implement-stories

name	implement-stories
description	Drives the implementation of one or more Scrum stories by dispatching to the correct downstream create-/update- skill per epic. Accepts a story ID, a comma-separated list, an epic ID (topo-sorted by Depends On), or a Sprint label (e.g. "Sprint 3"). Never edits story markdown itself — it only orchestrates the generators and surfaces their results. Use when the user asks to: - Implement STORY-NN-NNN or a list of stories - "Implement EPIC-02" / "Implement all of Sprint 3" - Build the next story in the backlog - Drive story-mode generation across several stories in one turn
argument-hint	[STORY-NN-NNN | EPIC-NN | 'Sprint N' | comma-list]
allowed-tools	Read, Grep, Glob, Bash, AskUserQuestion, Skill
context	fork

Implement Scrum Stories

You are the orchestrator for the existing create-* / update-* skills. Your job is to walk one or more stories through their generator in the correct order, stop on the first CRITICAL failure, and hand off to /developer-plugin:validate-stories when the batch finishes.

You do NOT edit story markdown files. You do NOT flip Status fields. You do NOT tick acceptance-criteria checkboxes. Those mutations belong to /developer-plugin:complete-stories after validation passes.

Workspace Discovery

Before any file operation, run the discovery helper and substitute the returned tokens into every path this skill reads, writes, or edits:

python3 ${CLAUDE_PLUGIN_ROOT}/skills/validate-stories/scripts/status_rollup.py --mode discover

The JSON output supplies {workspace_root}, {project_root}, {project_name}, {stories_dir}, and {learnings_queue}. The plugin is project-agnostic — never hardcode project or chapter names in edits.

Coding Patterns & Libraries Handbook (orchestrator-level)

Before dispatching to any sub-skill, the orchestrator runs the freshness check ONCE so every downstream generator inherits an up-to-date library cache without re-prompting:

PATTERNS_DIR=$(ls -d "{workspace_root}/inputs/code/v"* 2>/dev/null | sort -V | tail -1)
if [ -z "$PATTERNS_DIR" ] || [ ! -d "$PATTERNS_DIR" ]; then
  echo "CRITICAL: inputs/code/v*/ not found. Run /developer-plugin:refresh-libraries to initialize the library cache."
  exit 1
fi
LIBRARIES_FILE="$PATTERNS_DIR/LIBRARIES.md"
LAST_VERIFIED=$(grep '^last_verified:' "$LIBRARIES_FILE" | awk '{print $2}')
TODAY=$(date -u +%Y-%m-%d)
AGE_DAYS=$(python3 -c "from datetime import date; print((date.fromisoformat('$TODAY') - date.fromisoformat('$LAST_VERIFIED')).days")

If AGE_DAYS > 30, call AskUserQuestion once with options Refresh now / Proceed with cached versions / Cancel:

• Refresh now — invoke /developer-plugin:refresh-libraries, wait for completion, then continue dispatching.
• Proceed with cached versions — dispatch sub-skills with the cached versions; prepend a single stale-cache warning to the batch's final References trailer.
• Cancel — abort the batch; no stories run.

Do not let each sub-skill re-prompt — the orchestrator answers once for the whole batch. Downstream create-/update- skills skip their own freshness check when invoked from this skill (detect via an env var such as DEVPLUGIN_FRESHNESS_OK=1).

Read $PATTERNS_DIR/LIBRARIES.md in full so versions are available to any sub-skill that asks.

Dispatch (content-based classification — no config)

There is NO dispatch yaml. Story and epic numbers are not stable across projects — EPIC-02 may be ingestion in one project and something else in another. Each story is classified by its acceptance-criteria content:

python3 ${CLAUDE_PLUGIN_ROOT}/skills/validate-stories/scripts/status_rollup.py \
  --mode classify --story STORY-NN-NNN

The JSON output includes skill_kind — one of scaffold, dag, ingestion, pipeline, integration-test, deploy-validation, or unknown — plus confidence (high / low) and a reasons[] list citing the AC matches that produced the verdict.

