Code Discovery

Deep codebase research patterns for investigation and design workflows. Provides consistent approaches to exploring code, understanding recent changes, and finding precedents.

Core Principle

Understand the code before proposing changes. Evidence from actual implementation beats assumptions from documentation.

When to Use This Pattern

Always use for:

• Investigation work (finding what changed, what broke, when)
• Design work (understanding integration points, patterns)
• Research-driven analysis (precedents, similar features)

Pattern applies when:

• Feature touches existing code you haven't seen
• Need to understand recent changes (git log analysis)
• Looking for similar features to follow as precedent
• Need concrete file:line references for proposals

Git Log Analysis

Configurable Timeframes

Adapt git log timeframe based on investigation context:

# For suspected regressions: when did it break?
git log --oneline --since="{timeframe}" -- {path}

# Common timeframes:
# --since="2 weeks ago"  - Default for general investigation
# --since="1 week ago"   - Recent regression
# --since="1 month ago"  - Longer history for unclear issues
# --since="2026-01-15"   - Specific date from interview

Determine timeframe from context:

• Confirmed regression: User knows when it broke → use date
• Suspected regression: User reports "recently" → 1-2 weeks
• Unclear: No timeline info → 2-4 weeks for broader scan

Finding Relevant Commits

By file/directory path:

# Changes to specific file
git log --oneline --since="{timeframe}" -- path/to/file.py

# Changes to directory
git log --oneline --since="{timeframe}" -- src/module/

By commit message pattern:

# Find commits mentioning a keyword
git log --oneline --since="{timeframe}" --grep="{keyword}"

With file stats:

# Show files changed per commit
git log --oneline --stat --since="{timeframe}"

Reading Commit Details

# Full commit message and diff
git show {commit-sha}

# Just the commit message
git log -1 {commit-sha}

# Files changed in that commit
git show --name-only {commit-sha}

Code Reading Strategies

Use Context Files as Entry Points

Before reading code directly, consult .forge-context/context/ for structural information:

Context File	Use For
product.md	System overview and architecture
components.md	Component entry points

Why context files first? They summarize the project structure and point you to the right starting files.

Tracing Implementations from Interfaces

When you need to understand how something works:

1. Start with the interface (API endpoint, CLI command, public function)
2. Trace the call chain (what calls what, data flow)
3. Read critical paths (happy path + main error path)
4. Note assumptions (what must be true for this to work)

Example: Tracing a command

# 1. Find command definition (entry point)
grep -r "command.*{name}" src/

# 2. Read the handler function
# (use file:line from grep result)

# 3. Trace calls made by handler
grep -r "{function_name}" src/

# 4. Read implementations of called functions

Finding Similar Features for Precedent

When designing a new feature, find similar existing features:

By functionality:

# Find similar error handling
grep -r "raise.*Error" src/ | grep "{context}"

# Find similar API endpoints
grep -r "router.{method}" src/

By file naming patterns:

# Find related modules
find src/ -name "*{keyword}*"

Review similar feature:

• How did they structure it?
• What patterns did they use?
• What errors did they define?
• How did they test it?

Document precedent:

"Similar pattern exists in {file}:{line}. Follows approach X for Y. Will use same pattern for consistency."

Documentation Patterns

File:Line References

Always include file:line references for concrete findings:

Good:
The authentication check happens in auth/api.py:148

Bad:
The authentication happens somewhere in the auth module

Format: path/to/file.py:line or path/to/file.py:range

Why? Enables readers to verify findings and provides clickable links in many editors.

Confidence Levels for Findings

Not all findings are equally certain. Indicate confidence:

Level	Criteria	Example
High	Direct evidence in code/logs	"File X:Y raises AuthError"
Medium	Inferred from patterns	"Likely related to change Z"
Low	Possible based on architecture	"May involve component W"

Usage:

| Finding | Source | Confidence |
|---------|--------|------------|
| Auth check at api.py:148 | Direct code read | High |
| Related to recent token changes | Git log pattern | Medium |
| May affect mobile client | Architecture docs | Low |

Assumptions vs. Facts

Clearly separate what you observed from what you infer:

Facts (observed directly):

• "The git log shows commit abc123 changed file.py"
• "Function foo() at bar.py:42 raises ValueError when x < 0"

Assumptions (inferred, needs verification):

• "This likely broke when the auth refactor landed"
• "The mobile client probably uses this endpoint"

Mark assumptions explicitly:

• "Assumption: ..." (needs verification)
• "Hypothesis: ..." (testable claim)
• "Likely: ..." (inferred from context)

Integration with Other Skills

With evidence-based-analysis

Code discovery feeds evidence-based analysis:

1. Code discovery finds the facts (commits, code, patterns)
2. Evidence-based analysis structures findings for presentation
3. Both emphasize evidence over speculation

With systematic-debugging

For investigation work, code discovery supports debugging:

• Phase 1 (Root Cause): Git log + code reading to find cause
• Phase 2 (Pattern): Search for similar errors in codebase
• Phase 3 (Hypothesis): Code understanding informs predictions
• Phase 4 (Fix): Know what to change and why

Common Patterns and Anti-Patterns

Good: Progressive Exploration

1. Read context files first (project summary)
2. Use git log to narrow timeframe
3. Read specific files identified by log
4. Trace from entry points to implementations
5. Document findings with file:line references

Good: Concrete References

"The error handling in auth/api.py:148-150 catches all exceptions, masking the specific validation error from line 142."

Bad: Vague Assumptions

"The auth module is probably broken somewhere."

Bad: Skipping Context

Directly grepping the codebase without reading project context files first. Wastes time exploring the wrong areas.

Bad: Unbounded Exploration

Reading every file without a search strategy. Use git log and context files to narrow scope.

Summary

Phase	Tool	Purpose
Context	.forge-context/context/	Start at right place
Timeline	Git log	Find what changed and when
Deep dive	Code reading	Trace implementations
Precedent	Search patterns	Find similar features
Documentation	file:line refs	Verifiable findings

Key principle: Evidence-driven exploration. Read code, trace history, document findings with references. Avoid speculation — verify assumptions by reading the actual implementation.