From b4c2adc18e0b394bb7b8009bb610fa4f667c6d60 Mon Sep 17 00:00:00 2001 From: Juber Shaikh <40266375+CodeWithJuber@users.noreply.github.com> Date: Fri, 10 Jul 2026 14:21:24 +0400 Subject: [PATCH] Redesign public website --- landing/index.html | 1249 +++++++++----------------------------------- public/index.html | 4 +- 2 files changed, 237 insertions(+), 1016 deletions(-) diff --git a/landing/index.html b/landing/index.html index a8d2f35..2db6356 100644 --- a/landing/index.html +++ b/landing/index.html @@ -3,26 +3,21 @@ - forgekit — one brain for every AI coding agent + forgekit — cognitive infrastructure for AI coding agents - + - + - - - + - - - - -
-
- -
-

Cognitive substrate for AI coding agents

-

One brain for
every coding agent.

-

- forgekit compiles a single source of truth into native config for nine AI - coding tools — and gives them a shared, proof-carrying memory: every - lesson carries its evidence, earns trust only from real outcomes (tests, CI, - your reverts), decays without review, and merges across your team through - plain git. -

-
- $ - npm install -g @codewithjuber/forgekit - -
-
- How the memory works - Read the docs -
-
- 471 tests - 0 runtime dependencies - 14 tools, one config - MIT · Node ≥ 20 -
-
-
+ -
-
-
-
-
0.43 ms
-
“if I edit this, what breaks?” — blast-radius query
-
reports/benchmarks.md · this repo
-
-
-
118 ms
-
full pre-action gate: assumptions → routing → impact
-
reports/benchmarks.md · this repo
-
-
-
62.1 %
-
token cost saved by tiered routing vs always-premium
-
whitepaper prototype · live tokens
+
+
+
+
+

forgekit v0.10.0 · MIT · zero runtime dependencies

+

Give every coding agent the same working memory.

+

+ Forgekit turns one source of truth into native configuration for AI coding tools, then adds proof-carrying lessons, blast-radius prediction, reuse search, and pre-action guardrails so agents act from shared evidence instead of fresh guesses. +

+
+ Read the docs + See the workflow
-
-
3.5 M/s
-
confidence evaluations — trust math is never the bottleneck
-
reports/benchmarks.md · this repo
+
+ + npm install -g @codewithjuber/forgekit && forge init +
-

- measured, not asserted — every number regenerates from - npm run bench; targets are labeled as targets -

-
-
- -
-
-

01The root cause

-

Every agent failure is the same failure.

-

- Breaking callers it never looked at. Re-reading the same files every session. - Repeating the mistake you corrected last week. One root: the model is a - stateless next-token predictor eating flat text — no model of your - code, no memory of what it already tried. -

-
-
-
FACE A — NO WORLD-MODEL
-

It can't see consequences

-

- No live map of the codebase means no way to reason about what a change - breaks before making it. -

-

“If I change X, what else fails?”

-
-
-
FACE B — NO MEMORY
-

It can't accumulate judgment

-

- Each session re-derives your project from zero, then repeats the exact - correction you already gave it. -

-

“What did I already learn here?”

+
+
    +
  • Reuse candidates are searched before new code is written.
    • +
  • Impact atlas predicts likely files, tests, and blast radius.
    • +
  • Lessons carry provenance and gain trust only from oracle outcomes.
    • +
  • Guardrails compile to Claude Code, Codex, Cursor, Gemini, Aider, and more.
    • +
+
-
- -
-
-

02Proof-carrying memory

-

A memory with receipts.

-

- Most agent memory is a pile of notes that grows stale and lies with - confidence. In forgekit, every stored thing is a claim — and a claim - has to carry its own proof. -

-
-
    -
  1. - a -
    -

    Content-addressed

    -

    - A claim's id is the hash of its content. The same lesson learned on - two machines is the same claim — merging is set union, and - it cannot conflict. -

    -
    -
    2. -
  2. - b -
    -

    Evidence, not vibes

    -

    - Confidence moves only when an independent oracle speaks — a - test run, CI, the type checker, or you accepting or reverting a - change. Oracle weights live in code, so evidence can't be forged; - forge ledger verify recomputes every hash. -

    -
    -
    3. -
  3. - c -
    -

    Decays toward doubt

    -

    - Unreviewed knowledge slides back toward uncertainty — never to - “false,” never frozen at “true.” Stale lessons stop being injected - long before they can mislead. -

    -
    -
    4. -
  4. - d -
    -

    Merges through git

    -

    - The ledger lives in .forge/ledger/ as append-only, - union-merged files. git pull is the sync protocol. No server, - no accounts, no lock-in. -

    -
    -
    5. -
-
-
- - forge ledger show 64bf19 — a claim and its proof -
-
claim 64bf1958… kind lesson scope repo -body “Before editing computeTax, check its - callers and tests first.” - -evidence - ✓ confirm test.run w 0.8 ref run:9f31 - ✓ confirm human.accept w 0.9 ref pr:35 - ✗ contradict ci.run w 0.8 ref ci:7712 - -confidence val 0.71 (decays without review) -provenance alice · 2026-06-30 · task tax-rounding -merge union — conflict-free across teammates
-
-
+
+
118 msfull pre-action gate from repository benchmarks
+
0.43 msblast-radius lookup from benchmark reports
+
62.1%measured routing-stage cost saved
+
0runtime dependencies in package.json
-
+
-

03The loop

-

Think before. Verify after. Learn always.

