prism

README

Prism

Graph-ranked code context for AI coding agents.

Prism turns a task plus a few precise anchors into the code, callers, callees, tests, docs, and coverage gaps an agent needs to make a change safely.

Prism is not a better grep. Use rg/grep to find the first anchor. Use Prism to answer the follow-up questions that usually cost several file reads:

• What calls this?
• What does this call?
• Which tests define the contract?
• What else is in the blast radius?
• Which nearby exported functions have no direct test coverage?

The recommended agent path is now CLI text mode:

prism query "fix direct coverage gaps" --terms buildCoverageGaps --include graph,tests,coverage_gaps --format text
prism read internal/cli/commands.go --format text
prism lookup github.com/provasign/prism/internal/cli.invokeWithPersistentLedger --format text

--format text gives agents plain source-like context with short headers. Use lean or json only when automation needs structured output.

Grove is embedded in the Prism binary. There is no separate daemon, token, or grove_url setup in current releases.

---

Why Prism

Shell search gives pointers. Agents still have to chase those pointers by reading files, guessing test names, and manually reconstructing call paths.

Prism precomputes the project graph and lets the agent ask for relationships:

rg buildCoverageGaps internal/
  -> prism query "write tests for buildCoverageGaps" \
       --terms buildCoverageGaps \
       --include graph,tests,coverage_gaps \
       --format text

On this repository, five real maintenance scenarios were run both ways on 2026-06-07. Shell-only baselines used rg plus targeted sed reads; Prism used one CLI text command per scenario.

Scenario	Shell bytes	Prism CLI bytes	Context reduction
Init / CLI steering impact	19,970	12,818	35.8%
coverage_gaps precision	21,226	17,145	19.2%
CLI text/lean/json output formatting	15,820	14,198	10.3%
Session cache / savings ledger	33,134	19,922	39.9%
Release/version/install wiring	21,246	12,157	42.8%

The average reduction was 29.6% with one Prism command instead of 5-6 shell commands. The bigger correctness win is that Prism surfaces tests and coverage gaps proactively; shell-only workflows often discover those after CI fails.

Details: Real-World Prism CLI Benchmark.

---

How It Works

Task + anchor terms
      |
      v
Embedded Grove index
  - symbols
  - call edges
  - dependency edges
  - test edges
      |
      v
Prism ranking
  - graph distance
  - semantic similarity
  - recency
  - test relevance
  - edit frequency / learned weights
      |
      v
Budgeted text context
  - target symbols
  - callers/callees
  - tests
  - docs
  - coverage_gaps

Prism saves context by replacing repeated manual graph chasing with one graph-aware query. The CLI also persists its ledger and file-read cache between commands, so repeated unchanged file reads can return a short SHA pointer.

---

Installation

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/provasign/prism/main/install.sh | bash

# Windows PowerShell
irm https://raw.githubusercontent.com/provasign/prism/main/install.ps1 | iex

# Pin a version
VERSION=v0.5.6 curl -fsSL https://raw.githubusercontent.com/provasign/prism/main/install.sh | bash

The installer writes prism to ~/bin by default. Set INSTALL_DIR=/usr/local/bin or another directory to override.

Build from source:

make build
make test
make install

---

Quick Start

Run this once at the project root:

prism init .
prism index .

This writes:

• steering files such as AGENTS.md, CLAUDE.md, .cursorrules, .windsurfrules, .github/copilot-instructions.md, and others
• prism.yaml with the project profile and optional model settings

The generated agent instructions tell agents to use commands like:

prism query "trace the payment refund flow" --terms RefundPayment --include graph,tests --format text
prism query "find direct coverage gaps" --terms UpdatePayment,RequireScope --include graph,coverage_gaps --format text
prism read internal/payment/service.go --format text
prism lookup github.com/example/payflow/internal/payment.(*Service).RefundPayment --format text

Recommended agent workflow:

1. Locate the first anchor with rg, grep, or find.
2. Run prism query with the same anchor terms.
3. Use prism read for whole files only when needed.
4. Use prism lookup for one known function or method.
5. Treat coverage_gaps as a terminal structured output, not the start of manual cross-referencing.

CLI Reference

prism init [dir]
prism index [dir]
prism status [dir]

prism query <task> [dir] \
  --terms a,b,c \
  --include graph,tests,docs,coverage_gaps \
  --depth 2 \
  --format text

prism read <file> [dir] --format text
prism lookup <name> [dir] --format text
prism search <keyword> [dir] --format text
prism savings [dir]
prism compact [dir]
prism feedback --tool <name> --rating <0-5> [dir]
prism version

Output formats:

Format	Use
text	Default and recommended for agents
lean	Compact JSON without most metadata
json	Full metadata for tooling/debugging

---

Configuration

prism.yaml is intentionally small:

version: 1
profile: "default"

Optional keys:

model: "claude-sonnet-4-6"
grove_binary: "grove"
embeddings_backend: "tfidf"

Environment overrides include PRISM_MODEL, PRISM_PROFILE, PRISM_GROVE_BINARY, and PRISM_EMBEDDINGS_BACKEND.

---

Language Support

Prism delegates parsing and graph construction to embedded Grove.

Language	Extensions
Go	.go
TypeScript / TSX	.ts, .tsx
JavaScript / JSX	.js, .jsx, .mjs, .cjs
Python	.py
Java	.java
Rust	.rs
C / C++	.c, .h, .cc, .cpp, .hpp, ...
C#	.cs
PHP	.php, .phtml, ...

Markdown, YAML, JSON, shell scripts, Dockerfiles, Makefiles, SQL, GraphQL, and other non-code files are indexed as document symbols and can be requested with --include docs.

---

Benchmarks

Read benchmark numbers by transport:

• Prism CLI Real-World Benchmark: current recommended path; Prism CLI text mode vs shell-only on Prism itself.
• Payflow A/B Agent Benchmark: controlled baseline vs Prism benchmark, including the CLI text-mode round and coverage-gap precision fix.
• Agent Workflow Benchmark: historical MCP-focused workflow benchmark on Prism itself.
• Indexing/Read Benchmark: older performance run for indexing and repeated file reads.

Current practical summary:

• CLI --format text is the recommended default for shell-capable agents.
• Prism is strongest on graph/blast-radius/test/coverage-gap questions.
• Shell tools remain best for locating exact strings or filenames.

---

Troubleshooting

prism query returns nothing: run prism index . from the project root.

Agent still uses stale instructions: run prism init . and restart or reload the coding agent so it reads the updated steering files.

Wrong Prism binary: run command -v prism and prism version. Reinstall if the version is old.

macOS quarantine:

xattr -d com.apple.quarantine "$(which prism)"
codesign -f -s - "$(which prism)"