Stack

Stack

Thin launcher for the prompt-stack runtime at $HOME/digital/leviathan/plugins/prompt-stack.

Runtime ownership stays with the plugin. This skill teaches agents how to call it correctly and present results as a dashboard, not raw JSON.

Autopilot Rule

When the user says /stack next, treat it as an autopilot command, not a checkpoint request.

• Prefer the most recent active session for the current project before initializing a new stack.
• If an active session exists:
  1. run next or status to confirm the true active step
  2. finish that step
  3. run record
  4. immediately follow the CLI response
  5. continue until the stack completes or a real blocker occurs
• Do not stop after each successful step to ask for permission to continue.
• Send short progress updates, but keep the session moving.

Only stop for:

• missing required artifact or context that cannot be inferred from disk
• contradictory session state after re-checking with status
• failed record that cannot be corrected locally
• a true product decision that is not covered by existing plan/spec/proposal artifacts

Run Modes

Once launch is approved, do not ask again until completion or blocker.

• foreground = current agent/session drives the run in-process through the returned CLI contract
• background = detached execution managed by prompt-stack and lev exec
• chained runs continue across stacks without stopping
• blockers come from merged runtime + project + invocation policy

Hard Rules

• Respect the chosen run mode after launch approval.
• Never ask “should I continue?” after go.
• Stack session serialization stays single-writer per session.
• Git checkpoint serialization stays single-writer per repo through .lev/steering/checkpoint.lock.
• Use the prompt-stack checkpoint helper, not ad hoc repo-local scripts.

Checkpoint Mode

Use the canonical prompt-stack helper layer when multiple runs share one repo and need serialized git checkpoints:

cd $HOME/digital/leviathan

prompt-stack checkpoint-lock status --project-dir <path>
prompt-stack checkpoint-lock acquire --project-dir <path> --owner-tag <tag>
prompt-stack checkpoint-lock release --project-dir <path> --owner-tag <tag>

Rules:

• concurrent prompt-stack runs keep runtime state under .lev/steering/runs/<run-id>/...
• git checkpointing is serialized through .lev/steering/checkpoint.lock
• dirty worktrees are warning-only, not automatic blockers

Session Serialization

A stack session is single-writer state.

• Never run next, record, status, or validate in parallel against the same session.
• Always use this order:
  1. status or next
  2. do the step work
  3. record
  4. status or next again
  5. repeat
• After any interruption, timeout, or aborted turn, re-run status and trust the session file over memory.

Dashboard Mode (/stack list)

When the user says /stack list, do NOT dump raw JSON. Instead:

1. Run cd $HOME/digital/leviathan && bun plugins/prompt-stack/src/cli.ts list
2. Parse the JSON response
3. Format as a grouped dashboard — one line per stack:

━━━ CORE-LEV (9 stacks) ━━━

  handoff              (4 steps)  gather → distill → write → validate
  lev-align            (4 steps)  load north-star → inspect → classify drift → propose action
  plan-align-gates     (5 steps)  governance → refresh plan → align canon → gates → emit proposal
  research-propose-spec(3 steps)  research evidence → proposal synthesis → spec crystallization
  validate-emit-learn  (4 steps)  review artifacts → validate → emit → capture learnings
  work-deepen          (5 steps)  decompose → discover skills → parallel research → review → brief
  work-full-lifecycle  (9 steps)  governance → plan → align → gates → proposal → review → validate → emit → learn
  work-proof-loop      (3 steps)  handoff update → spec delta → validation report
  robot-readme-harden  (3 steps)  robot-mode hardening → cli error tolerance → readme sync

━━━ JEFF (5 stacks) ━━━

  jeff-delivery-loop   (10 steps) primer → ideation → synthesis → premortem → robot → cli → readme → review → e2e → commit
  jeff-ideation-risk   (4 steps)  idea wizard → multi-model synthesis → premortem → readme
  jeff-primer-quality  (4 steps)  deep primer → e2e validator → peer review → bug hunter
  jeff-quality-gauntlet(5 steps)  bug hunter → peer review → e2e → deploy verify → de-slopify
  jeff-robot-launch    (4 steps)  robot-mode → idea wizard → readme → git commit

━━━ SDLC (7 stacks) ━━━

  sdlc-deepen-plan     (3 steps)  decompose → parallel research → synthesize
  sdlc-exec-validate   (3 steps)  implement → fitness functions → route verdict
  sdlc-handoff-close   (4 steps)  verify handoff → extract decisions → approve → process
  sdlc-handoff-rollup  (4 steps)  scan unprocessed → pick one → extract all → process items
  sdlc-hygiene         (4 steps)  scan staleness → check alignment → propose updates → report
  sdlc-plan-to-beads   (5 steps)  extract items → create issues → review pass 1 → pass 2 → pass 3
  sdlc-spec-lifecycle  (5 steps)  validate completeness → check transition → quality gates → route → approve/iterate