-

- Six deterministic stages wrap every edit — no LLM calls inside the gates, so - they run in milliseconds and never hallucinate. -

-
-
- 1 · GATE - forge substrate -

Assumptions checked, task routed to the cheapest capable tier.

-
-
- 2 · CONTEXT - forge context -

Budgeted assembly; missing context becomes a computed question list.

-
-
- 3 · REUSE - forge reuse -

Verified artifacts served from cache instead of regenerated.

-
-
- 4 · IMAGINE - forge imagine -

Blast radius predicted; minimal test suite dry-run in a sandbox.

-
-
- 5 · VERIFY - forge verify -

Independent oracles judge the edit — types, tests, review.

-
-
- 6 · LEARN - forge ledger -

Outcomes update the confidence of every claim that informed the edit.

-
-
- ↺ outcomes flow back — the memory that informed a bad edit loses - trust automatically -
+
+

Platform

A cognitive substrate, packaged like developer tooling.

+

The site uses only repository-backed claims: README, CHANGELOG, package metadata, benchmark reports, and optional GitHub API counters on the status page.

-
-
- -
-
-

04Cross-tool

-

Write the rules once. Every agent obeys.

-

- One source of truth compiles to each tool's native config format — switch - agents without losing your project's accumulated judgment. -

-
- Claude Code - Codex - Cursor - Gemini CLI - Aider - Copilot - Windsurf - Zed - Continue +
+
01

Cross-tool config

Author rules once, then compile native config layers for the agents your team already uses.

+
02

Proof-carrying memory

Every durable lesson includes evidence, provenance, confidence, decay, and merge behavior.

+
03

Impact foresight

Before editing, Forge predicts affected files and tests so agents can plan smaller, safer changes.

+
04

Reuse-first retrieval

Agents are pushed toward existing code paths and prior fixes before inventing another solution.

+
05

Guardrails that run

Scope, assumptions, anchoring, cost, and secret-redaction checks are enforced by CLI hooks.

+
06

Team memory through git

Knowledge moves with the repository, not with a single chat transcript or vendor workspace.

-

- plus MCP config for Roo and VS Code — forge sync - regenerates everything from one file. -

-
-
-

05Team memory

-

Your whole team's mistakes, learned once.

-

- When a teammate's agent burns an hour on a dead end, yours shouldn't have to. - Shared memory is three commands — all of them ordinary git. -

-
-
- - team memory in three commands +
+
+
+

Workflow

+

Install once. Gate every meaningful agent action.

+

The CLI stays small and dependency-free, while optional live status data is fetched with timeouts, retries, jitter, and HTTP cache validators.

+
+
1

Initialize the substrate

Run the installer and generate project-local agent instructions, hooks, and memory stores.

+
2

Let agents plan with evidence

Forge assembles context, reuse candidates, assumptions, scope, and impact predictions before edits.

+
3

Promote only verified lessons

Tests and human corrections update memory; self-graded claims do not become trusted facts.

-
$ forge init # once — sets up .forge/ledger + union merge -$ git pull # teammate lessons arrive, conflict-free -$ forge cortex # their hard-won lessons now inject into - # YOUR agent's context — with provenance: - # forge ledger blame shows who learned what, when
+
+
+
forge substrate
+ 
$ forge substrate "Change auth validation and update tests"
+✓ loaded team rules from AGENTS.md
+✓ found reuse candidates before implementation
+✓ predicted blast radius and test targets
+! assumption requires citation before action
+✓ emits native config for agent tools
-
+
-

06The honest part

-

What it won't do.

-

- The project keeps a public honesty register — claims are labeled - measured, target, or known limit. These are the limits. -

-
    -
  • - limit -

    - MinHash is weak on very short specs. Two four-word tasks can look - unrelated. Fix: point FORGE_EMBED at any - embedding provider — retrieval upgrades to cosine, and silently falls - back if the provider dies. -

    -
    • -
  • - limit -

    - The impact atlas is approximate. Regex-built import graph: - measured 0.90 precision / 0.97 recall on this repo's hand-labeled cases - — good enough to gate, never claimed perfect. -

    -
    • -
  • - target -

    - The ~90% cost-reduction figure is a composition argument, not yet - an end-to-end measurement. Only the 62.1% routing stage is measured; - forge cost --stages reports measured factors - only and refuses to blend in estimates. -

    -
    • -
  • - design -

    - It never self-grades. Confidence moves only on oracle outcomes — - an agent claiming its own change works is worth exactly zero evidence. -

    -
    • -
+
+

Evidence

Professional, but honest about limits.

