#agents
49 posts
-
Testing the Loop That Runs While You Sleep
The 15-phase run found five bugs in kodr2's loop scripts, and four of them failed silently - the run looked healthy while it was already broken. So the scripts finally got tests: a scripted kodr on PATH, a real temp git repo, 31 cases in nine seconds. Then the drift guard turned out to need a guard of its own.
#ai#kodr#agents#automation#loops#testing -
Looping a One-Shot Harness
Claude Code sorts agent loops by trigger and stop condition. Where does a zero-dep, one-shot harness sit? It already has the per-iteration parts - the outer loop is a shell script. The one thing a shell can't fake is a fuzzy stop condition, so kodr2 gets a `goal` command: iterate until a read-only model judge says met. The first live run caught one real bug, the next caught three more, and a 15-phase build over two long runs caught five more - one of them a fix that regressed before it stuck.
#ai#kodr#local-models#agents#automation#loops -
The Harness Is the New IDE
Models are the force multiplier, but the harness is where the leverage lives. The IDE matters less every month.
#ai#agents#harness#tooling#kodr -
The Harness Was the Hard Part
I had kodr build a smaller version of itself - a one-shot local-model coding agent called kodr-jr - one phase at a time. It worked. But every lesson worth keeping was about the harness, not the model: a Stop hook that lied twice, and a 26B model that believed it completely.
#ai#kodr2#local-models#agents#harness#cli -
Dropping the TUI, Then Earning It Back
Kodr2 shipped on purpose without a TUI - "one job, done well." A few weeks in, the one-shot CLI was the thing making it unpleasant to use. The way back wasn't to re-add the feature; it was to build the boundary the old codebase never had - a reporter channel - and let the TUI fall out of it as just another consumer.
#ai#kodr2#local-models#agents#cli#tui -
Kodr2: Multiple Providers - OpenRouter, Ollama, and LM Studio
Kodr2 grows from an LM Studio-only harness into a three-provider one, adds OpenRouter reasoning support, and gets dogfooded against all three by driving the same tasks through itself and fixing what breaks.
#ai#kodr#kodr2#local-models#agents#testing -
Kodr2: Starting Over, On Purpose
Kodr grew to 264 phases and got hard to reason about. Kodr2 is a from-scratch rebuild - same idea, a cleaner foundation. The old repo stays up as an archive; this is where development continues.
#ai#kodr#kodr2#local-models#agents -
The Frontier: Real LSP, and Editing Brownfield
Phases 99-100 of kodr: a from-scratch Language Server Protocol client replacing four phantom registry entries, and a brownfield edit eval suite that turns "can it edit an existing codebase" into a number. Where the journey has reached - for now.
#ai#kodr#local-models#agents#tools#verification -
Ergonomics: Making kodr Pleasant to Use
Phases 94-98 of kodr: git-aware undo, the repo-map extracted into a publishable library, project config and init, sensible auto defaults, and an interactive apply prompt that kills the dry-run dead-end.
#ai#kodr#local-models#agents#cli#safety -
The Context Firewall
Phases 90-93 of kodr: hybrid skill-driven orchestration, where the implementer stops drowning in its own accumulating context and each file gets an isolated author that sees its siblings' interfaces but never their code.
#ai#kodr#local-models#agents#skills#automation -
Async Runs and Prompt Caching
Phases 85-87 of kodr: the web sketch grows into a task-shaped control plane with SSE, a conservative provider-aware prompt cache, and a context packer split into stable and volatile sections so a prefix cache has a chance.
#ai#kodr#local-models#agents#automation#prompt-engineering -
Subagent Orchestration: Divide Judgment, Not Work
Phases 80-83 and 89 of kodr: splitting planner, implementer and reviewer into isolated model conversations, giving each its own model, shrinking the handoffs, and making sure every stage inherits the real harness contract.
#ai#kodr#local-models#agents#prompt-engineering#context -
Hooks, Harder - and a Splash of Color
Phases 77-79 of kodr: command hooks promoted to a real user surface with a Stop control, a named hook lifecycle that follows the sandbox, and zero-dependency ANSI color for the TUI.
