v0.17.0 · macOS · Linux · Windows · Docker · Workers · MIT
Microfiche for Coding Agents
A local, cache-safe hybrid optical and 4-layer deterministic text context compiler.
Compresses verbose tool schemas, ANSI build spam, and multi-turn logs into dense
microfiche plates or byte-exact @toon deltas — keeping numbered reads
and diffs 100% editSafe.
npm install -g github:sriinnu/fiche
Optical Split Curtain & 6-Stage Signal Bus
Every request enters at 127.0.0.1:47821 (or offline via fiche export / fiche-mcp),
routes by model capability in readerTier(), and seals completed epochs behind an immutable KV prefix boundary.
// Turn 1..14 static tool schemas (14,200 tok repeated verbatim)
{"name":"read_file","description":"Read file from workspace with optional line range...","parameters":{"type":"object","properties":{"path":{"type":"string"},"start":{"type":"integer"}}}}
{"name":"run_command","description":"Execute shell command in workspace..."}
\x1b[33m⠋ Compiling 184 TypeScript modules...\x1b[0m\r\x1b[2K\x1b[33m⠙ Compiling 184 TypeScript modules...\x1b[0m\r\x1b[2K
src/core/proxy.ts(142,19): error TS2345: Argument of type 'Buffer' is not assignable to parameter of type 'Uint8Array'.
[11:42:01.004] GET /v1/messages 200 1.4ms
[11:42:01.112] GET /v1/messages 200 1.3ms
[11:42:01.240] GET /v1/messages 200 1.5ms (×48 repeated heartbeat rows)
--- a/src/core/proxy.ts
+++ b/src/core/proxy.ts
@@ -140,4 +140,4 @@
- return digestChunk(rawBody);
+ return digestChunk(new Uint8Array(rawBody));
@tscg.tools[2]{name,req}: read_file,path; run_command,cmd
[fiche:plate#01 312×120 spleen-5x8 · turns 1..14 sealed]
▸ TS2345 src/core/proxy.ts:142 Buffer -> Uint8Array
▸ drain[×48]: GET /v1/messages 200 <ms> (kept errors)
▸ casi_delta: 0 new lines in unchanged config blocks
/* editSafe guard: active diff & numbered reads UNTOUCHED */
--- a/src/core/proxy.ts
+++ b/src/core/proxy.ts
@@ -140,4 +140,4 @@
- return digestChunk(rawBody);
+ return digestChunk(new Uint8Array(rawBody));
Anthropic Fable 5 · Optical Tier
anthropic
Tokens: -43.6%
USD: -34.6%
Warm p50: 2.75 ms
Physical Microfiche Reader Bay
Explore how render.ts packs dense agent traces onto a 312-column Spleen 5×8
microfiche plate with a live 2.5× optical loupe, verbatim FactSheet anchors, and stage waterfall.
@toon + Drain + [fiche:image])
16-Model Pareto Frontier & Team Savings Calculator
Every figure is loaded directly from bench-results.generated.json
(snapshot 2026-11-21, 16-turn agent workload, editSafe: true on all 16 models).
Tokens Saved % vs Avg Transform Time (ms)
Dashed line traces the non-dominated Pareto frontier. Hover or tap any point to inspect KV cache hit, p95 latency, and wire mode.
| editSafe | ||||||||
|---|---|---|---|---|---|---|---|---|
grok-4.7 | xAI | 4-Layer Text | 44.5% | 40.8% | 90.9% | 1.3 | 2.5 | yes |
glm-5.3 | z.ai | 4-Layer Text | 44.6% | 40.1% | 90.9% | 1.4 | 2.3 | yes |
deepseek-v4-pro | DeepSeek | 4-Layer Text | 44.6% | 37.9% | 90.9% | 1.5 | 2.7 | yes |
gpt-6-astra | OpenAI / Codex | 4-Layer Text | 44.5% | 36.7% | 90.9% | 3.3 | 12.7 | yes |
gpt-6.2-astra | OpenAI / Codex | 4-Layer Text | 44.5% | 35.5% | 90.9% | 1.5 | 2.6 | yes |
claude-opus-5-5 | Anthropic | 4-Layer Text | 44.5% | 35.5% | 90.8% | 1.6 | 3.2 | yes |
kimi-k3 | Moonshot | 4-Layer Text | 44.6% | 35.0% | 90.9% | 1.6 | 3.1 | yes |
claude-fable-5-20260609 | Anthropic | Optical | 43.6% | 34.6% | 90.0% | 6.9 | 38.2 | yes |
gemini-4-carbon | Google / AGY | 4-Layer Text | 44.9% | 33.4% | 90.9% | 1.5 | 2.8 | yes |
gemini-4-hydrogen | Google / AGY | 4-Layer Text | 44.9% | 33.4% | 90.9% | 1.4 | 2.7 | yes |
gemini-3.1-pro-preview | Google / AGY | 4-Layer Text | 44.9% | 33.3% | 90.9% | 1.2 | 2.2 | yes |
gemini-3.8-flash | Google / AGY | Optical | 51.9% | 28.8% | 84.0% | 9.4 | 39.7 | yes |
gemini-4-argon | Google / AGY | Optical | 39.7% | 27.9% | 90.0% | 10.5 | 43.3 | yes |
gpt-5.6-sol | OpenAI / Codex | Optical | 35.1% | 5.0% | 80.9% | 10.0 | 29.1 | yes |
gpt-6.2-sol | OpenAI / Codex | Optical | 35.1% | -1.4% | 80.9% | 15.3 | 91.6 | yes |
gemini-4-neon | Google / AGY | 4-Layer Text | 44.9% | n/a | 90.9% | 2.0 | 5.3 | yes |
FicheBar.app — Zero-Token Menubar Telemetry
Native Swift 6 / SwiftUI menubar companion polling local 127.0.0.1:47821 with zero LLM tokens,
5 Apple HIG themes, weighted vs raw savings clarity, and 1-click third-party model controls.
2-Axis OS × Agent Configurator
Select your host environment and coding agent to generate verified loopback, TLS warp, shell-init, MCP, Docker, Cloudflare Workers, or offline export commands.
Loopback Proxy + Antigravity Telemetry Plugin
Route Antigravity traffic through the local fiche proxy or wrap the CLI with HTTPS CONNECT warp, and wire the zero-token PostInvocation hook.
Included everywhere (macOS · Linux · Windows / WSL · Docker · Workers)
- Loopback proxy daemon (
ficheon127.0.0.1:47821) & Warm Paper web dashboard - 4-layer deterministic text compiler (
TSCG,ANSI/CR,@toon,Drain,ACDC,CASI-Delta) - Optical microfiche PNG renderer (
Spleen 5×8+GNU Unifont+JetBrains Mono) - HTTPS CONNECT interceptor (
fiche warp -- <cli>) &fiche-mcp(stdio /:47823) - Offline generator (
fiche export) & fast reverse-tail CLI (fiche turn-report,fiche stats)
Native macOS integration (never required for core)
- FicheBar.app — Swift 6 / SwiftUI menubar monitor & controller
- GUI Desktop App Launcher —
fiche warp --app <Name>for Electron / native IDEs - launchd Service Manager —
fiche service installbackground daemon supervision - Shell Init Wrappers —
fiche shell-init --wrap(fclaude,fcodex,fiche-app)