/discovery — Full Discovery Phase Orchestrator

YOU MUST EXECUTE THIS WORKFLOW. Do not just describe it.

Strict Delegation Contract (default)

Discovery delegates to

/brainstorm

/design

/research

/plan

/pre-mortem

Skill(skill="<name>", ...)

Anti-pattern to reject: inlining

/research

/plan

/pre-mortem

../shared/references/strict-delegation-contract.md

--quick

--skip-brainstorm

--interactive

--auto

--no-scaffold

See

.agents/learnings/2026-04-19-orchestrator-compression-anti-pattern.md

DAG — Execute This Sequentially

mkdir -p .agents/rpi
detect bd and ao CLI availability

Run every step in order. Do not stop between steps.

STEP 1  ──  if not --skip-brainstorm AND goal is vague (<50 chars or vague keywords):
              Skill(skill="brainstorm", args="<goal>")
              Use refined goal for subsequent steps if produced.

STEP 1.5 ── if PRODUCT.md exists in repo root
              AND goal appears to be a feature or capability
              (not a bug fix, chore, or docs task — i.e., goal does NOT start with
               "fix", "chore", "docs", "typo", "bump", "update dep", "lint", "format"):
                Skill(skill="design", args="<goal> [--quick]")
                FAIL verdict? → output <promise>BLOCKED</promise>, stop (design is a blocking product-alignment gate).
              Skip silently if PRODUCT.md does not exist or goal is non-feature.

STEP 2  ──  if ao available:
              ao search "<goal keywords>" 2>/dev/null || true
              ao lookup --query "<goal keywords>" --limit 5 2>/dev/null || true
              Assemble ranked packet: compiled planning rules + active findings
              + unconsumed high-severity next-work items. Carry forward as context.
              For each returned learning, check applicability to the goal. If applicable,
              cite by filename and record: ao metrics cite "<path>" --type applied 2>/dev/null || true
              (orchestrator-owned: this knowledge retrieval is intentionally inline CLI,
               not a Skill() delegation. Do NOT expand into a separate /research call.)

STEP 3  ──  Skill(skill="research", args="<goal> [--auto]")
              Pass --auto unless --interactive. Output lands in .agents/research/.
              After: identify applicable test levels (L0-L3) for downstream /plan.

STEP 4  ──  Skill(skill="plan", args="<goal> [--auto]")
              Pass --auto unless --interactive.
              After: extract epic-id, auto-detect complexity from issue count
              (1-2 → fast, 3-6 → standard, 7+ → full) unless --complexity override.

STEP 4.5 ── if --no-scaffold is NOT set (alias: --no-lifecycle, deprecated)
              AND plan output contains new project/module creation
              (keywords: scaffold, new project, bootstrap, init, create module,
               new package, new service):
                detect language from plan context or existing project files
                Skill(skill="scaffold", args="<detected-language> <project-name>")
                Scaffold output becomes input context for pre-mortem.
              Skip if: --no-scaffold flag (or deprecated --no-lifecycle), no new project/module detected in plan.

STEP 5  ──  Skill(skill="pre-mortem", args="<plan-path> [--quick]")
              Use --quick for fast/standard. Full council for full.
              PASS/WARN? → continue to STEP 6
              FAIL?      → re-plan with findings, re-run pre-mortem (max 3 total)
                           Still FAIL after 3? → output <promise>BLOCKED</promise>, stop

STEP 6  ──  Write execution-packet.json (latest alias) + per-run packet archive
              to .agents/rpi/ and .agents/rpi/runs/<run-id>/ when run_id exists.
              Include plan_path, test_levels, ranked_packet_path, epic-id, complexity.
              ao ratchet record discovery 2>/dev/null || true
              Output <promise>DONE</promise>

That's it. Steps 1→1.5→2→3→4→5→6. No stopping between steps.

