synaps 0.3.3

# Changelog

All notable changes to this project will be documented in this file.

## [Unreleased]

## [0.3.3] — 2026-06-14

### Added
- **`night-city` theme** — a premium neon-noir palette (Cyberpunk / Blade Runner / Neuromancer accents on a near-black base). Strictly opt-in via `theme = night-city`.
- **Tool transcript overhaul** — per-tool colored gutter + glyph identity (bash / read / write / edit / grep / find / ls / subagent / ext), subtle per-tool panel backgrounds (input darker, output lighter), paired input→output ordering for parallel tool calls, and an `ok / timed out` status footer under each result. All themeable via new `tool_*` theme keys.

### Performance
- **Streaming redraw throttled ~60fps → ~10fps during active turns** (#131). The per-delta full `RenderModel` rebuild — which scales with session size — was the main-thread CPU cost; the spinner only needs ~10fps and deltas arrive faster than the eye reads. The idle/final frame still renders immediately, so end-of-turn state never lags.

### Fixed
- **Tool colors no longer leak across themes** — `tool_*` accents now derive from each theme's *own* palette (only `night-city` carries the neon set), and the no-config default theme is unchanged. Pick `dracula` and the tool gutters look like dracula, not Night City.
- **Thinking spinner could spin forever** on an empty thinking step — the placeholder is now dropped when the agent produces text/tools or the turn ends. A real model `…` is preserved via a sentinel so it's never mistaken for the placeholder.

## [0.3.2] — 2026-06-14

### Fixed
- **`synaps.log` was silently dead** — after the A3 split the tracing `EnvFilter` only enabled `synaps_cli=debug`, but the code now emits from `agent_core` / `agent_engine` / `agent_tui` targets. With `RUST_LOG` unset (the default) every log line was filtered out. Added the three crate targets to the filter.
- **Version header showed a bogus `<hash>-dirty` on source-tarball installs** — `build.rs` shelled out to `git` and trusted whatever repo it found; AUR/Homebrew extract the source *inside* the package's own git checkout, so it baked an unrelated commit + a spurious `-dirty`. Now it only trusts git when its repo root **is** the synaps workspace root; any packaged build (AUR / Homebrew / crates.io) shows `vX.Y.Z · release` instead. CI-built release binaries still carry the real commit.
- **Empty `end_turn` could misclassify as a silent-stop error in the default (telemetry-off) config** — `has_stop_reason` was derived from a telemetry-gated field. Now captured unconditionally. (#130 follow-up)
- **In-stream API error bodies reached the screen unsanitized** — a hostile/MITM endpoint could inject terminal escape sequences via an `error` event message. Error text is now sanitized like notices.
- **Mid-stream transport drops are now retried** through the unified retry budget instead of hard-failing the turn.
- **`build.rs` `-dirty` no longer triggers on untracked files** (`--untracked-files=no`); SSE line buffer gained an 8 MiB hard cap against a no-newline OOM; telemetry records now carry the real retry attempt number.

### Changed
- **crates.io publish job** runs with verification enabled and keys idempotency off the process exit code (not log grep), so a genuine publish failure fails the release loudly instead of silently skipping a crate.

### Docs
- User + contributor docs (README, AGENTS, in-app `/help`, and the 36-page wiki) cross-checked against the code and refreshed for the 0.2→0.3 / A3 workspace reality.

## [0.3.1] — 2026-06-14

### Fixed
- **`cargo install synaps` from crates.io was broken in 0.3.0** — after the A3 workspace split, `synaps-engine` embedded its help data via `include_str!("../../../assets/help.json")`, a file at the *workspace root, outside the crate*. `cargo publish` only packages files inside the crate, so the published `synaps-engine`/`synaps` 0.3.0 could not compile (`No such file: assets/help.json`). The asset now lives inside the crate at `crates/agent-engine/assets/help.json`, and the crates.io publish job runs **with verification** (no `--no-verify`) so this class of packaging bug is caught before the permanent upload. Binary installs (GitHub release, Homebrew, AUR, shell installer) were never affected — they bundle the full workspace. The broken 0.3.0 crates are yanked.