#ai#kodr#local-models#agents#safety#automation -
Three Sandboxes, Escalating Paranoia
Phases 60, 76 and 88 of kodr: an OpenShell upload-and-execute backend, an opt-in Docker command boundary, and an OpenShell worker mode that runs the whole harness inside the sandbox - three answers to "where does untrusted code actually run?"
#ai#kodr#local-models#agents#safety#security -
Knowing the Model, Shrinking the Conversation
Phases 69-70 of kodr: a model profile registry that treats capability as harness behaviour rather than scattered defaults, and deterministic session compaction that keeps a long conversation usable for a small local model.
#ai#kodr#local-models#agents#context#memory -
Skills Grow Up: Resources, Permissions, Code
Phases 66-68 of kodr: skills gain referenced resources and executable helpers, gated behind a shared approval contract - each step widening what a skill can do without widening it by accident.
#ai#kodr#local-models#agents#skills#security -
Install It, Plan It, Heal It
Phases 65, 71 and 72 of kodr: a controlled dependency install so generated apps can actually run, a plan-then-execute self-dev acceptance test (and the local tool-call bug it found), and a bounded self-healing repair loop that feeds real verification failures back to the model.
#ai#kodr#local-models#agents#automation#verification -
Inspection as Tools, and a Plan You Can Read
Phases 62-64 of kodr: structural inspection handed to the model as two narrow tools, the same inspection exposed to the human in CLI and TUI, and a deterministic inspection plan that hands a small model the likely edit path instead of asking it to derive one.
#ai#kodr#local-models#agents#tools#context -
Ranking and Budgeting What the Model Sees
Phases 59 and 61 of kodr: a deterministic ranked repo-map so the likeliest code lands first, and token-budget-aware packing so a small local model gets useful context without starving its own completion room.
#ai#kodr#local-models#agents#context -
Scratchpads That Survive, Stages That Verify
Phases 57-58 of kodr: a planning scratchpad that carries between runs so plan-then-execute works on a small model, and staged execution that refuses to call one giant local-model dump a finished app until verification actually passes.
#ai#kodr#local-models#agents#memory#verification -
kodr Edits kodr
Phases 54-56: three self-development trials where a local model edits the harness's own source. A scorecard of what broke, the harness fixes each failure forced, and the two-location bug that keeps catching small models.
#ai#kodr#local-models#agents#automation#testing -
Inspection: kodr Learns to Read Code
Phases 51 and 53 of kodr: a zero-dependency structural code index across four languages, then an optional registry that lets real language servers enrich it - so the model reads before it writes, without the harness taking on a parser stack.
#ai#kodr#local-models#agents#context#tools -
Contract Tests and a Web Channel Sketch
Phases 49-50 of kodr: pinning the channel boundary down with contract tests so CLI, TUI and web can never drift apart, then the smallest honest proof of a web UI - a local-only JSON server that is emphatically not a second harness.
#ai#kodr#local-models#agents#cli#testing -
A Terminal UI, and Why It's Boring on Purpose
Phases 45-48 of kodr: a line-oriented TUI, an apply-review loop, a heartbeat for slow local calls, and Markdown session export - all in service of one idea, that the UI is just an adapter over a single shared request flow.
#ai#kodr#local-models#agents#cli#safety -
Sessions: Write It Down, Pick It Up, Look Back
Phases 42-44 of kodr: a trilogy that turns one-shot runs into conversations - record the full transcript, resume it with --continue, then browse what you built. Plus the one-line bug that was quietly truncating every transcript.
#ai#kodr#local-models#agents#memory#context -
Five Bugs, a Real Diff, and the Token Bill
Phases 39-41 of kodr: a review that found five plausible bugs hiding behind green tests (including an SSRF redirect bypass), a zero-dependency unified diff worth reading, and token usage finally shown where you look.
#ai#kodr#local-models#agents#security#cli -
Evals, Scores, and Prompt Receipts
Phases 37-38 of kodr: a scored eval command so model regressions surface as a number instead of a squint, and prompt versioning so every run can be traced back to the exact prompt that produced it.
#ai#kodr#local-models#agents#verification#prompt-engineering -
The CSV Example That Fought Back
Phases 24-29 of kodr: an MCP client seam, then one stubborn CSV expense example that refused to generate cleanly and dragged streaming, patch-oriented repairs, a regeneration, and memory scopes out of the harness on its way.