Map skill_kind to the generator by literal convention (no lookup):

skill_kind	create skill	update skill	validate skill
scaffold	create-scaffold	update-scaffold	validate-scaffold
dag	create-dag	update-dag	validate-dag
ingestion	create-ingestion	update-ingestion	validate-ingestion
pipeline	create-pipeline	update-pipeline	validate-pipeline
integration-test	create-integration-test	create-integration-test (re-emit; idempotent)	validate-stories (story-scoped AC verifier)
deploy-validation	create-deploy-validation	create-deploy-validation (re-emit; idempotent)	validate-stories (story-scoped AC verifier)
unknown	— (see Phase 2 Step 3)	—	—

Story-type → skill-kind shortcut. When the story's Story Type metadata is integration-test or deploy-validation, route to the matching kind directly — the AC text for those story types is mostly runtime behaviour (no path tokens), so the classifier's path-matching will not find an owner. Use the story-type cell as authoritative for those two kinds.

If skill_kind == "unknown" OR confidence == "low", stop the batch and ask the user via AskUserQuestion, showing the reasons[] so they can override. Record any override as a learning (Phase 5).

A downstream generator MAY accept a STORY-NN-NNN argument directly (story mode) — forward the story ID verbatim when it does. Otherwise forward the LLD path (full mode) and tell the user only the covered story's deliverables will be reviewed from that run. Classification does not distinguish the two modes; detection is the invoked skill's responsibility.

Workflow

Phase 0: Resolve Target Set

FIRST — resolve the effective argument. Do this before any other Phase 0 step.

The argument passed via the Skill tool may not reach this forked subagent. The shared resolver script auto-discovers the workspace root and checks four sources in order; stop at the first non-empty hit:

1. $SKILL_ARG environment variable.
2. <workspace>/.skill-arg file — consumed (deleted after read).
3. The conversational argument supplied to the skill.
4. Auto-mode default — if $CLAUDE_AUTO_MODE=1 OR <workspace>/.auto-mode exists → resolve the latest backlog and implement the first un-Done epic (as returned by status_rollup.py in backlog order).
5. Only if ALL four above are empty, ask the user via AskUserQuestion.

You MUST NOT ask the user until sources 1–4 have been checked.

Mechanical resolver — execute exactly this bash block (no placeholders to substitute, no edits to the script invocation; the only line you fill in is CONV_ARG):

# Step 1 — capture the user's conversational argument. Substitute the
# bracketed text below with the EXACT message the user supplied after the
# skill name; if they supplied no message, leave it as an empty string.
CONV_ARG='<<EXACT_CONVERSATIONAL_TEXT_FROM_USER_OR_EMPTY_STRING>>'

# Step 2 — run the shared resolver. It auto-discovers the workspace from
# $PWD, so no {workspace_root} substitution is required. Output is two
# lines on stdout: the resolved value, then the source token.
read -r RESOLVED_ARG RESOLVED_SOURCE < <(
  bash "${CLAUDE_PLUGIN_ROOT}/scripts/resolve_skill_arg.sh" "$CONV_ARG" \
    | paste -sd' ' -
)

If $RESOLVED_ARG is __AUTO__, expand it into the first un-Done epic in the backlog and proceed without asking. If $RESOLVED_SOURCE is EMPTY, the resolver writes a diagnostic block to stderr listing every location it probed — surface that diagnostic to the user verbatim before asking via AskUserQuestion, so they can see exactly what was missing instead of guessing.

Hard rule: before declaring "no argument supplied", you MUST have actually invoked bash resolve_skill_arg.sh "$CONV_ARG" and quoted its output. Inventing a four-bullet checklist of "sources I checked" without running the script is forbidden — that pattern has caused multiple silent argument-loss bugs in this session.

