Gaia Process (SDLC Controller)

When to use

Use this skill for every request to ensure docs/code/CI/skills stay in sync and work is delivered with the required quality gates.

All rules, gates, roles, and MCP tool signatures live in AGENTS.md. This skill is the procedural workflow only.

Inputs

• User request (feature/bug/refactor/docs)
• Current repo state (unknown until Repo Explorer runs)
• Available MCP tools: tasks_*, memory_*, self_improve_* (see AGENTS.md §7)

Step 0 — Load context (fast)

1. Read AGENTS.md (canonical rules, roles, gates, MCP tools, proof, Definition of Done).
2. Call memory_recall(project) and self_improve_list() to load prior context and lessons.
3. Identify existing conventions (Make targets, CI workflows, test folders, docker-compose).

Step 1 — Repo Explorer FIRST (always)

Delegate to Repo Explorer (SKILL: repository-audit) for a compact "Repo Survey":

• Stack(s) and build system
• /docs presence + gaps
• docs ↔ code drift signal
• CI presence + status
• lint/format tooling
• tests (unit/integration/e2e) status
• docker-compose status (if HTTP API)
• Makefile presence and targets
• skill drift signal (skills vs reality)

Repo Explorer stores discovered conventions via memory_remember(project, key, value). Orchestrator owns the actual MCP task graph.

Step 2 — Hard blockers (resolve before feature work)

If any of the following are true, create blocking tasks and fix autonomously first:

• Docs ↔ code drift exists
• CI missing or failing
• Skill drift exists (skills don’t match repo reality)
• HTTP API without docker-compose (and request involves use cases)

Drift resolution direction:

• Decide case-by-case; if unsure, default to docs.
• If “code wins” implies behavior/use-case change → apply use-case gates.

Step 3 — Build the Task Graph (orchestrator supremacy)

Call tasks_create(project, title, requiredGates[]) for all work:

• Foundations (CI, lint/format, docker-compose, Makefile targets)
• Docs/spec (create/derive/update — use templates in /docs/)
• Implementation
• Tests (unit + integration/e2e as required)
• Manual regression (labels only)
• QA Gatekeeper review

In-flight discovery:

• If you uncover TODOs, gaps, or new risks: call tasks_create immediately.
• “No TODO left behind”: turn TODOs into MCP tasks or blockers[].

Step 4 — Decide gates per task (explicit)

For each MCP task, set required_gates[] explicitly per AGENTS.md §10 Canonical gate vocabulary.

If tests/regression cannot be run:

• call tasks_flag_needs_input(project, id, questions[]) to block on human input
• do parallelizable work, but completion stays blocked

Step 5 — Execute with delegation

Delegate by intent:

• Repo survey → Repo Explorer
• Architecture / doc structure → Architect
• Implementation → Developer
• Test authoring → Tester
• Independent verification → Quality Gatekeeper

Subagent requests must be tight (inputs + expected output + constraints) to avoid context bloat.

Step 6 — Completion proof (MCP-enforced, low-context)

Call tasks_mark_done(project, id, changedFiles[], testsAdded[], manualRegressionLabels[]):

• changedFiles[] (paths)
• testsAdded[] (paths)
• manualRegressionLabels[] (labels like curl, playwright-mcp)

Do not paste logs; store only paths/labels. Store discovered conventions via memory_remember(project, key, value). Log lessons learned via self_improve_log(project, suggestion, category).

Step 7 — QA Gatekeeper review (veto)

Before final completion:

• Ask Quality Gatekeeper to verify:
  • blockers empty
  • gates satisfied
  • CI present + green
  • docker-compose present if HTTP API
  • use-case triggers handled correctly
  • proof args are complete and consistent If vetoed → create/fix tasks until approved.

Step 8 — Final response discipline

End with one short paragraph:

• docs touched
• code touched
• tests added/updated (paths)
• manual regression labels performed
• how to run locally (Make targets)

References (read/consult)

• AGENTS.md (repo constitution — canonical rules, MCP tools, gates)
• .github/agents/ (role definitions)
• .github/skills/ (procedures)
• Makefile (canonical local commands)
• .github/workflows/ (CI gates)