DEVLOG.md

DEVLOG

2026-07-19

Work added:

• Added Rust doctor platform diagnostics for the Termux/Android fork-first direction: OS/arch, Termux detection, config path, shell, TTY/raw-mode availability, sandbox backend, and kernel-sandbox status now appear in JSON output.
• Marked Termux/Android as a degraded termux-contained backend instead of pretending Linux bubblewrap/seccomp semantics are available on Android/bionic.
• Documented the KiloCode fork-first Android strategy and added TODO items for a compatibility spike, source-mode validation, and Termux PTY/TUI checks.
• Reverted the raw TUI chrome to a pi-tui-default shape: Text output, Editor input, and CombinedAutocompleteProvider for slash/file completion instead of Axum-owned Session/Commands/Prompt panels and manual slash selection.
• Kept Axum-specific behavior only where necessary for command dispatch, provider/model commands, runtime output, and cancellation status.
• Updated TUI screenshot regressions for the simpler pi-tui rendering.
• Direction changed again by explicit user confirmation: start the Rust/clap/Ratatui migration and treat the pi-tui alignment as superseded.
• Added Phase1 Rust CLI crate under crates/cli: clap command surface for init, chat, tui, doctor, providers, modes, workflow, parallel, config-web, and run --auto; serde-compatible ~/.axum/config.toml; OpenAI-compatible reqwest provider calls with retry/backoff; scriptable doctor --json; five agent modes; and Phase2/3 placeholders that do not pretend Ratatui/tool runtime work is finished.
• Added Rust validation scripts to package.json and documented the migration shape in README/TODO.
• Continued into Phase2 with a first Ratatui TUI slice: session header, transcript pane, status/input pane, slash command suggestions, /tasks, /providers, /provider, /model, /mode, history navigation, cursor movement, paste/text insertion, Ctrl-Z undo, and Ctrl-K/Ctrl-U/Ctrl-Y kill-ring behavior.
• Continued into Phase3 with a first Rust tool sandbox slice: workspace path containment, read, exact-once precise edit, safe exec allowlist, read-only git allowlist, and shell-operator rejection; workflow skeleton now records sandboxed pwd evidence.
• Continued into Phase4 with distinct Code/Plan/Ask/Debug/Review system prompts wired into Rust chat, tui, and run --auto, so mode selection affects runtime behavior instead of only rendering labels.
• Continued into Phase5 with a first autonomous-run slice: run --auto now builds a noninteractive fail-closed system prompt with workspace/sandbox evidence and the safe tool policy instead of merely aliasing chat.
• Wired normal Ratatui prompt submissions into the Rust provider chat path and render assistant/error rows back into the transcript; streaming deltas remain a follow-up.
• Wired Rust ToolSandbox into the OpenAI-compatible provider loop: read, precise_edit, and safe_exec are advertised as tools, provider tool_calls are executed through the sandbox, and tool results are returned to the model for a follow-up completion.
• Added a Pi-style Plan/Now/Evidence/Result/Next/Issues trace around Rust provider/tool turns, exposed in JSON chat output and TUI transcript when /tasks is active.
• Added the first streamed provider-turn slice: axum chat --stream uses OpenAI-compatible SSE parsing for content deltas and writes deltas directly to stdout.
• Wired the Rust TUI to reuse streamed provider turns when launched with --stream, recording the assembled streamed assistant text and trace into the transcript; true in-frame incremental redraw remains a follow-up.
• Added a streamed tool-call delta accumulator for OpenAI-compatible SSE chunks, including split function.arguments; stream mode now executes one fully assembled streamed tool round through ToolSandbox and follows up with a non-streamed completion.

Validation:

• npm run build passes.
• npm test passes after snapshot updates.
• npm run pack:dry passes.
• cargo fmt --check, cargo test, and cargo build pass for the Rust crate after Phase5 auto-policy wiring.
• Rust TUI dry-run smoke renders expected header, command suggestions, and input line; provider response extraction is covered by unit test.
• Rust tool sandbox unit tests cover operator rejection, read-only git gating, workspace containment, precise edit, allowed cat execution, provider tool-call dispatch through sandbox tools, Pi-style trace rendering, OpenAI SSE content/tool-call delta parsing, and TUI stream flag rendering.