Resolution-source banner (mandatory, every run). As the first line of skill output, before any other work, print exactly which source supplied the target so the user can spot a lost arg immediately:

RESOLVED TARGET: <EPIC-NN | STORY-NN-NNN | Sprint N> (source: <SKILL_ARG | .skill-arg | conversational | __AUTO__>)

Whenever the source is __AUTO__ (sources 1–3 all empty), upgrade the banner to a hard warning:

⚠ AUTO-MODE FALLBACK: no explicit target found in $SKILL_ARG, .skill-arg,
  or conversational arg. Resolved to <EPIC-NN> (first un-Done epic in backlog order).
  If you intended a different target, abort now and re-invoke
  AFTER writing the target to {workspace_root}/.skill-arg, e.g.:
      echo "EPIC-02" > {workspace_root}/.skill-arg

Mandatory confirm (cannot be skipped). Phase 0 Step 3's AskUserQuestion gate runs on every invocation, including auto-mode. There is no "skip confirm" path. The resolution-source banner makes the auto-mode disagreement visible; the AskUserQuestion gate then lets the user redirect. Skipping the confirm to "save a turn" is a defect — the loop the gate prevents (silently dispatching to the wrong epic, which leaves no plan JSON for the intended one) is far more expensive than one extra prompt.

Caller contract. The Skill-tool argument frequently fails to reach forked subagents. Any caller (slash command, parent agent, hook) that invokes /developer-plugin:implement-stories <arg> programmatically MUST also echo "<arg>" > {workspace_root}/.skill-arg immediately before the Skill invocation. The skill consumes the file (deletes after read), so it is a one-shot. Without this, auto-mode silently picks the first un-Done epic — the banner + Step 3 confirm above are the safety net, not a substitute for the contract.

Auto-mode marker hygiene. $CLAUDE_AUTO_MODE=1 is the supported auto-mode trigger. The legacy {workspace_root}/.auto-mode file is honoured for backward compatibility but strongly discouraged: it is too easy to leave behind across sessions. If you find a .auto-mode file with no recent edits (older than the latest .skill-arg write), treat it as stale — emit a WARNING in the banner and ask the user whether to delete it before continuing.

Version lockfile (pinned upstream v{N})

Before any dispatch, run:

python3 ${CLAUDE_PLUGIN_ROOT}/scripts/resolve_versions.py --write

This creates {workspace_root}/outputs/dev-lock.yaml pinning the exact v{N} folder for each upstream (lld, stm, dqs, stories) for the duration of this batch. Every downstream create-/update- skill sources the same file — they never re-resolve ls -d outputs/*/v* | tail -1 on their own once the lockfile is present. Delete the lockfile to re-pin on the next batch.

Bootstrap gate (chicken-and-egg)

Check the discovery result:

DISCOVERY=$(python3 ${CLAUDE_PLUGIN_ROOT}/skills/validate-stories/scripts/status_rollup.py --mode discover)
NEEDS_BOOTSTRAP=$(echo "$DISCOVERY" | python3 -c "import sys,json; print(json.load(sys.stdin).get('needs_bootstrap', False))")

If needs_bootstrap == True, classify every story in the target set (via status_rollup.py --mode classify). If at least one classifies to scaffold, dispatch /developer-plugin:create-scaffold FIRST (passing the original target via $SKILL_ARG), wait for completion, then re-run discovery before continuing. create-scaffold handles the cookiecutter bootstrap as its first step.

If needs_bootstrap == True but NO story in the target set classifies to scaffold, stop with CRITICAL: project not bootstrapped — run /developer-plugin:create-scaffold first, or include a foundation story in the target set. This keeps the bootstrap gate tied to kind (scaffold), not to a specific epic number — any project's foundation epic, numbered however the reader's scrum-master emitted it, triggers the gate.

Step 1 — Parse the argument into an ordered story list.

