hashan-77/Stitch-QA-Code-Agent

Stitch QA Source Quality and Repair Assurance Intelligence Analyst - Code Agent

Source Quality Intelligence Analyst and Repair Assurance Intelligence Analyst are the two Agent 3 responsibilities in the Stitch QA workflow. Both are served under the original service name: Code Agent.

Agent 3 is intentionally split into two logical parts while keeping the same service identity. Part 1 reviews supported application source files for evidence-backed risks. Part 2 turns approved repair contracts into code-level implementation guidance without generating or applying unvalidated patches.

Role in Stitch QA

Part 1:
Project source files
        ↓
Source Quality Intelligence Analyst - Code Agent
        ↓
Source findings and source risk

Part 2:
Stitch Repair Contracts
        ↓
Repair Assurance Intelligence Analyst - Code Agent
        ↓
Target files, target symbols, change boundaries, verification guidance

Agent 3 supports both source-level review and repair-assurance guidance, but it does not automatically modify user projects.

Agent 3 Part 1: Source Quality Intelligence Analyst

The Source Quality Intelligence Analyst reviews supported application source files for deterministic and bounded AI-assisted source-quality signals.

Responsibilities

• —Discover supported application source files.
• —Exclude test files from application-source review.
• —Detect evidence-backed source risks.
• —Report findings with file paths, lines, severity, evidence, impact, and recommendations.
• —Preserve current knowledge flags.
• —Communicate limitations clearly when source review cannot prove full correctness.

Typical findings

• —Missing automated tests in source-only Python projects.
• —Potential unguarded divisor behavior.
• —Potential empty-collection divisor behavior.
• —Other evidence-backed source-quality risks supported by deterministic checks and validated reasoning.

Agent 3 Part 2: Repair Assurance Intelligence Analyst

The Repair Assurance Intelligence Analyst receives Stitch Repair Contracts from Agent 2 and converts them into bounded code-level guidance.

Responsibilities

• —Link guidance back to repair contracts.
• —Identify target files and target symbols.
• —Preserve Agent 2 change boundaries.
• —Explain implementation intent.
• —Provide targeted and regression verification guidance.
• —Avoid producing unvalidated patches.
• —Return NOT_REQUIRED when no repair contract exists.
• —Return NO_CODE_CHANGE_REQUIRED when contracts are environment-only.

Inputs

Agent 3 may receive:

• —Application source files.
• —Source discovery metadata.
• —Stitch Repair Contracts.
• —Runtime evidence summaries.
• —Source review findings.
• —Project profile information.

Outputs

Agent 3 may produce:

• —Source findings.
• —Source risk level.
• —Source release recommendation.
• —Repair assurance guidance.
• —Target files.
• —Target symbols.
• —Change boundaries.
• —Verification plans.
• —Patch validation status.
• —Current knowledge requirement flag.

AI Usage

Agent 3 uses deterministic validation as a safety layer and bounded local model reasoning as an optional support layer.

If model output is invalid, incomplete, ungrounded, or not contract-safe, Agent 3 falls back to deterministic validated behavior. This protects the workflow from unsafe or overbroad code recommendations.

Safety Rules

• —Do not automatically modify project source code.
• —Do not emit an unvalidated patch as a trusted repair.
• —Do not broaden Agent 2 repair boundaries.
• —Do not invent unsupported source findings.
• —Do not recommend code changes for environment-only contracts.
• —Do not claim complete correctness from source review alone.
• —Keep auto_apply false in repair-assurance outputs.

Typical Outcomes

Service Summary