━━━ OTHER ━━━

  knowledge-6r         (6 steps)  record → reduce → reflect → reweave → verify → rethink
  drift-detection      (5 steps)  positive drift → negative → undocumented → stale → alignment
  spec-driven-dev      (5 steps)  detect divergence → classify → push/pull → arbitrate → sync
  cass-cm-audit        (6 steps)  health check → memory scan → playbook audit → gap analysis → reindex → report
  research-worker      (3 steps)  orient → research rounds → synthesize

4. After the full list, add context-aware suggestions:

━━━ SUGGESTED FOR CURRENT CONTEXT ━━━

Based on: {describe what you detected — active handoff, case folder, recent work}

  (1) {stack-id}  — {why this stack fits right now}
  (2) {stack-id}  — {why}
  (3) {stack-id}  — {why}

  (s)how <id>  — Inspect a specific stack
  (r)un <id>   — Init a session immediately

Context detection rules:

• .lev/osint/ exists → suggest research-worker, work-deepen
• .lev/pm/handoffs/ has unprocessed handoffs → suggest sdlc-handoff-rollup
• Conversation mentions case/legal/research → suggest jeff-ideation-risk, research-propose-spec
• Recent code changes or implementation work → suggest jeff-quality-gauntlet, sdlc-exec-validate
• Session ending or handoff requested → suggest handoff, validate-emit-learn
• First time seeing the stacks → suggest work-deepen, jeff-ideation-risk, sdlc-exec-validate

Commands

cd $HOME/digital/leviathan

# List (then format as dashboard above)
bun plugins/prompt-stack/src/cli.ts list

# Inspect (show steps, schemas, no hidden prompts)
bun plugins/prompt-stack/src/cli.ts show <stack-id>

# Init session
bun plugins/prompt-stack/src/cli.ts init --stack <stack-id> --project-dir <path>

# Get active step prompt
bun plugins/prompt-stack/src/cli.ts next --session <id> --project-dir <path>

# Record step output (validates and advances)
bun plugins/prompt-stack/src/cli.ts record --session <id> --step <step-id> --report <path> --project-dir <path>

# Check progress
bun plugins/prompt-stack/src/cli.ts status --session <id> --project-dir <path>

# Validate completed session
bun plugins/prompt-stack/src/cli.ts validate --session <id> --project-dir <path>

Workflow

1. /stack list → dashboard with suggestions
2. User picks a stack (by number or name)
3. init --stack <id> → creates session
4. next --session <id> → reveals active step prompt
5. Agent does the work, writes report matching schema
6. record --session <id> --step <step-id> --report <path> → validates + advances
7. Repeat 4-6 until all steps complete
8. validate --session <id> → checks lifecycle integrity

For /stack next, steps 4-8 are one continuous autopilot loop. Do not pause between them unless blocked.

Skill Discovery Integration

Before executing the first step of any stack, run lev-skills <domain> to surface relevant skills from the skills-db. The domain is inferred from context:

• Stack family legal-osint → lev-skills legal
• Stack family sdlc → lev-skills sdlc
• Stack family jeff → skip (jeff prompts are self-contained)
• Active case folder detected → lev-skills with case domain keywords
• Unknown → lev-skills with the stack's title as query

Inject discovered skill names + one-line descriptions into the first step's context. Workers then know what specialized skills are available during execution.

Operator Notes

• Always use absolute paths for --project-dir and --report
• Session state lives in <project>/.lev/steering/sessions/
• record REJECTS invalid reports — fix and re-submit
• Report format: YAML frontmatter + required sections (varies by stack)
• If part of active $work handoff, keep planning there — stack is execution only
• If the stack is being used inside active $work, stack state is not the only artifact. You must also persist the corresponding disk-backed work template artifact:
  • planning / execution slice -> .lev/pm/plans/plan-*.md
  • research lane -> .lev/pm/reports/report-*.md
  • shaping / architecture lane -> .lev/pm/designs/*.md, .lev/pm/proposals/*.md, or .lev/pm/specs/*.md
• For large or multi-workstream implementation planning, route through writing-plans and split independent workstreams into separate plan files. Do not keep one thin coordination plan where the actual execution detail belongs in child plans.
• A good default: stack session for runtime state, work artifact for durable repo state, handoff for continuity. All three should agree.
• For subagent orchestration: plugins/prompt-stack/references/subagent-contract.md

References

• CLI: plugins/prompt-stack/src/cli.ts
• Catalog: plugins/prompt-stack/prompts/prompt-stack.prompts.yaml
• Subagent contract: plugins/prompt-stack/references/subagent-contract.md
• FlowMind path (optional): plugins/prompt-stack/flows/prompt-stack.flow.yaml