Setup Detail

State:

discovery_state = {
  goal: "<goal string>",
  interactive: <true if --interactive>,
  complexity: <fast|standard|full or null for auto-detect>,
  skip_brainstorm: <true if --skip-brainstorm or goal is >50 chars and specific>,
  epic_id: null,
  attempt: 1,
  verdict: null
}

CLI dependency detection:

if command -v bd &>/dev/null; then TRACKING_MODE="beads"; else TRACKING_MODE="tasklist"; fi
if command -v ao &>/dev/null; then AO_AVAILABLE=true; else AO_AVAILABLE=false; fi

Gate Detail

Discovery has two blocking gates.

• STEP 1.5 (design gate):

  FAIL

  blocks discovery immediately for feature/capability goals when

  PRODUCT.md

  exists.
• STEP 5 (pre-mortem gate): Max 3 attempts with plan→pre-mortem retry loop.
  • PASS/WARN: Store verdict, apply any required pre-mortem hardening back into the plan issues or file-backed task specs, then proceed to STEP 6.
  • FAIL: Log

    "Pre-mortem: FAIL (attempt N/3) -- retrying plan with feedback"

    . Re-invoke

    /plan

    with findings context, then re-invoke

    /pre-mortem

    . After 3 total failures: output

    <promise>BLOCKED</promise>

    , stop.

Step Detail

STEP 1 (brainstorm): Skip if

--skip-brainstorm

improve

better

something

somehow

maybe

.agents/brainstorm/

STEP 1.5 (design gate): Optional. Runs

/design

FAIL

PASS

WARN

STEP 2 (search history): Ranked packet assembly — match compiled planning rules, active findings from

.agents/findings/*.md

.agents/rpi/next-work.jsonl

STEP 3.1 (test levels): After research, determine L0-L3 applicability. External APIs/I/O → L0+L1+L2 min. Cross-module → add L2. Full subsystem → add L3. Record in

discovery_state.test_levels

STEP 4 (plan): After plan, record the exact

plan_path

bd list --type epic --status open

plan_path

.agents/rpi/execution-packet.json

tasklist

STEP 5 (pre-mortem): Pass the recorded

plan_path

/pre-mortem

STEP 5.5 (pre-mortem fix propagation): Before STEP 6, copy any required pseudocode fixes from the pre-mortem report into the affected plan issues or file-backed task specs. Workers read issue/task bodies, not the pre-mortem report.

STEP 6 (output): Write execution packet and phase summary per

references/output-templates.md

.agents/rpi/execution-packet.json

.agents/rpi/runs/<run-id>/execution-packet.json

run_id

plan_path

test_levels

ranked_packet_path

/crank

/validation

Flags

--auto

--interactive

/research

/plan

--interactive

--skip-brainstorm

--complexity=<level>

fast

standard

full

--no-budget

--no-scaffold

--no-lifecycle

--no-scaffold

Deprecation note: When Claude encounters

--no-lifecycle

on

/discovery

, treat it as

--no-scaffold

and mention the deprecation inline in the phase summary (e.g.,

"used deprecated --no-lifecycle, prefer --no-scaffold"

). This surfaces guidance in the RPI output without a runtime parser.

Quick Start

/discovery "add user authentication"              # full discovery
/discovery --interactive "refactor payment module" # human gates in research + plan
/discovery --skip-brainstorm "fix login bug"       # skip brainstorm for specific goals
/discovery --complexity=full "migrate to v2 API"   # force full council ceremony

Completion Markers

<promise>DONE</promise>      # Discovery complete, epic-id + execution-packet ready
<promise>BLOCKED</promise>   # Pre-mortem failed 3x, manual intervention needed

Troubleshooting

Read

references/troubleshooting.md

Reference Documents

• references/complexity-auto-detect.md — precedence contract for keyword vs issue-count classification
• references/idempotency-and-resume.md — re-run safety and resume behavior
• references/phase-budgets.md — time budgets per complexity level
• references/troubleshooting.md — common problems and solutions
• references/output-templates.md — execution packet and phase summary formats

See also: brainstorm, design, research, plan, pre-mortem, crank, rpi, scaffold