RepoBound v0.5.1

ContextForge has been renamed to RepoBound.

• Engine behavior is preserved in v0.5.1.
• npm package: @kallist/repobound
• CLI: repobound
• Skill: repobound
• Existing .contextforge state remains compatible.
• The old @kallist/contextforge package remains available but is deprecated; install @kallist/repobound for RepoBound.

ContextForge v0.5.0

ContextForge v0.5.0 is an Agent Experience & Launch release. It makes the existing repository-context engine easier to discover, inspect and use with coding agents.

What's new

• One canonical ContextForge Agent Skill for task context, local Git Review and context debugging.
• Task-first Studio 2.0 with Context / Why / History, exact context, explicit controls, semantic Diff and Replay.
• English and Chinese product READMEs, real fixture screenshots, launch assets and a public landing page.

Start here

Requires Node.js >=24.15 <25.

npm install -g @kallist/contextforge@0.5.0
contextforge studio

Install the Skill from the public repository in the project where it is wanted:

npx skills@1.5.26 add kallist/ContextForge --skill contextforge --agent codex --copy
npx skills@1.5.26 list

Landing page · npm package · English README · 中文 README · Agent integration

Preserved contracts

Compiler behavior, ranking, ranking weights and benchmark Gold are unchanged. V1 remains the public default; V2 remains internal. MCP still exposes exactly status, index, search and pack. Review uses the CLI. API and database semantics are unchanged.

ContextForge inspects the repository context it compiles. It does not expose hidden model context, guarantee complete impact analysis, or establish improved coding success, review accuracy or general token savings.

Release validation

• Exact main/tag commit: edfb3ad7c0780b198bcec742e5f4126111d6e674.
• 214 tests passed, zero failed/skipped; package smoke, release hardening and release check passed.
• Frozen benchmark: 720/720 case objects unchanged. Capsule conformance: 240/240 passed. Public 0.4.1 versus candidate engine parity passed.
• Fresh public npm 0.5.0 install with a new cache matched the audited tarball integrity. CLI, official MCP client, real Chromium Studio/Review journeys, five controls, four viewports, XSS/privacy and package leakage checks passed.
• Public GitHub main Skill installation, references, Codex-native discovery and explicit coding/Review/context-debug use passed. Cursor and Claude Code installer targets passed; their host runtimes and autonomous model performance were not tested.
• MODIFY, metadata-only DELETE, RENAME and rename round-trip indexing regressions passed against the public installed package. Deleted historical source is not reconstructed.
• Public Pages and GitHub-rendered English/Chinese READMEs passed browser checks. Pages used no unexpected remote runtime requests.

Main CI · Tag CI · Pages deployment

Both CI runs passed Node 24.20 on Ubuntu, Windows and macOS; minimum Node 24.15; full frozen benchmark/conformance; and Studio browser/installed-package checks. Local public acceptance used Windows and Chromium. Docker, non-Chromium browsers and screen-reader certification were not tested.

Registry tarball SHA-1: 2e98f42540f857a0994e8fb348954336c6458c10.

Registry integrity: sha512-/S8veLZRHt9MhDfAepy7Z54TqmE5nEkhq8XYaLqQvCvKKgNQ+qEsdWFUmCvmu6EWvJ7ipaPqG7x5Ww6XAlSduw==.

ContextForge v0.4.1

ContextForge v0.4.1 is a narrow correctness hotfix.

Fixed:

• tracked DELETE Review can now produce a validated metadata-only Review Capsule without reconstructing deleted source;
• rename round trips no longer leave stale generation-bound Git signals that block later durable indexing.

Preserved:

• no new Review modes;
• no MCP expansion;
• no ranking changes;
• no schema migration;
• no historical source archive.

Verification:

• 214 tests;
• 720/720 frozen compiler cases unchanged;
• 240/240 Capsule conformance;
• Windows, macOS, and Linux Hosted CI;
• packed and public-package CLI, Studio, and MCP smoke;
• original DELETE and rename-round-trip public UAT reproducers.

ContextForge v0.4.0 — Review Context

ContextForge v0.4.0 — Review Context

A diff shows changed lines. ContextForge Review traces supported surrounding code
and compiles an explainable Review Context under your token-estimate budget.

• Analyze a local base commit against tracked working-tree changes, with syntax-level
  changed symbols and ranges in supported JavaScript/TypeScript/Python sources.
• Inspect one-hop import, direct symbol reference, test, documentation and supported
  interface/implementation evidence. Structural facts and heuristics stay distinct.
• Inspect the dedicated Review Workspace; PIN, EXCLUDE, PREFER, FOCUS or RANGE a
  safe candidate, then save an immutable child Capsule and compare its context.
• Keep Review provenance in Capsule v2, source-light history, deterministic Explain,
  captured-entity Coverage/Lint, semantic Diff and current-source Replay verification.
• Use contextforge review or the local Studio. Normal CLI/MCP search and pack
  retain the V1 default; V2 ranking remains experimental.