2026-07-17

Work added:

• Reworked the workflow skeleton from basic pi-style events into a Codex-style Plan/Execute shape with request/tool hash anchors and bounded auto-fix loop metadata.
• Added a Kilo/Codex-inspired tool registry for precise local edit, safe command sandbox, and LSP symbol lookup surfaces.
• Added permission-gated runtime executors for precise local edits, safe allowlisted commands, and lightweight TypeScript symbol lookup.
• Added axum parallel as the first swarm/sub-agent planning surface, persisting planned fan-out tasks with a hash-anchor merge policy.
• Added TUI /parallel <goal> :: <task> | <task> so the interactive shell can plan swarm fan-out without leaving the session.
• Added pi-style child-task statuses and merge-review metadata so future managed swarm execution has runtime-owned state instead of TUI-owned strings.
• Added provider-layer safety guard helpers that correct unsupported message/tool-call shapes before OpenAI-compatible transport.
• Replaced the fake workflow-only execution path with a Codex-like TypeScript runtime split into protocol, event bus, session, turn loop, and tool-runner layers.
• Routed axum chat through the runtime session so model sampling can request gated tools, feed tool output back into the next sampling request, and emit turn lifecycle events.
• Added OpenAI Chat tool-call transport support while preserving retry behavior and tool-role message forwarding.
• Hardened OpenAI-compatible provider diagnostics: HTML/Cloudflare challenge responses are categorized (http 403 html challenge, HTML/base-URL mismatch, JSON API error, or transport failure), doctor probes the same runtime/tool-call request shape used by TUI, tool-call transport failures retry once without tools, key source is reported as literal/env/missing, and masked request previews are included; provider listing now reflects root-level provider_config. Added /tasks in TUI, Enter-to-complete for bare incomplete slash commands, clearer permission-denied event details, shell-like splitting for allowlisted safe_exec commands such as npm test, guarded workspace inspection commands (find, grep, cat, sed, head, tail, wc, etc.), TUI provider/profile listing plus /provider use <id|number> switching, Activity dashboard rendering for /tasks, semantic progress text, summarized redacted command/file evidence, human-readable Working elapsed time, latest-activity status text, raw pi-tui input without the legacy boxed Prompt panel, fixed prompt label, or decorative input prefix, Working status on the editor bottom border, plus repeated-denial fail-fast behavior.
• Routed TUI prompt execution through AxumRuntimeSession, projecting runtime events into the interface while model/tool execution now uses the same Codex-like loop as axum chat.
• Fixed OpenAI-compatible chatWithTools transport retry handling so a failed no-tools fallback does not prematurely stop the configured network retry budget; the fallback success warning path remains intact.
• Reworked the TUI runtime projection into a user-facing Activity view with Plan/Now/Evidence/Result/Next/Issues sections, visible intent, evidence meaning, command/file counters, blocked-action explanations, and optional tool-call description/intent fields while keeping raw events as the audit/debug layer.
• Added Codex-style visible assistant streaming for runtime turns: OpenAI-compatible streaming now supports tool-call deltas, the runtime emits assistant_message_delta, and the TUI renders assistant text as the primary live output while keeping Activity projection as the auxiliary /tasks/status view.
• Reworked the default TUI runtime surface from a single Axum dashboard box into transcript-style rows: user prompts, assistant stream text, tool work, warnings, and blocked actions render as message/work cells, while the Plan/Now/Evidence Activity dashboard is kept out of the primary reply path for /tasks/debug use.
• Aligned TUI runtime progress closer to Codex: user prompts now use ›, Working status no longer pulls dashboard/activity text, tool rows use pending/success/failure markers, provider warnings and turn failures render as distinct rows, and repeated tool-denial failures summarize without exposing raw ENOENT paths as the final assistant answer.
• Hardened TUI display-width helpers so clipping, wrapping, and slash-command padding use terminal cell width instead of JavaScript string length for wide/combining Unicode.
• Added a 51-column TTY regression with Chinese/wide-symbol input to guard against pi-tui exceeds terminal width crashes after the Termux fix.
• Replaced Axum's raw TUI input/cursor/history/paste implementation with the @earendil-works/pi-tui editor component while keeping Axum-owned slash commands and runtime dispatch.
• Reworked the TUI chrome into panelized Session/You/Axum/Commands/Prompt sections so raw TTY and fallback rendering share the same cleaner pi-style screen shape instead of loose string rows.
• Pulled latest KiloCode (Kilo-Org/kilocode 938919a) and started the ag rewrite as a bounded Kilo-aligned split: moved terminal text-width helpers into src/tui/text.ts, moved runtime-visible transcript/projection rendering into src/tui/runtime-view.ts, exported those modules from the package, and changed the empty runtime fallback from • Working… to • Working so runtime status no longer uses a speech-like ellipsis.