Argument grammar (shared with validate-stories and complete-stories):

Form	Example	Meaning
STORY-NN-NNN	STORY-02-002	single story
comma-list	STORY-02-001,STORY-02-002	ordered list — respect declared Depends On
EPIC-NN	EPIC-02	every child story, topo-sorted
Sprint N	Sprint 3	every story whose Sprint cell is Sprint N
__AUTO__ / (no arg)	—	auto-mode: next un-Done epic from EPIC-01; else ask via AskUserQuestion

Normalize the argument exactly as create-ingestion Phase 0 does: uppercase, zero-padded ("story 2 of epic 2" → STORY-02-002). Reject ranges (STORY-02-001..003) explicitly — they are not supported.

Step 2 — For EPIC-NN / Sprint N, expand to stories and topo-sort.

Use the helper:

python3 ${CLAUDE_PLUGIN_ROOT}/skills/validate-stories/scripts/status_rollup.py \
  --mode parse-epic --epic EPIC-02

Read the child_stories array. Build a DAG from each story's depends_on field and emit the stories in topological order. If a cycle is detected, stop with a CRITICAL: cycle detected: A → B → A.

For Sprint N, parse every story file under the latest {stories_dir}/v*/ (from discovery) and filter by sprint == "Sprint N", then topo-sort the same way.

Step 3 — Show the resolved list and confirm (mandatory).

This AskUserQuestion gate is the last guard before plans are persisted and sub-skills mutate the project. It runs on every invocation — no exceptions for auto-mode, no exceptions for $CLAUDE_AUTO_MODE=1. The prompt MUST surface (a) the resolved target with its resolution source (matching the banner above) and (b) the ordered story list, so a lost conversational arg becomes visible here, not after EPIC-01 plans overwrite EPIC-02 work:

RESOLVED TARGET: EPIC-02 (source: .skill-arg)

About to implement 3 stories in this order:
  1. STORY-02-001 — Per-Table YAML Ingestion Configs   (create-ingestion)
  2. STORY-02-002 — Generic Ingestion Runner           (create-ingestion)
  3. STORY-02-003 — SparkSubmitOperator Wrapper        (update-ingestion — file exists)

Proceed? (Yes / No / Edit order)

If the user picks Edit, ask which stories to keep or reorder, then re-confirm. If the user picks No, exit cleanly with no plan writes and no dispatches.

Phase 1: Pre-flight

For each target story, call:

python3 ${CLAUDE_PLUGIN_ROOT}/skills/validate-stories/scripts/status_rollup.py \
  --mode parse-story --story STORY-NN-NNN

From the JSON:

• If status == "Done" → skip (single-line log: STORY-NN-NNN already Done — skipped). Never re-dispatch.
• If epic_id's epic file has status == "Done" but this story is not Done → stale backlog. Stop and tell the user to resolve the drift before continuing.
• If the story file is not found → stop with the error from the helper.

Phase 1.5: Build and Persist the Story Plan

Every story gets a persistent execution plan — the single ledger that implement-stories, validate-stories, and complete-stories all read and enrich. One plan per story, located at {stories_dir}/v{N}/plans/STORY-NN-NNN.plan.json.

Before dispatching any sub-skill, build (or refresh) the plan:

python3 ${CLAUDE_PLUGIN_ROOT}/skills/validate-stories/scripts/status_rollup.py \
  --mode build-plan --story STORY-NN-NNN --save

The plan captures:

• tasks[] — one per (skill, paths) group from extract-deliverables, in dispatch_order, with a depends_on chain so later tasks wait for earlier ones.
• acceptance_criteria[] — every AC line linked to the task(s) whose deliverable paths cover its backtick-quoted tokens. Behavioural ACs (no tokens) link to every task so any validator can attempt coverage.
• status: planned initially; walks planned → in_progress → implemented → validated → done over the three phases; or failed at any step.

If the scrum-master re-runs and an AC or deliverable changes, re-running build-plan --save overwrites the plan and increments plan_version. Prior task outcomes are NOT preserved across re-builds — callers that care should compare plan_version.

