Code Context - Task 0.3.2: Agent Code Context Injection

[bordumb]

Task 0.3.2: Agent Code Context Injection

Title: Include relevant code changes in agent prompts

Description:
When generating hypotheses, the agent should have access to recent code changes affecting the investigated asset. The hypothesis prompt should include a "Recent Changes" section with commit summaries.

Why: An agent investigating "why did nulls spike?" needs to know "yesterday's deploy added a LEFT JOIN that produces nulls for unmatched rows."

Acceptance Criteria:

• Hypothesis generation prompt includes code changes section
• Changes filtered to: last 14 days, affecting investigated asset or upstream
• Each change shows: commit hash, author, message, affected files, diff summary
• Maximum 5 most relevant changes included (sorted by recency × relevance)
• Agent can generate hypothesis category "code_change" with commit reference
• Evidence can link to specific commit: "Commit abc123 introduced LEFT JOIN"
• Final synthesis includes "Related Code Changes" section if relevant

Key Design Notes:

• Relevance scoring: exact asset match > upstream asset > path pattern match
• Diff summary: use LLM to summarize large diffs (separate prompt)
• Store commit reference in hypothesis metadata for linking

Key APIs:

• New internal function: get_relevant_code_changes(asset_id, lookback_days)
• Modify hypothesis prompt builder

Dependencies:

• Task 0.3.1 (git integration)

Risks + Mitigations:

• Risk: Too many changes overwhelm context → Mitigation: Strict relevance filtering, summarization
• Risk: False correlation (unrelated changes) → Mitigation: Agent trained to evaluate causality

Effort: M (4-5 days)

Designation: OSS

---

Task 0.3.3: PR/Commit Link in Investigation Output

Title: Deep link to source code in investigation results

Description:
When an investigation identifies a code change as the root cause, the output should include a direct link to the PR or commit in GitHub/GitLab/Bitbucket.

Why: "The root cause was commit abc123" is good. A clickable link to the PR with context is better.

Acceptance Criteria:

• Investigation synthesis includes links to relevant commits
• Links formatted for web (https://github.com/...) and CLI (commit hash + URL)
• PR link preferred over commit link when available
• Link appears in: CLI output, notebook output, markdown export, API response
• Metadata includes: PR number, PR title, PR author, merge date
• Support for GitHub, GitLab, Bitbucket URL formats

Key Design Notes:

• Store PR metadata during git sync (requires GitHub API calls)
• URL templates per provider

Key APIs:

• Extend code_changes table with pr_number, pr_url, pr_title
• Modify investigation response schema

Dependencies:

• Task 0.3.1, 0.3.2

Risks + Mitigations:

• Risk: PR lookup adds latency → Mitigation: Async enrichment during sync

Effort: S (2 days)

Designation: OSS