hologram

Hologram gives coding agents a compact map of a repository before they start exploring it. The map keeps exact source paths, public callables, named fields, relationships, project-internal calls, constants, useful private names, and pointers to tests, tools, and benchmarks in the context files agents already read.

There is no LLM, query step, or separate index service. Generation runs locally and is deterministic for the same sources, Hologram build, settings, runtime, and parser versions.

Quick start

Hologram requires Python 3.11 or newer. For a polyglot repository, install it with all tree-sitter grammars:

python3 -m pip install "hologram-map[grammars]"
hologram init --root .

init builds the map, embeds it in the repository's agent context files, and installs hooks that refresh it after commits, merges, and checkouts. Content outside Hologram's managed block is left untouched.

Python-only repositories can use the dependency-free base package:

python3 -m pip install hologram-map

You can also download the single-file hologram.pyz from the latest release:

python3 hologram.pyz init --root .

If a parser is missing, an interactive run can create a nearby .venv and install only the grammars it needs. Non-interactive runs print the equivalent install command instead.

What the map looks like

An abridged Java map looks like this:

# hologram ·.java
· C/R/I{fields} · f(args):Ret > calls · ×0=unused · !E=throws · p{a,b}s=pas,pbs · ←A|B=implementors
src
 App(C) ×0
  main(args) ×0 > PricingEngine,evaluate,{Order,Item}Id.of
 engine
  PricePort(I) ←PricingEngine
   quoteFor(order):Quote ×0
  PricingEngine(C{basePrices})
   evaluate(order,items):Quote !UnknownItem > UnknownItemException,Quote
  Quote(R{order,totalCents})
 ids
  {ItemId,OrderId,UserId}(R{value})
? tests
 src/test
  PricingEngineTest
· 186 LOC · input 1,051 · output 233 tokens · state 0123456789ab

The second line is generated with the map and explains its applicable core notation. Section-specific layout such as ? tests is described below. The essentials are:

• The tree mirrors the repository and names exact source files. Similar files may be grouped losslessly with braces.

• A title line such as # hologram ·.py declares the extension the map would otherwise repeat on every leaf: a bare leaf carries it, and a leaf that states an extension has exactly the one it states. Grouped landmarks ({ItemId,OrderId}(R{value})) take the declaration too. A node with other nodes below it is a directory; otherwise it names a file. That rule reads a directory whose files are all grouped landmarks as a file — about one node in fifty on the corpora it was measured against — but never silently: the reconstructed path simply does not exist, and the directory does. The declaration is made only where it costs less than the repetition, so small or mixed corpora carry none.

  # hologram ·.py
  · C/R/I{fields} · f(args):Ret > calls · ×0=unused · !E=throws
  app
   main() ×0 > price_order,OrderId,ItemId
   price_order(order,items):int
  models
   ItemId,OrderId,UserId(R{value})
  

  Here app is app.py and models is models.py; both are files, because neither has a node below it and neither names a file of its own.

• f(args):Ret > calls shows a callable, its return type, and retained internal calls. ← shows implementors. A call target is named relative to the caller: it carries only what the caller's own path, file and owner do not already say, so total is in the same class, order.py:Order.total is a sibling file, and a full path means another part of the tree.

• ×0 means no static project reference was found; !E means a callable throws. These are navigation hints, not correctness claims. Code a framework constructs or calls — routes, listeners, DI components and @Bean factories, ORM entities, lifecycle callbacks — is exempt, because static fan-in cannot see the container.

• Deselecting params keeps each signature's arity and drops only the argument names: place(order,items) becomes place(_,_), which is the same placeholder the map already uses for an argument no extractor could name.

• ? tests uses the same path-compressed tree as the source section and retains a compact, reconstructable landmark for every detected test file, so an agent can see where coverage lives before adding more. It names files, not cases: case names were the most expensive thing in the map — a third of it on a large corpus — for a fact that is read in the file the landmark points at. * marks reusable test support: helper classes, declared fixtures, and helper functions another test file uses, because those are what another test consumes without opening anything. The index names no call targets either. Separate tool/benchmark landmarks stay compact.

• The footer carries freshness and saved settings, plus what the map cost: input is the estimated token count of the scanned sources, output the estimated count of the map itself — the compression the block bought, stated inside the block. Both use the ceil(characters / 4) estimate the budget uses.

Common commands

build and init remember their settings in the map itself:

• --lang java,python limits extraction; clear it with --lang all.
• --target AGENTS.md selects context files; clear it with --target all.
• --budget 8000 sets an estimated digest-token target; clear it with --budget 0.
• --features calls,tests selects which fact classes the map carries; clear it with --features all.