Phase 2: Dispatch (multi-skill, path-based, plan-driven)

This phase is path-based, not kind-based. A single story whose ACs name deliverables under two skills' domains (e.g. an ingestion module AND the DAG task that wires it in) dispatches to BOTH skills in sequence, each scoped to the paths it owns. No skill ever sees paths outside its domain.

The routing registry is {workspace_root}/inputs/code/v*/DELIVERABLE-OWNERS.yaml — a reader's extension point for re-routing path classes (e.g. swapping Airflow for Dagster).

Step 1 — Extract deliverables for each story.

For each story in the resolved list:

python3 ${CLAUDE_PLUGIN_ROOT}/skills/validate-stories/scripts/status_rollup.py \
  --mode extract-deliverables --story STORY-NN-NNN

The JSON output contains:

• paths — every backtick-quoted path from the AC (placeholders stripped).
• by_skill — paths grouped by owning skill from the registry.
• dispatch_order — the order to invoke those skills (from the registry's dispatch_order: typically scaffold → ingestion → dag → pipeline).
• unmatched — paths no glob matched (added to fallback_kind's group via the classifier).
• fallback_kind — the classifier's kind, used when paths is empty (purely behavioural ACs) or for unmatched paths.

Step 2 — Fall-back for behavioural stories.

If by_skill is empty, the story's ACs have no path tokens (behavioural story). Dispatch the single skill named by fallback_kind (from the classifier) with STORY-NN-NNN and no .skill-paths file — the sub- skill will read the story's AC and handle it in its historical single-skill mode.

If fallback_kind is also missing / unknown, stop the batch:

Cannot route STORY-NN-NNN. No deliverable paths in its AC and the
classifier returned `unknown`. Tighten the AC text to reference concrete
paths (e.g. `airflow/configs/foo.yml`, `src/<project>/bronze/...`) or
override with "treat STORY-NN-NNN as {scaffold|dag|ingestion|pipeline}".

Record any user override in the learnings queue (Phase 5).

Step 3 — Per-group dispatch (plan-driven).

Read the plan's tasks[] (built in Phase 1.5). For each task in order:

Before dispatching a task, update the plan:

• set tasks[i].status = "in_progress" and stamp started_at
• if this is the first task, bump plan status = "in_progress"

Persist the plan after every status change so a mid-batch interruption leaves a truthful ledger. Use a simple Python one-liner:

python3 -c "
import json, sys
from datetime import datetime, timezone
p = json.load(open('$PLAN_PATH'))
for t in p['tasks']:
    if t['id'] == '$TID':
        t['status'] = '$NEW_STATUS'
        if '$NEW_STATUS' == 'in_progress' and not t['started_at']:
            t['started_at'] = datetime.now(timezone.utc).isoformat()
        if '$NEW_STATUS' in ('done','failed'):
            t['finished_at'] = datetime.now(timezone.utc).isoformat()
json.dump(p, open('$PLAN_PATH','w'), indent=2)
"

Then dispatch:

1. Write the scoping files:

   echo "STORY-NN-NNN"            > {workspace_root}/.skill-arg
   printf '%s\n' "${paths[@]}"    > {workspace_root}/.skill-paths
   

   The sub-skill reads both and treats .skill-paths as the ONLY paths in scope for this invocation. Both files are consumed (deleted) by the sub-skill after reading.

2. Pick create- vs update-** using the same rule as before, but scoped to THIS group's paths:

   • status == "To Do" AND none of this group's paths exist on disk → create-<skill>.
   • status == "In Progress", OR any of this group's paths already exist → update-<skill>.
   • Split (some present, some absent) → AskUserQuestion with the present/absent lists.

3. Invoke via the Skill tool and capture output.

4. CRITICAL handling (intra-story): if this dispatch reports CRITICAL, stop this story (do not dispatch the subsequent groups for the same story — they likely depend on this one). Continue with the next story in the batch unless --stop-on-critical-batch is passed.

