full-sdc

full-sdc (lean EXECUTE)

Orchestrator that runs the SDC for one story. Atomic skills + task files do phase work; CLI owns plan/verify/progress.

Not in scope: hub full-sdc (~2k LOC), worktree product registry, hub conductor adapters, product harvest trees, product deploy hosts.

Invocation

/full-sdc {story-path} [yolo|interactive]
# or
aiox sdc plan {story-path} --mode yolo

Default mode: interactive.

Automated/yolo dispatch additionally requires a positive explicit AIOX_MODEL_BUDGET_CEILING_USD; the story path and full child intent are validated before every model call.

Before each direct subagent/model spawn, materialize the exact outgoing prompt and context in isolated files and run this mandatory mechanical gate:

aiox sdc preflight "{story-path}" \
  --task "{phase-task}" \
  --budget-usd "$AIOX_MODEL_BUDGET_CEILING_USD" \
  --intent-file "{exact-child-intent-file}" \
  --context-file "{exact-child-context-file}"

Continue only on exit 0. Exit 5 is a governance rejection and the model or subagent must not be invoked. Pass the same bytes from these files to the child; never rebuild or enrich the prompt after preflight.

CLI (mechanical — always use)

# 1) Init / resume durable state
aiox sdc plan {story-path} --mode {yolo|interactive}

# 2) What to run next
aiox sdc next {story-path}

# 3) After each phase completes
aiox sdc verify {story-path} {phase} --mark

# 4) Inspect
aiox sdc status {story-id}

State file: .aiox/sdc/{story-id}/state.json (runtime, gitignored under .aiox/).

Phases: validate → develop → review → apply_qa_fixes (loop) → close

Phase map

Skill SOT: .aiox-core/development/skills/<name>/SKILL.md

EXECUTE loop (orchestrator MUST follow)

0. aiox sdc plan {story} --mode {mode}
1. LOOP:
   a. aiox sdc next {story}  → phase + skill path
   b. IF no next phase → DONE (status completed)
   c. Load skill SKILL.md + its task SOT without executing phase work
   d. Materialize the exact payload for inline or spawned phase execution
   e. Run `aiox sdc preflight` over that exact payload; HALT on failure
   f. Execute inline or spawn the phase only after preflight succeeds
   g. IF yolo: autonomous; IF interactive: report + pause on decisions
   h. aiox sdc verify {story} {phase} --mark
   i. IF phase=review and verdict FAIL:
        IF mode=interactive and no explicit fix approval:
             report FAIL and pause (do not invoke apply-qa-fixes)
        IF mode=yolo or explicit fix approval:
             CLI selects apply_qa_fixes automatically
             run apply-qa-fixes skill → verify apply_qa_fixes --mark
             IF apply_qa_fixes verify FAIL:
                  HALT and escalate human without returning to review or incrementing qgIterations
             IF apply_qa_fixes verify PASS:
                  CLI returns to review and increments qgIterations
                  IF the third re-review fails → HALT and escalate human
   j. ELSE IF verify FAIL → HALT (do not advance)
   k. ELSE continue loop
2. QA sets Done on PASS/CONCERNS/WAIVED; close-story only finalizes bookkeeping
3. Push/PR: hand off @devops — never push from this skill

Grok / multi-agent dispatch (preferred when available)

For each phase, spawn the matching persona (or run inline if spawn unavailable):

Before dispatch, declare the shared budget ceiling, bind {path}, materialize the exact child prompt/context, and run the aiox sdc preflight command above against those exact files. Pass those same bytes to the child; never rebuild or enrich the payload after preflight. After each subagent returns: run aiox sdc verify … --mark in the main session (orchestrator owns the lock).

Sequence Lock (soft + CLI)

1. Phases in order. No N+1 until verify of N passes (or skip).
2. Only QA review sets Status Done.
3. Close requires the approved QA verdict and an already-Done story.
4. QG loop max 3 (maxQgIterations in state).
5. Anti-self-validation: executor ≠ quality_gate.

Post-phase verification (CLI-enforced)

Modes

Explicitly non-ported

• Worktree auto-spawn / registry / GC
• Full auto-ACK matrix (v1 = CLI state + checklists)
• Product deploy targets

Strip checklist

• Skills invoke tasks only
• CLI First progress/verify
• No product harvest trees in protocol