Other useful options include --warn-tokens N, review --brief K, review --quiet-if-clean, and uninstall --keep-blocks. Run hologram <command> --help for the full CLI.

Choosing what the map carries

A budget decides how much fits. --features decides what is eligible in the first place — useful when a fact class is noise for your repository rather than merely expensive.

The package tree, type headers, and public signatures are the map's identity and always render; --features none leaves exactly those. A deselected class is absent at every level, so a budget can never restore it. The legend states only notation the map actually uses, so it shrinks with your selection.

hologram build --root . --interactive

--interactive (on build and init) prices a selection against the full map before writing it. It needs a terminal and fails with the equivalent --features flag otherwise, so hooks and CI never block on a prompt. Note that dropping calls can make a map larger: a private helper named only by a retained call chain has to reappear in the private inventory once that chain is gone.

Review changes

hologram review [REV] looks for near-duplicate callables, repeated test coverage, new public symbols with no static references, tests that name removed code, public API drift, and additions that appear misplaced. Findings are deterministic heuristics and advisory: a successful review still exits zero. The post-commit hook runs the same review against the previous commit.

Review scans Git-indexed files. Use git add -N path/to/file to include a completely untracked addition. JSON output contains project paths and symbols, so treat it as repository-derived data.

Token budgets

When a full map exceeds --budget N, Hologram starts with a compact semantic floor: retained business types, fields, and top-level signatures with exact file ownership, plus external entrypoints and tool/benchmark orientation. The test index is optional in full: at the floor it is absent entirely, and above it each test file landmark, suite/case label, and test-to-business edge is a separately restorable fact — a restored label always brings back the file landmark it renders on. Hologram then restores ranked whole facts, prioritizing tested and cross-file paths, widely used APIs, and breadth across files.

Below that floor is one more level, reached only when the semantic floor itself cannot fit: a structure-only map that states the same facts in project vocabulary alone. Return and parameter types, decorators and route paths, !throws, the ~N/×0 markers, {field} lists and type relations are language and framework words, and all of them go; the source tree, type names, and function names with their parameter names stay. On this repository that is 714 tokens down to 460. Ranked facts still compete for whatever slack remains above it. If even that cannot fit, Hologram warns and emits the smallest complete candidate instead of cutting facts in half.

The budget applies to the digest. hologram stats separately reports the wrapper, coaching text, and total managed-block estimate. --warn-tokens checks that managed block for build/init, while print checks its printed digest. Estimates use ceil(characters / 4) for deterministic planning; they are not tokenizer counts from a particular model.

Language support

• Application code: Java, Python, TypeScript, JavaScript, TSX/JSX, C#, Kotlin, Go, Rust, C, C++, PHP, Swift, Scala, Ruby, and Lua.
• Components and web assets: Angular, React, Vue, Svelte, HTML, and CSS.
• Project files: Bash/zsh scripts, Helm charts, and Makefiles.

Extraction depth varies by language. Where supported, maps retain types, fields, signatures, relations, resolved calls, constants, throws, routes, annotations, component usage, and framework entrypoints. Python uses the standard library's ast; Helm and Make support are also built in. Most other languages use optional tree-sitter grammars.

Context files and freshness

When --root is a Git worktree root, Hologram scans indexed source files. Otherwise, it walks the tree while pruning hidden, generated, vendored, and fixture directories.

Hologram recognizes the instruction files used by Claude Code, Codex, opencode, Jules, Zed, Amp, Gemini CLI, Qwen Code, Aider, GitHub Copilot, Cline, Cursor, Windsurf, Roo Code, JetBrains Junie, Continue, and Kiro. Auto-detection updates supported files that already exist; supported rule directories receive one managed rule file. If no target exists, Hologram creates CLAUDE.md. Use --target when you want an explicit destination.

The state stamp covers source content and Hologram's generator code. It does not fingerprint the Python runtime or installed grammar versions, so rebuild after upgrading that toolchain even if check still reports fresh.

Limits

Hologram is static context, not proof:

• ×0 means no static project reference was observed, not that code is safe to delete.
• A test edge means a test references a symbol, not that the behavior is correct.
• Test inventories are declaration-based. Cases named only through strings, framework DSL calls, or macros are not extracted as function/method names.
• Function bodies are summarized rather than embedded.
• Extraction depth varies, and review findings can produce false positives.

Development

Run the dependency-free test profile:

python3 tools/run_tests.py --profile core

With every optional grammar installed, run the complete profile with no allowed skips:

.venv/bin/python tools/run_tests.py --profile full

See CHANGELOG.md for release details. The benchmark guide describes matched map/control experiments and their privacy boundary.

License

MIT — see LICENSE.