Review does not detect bugs, approve PRs, guarantee complete impact or replace a
reviewer. Deleted source is metadata only; untracked files must be staged. The
separate small scenario evaluation is synthetic and does not measure agent success.

See the product guide,
architecture decision and
evaluation for precise limits and reproducible evidence.
Publication status must be verified through the release report and public registry.

Verified release evidence

• PR #13 merged at 72a70a11ffb87749bd39a2ed80300fe47d33f5de; final PR and post-merge main CI passed all six jobs.
• Full regression: 208 passed, zero failed/skipped. Historical compiler: 720/720 unchanged cases; Capsule conformance: 240/240.
• Public npm @kallist/contextforge@0.4.0 is latest. A fresh public-registry cache installation passed CLI, official-client MCP, parser WASM, installed Studio and Review Chromium journeys, four Review viewport widths, privacy/XSS checks and package leakage checks.
• Public tarball SHA-1: f0db149f3a70714416f8ae346e25344199697696.
• Public tarball integrity: sha512-dt3pSnIEIpabvqudQo35sjOoP9T3XadXMpc0Oltlo0nmh0jJaWkT61tOpa8JDfNp8esueRnSwyQZ9Wuuh41fOQ==.

The small Review evaluation retains only 2/9 and 5/9 related caller paths at 1K and 3K budgets. These losses remain documented. Coding/review agent success, Docker, whole-program runtime impact and large-project performance were not measured or tested. Normal compiler default remains V1.

ContextForge v0.3.0

ContextForge v0.3.0

Compile what your coding agent sees. Explain it. Control it. Replay it.

V0.3 turns a context compilation into a local, inspectable lifecycle. Context
Capsules record selected and dropped candidates, source identities, evidence,
strategies and exact payload hashes. Compressed local history lets you revisit them
without silently turning your repository into a source archive.

Start contextforge studio --repository ./repo to inspect a proposal in the local
browser UI. Semantic controls let you PIN, EXCLUDE, PREFER, FOCUS or request a RANGE.
Recompilation creates a new Capsule and immediately compares before/after decisions,
ranges, evidence, roles and token contributions. Hard budgets and source safety remain
compiler invariants; impossible or stale controls fail explicitly.

Replay clearly separates offline recorded inspection from actual reproduction
verification. The product cannot restore old source bytes from source-light history.
Coverage and deterministic lint describe captured facts and tradeoffs; they do not
invent a context-completeness score or infer benchmark Gold.

The CLI exposes history, replay, diff, coverage and recompile alongside existing
pack/explain commands. The four MCP tools and public contextforge-v1 compiler
default remain unchanged. V2 remains internal/experimental with mixed measured
evidence. No cloud backend, account, telemetry, new agent runtime or V0.4 Review
product was added. AGENT SUCCESS NOT MEASURED.

Requires Node >=24.15 <25. Frontend assets and the five WASM parser assets ship
inside the existing @kallist/contextforge package. See the
product guide for exact contracts and limitations, and
validation evidence for measured gates. A normal public release
requires green post-merge main CI, immutable annotated tag, successful npm publish,
public-registry fresh install/product smoke and only then GitHub Release creation.

Verified release evidence

• Release commit: c51a7d3ac4e92d44ea7fe90df5f09f6dedf42af4.
• PR #12: final HEAD passed all six jobs in both PR and push CI.
• Post-merge main CI: all six jobs passed, including the full frozen benchmark and Capsule conformance.
• Public npm @kallist/contextforge@0.3.0, dist-tag latest, 214 files, five WASM assets, three Studio assets.
• Public registry fresh-cache installation passed CLI, official-client MCP, Capsule/history/Replay/Diff/Coverage/recompile, and a real installed-package Chromium workflow including stored XSS and restart checks.
• Published integrity: sha512-OhLzfCngDkJZ9xy5scBYdi7khln8GHbfPQdwaXU981bGcGaq7KTRaRWGINXrWWY6UPXYoPH8RcE7BeHT5Hf8uA==.
• Public default remains contextforge-v1. Agent success is not measured. V0.4 is not started.

ContextForge v0.2.0

ContextForge v0.2.0

What changed

• Deterministic TaskAnalysis, hybrid Retrieval V2 evidence fusion and lexical symbol ownership.
• Bounded transient caller, test and implementation relationship evidence.
• Experimental ContextPlan and plan-aware Pack V2, with exact final budget verification and bounded request-local fragment reuse.
• Final frozen six-system evaluation, failure taxonomy and release evidence checks.