## [0.3.0] — 2026-06-13

### Added
- **Workspace architecture split (A3)** — the monolith is now three library crates (`agent-core`, `agent-engine`, `agent-tui`) behind the `synaps` binary. Cleaner boundaries, faster incremental builds, independent test surfaces. No user-facing CLI changes.
- **Dedicated render thread** — the terminal is driven from its own thread off an immutable `RenderModel` snapshot, decoupling draw latency from the event loop and closing a class of resize/teardown races. The signal watchdog is retired now that rendering is isolated.
- **`/stats` command** — per-session token + cost receipt with a cache-split footer.
- **`/context` command** — inspect live context-window usage.
- **Configurable prompt-cache TTL (`cache_ttl` config key)** — `5m` (default, unchanged behavior), `1h`, or `hybrid` (1h on the stable tools/system prefix, 5m on the per-turn message tail). Long-gap sessions (> 5 min between turns) can keep their cache warm instead of paying full input price every turn. Invalid values fall back to 5m with a boot warning. Spec: `docs/specs/cache-ttl-spec.md` (v4)
  - Pricing is TTL-aware: 1h cache writes billed at 2.0× input (5m stays 1.25×); when the API reports no 5m/1h split, aggregate writes are billed at the 5m rate (fail-cheap)
  - `extended-cache-ttl-2025-04-11` beta header sent only on API-key auth — 1h is live-verified working bare over OAuth, whose beta set is left untouched
  - Silent-downgrade detector: if 1h is configured but the account never honors it, a one-time-per-session notice fires; the configured mode is never auto-flipped
  - RPC protocol: `agent_end.usage` gains optional `cache_creation_5m` / `cache_creation_1h` fields (omitted when unknown — additive, existing consumers unaffected); same split flows through server-mode `Usage` messages and subagent aggregation
- **`/model <name>` and `/thinking` persist to config** — in-session changes survive restarts.
- **Paste while streaming** — input can be pasted mid-turn without dropping characters.
- **Build commit in the TUI header** — the version now shows the exact git commit the binary was built from (e.g. `v0.3.0 · 88e08e0`), with a `-dirty` marker for builds from a modified tree.

### Changed
- **Honest billing** — reported cost now reflects exactly one Usage event per request and TTL-aware cache-write pricing. Figures will differ from 0.2.1 (more accurate, not more expensive).
- **Signal shutdown overhaul** — SIGINT/SIGTERM handling rebuilt around the real root causes: clean teardown ordering, bounded budgets, and no watchdog.

### Fixed
- **The "stopping" bug (#130)** — in large conversations the agent could silently end a turn after running tools, never landing the synthesizing reply. Anthropic delivers some failures (e.g. `overloaded_error`) as an in-stream `error` event under a 200 response; the SSE parser had no `error` arm and dropped it, leaving an empty stream that was treated as a clean end-of-turn. The parser now surfaces these as visible, actionable errors; an empty model response is never swallowed; and the retry budget is unified so HTTP 5xx, transport drops, and in-stream errors all draw from one policy (429 keeps its dedicated budget). Prompt caching is unaffected — retries re-send the byte-identical request, hitting cache.
- **Cache-TTL usage split was read from the wrong SSE event** — live `message_delta` frames carry only the aggregate; the 5m/1h breakdown arrives on `message_start`. The split is now captured at `message_start` with a delta-arm fallback, restoring telemetry split keys, the downgrade detector, and correct 1h write pricing in streaming sessions.
- **Signal delivery + shutdown** — busy-loop and unkillable-process on a dead PTY fixed (three-layer); SIGINT clean-exit path restored.
- **Toast renderer panic** on small-terminal resize (`min > max`).
- **Idle logo gradient** frozen until the first keystroke.
- **Retry backoff** is now header-aware (honours `retry-after`/reset) with a higher 429 budget.
- **`/model`, `/thinking`, `/compact` side effects** now run on the live engine dispatch path.
- **Server-mode clients** now receive `Notice` events.

