Runtime·The Rust core
It reads your code
before the model does.
Runtime is the Rust engine under every Forge task. It parses your project, links every symbol, sends the model only what a change needs, and runs your tests in a sandbox before it shows you a diff.
- contexttree-sitter map · 3 files read
- sandboxnetwork none · root read-only
- editsrc/store.ts +15 −4
- verify31 of 31 tests
- cache$0.0012, not $0.0248
- review3 files · waits for you
- Network off by default
- Tests decide
- Nothing applied unasked
How a run works
Seven steps,
one engine.
Every task in Forge goes through the same pipeline, from a copy of your project to a diff that passed. Click a step, or let it play.
Copy the project into a sandbox volume
The task gets its own copy of your project in a Docker volume. Your real folder is never mounted, and secrets, .git and node_modules are left out.
Parse
Every file becomes
a syntax tree.
Tree-sitter turns each file into a concrete syntax tree. The engine walks it and records every definition, from functions and classes to types, constants and imports, with its exact byte range.
- IncrementalA file is parsed again only when its BLAKE3 hash or grammar version changes.
- ConsistentUpdates are published as whole generations, so a query never sees half an old index and half a new one.
- BoundedUp to 20,000 files per workspace, 4 parses in parallel, 1 MiB and 2 s per file. Anything skipped is reported.
- HonestFiles with syntax errors keep the symbols that were found and are marked as such.
import { db } from "./store"; export function streakFor(habit: Habit, today: Day) { const days = db.checkins(habit.id); let n = 0, d = today; while (days.has(d)) { n++; d = prev(d); } return n; } export interface Habit { id: string; name: string }
- program
Hover a line or a node. Definitions become symbols with a byte range, a line range and an id: BLAKE3(workspace, path, kind, name, ordinal).
Link
Every symbol, linked
to what it touches.
Language servers, rust-analyzer and typescript-language-server, run in containers with no network on a copy of your files and answer definitions, references and incoming calls. The index adds what contains what.
- Labelled evidenceEvery edge says how it is known. A reference is never presented as proof that a call happens at runtime.
- Never staleAnswers are tied to the document version and a digest of the files; if either changed, the answer is thrown away.
- Bounded queriesNeighbourhoods go 3 levels deep, 200 nodes, 12 edges per node, so a busy helper can’t eat the budget.
- FallbackNo server? Same-name definitions and word matches stand in, labelled HEURISTIC.
Select
Only what the task needs,
with the reasons written down.
The context engine starts from what you named and where tests fail, follows the graph a short way, and scores every candidate with written reasons. Those reasons go into the prompt too.
- SeedsPaths and names in your request or the test output, and error locations from tracebacks and stack traces.
- Whole or signatureA definition goes in whole or as a labelled [signature], never as half a body.
- Budget10% of the model’s context window, between 4,000 and 20,000 tokens. What doesn’t fit is listed, not dropped silently.
- On by defaultThe pre-registered benchmark found it no worse than the baseline on any held-out task, so it is the default.
- streakForsrc/streak.tsnamed in the requestwhole≈ 82 tok
- Habitsrc/streak.tstype the target useswhole≈ 20 tok
- streak testssrc/streak.test.tstest that names the targetwhole≈ 263 tok
- db.checkinssrc/store.tscalled by the target[signature]≈ 31 tok
- exportCsvsrc/export.tssame directory, not usednot included0 tok
- Chartsrc/ui/Chart.tsxno link to the tasknot included0 tok
Illustrative task. Every segment is stored with its path, byte range, BLAKE3 digest and reason, and the prompt block is re-rendered from the manifest before the call.
Cache
Pay once for the part
that doesn’t change.
Tools, system prompt and task context form a stable prefix, with no timestamps and nothing random. One cache marker sits on it; a second moves to the newest message every turn, so each turn reads the one before.
- Provider-specificMarkers go only where they were measured to work: Anthropic models via OpenRouter. Other APIs never get the field.
- Same contentMarking doesn’t change a word of the prompt. A cache miss costs money, never correctness.
- Counted properlyCache reads and writes are tracked apart; cost is the provider’s own figure.
Bar widths are illustrative. The prices below are measured.
Measured live, 26 Sep 2026 · anthropic/claude-opus-5.5 via OpenRouter · ADR 0010
Verify
Tests decide,
before you see anything.
Checks are run by the engine, not suggested by the model. Results are compared test by test with a baseline taken before the change, so a regression is named, and an old unrelated failure isn’t blamed on it.
- 1Related tests
Test files that changed, or that name the changed module, run first as a fast check. If they fail, repair starts without waiting for the full suite.
- 2The whole suite
Your project’s own test command. It is always the final check; the targeted run never replaces it.
- 3Independent acceptance
If the change edited or deleted a test file, the original tests are put back and the suite runs again. Edited tests that pass while the originals fail give UNVERIFIED, never success.
A container with
nothing to lose.
The sandbox is a Docker container on your Mac, created per operation and recorded in a ledger so it can be recovered after a crash. If no runtime is installed, Forge says so; it never falls back to your shell.
Review
A diff that passed,
and how it got there.
You get the change, the test result and the context manifest: what the model saw and why. You pick the model for every conversation, and it is bound when the task starts.
- No silent fallbackIf the model is missing or can’t use tools, the task isn’t created.
- No model swapsA streamed answer from a different model than the one bound is rejected.
- Bounded loopUp to 40 turns, 16 tool calls per turn, 2 repair turns. Extra calls are reported, not run.
export function streakFor(habit, today) {- let n = 0, d = today;+ let n = 0;+ let d = days.has(today) ? today : prev(today); while (days.has(d)) { n++; d = prev(d); }
Privacy
What leaves your Mac,
and what never does.
Leaves your Mac
- Your message and the context segments listed in the manifest
- Tool results the agent needs for the next turn
- Sent from the Rust core to the model you picked, over HTTPS
Stays on your Mac
- Your project folder; the sandbox works on a copy in a local volume
- The provider key, in the macOS Keychain; the interface never sees it
- The syntax index and symbol graph, in a local SQLite database
Measured
Measured,
not promised.
A pre-registered benchmark: 28 tasks, hidden tests the agent never saw, one model (Claude Opus 5.5 via OpenRouter), one machine.
- −49%Cost per accepted taskcontext engine + prompt cache vs baseline · $0.0405 → $0.0205 · 240 runs
- −18%Turns to finish a taskcontext engine vs baseline · 3.05 → 2.50 on average
- −95%A repeated request, cache warm$0.0248 cold → $0.0012 warm
- 312/312Runs accepted by hidden testsevery arm, both sets, in a container with no network