Sem

sem

Use the pi-sem tools as a semantic lens, not as a universal replacement for raw git diff.

Default decision tree

Choose the smallest useful tool first:

1. Focused understanding of one entity → sem_context

• Best for a single function, method, class, block, or config section
• Prefer this before reading a whole large file

1. Blast radius / affected tests / hidden dependents → sem_impact

• Use when reasoning about what could break
• Prefer scope=tests for test selection
• Prefer scope=all when validating broader impact

1. Structural inventory of a file → sem_entities

• Use before drilling into a suspicious file
• Good for large files and mixed code/config files

1. What changed across a commit/range/working tree → sem_diff

• Use for semantic summaries, entity counts, and review overviews
• Do not default to it when you only need exact patch details or a single entity

1. History / ownership of an entity → sem_log, sem_blame

• Use for regressions, archaeology, and ownership questions

Review workflow

For commit / branch / PR review:

1. Run sem_diff once for a semantic overview
2. Pick the riskiest changed entities
3. Run sem_impact on those entities
4. Run sem_context on the suspicious ones you need to understand deeply
5. Confirm final findings with raw git diff, read, or direct file inspection before citing line numbers

For snapshot / folder review:

1. Start with sem_entities
2. Use sem_context on the most relevant entities
3. Use sem_impact only after you identify something suspicious

Important caveats

• sem diff --format json is not always smaller than raw git diff
• sem may under-cover tests, assets, generated files, or non-semantic glue code
• Do not cite sem output alone as final evidence for line-level review comments
• If semantic coverage looks incomplete, fall back to raw git diff, read, grep, and file inspection

Good prompts / tool choices

• “What changed in this commit?” → sem_diff
• “What tests are affected by this function?” → sem_impact with scope=tests
• “Give me focused context for this class without reading the whole file.” → sem_context
• “What is in this Terraform file?” → sem_entities
• “How did this function evolve?” → sem_log

Anti-patterns

Avoid these habits:

• Running sem_diff repeatedly when sem_context would answer the question faster
• Using only raw git diff for entity counts or blast radius questions
• Treating sem_impact as infallible; verify surprising results with file reads and grep
• Writing review findings from semantic summaries without checking the actual changed code

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: tomsej
• Source: tomsej/pi-ext
• License: MIT

Install and usage instructions live in the source repository linked above.