### Performance
- **Boot:** `--continue` no longer parses every session file (~11s → ~1s); the skills fs-walk moved off the async runtime; `plugins.json` read hoisted out of the extension-load loop.
- **Sessions:** `/sessions` reads headers + recent-only instead of parsing 76MB; `save_chain`'s 76MB collision scan dropped.
- **Runtime:** request body serialized once across retries; `sanitize_thinking_blocks` O(N²) → single tail pass.
- **TUI:** dirty-checked widget events killed the ~30% idle CPU burn (#119); heavy `App` structures Arc-projected; `/extensions` audit bounded to a tail read.

## [0.2.1] — 2026-06-12

### Fixed
- **AUR build failure on LTO-enabled systems** — the rustls migration (0.2.0) introduced `ring`, whose C/asm objects are GCC-LTO-incompatible with `rust-lld`: makepkg's `-flto=auto` produced bitcode the linker can't read, failing every `ring_core_*` symbol. PKGBUILD now sets `options=(!lto)`; stale `openssl` runtime dependency dropped (TLS is rustls since 0.2.0). Binaries unchanged from 0.2.0.

## [0.2.0] — 2026-06-12

### Performance
- **Typed SSE pipeline — per-token DOM allocation eliminated** — the streaming hot loop no longer builds a `serde_json::Value` per event
  - New `runtime/sse.rs::SseLineBuffer` — zero-copy line buffering: lines borrowed as `&str`, memchr (SIMD) newline search, O(1) read-cursor advance with periodic compaction (was two copies + an O(n) drain per line); UTF-8 splits across chunk boundaries handled by construction, with an exhaustive every-offset re-chunking test suite
  - Typed `AnthropicEvent` wire model with `Cow<'a, str>` borrowing — text deltas borrow from the line buffer on the escape-free fast path, own only when JSON escapes force decoding
  - `ParseState` + `process_event` consolidate the three parse sites (main loop, tail-flush, end-of-stream finalize) into a single write path — duplicate-site drift now structurally impossible; 15 behavioral seam tests pin the contract
- **Dirty-flag render loop + streaming redraw coalescing** — the TUI burned a full core during streaming (unconditional full-frame draw on every 16ms tick and every SSE delta); `needs_redraw` flag draws only on actual state change, idle costs zero draws
- **Streaming raised 30→60fps** — redraw coalescing throttle tightened now that frames are cheap
- **Draw-path allocation elimination** — viewport `Paragraph` moves the line Vec instead of deep-cloning every frame; ASCII-art rendering collapses same-style runs from one Span-per-char to one Span-per-line; per-frame version `format!` → `concat!` const
- **`mem::take` at session resume** — `rebuild_display_messages` no longer deep-copies the entire message history (potentially MBs) to dodge the borrow checker
- **Cache strategy: sliding-4 → single-last** — one stationary `cache_control` marker on the last message replaces ~57 lines of sliding-breakpoint logic
  - Bench evidence: equivalent hit rate (96–97% both strategies); live-verified 99% hit on warm turn
  - Prefix-invalidation bug class (mutating old markers) eliminated by construction; 5 new unit tests on a previously untested function

### Fixed
- **SSE double-emit on partial final line** — when the final `content_block_stop` arrived without a trailing newline, the tail-flush pushed the block but left `in_thinking`/`in_tool_use` set, so the end-of-stream flush pushed it again — duplicated content block sent to the model on the next turn
- **PTY zombie reaping** — `PtyHandle::Drop` killed the child but never `wait()`ed; every timed-out or dropped shell session left a zombie. Now reaped with bounded-retry `try_wait()` (all other child-process sites already used `kill_on_drop`)
- **Extension crash-loop health lied** — `restart_count` reset on successful handshake, so an extension that initialized fine and died on the next request never hit the exhaustion limit and reported Running. Consecutive-failure counter now resets only when the restarted process actually serves a call; new `total_restarts` (never reset) feeds health so recovered extensions report Degraded
- **Integration test rot from `env_clear`** — security-hardened extension spawns silently broke 20+ test fixtures parameterized via env vars; fixtures moved to argv parameters, plus poison-cascade and parallel-env-race cleanup — full suite green, 3 consecutive runs
- **Hardcoded Claude Code identity removed from API preamble** — replaced with a configurable `identity` config key + proper SynapsCLI default
- **Dotted (namespaced) config keys exempt from unknown-key warnings**