5. Append one row to the implementation log per dispatch:

   STORY-02-004 | create-ingestion | 2 files created     | OK
   STORY-02-004 | update-dag       | 1 file patched      | OK
   STORY-02-006 | create-ingestion | 1 file created      | OK
   

After the sub-skill returns, parse its output to populate tasks[i].files_created, files_updated, artifacts, and any critical_findings. Then set tasks[i].status = "done" (or "failed" if CRITICAL) and save.

When every task has status == "done", set plan status = "implemented". If any task status == "failed", set plan status = "failed".

Step 3.4 — Post-dispatch handbook validator gate (tight loop).

The AC verifier in Step 3.5 only checks what each story's Verification block names. Pattern-handbook rules (env-driven paths, hard imports, required test modules, etc.) live in the matching validate-* skill and are invisible to the AC verifier. To catch them at write-time rather than at end-of-batch, run the handbook validator immediately after each create-/update- dispatch and treat its CRITICAL output as a story failure.

Validator scripts (callable directly, no Skill fork required):

skill_kind	Validator script (relative to plugin root)
ingestion	skills/validate-ingestion/scripts/validate_ingestion.py
dag	skills/validate-dag/scripts/validate_dag.py
scaffold	SKILL.md-only — invoke /developer-plugin:validate-scaffold via Skill
pipeline	SKILL.md-only — invoke /developer-plugin:validate-pipeline via Skill

For ingestion dispatches:

python3 ${CLAUDE_PLUGIN_ROOT}/skills/validate-ingestion/scripts/validate_ingestion.py \
  --project-root <project-root> 2>&1 | tee /tmp/handbook-<story>-<task>.txt

For dag dispatches (positional path; use --all to recurse the project):

python3 ${CLAUDE_PLUGIN_ROOT}/skills/validate-dag/scripts/validate_dag.py \
  <project-root> --all 2>&1 | tee /tmp/handbook-<story>-<task>.txt

Parse the last line: Result: PASS or Result: FAIL.

• Result: PASS — set tasks[i].handbook_validator = "pass"; continue to Step 3.5 (AC verification).
• Result: FAIL — extract every line under CRITICAL: <n> issue(s) up to the next blank line. For each:
  1. Append the line (verbatim) to tasks[i].critical_findings.
  2. Set tasks[i].handbook_validator = "fail" AND tasks[i].status = "failed" (overrides the prior done).
  3. Set plan status = "failed".
  4. Halt the story (same semantics as an AC inline-FAIL — do not run Step 3.5 for this task, do not dispatch subsequent groups for the same story). Continue with the next story in the batch unless --stop-on-critical-batch is passed.
  5. Log the row: STORY-NN-NNN | <skill> | handbook FAIL: <n> CRITICAL | STOPPED.

For scaffold and pipeline dispatches (no callable script), the fallback is to dispatch the matching /developer-plugin:validate-<kind> skill via the Skill tool and parse its final summary line the same way (Result: PASS|FAIL + CRITICAL extraction). Record the same fields on the task. When neither a script nor a Skill is available, set handbook_validator = "skipped" and emit one trailer-line warning per story affected.

Persist the plan JSON after the validator runs. This gate runs before the AC verifier so a handbook FAIL halts the story without waiting for the AC pass — most handbook violations would otherwise leak past the AC verifier and only surface during the end-of-batch validate-stories run.

Step 3.5 — Post-dispatch AC verification (tight loop).

