ai-fob 1.7.0 → 1.7.2

package/assets/pi/prompts/build-phase-pi.md

@@ -29,31 +29,73 @@ command with no arguments on its own line:
 4. `/skill:agent-browser` — Browser Tool Constraint (NEVER macOS `open`;
    ALWAYS `agent-browser open`).
 
-Derive identifiers from $1: task slug = basename of the spec parent directory
-minus the leading `NN_` prefix (e.g. `specs/17_pi-build-phase/hl-plan.md` ⇒
-`pi-build-phase`). `phaseKebab` = the HL plan's phase-$2 section heading
-(lowercase, hyphenated). If either is unavailable, ask the user.
-
-Call the `state` tool three times with arguments shaped like:
+Derive identifiers from $1:
+- `SPEC_DIR` = `dirname($1)` — the actual on-disk directory, numeric prefix
+  preserved (e.g. `specs/01_kanban-core-ui-schema/HL_PLAN.md` ⇒
+  `specs/01_kanban-core-ui-schema/`).
+- `task` (slug) = basename of `SPEC_DIR` minus the leading `NN_` prefix
+  (e.g. `01_kanban-core-ui-schema` ⇒ `kanban-core-ui-schema`). The slug
+  is what STATE.md uses internally.
+- `PHASE_DIR` = `${SPEC_DIR}/phase-$2/` — where your artifacts live.
+- `phaseKebab` = the HL plan's phase-$2 section heading (lowercase,
+  hyphenated). If any of these is unavailable, ask the user.
+
+Scaffold the Phase block in STATE.md if needed (idempotent — leaves any
+existing Phase block intact):
 
 ```json
-{ "operation": "mark_step_start", "task": "<slug>", "phase": $2, "step": 0 }
 { "operation": "init_phase", "task": "<slug>", "phase": $2, "phaseKebab": "<derived>" }
-{ "operation": "reconcile", "task": "<slug>", "phase": $2 }
 ```
 
-After `reconcile` returns, read `details.reconcile.entryStep` (0..6) as the
-starting step. If `entryStep > 0`, present `details.reconcile.table` to the
-user as a banner (echo $@ for context) and resume at `entryStep`. Do NOT
-re-execute any step `< entryStep`. The reconciler is the single source of
-truth for resume.
-
-When Step 0 is complete:
+**Resume detection — YOU do it, not a tool.** Trust your eyes over any
+classifier. Read `specs/STATE.md` and inspect `PHASE_DIR` on disk:
+
+1. Read `specs/STATE.md`; locate the `Phase $2` block under your task. Note
+   which Step markers are `[x]`, `[~]`, or `[ ]`.
+2. List `PHASE_DIR` (use the `bash` tool: `ls -la PHASE_DIR`). Note which
+   files are present.
+3. Canonical artifact ↔ Step mapping (from `phase-build-workflow` skill):
+   Step 1 → `explorer_findings.md` AND `docs_research.md`
+   Step 2 → `plan_V1.md`
+   Step 3 → `plan_validation_report.md`
+   Step 4 → `build_report.md`
+   Step 5 → `build_validation_report.md`
+   Step 6 → `phase_completion_report.md`
+4. Compute the resume entry step (`ENTRY_STEP`) as the lowest-numbered
+   step that is NOT already complete. For each step 1..6, "complete"
+   means BOTH:
+   - STATE.md marker is `[x]`, AND
+   - the canonical artifact(s) for that step are present in `PHASE_DIR`.
+5. Disagreement handling:
+   - STATE `[x]` AND artifact present → the step is done. Skip past it.
+     Do NOT re-run it. Do NOT modify its marker.
+   - STATE `[x]` AND artifact missing on disk → STOP and ask the user.
+     Show them the inconsistency (which step, which file expected,
+     directory listing). Do NOT silently reset markers.
+   - STATE `[~]` or `[ ]` → that's the `ENTRY_STEP`.
+6. Echo a one-line banner: `Resuming at Step <ENTRY_STEP> — Steps 1..<ENTRY_STEP-1> verified complete (STATE markers [x] AND artifacts present in PHASE_DIR).`
+   Then echo `$@` for any user context.
+
+**Step 0 marker handling — do not touch STATE on a resume.** If
+`ENTRY_STEP > 1` (this is a resume from a prior run), Step 0 was already
+completed previously — leave its marker `[x]` alone and proceed directly
+to `ENTRY_STEP`. Do NOT call `mark_step_start` or `mark_step_complete`
+for Step 0.
+
+If `ENTRY_STEP == 1` (this is a fresh phase, Steps 1..6 all `[ ]`), bracket
+this Step 0 preparation with the usual markers:
 
 ```json
+{ "operation": "mark_step_start",    "task": "<slug>", "phase": $2, "step": 0 }
 { "operation": "mark_step_complete", "task": "<slug>", "phase": $2, "step": 0 }
 ```
 
+(The legacy `state.reconcile` operation still exists in the `task-state`
+extension for diagnostic use, but is NO LONGER CONSULTED for resume.
+Tool-codified verdicts proved less reliable than direct inspection: the
+agent reading STATE.md and listing the phase directory is the source of
+truth.)
+
 # Step 1 — Research (parallel: phase-explorer + phase-docs-researcher)
 
 Mark step start (step 1) via `state`.
@@ -858,7 +900,9 @@ summary as your final assistant message. Include:
 
 Do NOT:
 
-- Skip Step 0 (reconcile is mandatory; resume detection drives the entry).
+- Skip Step 0 — even on resume you must verify STATE.md ↔ PHASE_DIR
+  yourself before proceeding (read STATE.md, list the phase directory,
+  decide ENTRY_STEP; do NOT delegate the decision to `state.reconcile`).
 - Bypass either validator — `phase-plan-validator` and `phase-build-validator`
   MUST be called.
 - Continue past 3 fix cycles in either loop — escalate to the user instead.
@@ -887,5 +931,6 @@ Do NOT:
 # Execute this now
 
 Start by loading the four skills, then call `state` with
-`operation=init_phase, task=<derived>, phase=$2, phaseKebab=<derived>`, then
-`operation=reconcile` to detect resume, then proceed step by step.
+`operation=init_phase, task=<derived>, phase=$2, phaseKebab=<derived>`,
+then inspect `specs/STATE.md` and `PHASE_DIR` yourself to identify
+`ENTRY_STEP` (the resume entry), then proceed step by step from there.

package/manifest.json

@@ -1,5 +1,5 @@
 {
-  "version": "1.7.0",
+  "version": "1.7.2",
   "presets": {
     "coding": {
       "description": "Research-driven coding workflow",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ai-fob",
-  "version": "1.7.0",
+  "version": "1.7.2",
   "description": "Deploy research-driven AI coding assistant assets (skills, agents, commands) into your projects",
   "bin": {
     "ai-fob": "bin/install.js"