### Added
- **claude-fable-5 support** — known-models entry with adaptive thinking, pricing ($10/$50 per MTok), and 1M context opt-in
- **Telemetry module** — structured per-request API records: `TelemetryLevel` (off/basic/full), usage with cache TTL breakdown, rate-limit headers, cache-diagnosis records; JSONL writer to `~/.cache/synaps/api-log.jsonl` (0600 + `O_NOFOLLOW`), wired into config keys and the Anthropic SSE stream
- **Cache strategy benchmark suite (`bench/`)** — 21 tool-heavy questions with deterministic outcomes across 4 swappable breakpoint strategies (none / single-last / last-3 / sliding-4); per-turn JSONL logging of tokens, cache read/write, hit %, cost, latency; `compare.py` for side-by-side runs
- **`synaps completions <shell>`** — shell completions via clap_complete (bash, zsh, fish, elvish, powershell)
- **First-run banner** — boot with no OAuth, no `ANTHROPIC_API_KEY`, and no provider keys shows a getting-started banner with the three setup paths instead of a silent TUI

### Changed
- **reqwest 0.11 → 0.12 — duplicate HTTP stack collapsed** — reqwest 0.11 was the sole consumer of hyper 0.14/http 0.2/h2 0.3 alongside axum's hyper 1.x; the old stack (plus native-tls/OpenSSL) is gone, TLS is now rustls with native root certs — explicit in the manifest, smaller binary, zero source changes
- **Humanized API + network errors** — raw JSON dumps and reqwest debug strings replaced with actionable text (529 → "overloaded, wait", 401 → "run synaps login", context overflow → "run /compact"); retry notices ("⏳ API error, retrying…") are now display-only system lines instead of fake assistant content persisted into session history
- **Input recovery on stream error** — when a stream dies after retries, the user's popped message is restored to the input box instead of silently destroyed
- **Config parse warnings with did-you-mean** — unknown keys suggest corrections via levenshtein (`modle` → did you mean `model`?); unparseable values warn and fall back to defaults instead of silently misbehaving
- **Login/status polish** — token-refresh errors now say `synaps login` (was a nonexistent command); `synaps status` refreshes the OAuth token before the request instead of 401ing on expired tokens; `--help` rewritten with real about text and `--profile` documented