Stable default

V1 remains the public/default CLI and MCP Search/Pack behavior. Node.js support remains >=24.15 <25. Package identity is @kallist/contextforge; executable is contextforge.

Experimental V2

The selected candidate is contextforge-v2-plan-pack, retained INTERNAL_ONLY. There is no public V2 flag or MCP strategy field. ADR-009/010/011 are accepted as maintained experimental architecture, not public-default promotion.

Benchmark

The frozen 24-task/four-repository/five-budget comparison completed 720 cases with zero estimator-budget violations. All historical final case objects were reproduced. At 8K, required-symbol recall is 0.886364 versus V1's 0.863636, Overall Gold is 0.899802 versus 0.870866, and required-file recall is unchanged at 0.916667. All 19 V1 successes remain successful. Pack V2 retains 36/39 pre-Pack-present required symbols versus 35/39 for the relationship candidate with Pack V1.

The same-run Windows/Node 24.20 measurements give a fixed-corpus median total ratio of 1.478582x V1 (within the 1.5x guardrail, with limited headroom) and a synthetic 1,000-file ratio of 1.084919x. These are environment-specific measurements, not an SLA.

Important limitations

Cross-budget evidence remains mixed: weaker 2K symbol recall, weaker 16K/32K file recall than the V2 Pack V1 alternatives, and weaker precision outside 8K versus the relationship alternative. Task-aware planning is not proven stably superior. The strict .917 file and .900 symbol design thresholds are not met. No universal superiority or general token-savings claim is made.

Not measured

Coding-agent success was not measured. Docker, third-party agent-host integration, and external very-large repository generalization were not tested. V0.3 was not started.

See final evaluation and release report for evidence, provenance and release gates.

Release closeout — verified 2026-09-04

The release was completed from merged main 5741d72791c1c8fb2fa3fb6faa223fa0257f02fc, following release PR #10. No product, dependency, benchmark or audit-configuration changes were made during audit recovery.

Audit recovery and CI

The original post-main run failed when the public npm Security Audit endpoint returned HTTP 503, so tagging and publication stopped. At 2026-09-04 09:03:42 UTC, the official npm status endpoint reported Security Audit, Package publishing and Package installation operational; this was supporting evidence, not proof of request success.

One authorized no-code retry was performed: run 33854962429, attempt 2. All five mandatory post-main jobs passed. The audit returned zero vulnerabilities at every severity (info, low, moderate, high and critical), across 101 dependencies. The same job completed all 720 benchmark cases and the release-evidence check. Classification: TRANSIENT_NPM_AUDIT_SERVICE_503.

Existing main local release:check passed, including 172/172 tests, zero skips, package/MCP smoke and release hardening. Earlier local audit timeouts remain historical NOT VERIFIED results; the successful mandatory Hosted audit closes the release gate.

Immutable tag and published package

• Annotated tag: v0.2.0.
• Tag object: f90eb3b7eb8832a9c5b0a5aa0eeb7d8fcf89a8aa.
• Peeled target: 5741d72791c1c8fb2fa3fb6faa223fa0257f02fc, verified locally and remotely.
• Public package: @kallist/contextforge@0.2.0; registry latest verified as 0.2.0.
• License: MIT; Node engine: >=24.15 <25; bin: contextforge -> dist/cli/main.js.
• Reviewed tarball: 169 files, 663755 packed bytes, 4840887 unpacked bytes, five WASM assets.
• Public registry SHA-1: 7cc93872b29c2416bba40f3c312662977123d06a.
• Public registry integrity: sha512-PH3SvNVBYk82DeJG0/zRavFd5DJJGfPH7EwCgr3AetxKBzOQIplCyYyq1N3t+FC4PvmWz210ex72tRkjWK7iXA==.

The public registry metadata and the fresh installation lockfile integrity match the reviewed tarball. Publication used the exact reviewed tarball from the validated main checkout, following a successful final publish dry-run and legitimate npm authentication.

Fresh public installation

npm run package:smoke -- --public-registry passed on Windows x64 / Node 24.20.0. The consumer used a new temporary project and fresh cache, installing the exact version from https://registry.npmjs.org, not a local tarball.

Verified CLI --help, --version (0.2.0), map, index, inspect, graph, search and pack; packaged WASM parsers and SQLite; MCP initialization, status, index, search, pack and shutdown. The existing standalone CLI has no status command: status is an MCP tool, and no new CLI command was added. CLI/MCP ranking remains contextforge-structural-v1, and packing remains contextforge-pack-v1; public default remains V1.

