All 32 stages on one spine, in graph order. Every card names what it reads, what it writes, the agent that leads it, and the condition that decides whether it runs at all.
⓪ Initialization — bootstrap workspace & state
workspace-scaffold
create intent record directory
0.1
Outputs
intent record dir
orchestrator
ALWAYS · auto-proceed (no gate)
workspace-detection
scan project type & tech stack
0.2
Outputs
aidlc-state.md
orchestrator
ALWAYS · auto-proceed (no gate)
state-init
initialize audit & state machine
0.3
Outputs
aidlc-state.mdaudit/ shards
orchestrator
ALWAYS · auto-proceed (no gate)
① Ideation — validate initiative, capture intent, define scope
intent-capture
frame the initiative
1.1
Outputs
intent statementstakeholder map
product-agentarchitect (support)
ALWAYS
market-research
competitive analysis, build-vs-buy
1.2
Outputs
competitive analysisbuild-vs-buy
product-agent
CONDITIONAL — skip for bugfix, refactor, infra, security-patch
feasibility
constraints & RAID log
1.3
Outputs
feasibility assessmentconstraint registerRAID log
architect-agentaws-platform, compliance
CONDITIONAL
scope-definition
define boundaries & backlog
1.4
Outputs
scope definitionintent backlog
product-agentdelivery (support)
ALWAYS
team-formation
mob composition & capacity
1.5
Outputs
team assessmentmob composition plan
delivery-agent
CONDITIONAL — skip for poc, bugfix, refactor
rough-mockups
wireframes & user flows
1.6
Outputs
wireframesuser flowsconcept deck
design-agentproduct-lead
CONDITIONAL — skip if no UI
approval-handoff
decision log & proceed gate
1.7
Outputs
initiative briefdecision log
delivery-agentproduct (support)
ALWAYS
▸ Verification Gate: Ideation → Inception
② Inception — elaborate requirements, design architecture, plan delivery
reverse-engineering
analyze existing codebase
2.1subagent
Inputs
existing codebase
Outputs
business-overview.mdarchitecture.mdcode-structure.md+6 RE artifacts
After delivery-planning produces the bolt-plan.md, stages 3.1–3.5 execute as Bolts. Bolt 1 is the walking skeleton (always gated). Then the ladder prompt fires once: continue autonomously or gate every Bolt. Independent Bolts run in parallel batches. Stages 3.6–3.7 run once at the end across all units.
build-and-test
run full test suite & quality report
3.6
Inputs
all generated code
Outputs
test resultsquality report
quality-agentdevsecops (support)
ALWAYS — once at end of all Bolts
ci-pipeline
CI config & quality gates
3.7
Inputs
code artifactstest results
Outputs
CI configquality gates
pipeline-deploy-agent
CONDITIONAL — once at end
▸ Verification Gate: Construction → Operation
ADVERSARIAL REVIEW Reviewers are verifiers (v2.4)
Every declared reviewer runs under an explicit adversarial contract: its job is to refute the artifact, not confirm it. Findings must ground in machine-checkable evidence (validation tools, tests, acceptance criteria). A PreToolUse hook enforces per-unit read-scope bounds deterministically — reviewers cannot reach into sibling units' construction paths.
AGENT TIERS Tiered delegation model (v2.3)
Agents declare a work-shaped tier: instead of raw model pins. Judgment (9 agents: architect, aws-platform, compliance, composer, design, developer, devsecops, product, quality) — inherits session model. Balanced (2 reviewers: product-lead, architecture-reviewer) — Sonnet-class. Templated (3 planners: delivery, pipeline-deploy, operations) — Sonnet + reduced effort. Teams can cap tiers via tier_cap: in memory frontmatter.
PLUGIN MECHANISM Extensible stage graph (v2.3)
Plugins add stages, scopes, agents, and knowledge from plugins/<name>/. A compose hook merges the plugin into an install — new stages join the graph, contributions extend existing stages additively. select-plugins controls which plugins are active. Core is the implicit aidlc plugin and can itself be deselected for plugin-only installs.