little-coder

Origin: itayinbarr/little-coder — Itay Inbar, 2026. npm package on Node, Apache-2.0, with the whitepaper Honey, I Shrunk the Coding Agent.

Loop type: ReAct-style native tool-calling loop, inherited unmodified from pi and instrumented per turn by 20 lifecycle-event extensions.

Primary surface: terminal (pi’s interactive TUI, one-shot, and --mode rpc) against a local llama.cpp or Ollama server on consumer hardware.

Chimera primitive: chimera/shrew/ (verified at 2507d0c).

little-coder does not ship its own agent loop. pi is a plain dependency in package.json — the agent loop, provider API, TUI, session tree, and compaction all come from it. Everything little-coder-specific lives in .pi/extensions/ (20 TypeScript extensions hooking pi’s lifecycle events), skills/ (30 markdown files the extensions inject on demand), and a Python benchmarks/ harness. The thesis: small local models already know how to code; what they need is a scaffold matched to their capability ceiling — fewer tools, enforced edit discipline, per-turn just-in-time guidance, and hard budgets on reasoning and turns.

Loop

flowchart TD
    Start([User prompt]) --> Assemble["Assemble per-turn system prompt:<br/>AGENTS.md + tool skill cards + algorithm cheat sheets"]
    Assemble --> Model["Model turn<br/>(llama.cpp / Ollama, OpenAI-compatible)"]
    Model --> Think{"Thinking tokens<br/>over budget?"}
    Think -- yes --> Retry["Abort turn, disable thinking,<br/>queue correction follow-up"] --> Assemble
    Think -- no --> Parse{Tool call present?}
    Parse -- "text-mode leak" --> Repair["output-parser repairs fenced /<br/>XML-ish / bare-JSON tool calls"] --> Exec
    Parse -- structured --> Exec["Execute tool<br/>(write-guard, permission-gate,<br/>checkpoint snapshot first)"]
    Parse -- none --> Done([Final answer])
    Exec --> Quality{"quality-monitor: empty /<br/>hallucinated tool / loop?"}
    Quality -- issue --> Correct[Inject correction follow-up] --> Cap
    Quality -- clean --> Cap{"turn-cap exceeded?"}
    Cap -- yes --> Abort([Run aborted])
    Cap -- no --> Assemble

Tool Set

Tool	Purpose	Notable Constraint
Read	File contents with line numbers	pi built-in
Write	Create files	Refuses if the file already exists (write-guard); the error returns the exact Edit call shape for the same path
Edit	Exact-match string replacement	old_string must match exactly, whitespace included; replace_all for repeated matches
Bash	Shell execution	30 s default timeout; permission-gate whitelists read-only commands (ls, cat, git log/status/diff, …)
Grep / Glob	Content search / filename patterns	Grep from pi; Glob added by the extra-tools extension
WebFetch / WebSearch	URL fetch / DuckDuckGo search	extra-tools extension
ShellSession (Cwd, Reset)	Persistent shell	tmux-proxy and subprocess backends; mounted for Terminal-Bench runs
BrowserNavigate/Click/Type/Scroll/Extract/Back/History	Playwright browsing	Mounted per benchmark (GAIA)
EvidenceAdd / EvidenceGet / EvidenceList	Per-session evidence store	1 KB snippet cap; survives compaction via evidence-compact

Prompt Strategy

• Base: pi’s ~1,000-token system prompt plus a 61-line AGENTS.md project prompt that pi auto-discovers. AGENTS.md states the runtime invariants (Write refuses on existing files, 30 s bash timeout) in the same terms the runtime enforces.
• Per-turn assembly: skill-inject appends a ## Tool Usage Guidance block — tool cards selected by error recovery > recency > intent priority under a configurable token budget. knowledge-inject appends ## Algorithm Reference — cheat sheets scored against the problem statement by keyword (weight 1.0) and bigram (weight 2.0) matching with an inclusion threshold. The prompt tells the model to trust these blocks because they were selected for the current turn.
• Edit format: exact-match search-replace. Write is for new files only — “prefer Edit over Write” is promoted from guidance to an enforced invariant.
• Per-model profiles: .pi/settings.json carries model_profiles[<id>] (temperature, max_tokens, thinking_budget, skill/knowledge token budgets) plus benchmark_overrides for Terminal-Bench and GAIA runs.
• No few-shot examples; instructive prose only.

Context Strategy

• Session persistence and auto-compaction come from pi’s session tree; the evidence-compact extension re-injects the evidence store after compaction so collected citations survive.
• Injected skill and knowledge blocks are re-scored and replaced every turn, not accumulated — their context cost is bounded by the two token budgets.
• The effective window is pinned at serving time: the canonical llama.cpp incantation serves 16k context with experts in system RAM and only attention plus KV cache on GPU, so context length is the dominant VRAM consumer. A ~22 GB quantized MoE checkpoint runs within 8 GB of VRAM this way.
• benchmark_overrides can raise context_limit per benchmark.