After each sub-skill returns and the plan is saved, run the AC verifier for this story before dispatching the next task in the same story:

python3 ${CLAUDE_PLUGIN_ROOT}/../scripts/verify_acs.py STORY-NN-NNN --json

Parse the JSON output (shape: a list of one story object with has_verification, overall, and acs[]):

• has_verification: false — add one trailer line for the run summary: STORY-NN-NNN: no Verification block — inline validation skipped. Then continue as normal.
• For each AC object with status == "FAIL" (any non-manual check that failed):
  1. Determine which task to attribute the failure to:
     • If the AC has non-empty task_ids in acceptance_criteria[] and they include the just-completed task → that task fails.
     • If task_ids is empty OR doesn't reference any task in this story → attribute to the most recently completed task in this story (tasks[-1] whose status == "done"). Empty task_ids is a heuristic-population artifact, NOT a permission to ignore the FAIL.
  2. Append the first check.detail of the first failing check to tasks[i].critical_findings.
  3. Set tasks[i].validator_status = "fail" AND tasks[i].status = "failed" (overrides the prior done).
  4. Treat as CRITICAL for halt semantics: set plan status = "failed", stop this story, continue the batch. Log the row: STORY-NN-NNN | <skill> | AC<N> inline-FAIL | STOPPED.
• For status == "PASS" → set tasks[i].validator_status = "pass".
• status == "INDETERMINATE" never halts and never sets validator_status — leave the field as "pending".

Persist the plan JSON after updates. This tight loop surfaces gaps at the moment they appear rather than deferring them to an end-of-batch validate-stories run.

Step 4 — Verify no ROUTE-TO escapes.

A sub-skill invoked WITH a .skill-paths file should never emit ROUTE-TO: (every path it received was already vetted as owned). If the output contains ROUTE-TO:, record a learning — the registry has a gap for that path pattern, and a future apply-learnings run should propose adding the glob.

Phase 3: Execute

For each story in order:

1. Invoke the chosen skill via the Skill tool. Capture its final Output Summary text.

2. If the skill reports any CRITICAL failure (visible in its output), stop the batch. Do not continue to subsequent stories — later stories in a topo-sorted list typically depend on the one that just failed.

3. Append one row to your running implementation log:

   STORY-02-002 | create-ingestion | 1 file created | OK
   STORY-02-003 | update-ingestion | 0 CRITICAL, 1 WARNING | OK
   STORY-02-004 | create-ingestion | 2 CRITICAL        | STOPPED
   

Phase 4: Hand-off

Print the final table. Immediately above the next-step line, emit two inline-validation summary lines — one for the Step 3.4 handbook validator verdicts, one for the Step 3.5 AC verifier verdicts — tallied across all stories in the batch:

Handbook validator: <N> PASS, <M> FAIL (<story>:<critical-headline>, ...), <K> SKIPPED (no script or skill)
Inline AC verification: <N> PASS, <M> FAIL (<first-failing-story AC#>, ...), <K> SKIPPED (no verification block)

Then end with the exact next-step line:

Next: /developer-plugin:validate-stories <same-argument>

Do NOT run validate-stories yourself — the user chooses when to validate (they may want to inspect the generated diff first). The inline summary above is a heads-up, not a replacement.

Phase 5: Learnings

If the user corrected your classification (e.g. "no, STORY-02-010 is an ingestion test, not scaffold") or your topo-sort ordering, append a JSONL line to the {learnings_queue} returned by the discovery helper:

QUEUE=$(python3 ${CLAUDE_PLUGIN_ROOT}/skills/validate-stories/scripts/status_rollup.py \
  --mode discover | python3 -c "import sys,json; print(json.load(sys.stdin)['learnings_queue'])")
echo '{"skill": "implement-stories", "date": "'$(date +%F)'", "correction": "<quoted user correction>", "pattern": "<generalized rule>", "status": "pending"}' >> "$QUEUE"

The existing PostToolUse check-learnings-queue.py hook will remind you to run /developer-plugin:apply-learnings at session end.

