jeo-code v0.4.1 — TUI card polish

Patch release on top of v0.4.0 (gate: tsc clean, 1,253 tests green).

• gjc card parity: ⟦Ctrl+O for more⟧ clip hint on truncated cards, code highlighting in card bodies, and Ctrl+O now dumps the FULL tool output (not just the last reply)
• Clip hint also covers summarize-stage markers (live-verified in tmux)
• Done-time todo reconciliation gate: stale Todos can no longer survive a finished turn — the final checklist always reflects what actually completed

jeo-code v0.4.0 — verified TUI, resilient engine, batch input

Everything since v0.3.0, verified live against the real binary (tmux PTY + 1,247-test suite, tsc clean).

TUI

• joc-ref transcript parity: Todo Write tree-checklist cards (☑ + strikethrough as items complete), line-numbered Write previews (1│ ... + ... N more lines), agent-name reasoning blocks, tree-connector skill detail
• DNA Claw identity: twist-frame helix animation + flowing emerald→cyan→violet gradient on live forge cards and the status box (live cards only — scrollback stays static)
• Renderer corruption fix: insertAbove now EL-clears uncovered rows eagerly — no more bottom-margin anchor desync devouring flushed cards
• Final report format: styled markdown replies (accent headings / bold / inline code), Todos completion receipt after the answer

Input

• Bracketed-paste batch input (prompt_toolkit contract): a multi-line paste runs one command per line, in order, with ▶ pasted command (+N queued) echo, ESC/cancel flush, and the trailing partial left editable in the box
• Mid-turn typing is queued and folded into the next prompt's prefill — fresh input never lost, never auto-executed out of order
• ESC cancel, double-Ctrl-C graceful exit, stdin-EOF safe REPL

Engine

• Refusal recovery: stop_reason=refusal no longer kills the turn — plain resend → context reset (per Anthropic's documented contract) → actionable guidance incl. OAuth→API-key hint
• Loop termination: StepBudget novelty rule + finite dynamic cap (600) — progressing turns extend, cycling turns terminate promptly
• Context overflow recovery, mid-turn context guard, per-message BPE token caching (×ばつ faster ralph subagent steps), bounded memos — no cumulative memory growth

Workflows & providers

• Native spec-first chain: deep-interview → ralplan → approve → team → ultragoal with gate verdicts
• task subagents (executor/planner/architect/critic) with per-role model/step pins
• Antigravity OAuth, live model discovery repairs, 429/limit resilience, Anthropic prompt caching
• Session persistence + compaction + instant /exit (detached memory distill), jeo session/state/update, MCP server, $skill invocation cards

OAuth authentication and a local (offline) provider, on top of the real coding agent.

Install

bun install -g github:akillness/jeo-code

New in v0.3.0

• jeoc auth — OAuth credential storage (login/status/logout → ~/.jeoc/auth.json, chmod 600). OAuth access tokens are sent as Authorization: Bearer (Anthropic + anthropic-beta: oauth, Gemini, OpenAI) and take precedence over API keys. Env tokens supported (ANTHROPIC_OAUTH_TOKEN / CLAUDE_CODE_OAUTH_TOKEN / GEMINI_OAUTH_TOKEN / OPENAI_OAUTH_TOKEN).
• Local provider ollama — native /api/chat, keyless, default model qwen2.5:0.5b. Runs the agent fully offline.
• jeoc doctor now verifies the ollama server + model pull and reports authMode readiness.
• Credential model: authMode = local | apikey | oauth | none.

Verified live: a local ollama/qwen2.5:0.5b agent run emitted a real write_file tool call that wrote a 12-byte file to disk — the full tool-calling loop runs offline. 29/29 tests, green CI, zero runtime deps.

See docs/09-auth-oauth-local.md.

Real multi-turn LLM coding agent, verified live.

Install

bun install -g github:akillness/jeo-code
jeoc setup --provider gemini # default model: gemini-2.5-flash
export GEMINI_API_KEY=...
jeoc agent "write sum.py with add(a,b), run it, report output"

New in v0.2.1

• Verified live: jeoc agent (Gemini gemini-2.5-flash) wrote a file, ran it with python3, and reported the output — real multi-turn tool-calling.
• jeoc setup — one-shot provider/model onboarding with API-key detection.
• jeoc models [--live] — known model registry + live Gemini ListModels (37 models for a valid key).
• Default Gemini model is now gemini-2.5-flash (gemini-2.0-flash free tier is often 0-quota).
• 19/19 tests, green CI, zero runtime deps.

jeoc is now a working LLM coding agent.

Install

bun install -g github:akillness/jeo-code
jeoc --version # 0.2.0

New: jeoc agent + jeoc config

• jeoc agent — tool-calling turn loop (bash/read_file/write_file/list_dir), mirrors gjc agent-core.
• jeoc config — provider+model config (.jeoc/config.json + env), secrets masked.
• Providers: gemini, anthropic, openai (real HTTP via fetch) + mock (hermetic, scriptable).

jeoc config set provider gemini && export GEMINI_API_KEY=...
jeoc agent "add a hello() to util.ts and run tests"
jeoc agent "..." --provider mock # no key needed

Also retains jeoc autopilot (autopil×ばつautoresearch ratchet) and jeoc ledger.
gjc internals mapped by subagents into docs/05–07. 16 passing tests + green CI. Zero runtime deps.

First release of the jeoc CLI (gajae-code rebrand study).

Install

bun install -g github:akillness/jeo-code
jeoc --version # 0.1.0

Features

• jeoc autopilot — autopilot end-to-end build loop hardened with the autoresearch ratchet: frozen evaluator, mandatory baseline, one change per step, keep-if-improved / revert-otherwise (min|max|gate), append-only log, convergence/stop, operator --on-revert hook.
• jeoc ledger — cross-plan append-only ledger (register → review → checkpoint → sweep → status); verified = review CLEAR + all goals complete + ≥1 sweep.
• Zero runtime dependencies (Bun + Node stdlib). 7 passing end-to-end tests + CI.

Rebrand map: gjc→jeoc, gajae-code→jeo-code, .gjc/→.jeoc/.