Termination Heuristic

• Natural stop: an assistant turn with no tool call ends the run (pi’s loop semantics).
• turn-cap: counts turn_start events per run and calls ctx.abort() over the cap. Caps are per-benchmark settings — unbounded for Aider Polyglot, bounded for Terminal-Bench and GAIA — or LITTLE_CODER_MAX_TURNS.
• thinking-budget: when thinking tokens exceed the budget, the in-flight turn is aborted and retried with thinking disabled, plus a follow-up nudging the model to commit to an implementation. Turn-level, not run-level.
• quality-monitor never stops the run; it converts failure signatures (empty response, hallucinated tool name, repetition loop) into corrective follow-up messages that restate the user’s original goal.

Notable Quirks

• Write refuses on existing files — the upstream calls this “the whitepaper invariant” and credits it as the single highest-leverage mechanism. The refusal error contains the exact Edit call shape, so the failure itself teaches the recovery.
• Nothing forks pi. All 20 mechanisms are auto-discovered extensions; each can be deleted or disabled per deployment via .pi/settings.json.
• Small-model extensions auto-disable when a large or cloud model is active, so the same install hosts both.
• The mechanisms survived a substrate swap: v0.0.x was Python on a different agent substrate (a ClawSpring derivative); v0.1.0+ rebuilt the same adaptations as pi extensions.
• The benchmark harness is dev-only Python (excluded from the npm package) and drives the agent over pi --mode rpc.
• Hardware target is a consumer laptop, not a GPU server — the README’s reference runs used 8 GB of VRAM with the MoE expert-offload trick.

In Chimera

Shrew (chimera/shrew/, alias chimera tiny) replicates the posture with the same layering: shrew builds on weasel — Chimera’s replica of pi — exactly as little-coder builds on pi, so substrate improvements flow through.

Adopted:

• Small-model defaults: qwen3.6-35b-a3b via llama.cpp, --max-steps 30, --allowed-tools Read,Write,Edit,Bash (cli.py).
• Curated skill markdowns under chimera/shrew/skills/{knowledge,protocols,tools}/ with frontmatter (name, description, triggers) and ~/.shrew/skills/ plus project overlays.
• Scaffold-fit prompt wrapping below 13B active parameters (extensions/scaffold_fit.py) and tool-list trimming below 9B (extensions/tool_filter.py).
• MoE-aware context sizing as code (extensions/moe_offload.py: a MoEModelProfile catalog and compute_optimal_context_window() driven by --vram-gb).
• Text-mode tool-call repair (output_parser.py: fenced tool blocks, <tool_call> wrappers, bare JSON, Python call shorthand).
• Quality monitoring (quality_monitor.py: empty response, hallucinated tool name, self-correction language, repetition loop, plus a correction-message builder).
• Per-turn skill re-ranking (skill_injector.py: error > recency > intent scoring, top-K bodies injected at a system-prompt marker).
• Per-model profiles with benchmark_overrides (model_profiles.py, reading ~/.chimera/shrew/settings.json).
• Checkpoint-before-write and bash permission-gate extensions; Aider Polyglot, GAIA, harbor, and terminal-bench adapters under benchmarks/.

Diverged:

• Packaging: a Python module inside chimera-run, not an npm package on Node.
• The provider chain probes live — llama.cpp /health first, then vLLM, SGLang, and Ollama, before any cloud key — where the upstream routes by environment variable without probing and supports llama.cpp and Ollama only.
• Extensions are stdlib-only pure functions composed as hooks, not TypeScript lifecycle-event subscribers.
• Skill content leans toward scaffold discipline and a larger protocol set, where the upstream leans toward per-tool cards and algorithm cheat sheets.
• Shrew inherits Chimera substrate the upstream lacks: cooperative cancellation, event-sourced sessions shared with the sibling CLIs, and secret redaction.

Surface-by-surface status: parity matrix.

References

• Upstream repo: github.com/itayinbarr/little-coder (Apache-2.0). Read firsthand for this page: README.md, AGENTS.md, and the .pi/extensions/ sources (turn-cap, skill-inject, thinking-budget).
• Whitepaper: Honey, I Shrunk the Coding Agent (Substack, 2026).
• Substrate: pi (cited by the upstream as badlogic/pi-mono) — replicated in Chimera as weasel.
• Replicated in Chimera at commit 2507d0c: chimera/shrew/ — CLI, REPL, providers, 12 extension modules, 21 skill markdowns, 4 benchmark adapters.
• Shrew parity matrix — GREEN/YELLOW/RED per surface.