Skip to content

Instantly share code, notes, and snippets.

View Eigenwise's full-sized avatar

Kenny Vaneetvelde Eigenwise

View GitHub Profile
@Eigenwise
Eigenwise / README.md
Created October 11, 2026 09:17
Codebase Mapper: stale-but-valid path comparison with two fresh source readers

A saved map with a valid, inactive handler path

11 October 2026. One pair of fresh Claude Code sessions using the same small booking service and the same source-tracing question.

Both sessions found the handler used by the running service. In the stale-map case, the reader inspected the actual startup import, identified src/http/booking-app.cjs as the active handler and explicitly reported that the map still named src/http/app.cjs. The refreshed-map case identified the same active handler. This pair did not reproduce silent reliance on the old path.

Both case checkouts had the same Git history, including the recent summary “Relocate running handler and retain compatible prior module.” Both final answers refer to that relocation summary; case A quotes it fully and case B quotes its opening. This cue could encourage checking the other module. The common prompt also explicitly asks the reader to trace startup and cite source evidence. The result concerns this guided source-navigation task; it is

@Eigenwise
Eigenwise / README.md
Last active October 7, 2026 09:39
Codebase Mapper: a moved HTTP entry point and fresh source navigation

A Moved HTTP Entry Point

The booking service's startup file moved from src/server.cjs to src/http/server.cjs. After a manual Codebase Mapper refresh, a fresh Claude Code session followed the map back to the current startServer function.

Try the Example

Clone this Gist, or download its files into one directory. Open that directory in a terminal with Node.js 22 or later:

node unpack.mjs
@Eigenwise
Eigenwise / README.md
Created October 6, 2026 14:01
Claude Code 2.1.291 manual compaction: resumed versus one streaming process

A small manual-compaction check on Claude Code 2.1.291

I checked a resumed CLI session and a single streaming process on 6 October 2026. The requested opus alias resolved to claude-opus-5-5 in both actual receipts; effort was low.

One case per shape, adapted from Rulestack's experiment:

| Actual evidence | Separate processes with --resume | One streaming process |

@Eigenwise
Eigenwise / README.md
Created October 5, 2026 09:38
Sidequest: verified saved-comment recovery across two fresh authenticated Claude Code sessions (synthetic fixture, hooks disabled)

Sidequest: continue a saved task in a fresh Claude session

Eigenwise's own synthetic demonstration, run on 5 October 2026. No customer data.

Scope: two authenticated Claude Code processes used Sidequest's real MCP tools to save and recover a checkpoint comment. This run used a temporary source plugin and an isolated board. Hooks were disabled. Actual context compaction, automatic recovery, marketplace installation and dispatch were not tested.

What happened

The fixture accepts delivery webhook events. Its bug was counting the same event twice. Different events for the same shipment must still count separately.

@Eigenwise
Eigenwise / README.md
Created October 4, 2026 10:17
Synthetic agent completion records with real revision-bound test receipts

An idle job, a verified revision, and a review are three different things

Synthetic example with real local tests. The task, device data and OpenCode job metadata are mocked. The code changes, demo Git revisions and Node test receipts are actual local artifacts. No OpenCode process, provider call, customer task or integration ran.

The example fixes a small fleet-risk selector: API scores may be numbers or decimal strings. The selector should ignore malformed scores, preserve the input, and handle ties consistently.

The records

  • Completion record: the fixed candidate passes all seven tests. Its simulated runtime is idle; its separate review is still pending.
  • Changed candidate: a later commit intentionally reverses the comparison. The earlier passing receipt is stale for that revision, and a fresh run fails three tests.
@Eigenwise
Eigenwise / saved-task-handoff-template.md
Last active October 3, 2026 23:14
Saved-task handoff checklist: resume agent work after context compaction or a new session

Saved Task Handoff

Copy this into the saved task's description or comments. Fill in facts as the work proceeds.

Outcome and Scope

  • Desired result:
  • Acceptance conditions:
  • Files/project affected:
  • Constraints and explicit exclusions:
@Eigenwise
Eigenwise / quickstart_openai.py
Created February 27, 2026 07:44
Quickstart: OpenAI API directly with structured output + timings
import time
t0 = time.perf_counter()
import json
from pydantic import Field
from openai import OpenAI
from pydantic import BaseModel
from dotenv import load_dotenv
def elapsed(label: str, since: float) -> float:
@Eigenwise
Eigenwise / quickstart_instructor.py
Created February 27, 2026 07:44
Quickstart: Instructor (no Atomic Agents) with structured output + timings
import time
t0 = time.perf_counter()
from pydantic import Field
from openai import OpenAI
import instructor
from pydantic import BaseModel
from dotenv import load_dotenv
def elapsed(label: str, since: float) -> float:
@Eigenwise
Eigenwise / quickstart.py
Last active February 26, 2026 10:28
Atomic Agents Quickstart
import time
t0 = time.perf_counter()
from pydantic import Field
from openai import OpenAI
import instructor
from atomic_agents import AtomicAgent, AgentConfig, BasicChatInputSchema, BaseIOSchema
from atomic_agents.context import SystemPromptGenerator, ChatHistory
from dotenv import load_dotenv
@Eigenwise
Eigenwise / dynamic_orchestration_v2.py
Created July 24, 2025 18:54
Atomic Agents - Even More Dynamic Orchestration Example
import os
from typing import Union, List, Type
import instructor
import openai
from pydantic import Field
from rich.console import Console
from dotenv import load_dotenv
# Load environment variables
load_dotenv()