+

The strongest design choice is restraint: measured numbers are labeled as measured, targets stay targets, and generated pages document their sources.

+
+
+
↯

Approximate impact atlas

Import-graph predictions are useful for gating and planning, not presented as perfect dependency analysis.

+
%

Cost claims stay scoped

The page shows the measured routing-stage savings rather than blending measured data with aspirational targets.

+
-
-
-

Get started

-

Sixty seconds to a substrate.

-
- $ - npm install -g @codewithjuber/forgekit && forge init - +
+
+
+

Data sources

+

No mock marketing data.

+

Every metric and product claim on this site comes from local repository files or the optional public GitHub repository endpoint used by the generated status page.

-
- - Star on GitHub - - Read the whitepaper +
+
    +
  • package.json
    • +
  • README.md
    • +
  • CHANGELOG.md
    • +
  • reports/benchmarks.md
    • +
  • https://api.github.com/repos/CodeWithJuber/forgekit
    • +
- diff --git a/public/index.html b/public/index.html index 6131266..1c25158 100644 --- a/public/index.html +++ b/public/index.html @@ -39,7 +39,7 @@ code{font:500 13px var(--mono);background:var(--surface-2);border:1px solid var(--line);border-radius:var(--r-s);padding:0 8px} footer{padding:32px 0;color:var(--faint);font-size:14px} @media(max-width:800px){.grid{grid-template-columns:1fr}} -
forgekit

@codewithjuber/forgekit · v0.8.0 · Node >=20

Live status, straight from the repository.

One brain for every AI coding agent — the cognitive substrate every frozen model is missing (proof-carrying memory, impact foresight, enforced guardrails), authored once and delivered as native config to Claude Code, Codex, Cursor, Gemini, Aider, and more.

Install in 60 seconds Read the docs

MIT license0 runtime dependenciesclaude/forge-work-system-setup-p26ka5 @ 26586e2live GitHub stats disabled
0.43 ms
blast-radius lookup

Measured from this repo's benchmark report, not a marketing placeholder.

reports/benchmarks.md

118 ms
pre-action gate

Assumptions, routing, reuse, context, impact, scope, and anchoring.

reports/benchmarks.md

62.1%
cost saved

Documented from the white-paper prototype and exposed by Forge cost reports.

whitepaper prototype

Quickstart

npm install -g @codewithjuber/forgekit +
forgekit

@codewithjuber/forgekit · v0.10.0 · Node >=20

Live status, straight from the repository.

One brain for every AI coding agent — the cognitive substrate every frozen model is missing (proof-carrying memory, impact foresight, enforced guardrails), authored once and delivered as native config to Claude Code, Codex, Cursor, Gemini, Aider, and more.

Install in 60 seconds Read the docs

MIT license0 runtime dependencieswork @ d13b19clive GitHub stats disabled
0.43 ms
blast-radius lookup

Measured from this repo's benchmark report, not a marketing placeholder.

reports/benchmarks.md

118 ms
pre-action gate

Assumptions, routing, reuse, context, impact, scope, and anchoring.

reports/benchmarks.md

62.1%
cost saved

Documented from the white-paper prototype and exposed by Forge cost reports.

whitepaper prototype

Quickstart

npm install -g @codewithjuber/forgekit forge init forge doctor -forge substrate "Change auth validation and update tests"

Latest repo changes

    Benchmark sections indexed: 3 · benchmarks file updated 2026-07-08.

    Data Sources

    No mock data is used. This page is regenerated from repository files during CI (generated 2026-07-08T18:42:41.820Z from 26586e2). Enable BUILD_PAGES_LIVE=1 to refresh public GitHub counters with ETag/Last-Modified caching.

    • package.json
    • README.md
    • CHANGELOG.md
    • reports/benchmarks.md
    • https://api.github.com/repos/CodeWithJuber/forgekit (optional, no auth, only when BUILD_PAGES_LIVE=1)
    WCAG-minded semantic HTML, keyboard focus, responsive 320px–1920px+, and reduced-motion-safe. Same design tokens as the landing page — forge uicheck design gates both.
    \ No newline at end of file +forge substrate "Change auth validation and update tests"

    Latest repo changes

      Benchmark sections indexed: 3 · benchmarks file updated 2026-07-10.

      Data Sources

      No mock data is used. This page is regenerated from repository files during CI (generated 2026-07-10T10:12:16.159Z from d13b19c). Enable BUILD_PAGES_LIVE=1 to refresh public GitHub counters with ETag/Last-Modified caching.

      • package.json
      • README.md
      • CHANGELOG.md
      • reports/benchmarks.md
      • https://api.github.com/repos/CodeWithJuber/forgekit (optional, no auth, only when BUILD_PAGES_LIVE=1)
      \ No newline at end of file