](./phase-transition.md) bootstrap.
+## Then — checkpoint instruction + HARD STOP
+
+After writing the three files:
+
+1. Output this exact message to the user:
+   ```
+   Context at 65%+ — checkpoint saved.
+   Handoff: .e2e-engineering/handoff-<phase>-<timestamp>.md
+
+   Resume (manual):
+     1. /clear    ← reset context
+     2. /e2e-engineering    ← fresh session reads handoff automatically
+
+   <e2e-checkpoint handoff=".e2e-engineering/handoff-<phase>-<timestamp>.md" />
+   ```
+   (substitute actual handoff path in BOTH the Handoff line and the signal)
+2. **HARD STOP** — process NO further messages in this session. Any further user message gets one reply: "Checkpoint saved — `/clear` then `/e2e-engineering` to resume."
+
+> **Unattended automation (AFK wrapper):** `scripts/afk.ps1` detects `<e2e-checkpoint />` and restarts automatically. Run `.\scripts\afk.ps1` after gate 1 to enable AFK mode. Supports claude (default), opencode, codex via `-AI` param. (ADR 0005)
+
+The fresh session runs [phase-transition](./phase-transition.md) bootstrap when `/e2e-engineering` is invoked.
 
 ## Red flags (stop)
+- Checkpointing mid-task instead of waiting for fan-in boundary.
+- Treating 65% as an immediate hard stop — it's a "prepare to stop" signal.
 - Hand-writing the handoff instead of generating from state files (drift).
-- Checkpointing mid-subagent (checkpoint at fan-in boundaries, between slices — not inside a slice's TDD).
 - Writing the handoff in prose (it's caveman:ultra).
+- Continuing to process user messages after outputting the stop message.
+- Telling user to start fresh without providing the exact handoff path.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "e2e-engineering",
-  "version": "1.0.0",
+  "version": "1.1.0",
   "description": "Master engineering orchestrator skill — idea to passing E2E across pre-implementation, implementation, and post-implementation phases. Installs into Claude Code (full fidelity), or as a portable AGENTS.md / Cursor rules variant (sequential).",
   "bin": {
     "e2e-engineering": "bin/install.js"