Actual installed contents passed the leakage scan. OS process inspection found zero matching smoke processes before and after the run, with zero orphans. Temporary smoke resources were cleaned up.

This is public packaging/transport evidence, not coding-agent success. Third-party coding-agent hosts, agent-success quality, Docker, external very-large repository generalization and Pack V2-specific OS-process symlink/concurrency races remain NOT TESTED. The frozen mixed-evidence candidate decision and unmet strict recall targets above remain unchanged. V0.3 NOT STARTED.

ContextForge v0.1.1

ContextForge v0.1.1 Release Notes

ContextForge v0.1.1 is the first public npm release of the local-first, task-aware context compiler for coding agents. The only material release correction is the npm distribution identity:

contextforge
→
@kallist/contextforge

npm rejected the unscoped contextforge name under its package-name similarity policy. The immutable v0.1.0 tag remains the technically completed release candidate; its public npm publication did not complete. Production behavior changed: NO.

Install

Use Node >=24.15 <25, then install the scoped public package:

npm install -g @kallist/contextforge
contextforge --version

Release-specific verification may use npm install -g @kallist/contextforge@0.1.1. The CLI executable remains contextforge.

What ContextForge does

ContextForge safely maps a repository, parses JavaScript/TypeScript/TSX/Python with packaged WASM grammars, publishes complete SQLite index generations atomically, derives bounded structural signals, retrieves explainable task candidates, and builds deterministic Markdown under a hard ContextForge estimator budget. A local stdio MCP server exposes repository-bound status, index, search, and pack tools through the same application use cases as the CLI.

The production strategies remain contextforge-structural-v1, contextforge-pack-v1, and contextforge-generic-v1 version 1.0. V0.1 does not include embeddings, vector search, LLM reranking or compression, HTTP/remote MCP, OAuth, multi-root MCP, file watching, auto-indexing, agent outcome learning, or a Web UI.

Supported environment and distribution

The declared runtime is Node >=24.15 <25. Release CI validates Node 24.15 on Ubuntu and Node 24.20 on Ubuntu, Windows, and macOS. The package uses Node's built-in SQLite and packaged WebAssembly assets; it has no native compilation or install/postinstall script. SQLite WAL state must remain on a local filesystem; UNC/network filesystems are unsupported.

Benchmark evidence

The frozen contextforge-benchmark-v1 uses 24 tasks, three systems, five budgets, and dataset hash 75685087e8392840b4bb61ae19cef9ccb92df66844e79e0e87a0b7c6437a0002.

On benchmark-v1 at 8K, contextforge-v1 measured 0.917 required-file recall, 0.864 required-symbol recall, and 0.193 Gold-range precision. At matched required-symbol recall on paired tasks, macro mean token reduction was 1.5% versus lexical-full-file-v1 and 7.8% versus structural-full-file-v1.

The desired 60% token-reduction target was not met. Structural ranking did not show an aggregate advantage over lexical ranking on benchmark-v1, and no general “better than grep,” broad token-savings, coding-agent-success, or large-repository claim is supported. This finite self/curated benchmark has no external-human Gold review. No retrieval, ranking, packing, estimator, Gold, or benchmark result changed between v0.1.0 and v0.1.1.

Security and trust model

ContextForge runs with the invoking user's filesystem permissions and returns selected repository content to the local requesting CLI user or MCP host. Repository content can contain malicious instructions. ContextForge does not execute repository source, expose shell or arbitrary-file MCP tools, upload repository data, start a network listener, or collect telemetry. Sensitive-path exclusion is not a universal inline-secret detector; users must review generated context before forwarding it outside their trust boundary.

Index generations are activated in one SQLite transaction. Competing writers are bounded, readers retain complete snapshots, and release hardening exercises forced process termination before activation, integrity checks, recovery reindexing, corrupt databases, and unsupported future schemas.

Upgrade and migration

Schema version 1 indexes are migrated to version 2 when a new generation is built. Unknown future schemas fail closed without mutation. A corrupt index is recoverable local derived state: remove .contextforge/index.sqlite and run contextforge index again. Repository source is not stored in the database.

Known product limitations

• Generic token estimates are not model-tokenizer counts.
• Pure synonyms, semantic paraphrases, spelling variation, and cross-language task-to-code mapping are weak.
• TypeScript path aliases/package exports and advanced Python import behavior are incomplete.
• One MCP process binds to one repository; there is no auto-index or remote transport.
• Inline secrets inside otherwise eligible source are not guaranteed to be detected.
• Large external production repositories, real Codex/Claude Code/Cursor host setup, network filesystems, model-specific tokenizers, agent task success, and patch quality remain untested.