### Earlier in this cycle
- **Engine refactor + `chat` headless mode** — `daemon`, `run`, and `client` subcommands deleted; `chat` is now fully-featured headless mode with MCP, extensions, skills, sessions, compaction, and the event bus (same engine as the TUI, stdin/stdout rendering)
  - New `engine/` module: `setup.rs` (boot), `commands.rs` (headless slash commands), `stream.rs` (StreamEvent consumer), `session.rs` (ConversationState)
  - `chatui/` module deleted — `tui/` is the sole frontend
  - New `pricing.rs` — centralized Anthropic pricing (Opus/Sonnet/Haiku) with cache billing support
  - New `harbor/synaps_agent.py` — Terminal-Bench integration agent
  - Clippy CI added (`cargo clippy --all-targets` gates PRs)
  - Published on crates.io as `synaps` (v0.1.x); available via `cargo install synaps`, AUR (`yay -S synaps`), Homebrew, and GitHub Releases
- **Path B Phase 6 — The Big Move (extension contracts)** — synaps-cli core no longer contains whisper-specific code; the local-voice plugin owns it
  - Deleted `src/voice/{models,download,rebuild}.rs` (whisper.cpp catalog, HuggingFace downloader, cmake-rebuild orchestrator)
  - Slimmed `src/voice/discovery.rs` from ~400 LOC to ~190: only `discover()` / `discover_in()` / `DiscoveredVoiceSidecar` remain; `read_build_info`, `detect_host_backend`, `probe_*` are gone (callers use the Phase 5 `info.get` cache instead)
  - Deleted in-core `/voice models|help|download|rebuild` action arms and parsers — those subcommands now route through the plugin via the Phase 2 interactive command contract; missing-plugin case errors with a clear message
  - Deleted in-core whisper download UI: `App.{download_progress, download_filename, download_rx, voice_download_in_flight, model_browser_selected, cached_voice_compiled_backend}` plus `start_download` / `on_download_progress` / `on_download_complete` and the `download_change` event-loop arm; the sticky progress widget now exclusively renders generic `extensions::active_tasks::ActiveTasks` (Phase 3)
  - Deleted `EditorKind::{ModelBrowser, WhisperModelPicker}`, `ActiveEditor::ModelBrowser`, `ModelBrowserRow`, `model_browser_rows`, `whisper_model_options`, `render_model_browser`, `render_download_progress_line`, plus the `voice_stt_model` / `voice_stt_backend` / `voice_language` setting definitions — the plugin replaces them via the Phase 4 settings categories contract
  - **Migration shim:** on first launch, legacy global voice config keys (`voice_stt_model_path`, `voice_stt_model`, `voice_stt_backend`, `voice_language`) are copied into the plugin namespace `~/.synaps-cli/plugins/local-voice/config` if the destination is empty; legacy keys remain in place for safety. Reads at runtime prefer the plugin namespace and fall back to the legacy global key with a deprecation warning.
  - Net change: `~−2,400` LOC removed from synaps-cli core; behaviour preserved when the local-voice plugin is installed
- **Phase 2 — Extensions Capability Platform** — extensions are now first-class citizens
  - Slash command extension contract (`commands` in plugin manifest)
  - Settings panel extension contract (`settings` in plugin manifest)
  - Keybind extension contract (User-tier binds via plugin manifest)
  - Model-provider extension contract — extensions can register OpenAI-compatible providers as full agentic backends with tool-call support
  - Capability discovery + permission gating per extension
  - Slices P–W complete; 836 tests green (682 lib + 154 bin)
- **Voice dictation** — toggle-driven mic capture wired to the input buffer
  - `/voice` slash command with subcommands: `models`, `download <id>`, `help`, `rebuild [backend]`
  - Configurable toggle key — defaults to F8; supported: F8, F2, F12, C-V, C-G (Ctrl+Shift and Ctrl+Alt are unreliable in terminals)
  - Toggle-only flow: press once → start listening (🎤 listening pill), press again → stop, transcribe, append to input
  - VAD-aware: final transcripts during an armed session insert text without disarming