#ai#local-models#agents#mcp#memory#kodr -
inspect-json: a Zero-Dependency Lens for AI Logs
A small CLI I built to read LLM request and response payloads the way I actually want them - useful parts first, JSON-in-JSON decoded, zero dependencies. Now public on npm.
#ai#observability#agents#cli#nodejs -
Two More Apps, and a Policy Gate
Phases 21-23 of kodr: generating a Markdown blog and a notes API as harness trials - each one shakes out a real bug in kodr itself - plus a permission policy gate that builds on the hooks layer.
#ai#local-models#agents#security#testing#kodr -
A Deterministic Layer Around a Non-Deterministic Model
Phase 20 of kodr: pre_tool_use hooks - deterministic callbacks that can observe, mutate, or block a tool call before it runs, so policy lives in code instead of in a prompt the model might ignore.
#ai#local-models#agents#security#kodr -
A Task List the Harness Can Read
Phase 19 of kodr: a small task-plan primitive so a run can say what it thinks is done, blocked, or still pending. Early days - there is a fair bit still to come here.
#ai#local-models#agents#automation#kodr -
Autonomy, but on a Leash
Phase 14 of kodr: repeating a run for as long as it is useful, with a hard cycle count and explicit stop words so it never runs away.
#ai#local-models#agents#automation#kodr -
Exercising the Harness
Phases 15-18 of kodr: a local install, replay and model comparison, a security-review hardening pass, and using a real generated app to shake out bugs.
#ai#local-models#agents#testing#kodr -
One Repair, Then Stop
Phase 13 of kodr: when verification fails, let the model try to fix it exactly once - and no more.
#ai#local-models#agents#automation#kodr -
Limits Before Tools
Phase 11 of kodr: giving the model real tools, but wrapping every one of them in a budget first.
#ai#local-models#agents#tools#kodr -
Drawing the Workflow Before Hiring the Agents
Phase 12 of kodr: staged multi-agent coordination, modelled as plain deterministic data before a single extra model call is added.
#ai#local-models#agents#automation#kodr -
The First Full Coding Loop
Phase 10 of kodr: the moment all the careful little modules connect into a real prompt-to-patch loop.
#ai#local-models#agents#automation#kodr -
Running Checks Without Handing Over a Shell
Phase 09 of kodr: a verification runner that allowlists a handful of commands instead of giving the model a shell.
#ai#local-models#agents#safety#verification#kodr -
Skills, but Just the Markdown
Phase 07 of kodr: SKILL.md files as reusable instructions, with the executable runtime deliberately left out.
#ai#local-models#agents#skills#kodr -
Letting a Model Write Files Without Losing the Plot
Phase 08 of kodr: a path jail, dry-run diffs, and timestamped backups - the gate that sits between model output and your filesystem.
#ai#local-models#agents#safety#security#kodr -
Context Is Just an Input
Phase 06 of kodr: build the workspace context for a prompt deterministically, then let yourself look at it before it ever hits the model.
#ai#local-models#agents#context#prompt-engineering#kodr -
Defensive JSON Extraction
Local models love wrapping JSON in prose, fences, and broken escapes. Here is the little parser that survives them.
#ai#local-models#agents#parsing#kodr -
Learning by Commit
A technique for learning complex things: get an agent to build it, then learn by watching it work
#ai#local-models#learning#agents#kodr -
Probing Local Models and Building the Test Rig
Phases 02-04 of the kodr learning repo: connecting to LM Studio, faking a model server for tests, and capturing prompt run artifacts
#ai#local-models#learning#agents#lmstudio#testing#kodr -
Agentic Diagrams has slides now
Slides, flow sequence animations, AI Architect, and a beta tester callout for Agentic Diagrams.
#ai#agents#tools -
Agentxiety
The persistent, low-grade anxiety experienced when your AI agent subscription sits idle while you engage in non-productive human activities
#ai#agents#humour -
Agentic Diagrams
Using React Flow to visualise complex agentic systems - animated flows as a GIF
#ai#agents#tools -
Agentic Diagrams
Using React Flow to visualise complex agentic systems
#ai#agents#tools