---ci---
project: ci
phase: 1
milestone: v0.8
status: in_progress
decisions:
- id: D-022
decision: Validate BackendResult at boundary with Zod schema
rationale: External backend output is untrusted; runtime validation prevents corrupt commit streams
confidence: 0.92
- id: D-023
decision: opencode parseResult returns success:false on malformed JSON
rationale: Silent success:true on parse failure masks backend errors; fail loudly instead
confidence: 0.95
requirements:
covered: [FIX-02, FIX-03]
---/ci---
FIX-02: Add Zod BackendResultSchema and validateBackendResult() in
backends/types.ts. backendResultToAgentResult() in base.ts now validates
before passing through. Invalid results produce success:false with error
detail. Path traversal protection: artifact paths with '..' or leading '/'
are rejected.
FIX-03: opencode.ts parseResult() no longer defaults to success:true when
JSON parsing fails entirely. Both the inner parse error and the no-JSON
match case now return emptyBackendResult() with descriptive error messages.
---ci---
project: ci
phase: 1
milestone: v0.8
status: in_progress
decisions:
- id: D-021
decision: 6-phase wave-ordered vertical slices for v0.8
rationale: Each phase independently demoable; critical fixes first
confidence: 0.90
requirements:
covered: [FIX-01, FIX-06]
---/ci---
FIX-01: Replace 5 hardcoded phase=1 literals in orchestrator.ts mechanical
execution path with this.pipelineState!.current_phase. The orchestrator
correctly tracks current_phase but commits always embedded literal 1.
FIX-06: Replace getDecisions() redundant double-fetch with single
getRecentCommits(50) call, delegating to existing getDecisionsFromCommits().
Old code called getRecentCommits(50) once per grep match entry (O(N*M)
when it should be O(1)).
---ci---
phase: 2
milestone: v0.6
status: execute
decisions:
- id: D-001
decision: Pass GitAgentContext to agents instead of bare AgentContext
rationale: Agents need git-native context (gitContext, gitBranch, ciFiles, milestone) to operate autonomously
confidence: 0.95
- id: D-002
decision: Implement multi-phase iteration with totalPhases derived from ROADMAP.md
rationale: Milestones can span multiple phases; orchestrator must advance through all of them
confidence: 0.90
- id: D-003
decision: Add executeStageWithRecovery with retry + plan revision + escalation
rationale: Robust error recovery requires multiple fallback levels before giving up
confidence: 0.85
- id: D-004
decision: Add timer-to-escalation mapping in EscalationProtocol for proper cleanup
rationale: resolveEscalation must clearTimeout for the corresponding timer to prevent resource leaks
confidence: 0.90
- id: D-005
decision: Add dispose() to EscalationProtocol called in orchestrator finally block
rationale: Ensures all timers are cleaned up on orchestrator exit regardless of outcome
confidence: 0.95
- id: D-006
decision: Add mechanical TEST stage fallback running npm test via execSync
rationale: When no backend is available, tests can still be run mechanically
confidence: 0.85
---/ci---
Add IntelligenceBackend abstraction with two categories:
- LLMBackend (OllamaLocal, OllamaCloud): CI runs tool loop, provides tools, constructs prompts
- AgentBackend (Opencode): agent runs own tool loop, CI serializes request
Refactor all 18 agents from hardcoded stubs to persona loaders that delegate
to the active backend or fail honestly when no backend is available.
Refactor OrchestratorAgent.executeStage() from monolithic switch to agent
delegation via STAGE_AGENT_MAP for intelligent stages (research, plan, execute,
verify), with mechanical stages (specify, clarify, complete) staying inline.
Wire CLI commands with --backend flag and auto-detection (opencode →
ollama-local → ollama-cloud). Harden rollback/ship with real git operations.
No command returns fake success.
Implements the full PRD for CI - a fully autonomous AI-driven software
engineering harness derived from Learnship's architecture.
Core components:
- CI Orchestrator agent with autonomous pipeline (SPECIFY → CLARIFY →
RESEARCH → PLAN → EXECUTE → VERIFY → COMPLETE)
- Decision Engine with confidence thresholds (high/medium/low)
- Clarify Phase with question budget and default acceptance
- Escalation Protocol with timeout auto-proceed
- Audit Trail system (.ci/audit/) for post-hoc review
- Error Recovery with retry, plan revision, and rollback
18 agents (all Learnship agents + Orchestrator):
- Autonomous behavioral modifications per PRD §7.1
- Agent registry with factory pattern
11 CLI commands:
- ci init, ci run, ci quick, ci debug, ci verify
- ci review, ci status, ci audit, ci clarify
- ci rollback, ci ship
4-layer verification system:
- Structural, Behavioral, Security, Code Quality
3 autonomy levels: full, supervised, guided
Compatible with Learnship artifact schemas (.planning/)
Jon Chery
grimacing