Output Summary

End every run with this table (one row per target story):

Target: EPIC-02  |  Resolved: 3 stories  |  Dispatched: 3  |  Stopped: 0

Order | Story         | Skill              | Result
    1 | STORY-02-001  | create-ingestion   | 13 configs created, validate-ingestion PASS
    2 | STORY-02-002  | create-ingestion   | 1 module created,  validate-ingestion PASS
    3 | STORY-02-003  | update-ingestion   | 1 module edited,   validate-ingestion PASS

Next: /developer-plugin:validate-stories EPIC-02

Edge Cases (must handle)

• Lowercase arg (story-02-002) → normalize to uppercase silently.
• Circular Depends On → detect during topo-sort, stop with CRITICAL.
• Story already Done → single-line skip log; never re-dispatch.
• Multiple backlog files → helper picks the lexicographically latest non-.bak.
• AskUserQuestion unavailable → default to the conservative choice (update-* when deliverable exists, else create-*) and log the decision.
• Generator fails mid-batch → stop, print log so far, tell the user which story to fix before re-invoking.

Learnings & Corrections

No learnings recorded yet.

Learnings & Corrections

Active Learnings

• L-001 (2026-05-11): Always: When a forked skill returns asking for a target, the orchestrator should retry with the conversational arg explicitly OR detect that .skill-arg was not consumed (file still present after invocation) and surface a CRITICAL halt rather than silently moving on.
• L-002 (2026-05-11): Always: When story AC glob conflicts with LLD/project convention, halt and recommend scrum-master:update-stories rather than rename project-wide
• L-003 (2026-05-11): Always: When a path appears under a glob pattern like contracts/{table}.yml inside an AC line that describes ingestion behaviour, treat it as a reference dependency, not a scaffold deliverable; DELIVERABLE-OWNERS.yaml should distinguish contracts paths only when the AC explicitly says create/update/author/define.
• L-004 (2026-05-11): Always: When AC Verification uses grep_count equals:N where N == file count, but each file legitimately contains multiple matches, the verification spec is wrong; flag to scrum-master for spec correction rather than reduce real rule content.
• L-005 (2026-05-11): Always: When extract-deliverables sees a backtick-quoted path that already exists AND is referenced in an AC as a source/input (e.g. "uses StructType from contracts/{table}.yml"), do not emit a scaffold task. Only emit a task for the skill that owns the AC verb ("adds", "writes", "calls", "creates").
• L-006 (2026-05-11): Always: When the handbook validator (Step 3.4) reports CRITICAL issues, intersect each finding against the current task's paths/groups before halting. CRITICALs that name paths owned by sibling stories (not in this task's .skill-paths) should be recorded as out-of-scope and not halt the current story. Halt only when a finding cites a path the current task was responsible for.
• L-007 (2026-05-11): Always: When the handbook validator runs after a story dispatch, filter its CRITICAL findings to ones touching the current story task paths before halting. Project-wide deficiencies that belong to later, not-yet-dispatched stories must not halt the current story.
• L-008 (2026-05-11): Always: When a story routes to a skill that immediately ROUTE-OUTs based on its hard rules, the registry has a gap. Either (a) extend a skill to own that path class explicitly, or (b) add a new specialized skill (e.g. create-liquibase). Until then, orchestrator should detect double-route-out and either fall back to direct edits or block the batch with a clear error rather than thrashing between skills.
• L-009 (2026-05-11): Always: Stories whose deliverables live under tests/integration/, docs/runbooks/, or deploy-validation scripts have no current owning skill. Either: (a) add an integration-test domain to update-ingestion or update-dag, (b) introduce a dedicated create-integration-test / create-runbook skill, or (c) flag such stories as manual at backlog-generation time so implement-stories does not attempt to dispatch them.
• L-010 (2026-05-11): Always: Step 3.4 handbook validator should distinguish story-scope failures from project-wide pre-existing failures. Consider scoping the validator to the dispatched .skill-paths set, or whitelist pre-existing findings via a baseline file, so a story does not get halted by failures unrelated to its deliverables.
• L-011 (2026-05-11): Always: When adding new skill folders to a plugin mid-session, document the reload step (e.g. /plugin reload or marketplace refresh) in the create-role-plugin / skill-scaffolding handbook. implement-stories should surface this with a clear actionable error rather than silently failing dispatch when a fallback_kind resolves to an existing-but-unregistered skill.