Validation:

• npm test passes after workflow/tool/swarm/runtime/provider-diagnostic/TUI-runtime guard changes.
• npm run pack:dry passes; package dry-run includes the new dist/runtime/events.js, protocol.js, session.js, tool-runner.js, and turn.js files.
• npm test and npm run pack:dry pass after the Codex-style TUI progress/error display alignment.
• npm run build passes after the first Kilo-aligned TUI/runtime boundary split; full test and pack dry-run must pass before pushing this change.
• npm test and npm run pack:dry pass after splitting slash command registry/rendering and provider model-picker hydration/switching out of src/cli.ts.

Work added:

• Moved TUI slash command definitions, matching, completion, selected-row clamping, and command-palette rendering into src/tui/slash-commands.ts.
• Moved TUI model list hydration/fetching, dedupe, rendering, and model switching into src/tui/model-picker.ts.
• Added package exports for ./tui/slash-commands and ./tui/model-picker.
• Reduced src/cli.ts further so it imports command/model surfaces instead of owning their registries and picker logic.

2026-07-13

Direction changed toward a pi-aligned CLI with a TypeScript user-facing layer.

Work added:

• TypeScript CLI entrypoint under src/.
• OpenAI Chat Completions / OpenAI-compatible provider path for axum chat.
• Node shim now runs the TypeScript CLI directly.
• Mock HTTP regression test proving axum chat sends /v1/chat/completions requests with model, system/user messages, and bearer auth.
• Removed the old Rust workspace and prebuilt artifacts to avoid large disk usage and keep the project TS-only.
• OpenAI-compatible provider requests now retry transient transport/upstream failures 8 times by default with bounded 500–1500ms jitter; legacy fixed retry_delay_ms remains compatible, and /models fetches use the same retry policy.
• Added user-level TOML config loading (~/.axum/config.toml, AXUM_CONFIG, or --config) so provider settings can live outside project directories.
• Prepared npm packaging as axum-agent with committed dist output, axum binary, package exports, package metadata, MIT license, and dry-run pack script.
• Added a lightweight pi-style terminal UI command, axum tui, with a boxed terminal layout and --dry-run preview mode.
• Refined axum tui toward a Codex-like terminal layout with a top status line, conversation stream, and bottom message input.
• Fixed no-prompt axum tui so it stays open for interactive input instead of rendering once and exiting.
• Added Codex-like TUI behavior: dynamic terminal width, TTY alternate-screen redraws, and --no-alt-screen for scrollback-preserving sessions.
• Replaced the fake message box with raw TTY input rendering so typed characters appear inside the message frame before Enter sends.
• Moved TUI shortcut help below the message frame and added an in-frame cursor marker so the active input area is visible.
• Added an npm postinstall hook that creates ~/.axum/config.toml with an OpenAI-compatible template when missing, without overwriting existing config.
• Reworked axum tui empty state toward a compact Codex-like startup card with version/model/directory/permissions, a short tip, an inline prompt, bottom status, a dynamically updating Codex-like working timer, no placeholder/cursor rendered without user input, no user/assistant labels in the conversation body, and OpenAI-compatible SSE streaming updates while answers arrive.
• Added TUI model discovery and switching: config can provide models = [...], TUI defaults to the first configured/fetched model, OpenAI-compatible /models is used when the list is omitted even if a current model is already configured, /provider url and /provider key can save missing provider settings from inside TUI and immediately show the refreshed model list when fetch succeeds, /provider models lists models and /provider model <id|number> switches models by number/id, the empty input keeps an active cursor, ←/→ move the cursor for in-line edits, typing / shows a two-column command list with a selected item that Tab completes, and ↑/↓ recalls previous inputs when the slash command list is not active.
• Moved raw TTY rendering/input onto @earendil-works/pi-tui and added xterm-backed TUI screenshot snapshot tests for the slash command palette and remote model list layout.
• Reworked the TUI chrome toward Kilo/kilocode’s compact command/model picker style: tight title metadata, no fake tip/card, unified ▸ selection gutter, and updated screenshots as regression snapshots.
• Restored /model as the dynamic provider model command: /model fetches and lists current provider models, /model <id|number> refreshes then switches/saves, while /provider model <id> remains a custom-id persistence path.