  - Settings → Voice category: toggle key, language cycler (14 langs), STT model, STT backend
  - `voice_language` cycler: `auto / en / es / fr / de / it / pt / nl / ja / zh / ko / ar / hi / ru`
  - Hot-reload keybind on settings save — no restart required
  - Kitty keyboard protocol enabled for reliable F-key + modifier capture
  - Voice sidecar lives in the `local-voice-plugin` (synaps-skills repo) — synaps core stays voice-free
- **Whisper model manager** — discover, download, switch models in-app
  - 10-entry catalog hard-coded in `src/voice/models.rs`: tiny / base / small / medium / large-v3 / large-v3-turbo + `.en` variants, with real SHA256s from HuggingFace
  - `/voice models` renders an installed/uninstalled table
  - `/voice download <id>` — streaming download with atomic `.partial → rename` install, SHA256 verification, cancellable, single-in-flight
  - `EditorKind::ModelBrowser` — Settings → Voice → STT model browses the catalog inline, Up/Down to nav, Enter installs-or-selects
  - Inline footer download progress while a model fetches
  - Models live under `~/.synaps-cli/models/whisper/`
- **Whisper backend selection** — pick CPU / CUDA / Metal / Vulkan / OpenBLAS or auto-detect
  - `voice_stt_backend` cycler in Settings → Voice
  - `auto` probes the host (`nvcc`, `vulkaninfo`, `pkg-config`, target-os) and picks the best available
  - "Current build: <backend>" annotation surfaces the sidecar's compiled feature set
  - ⚠ rebuild annotation when the selected backend differs from the compiled one
  - `/voice rebuild [backend]` invokes `local-voice-plugin/scripts/setup.sh --features <backend>` and streams build output as System rows
  - Sidecar reports its build via `--print-build-info` (JSON: `{backend, features, version}`)
- **Chat UI rendering polish**
  - System / Error messages now split on `\n` and word-wrap on each sub-line (same path as User / Text)
  - Tab-aware soft wrap: continuation rows indent under the last-tab anchor column (4-col tab stops); safety clamp falls back to leading-space indent if the anchor exceeds 60% of width
  - Fixes `/voice help`, `/chain ls`, `/keybinds`, plugin command output rendering
- **Extension System** — process-based JSON-RPC hooks (before_message, before_tool_call, after_tool_call, on_session_start, on_session_end)
  - Tool-specific filtering: hooks can target specific tools
  - Context injection via HookResult::Inject
  - Permission-gated with 6 permission types
  - Drop-in plugin support
  - Silent loading (only show failures)
- **Watcher notify_inbox hook** — event bus notification on agent completion
  - Config option: `[hooks] notify_inbox = true` in watcher agent config
  - Drops JSON events into `~/.synaps-cli/inbox/` when agents complete
  - Works in both 'once' and 'deploy' modes
  - Event payload includes agent name, session number, elapsed time, exit code, timestamp
- **Table rendering improvements**
  - Cells wrap instead of truncating with ellipsis
  - Inline markdown (bold/italic/code) rendered in table cells
  - Smarter column shrinking — preserves narrow columns
- **Structural refactoring** — improved code organization
  - `catalog.rs` → `catalog/` directory with per-provider modules
  - `palettes.rs` → `palettes/` directory with per-palette files
  - `tools/subagent*.rs` → `tools/subagent/` directory
  - `tools/tests.rs` → distributed inline test modules per tool
  - `runtime/api.rs` split into `api.rs` + `api_sync.rs` + `request.rs`
  - `chatui/mod.rs` helpers extracted to `helpers.rs` + `lifecycle.rs`
- **UTF-8 char boundary panics** — multiple fixes for multi-byte character handling
  - Hook output truncation now finds valid char boundaries
  - Bash output highlighter fixed for emoji/CJK characters
  - Additional edge cases found by PR review
- **Hook system improvements**
  - Handle HookResult::Block in before_message hook
  - Set tool_runtime_name in all before_tool_call hook paths
  - Track truncation explicitly in bash tool instead of string matching