• Added raw TUI paste handling for Shift+Insert/bracketed paste and multi-character input chunks, with a TTY regression covering pasted prompt submission.
• Capped long TUI model-list rendering so oversized provider model lists keep the first entries, current model, and hidden-count markers visible instead of pushing earlier models off-screen.
• Split raw TUI transient Working status from persistent output so command results/model lists are not replaced by the Working indicator while a request is running; raw pi-tui now places the Working indicator on the editor bottom border.
• Reworked the runtime dashboard into a task-first status panel with current task, progress, results, and issues; raw tool events and long command fragments are summarized as evidence instead of taking over the main view.
• Added a default concise AxumAgent system prompt for chat and TUI provider calls so short or ambiguous inputs stay natural and brief instead of turning into encyclopedia-style explanations; explicit --system still overrides it.
• Added configurable provider request timeouts (request_timeout_ms, --request-timeout-ms, AXUM_OPENAI_REQUEST_TIMEOUT_MS) with a 10-minute default and 0 disable option so long requests fail predictably instead of hanging indefinitely.
• Changed raw TUI busy-state interrupt handling so Esc/Ctrl-C cancels only the active provider request and returns to the prompt instead of exiting the TUI during a request.
• Added one-line provider setup via config provider_config = "<base_url> <api_key|env:VAR> <model>" and TUI /provider set <url> <key> <model>.
• Added axum config-web, a temporary local web page for editing provider URL/key/model in the same config file.
• Hardened config-web so env-referenced API keys stay as env:... and raw stored keys are not echoed into the page; leaving the key blank preserves the existing key.
• Added axum doctor for provider config and /models connectivity checks, including --json output for scripts.
• Added axum init for safe manual config creation with one-line provider setup and no overwrite unless --force is used.
• Removed install-time config mutation; npm install no longer writes ~/.axum/config.toml, and users initialize explicitly with axum init.
• Removed the now-unused bin/create-default-config.js helper from the shipped package surface.
• Added axum --version / -v for installed CLI identification.
• Added --provider <id> for chat, TUI, and doctor so users can select a configured provider without editing the default.
• Added axum providers / --json to discover configured provider ids without exposing literal API keys.
• Documented the release checklist and corrected install/config wording to make axum init the explicit first-run step.
• Reworked top-level help around the product flow (init → doctor → tui) and separated config-web options from chat options.
• Added GitHub Actions CI for npm test and npm run pack:dry on push/PR to main.
• Added the first Kilo-inspired shell layer without adding a KiloCode runtime dependency: axum modes exposes build/plan/debug mode profiles, while axum workflow maps shell prompts into pi-style events, permission gates, and project-local checkpoints under .axum/state.
• Refined axum workflow output into a compact Unicode stage view: no artificial shell:/runtime: labels, intermediate setup is folded by default, and --verbose expands the hidden steps.

Validation:

• npm test builds TypeScript, checks generated JS, runs the OpenAI-compatible mock CLI regression, and compares TUI screenshot snapshots.

2026-07-12

Continued SakuraByteCore/AxumAgent from a clean clone. This section documents the old Rust prototype baseline before the repository was simplified to TypeScript-only on 2026-07-13.

Baseline:

• npm test passes.
• cargo check --workspace passes with one existing dead-code warning in axum-server.
• cargo test --workspace passes, but the original repository had no effective unit tests.

Work added:

• Root README documenting project shape, checks, CLI examples, and current safety boundary.
• TODO tracking next decisions and hardening gaps.
• Unit tests for policy, tools, and runtime basics.

Validation:

• node bin/axum.js validate --spawn-server starts a managed server and returns ok.
• node bin/axum.js run "echo hello" --spawn-server --max-steps 2 completes the demo runtime loop through echo tool and finish action.

Rename validation:

• Replaced source/package/bin/crate naming from the old name to axum.
• Rebuilt Linux and Windows artifacts from renamed sources.
• Removed stale Android prebuilts because this host lacks aarch64-linux-android-clang; Android artifacts must be rebuilt with NDK.