### Added
- **Multi-Provider Runtime** — OpenAI-compatible provider engine
  - 17 providers: Groq, Cerebras, NVIDIA NIM, OpenRouter, Google AI Studio, DeepInfra, HuggingFace, Fireworks, Hyperbolic, Scaleway, SiliconFlow, Together, Chutes, Codestral, Perplexity, OVHcloud, + Local (Ollama/LM Studio/vLLM)
  - 55+ models cataloged with tier ratings (S+/S/A+/A/B)
  - `provider/model` shorthand: `/model groq/llama-3.3-70b-versatile`
  - SSE streaming with tool calling across all providers
  - StreamDecoder with HashMap-based tool call accumulation
  - Anthropic↔OpenAI message/tool translation layer
  - Provider router in `api.rs` — model prefix routes automatically
  - Config: `provider.<name> = <key>` in `~/.synaps-cli/config`
  - Env var fallback: `GROQ_API_KEY`, `CEREBRAS_API_KEY`, etc.
  - Local model support: `provider.local.url = http://localhost:11434/v1`
- **`/ping` command** — health-check all configured provider models
  - Non-blocking, results stream in live as each model responds
  - Shows ✅/❌/⏳/🔒/⌛ with latency per model
  - Results cached for model picker display
- **Settings → Providers** — TUI panel for API key management
  - View all 17 providers with key status (✅ set / ⬚ not set)
  - Enter to set/update key, d/Del to clear
  - Key masking (last 4 chars only)
  - Local endpoint URL editing
  - Press `p` to ping all models
  - Scrollable list with overflow indicators
- **Settings → Model Picker** — expanded to show provider models
  - Models grouped by provider with headers (── Groq ──, etc.)
  - Only shows providers with keys configured
  - Header rows are unselectable (skip on navigation)
  - Health status shown when ping data available
- **App starts without Anthropic credentials** — users with only provider keys can boot the TUI and use free models
- **Session Naming** — `/saveas <name>` aliases for sessions
  - Name format `[a-z0-9-]{1,40}`, validated and collision-checked
  - `/saveas` (no arg) clears the name
  - `synaps --continue <name>` resolves session names
  - `/sessions` shows `[@name]` tags on named sessions
- **Chain Naming** — `/chain name/list/unname` bookmarks for compaction lineages with auto-advance
  - `/chain name <name>` bookmarks the current session's lineage
  - `/chain list` shows all named chains (`*` marks active)
  - `/chain unname <name>` removes a bookmark
  - `/chain` (no args) shows lineage + "bookmarked by: @name" if present
  - Chain pointers stored at `~/.synaps-cli/chains/<name>.json`
  - On `/compact`, chain pointers auto-advance to the new session
- **Unified Session Resolution** — `resolve_session()`: chain name → session name → partial ID, used by `--continue` and `/resume`
  - Shared by `synaps --continue`, `/resume`, and server `--continue`
  - Resolution path surfaced as a system message (`↳ resolved via chain 'foo'` / `↳ resolved via name 'bar'`)
  - `--continue` value_name changed from `SESSION_ID` to `NAME_OR_ID`
- **Event Bus** — universal message ingestion for agent sessions
  - `synaps send` CLI command with atomic file writes
  - inotify inbox watcher via `notify` crate (spawn_blocking, non-blocking)
  - Priority EventQueue with severity ordering (Critical→front, High→after)
  - `tokio::sync::Notify` for instant TUI wake on event push
  - Events auto-trigger model turns when agent is idle
  - Events buffer during streaming via `pending_events`, flush on completion
  - Styled TUI event cards with severity icons (🔴🟠🟡🔵)
  - XML-wrapped event format with prompt injection hardening
  - 256KB file size cap, symlink guard, 0700 permissions
- **Reactive Subagents** — dispatch, poll, steer, collect
  - `subagent_start` — spawn and return handle_id immediately
  - `subagent_status` — non-blocking progress snapshot
  - `subagent_steer` — inject guidance mid-flight via steering channel
  - `subagent_collect` — non-blocking result check
  - `subagent_resume` — restart timed-out agents (stub)
  - `SubagentHandle` with shared `RwLock<SubagentState>`
  - `SubagentRegistry` with cleanup_finished on stream Done
  - Abort cancels all running reactive subagents
  - Thread handles stored for graceful shutdown
  - 13 unit tests for handle + registry
- **Chain Sessions** — `/compact` creates linked child session
  - Old session preserved on disk with `compacted_into` forward link
  - New session starts with `parent_session` back link
  - `/chain` command walks the session lineage
  - System prompt included in compaction summary
  - Configurable compaction model (defaults to Sonnet)
  - ModelPicker for compaction model in settings
  - Non-blocking compaction with spinner animation
  - Compacted summary hidden from TUI display
  - Message queuing during compaction
- **`respond` tool** (stub — returns honest failure until wired)
- **`send_channel` tool** (stub — returns honest failure until wired)
- **`/status` command + `synaps status` subcommand**: check account usage (5-hour, 7-day, Sonnet) with progress bars and reset countdowns. Hits OAuth usage API.
- **`/compact` slash command**: summarize & compact conversation history when context gets long
  - Structured checkpoint format (goals, progress, decisions, file ops, next steps)
  - Iterative compaction — re-compacting merges new work into existing summary
  - File operation tracking (read/write/edit paths preserved across compactions)
  - Custom focus instructions via `/compact <focus>`
  - Uses dedicated low-effort API call (no tools, summarization system prompt)
- **`context_window` setting wired to API**: `200k` (default) omits beta header; `1m` sends `context-1m-2025-08-07` on supported models (Opus 4.6+, Sonnet 4.x); previously was UI-only display cap
- **Claude Code marketplace compatibility**: probe both `.synaps-plugin` and `.claude-plugin` layouts, `${CLAUDE_PLUGIN_ROOT}` substitution in skill bodies
- **Plugins subdir sources and cascade uninstall**: install from subdir-based plugin repos, cascade-remove plugins when their marketplace is deleted
- **Settings → Plugins marketplace overlay**: "Open Plugin Marketplace" action row in Settings, opens plugins modal as nested overlay
- **Hidden binary**: GamblersDen bundled as `hidden` binary alongside `synaps`
- **Single binary architecture**: all binaries consolidated into `synaps`
  - `synaps` (no args) = TUI (was `chatui`)
  - `synaps chat` = fully-featured headless mode (MCP, extensions, skills, sessions)
  - `synaps server` = WebSocket API (was `server`)
  - `synaps agent` = headless worker (was `synaps-agent`)
  - `synaps watcher` = supervisor (was `watcher`)
  - `synaps login` = OAuth (was `login`)
  - `synaps send` = push events into a running session
  - `synaps status` = check account usage
  - (Removed: `run`, `client`, `daemon`)
- **Live theme preview in /settings**: scroll themes to preview, Enter confirms, Esc reverts
- **Theme hot-reload**: `/theme <name>` applies instantly without restart (ArcSwap)
- **Settings picker scroll**: theme/model picker scrolls with cursor
- **Adaptive thinking for Opus 4.7+**: `{type: "adaptive", display: "summarized"}` with effort mapping (xhigh/high/medium/low/adaptive)
- **Model-agnostic context window**: bar denominator adapts per-model (1M Opus 4.7, 200K Sonnet/Haiku)
- **Per-turn context tracking**: usage bar shows actual request size, not cumulative cost
- **Effort parameter**: thinking depth on adaptive models controlled via `output_config.effort`
- **"adaptive" thinking option**: new cycler value in `/settings` — lets model decide thinking depth
- **Tab-cycle for slash commands**: `/s` + Tab cycles through sessions → settings → system
- **Streaming command guard**: known slash commands no longer leak into model stream as steering
- **Usage log opt-in**: `SYNAPS_USAGE_LOG=1` writes to `~/.cache/synaps/usage.log` (0600, O_NOFOLLOW)
- **Opus 4.6 in model picker**: was missing from `/settings` model list
- **`settings` + `plugins` in tab-complete**: were missing from `BUILTIN_COMMANDS`
- **Plugin Keybinds** — plugins declare custom keyboard shortcuts in `plugin.json`
  - `KeybindRegistry` with parser, matching, and conflict resolution
  - Key notation: `C-x` (Ctrl), `A-x` (Alt), `S-x` (Shift), `F1`–`F12`, special keys
  - 4 action types: `slash_command`, `load_skill`, `inject_prompt`, `run_script`
  - User overrides via `keybind.*` in config — override or disable plugin binds
  - Priority: core > user > plugin (core binds never overridable)
  - `/keybinds` command shows all registered binds with source attribution
  - 23 unit tests for parser, registry, conflicts, overrides
- **Plugin Agent Namespaced Resolution** — `subagent(agent: "dev-tools:sage")` resolves plugin agents via `plugin:agent` syntax
  - Searches `plugins/<plugin>/skills/*/agents/<agent>.md`
  - Input validation, path traversal protection, ambiguity detection
  - Updated error messages to mention `plugin:agent` syntax
- **Mouse Text Selection & Clipboard** — left-click drag to select text in chat area
  - Right-click with selection → copy to system clipboard
  - Right-click without selection → paste from clipboard
  - Singleton clipboard thread (no thread-per-copy accumulation)
  - Paste suppression with 150ms TTL (prevents terminal double-paste)
  - Selection highlight via `Block::inner()` computed content rect
  - Theme-aware highlight color

### Fixed
- **Empty thinking/text blocks rejected by API** — fixed two related 400 errors that could permanently brick a session:
  - `messages.N.content.M.thinking: each thinking block must contain thinking` — streaming accumulator (`runtime/api.rs`) was pushing thinking content blocks even when no `thinking_delta` arrived (only a signature, or stream cut off). Now skipped at all three accumulation sites (mid-stream, tail buffer, post-loop fallthrough).
  - `messages: text content blocks must be non-empty` — defensive sanitizer (`runtime/helpers.rs::sanitize_thinking_blocks`) added; runs on every outbound API call from both `call_api_stream_inner` and `call_api`. Strips empty `thinking`, `redacted_thinking.data`, and `text` blocks; drops assistant messages whose entire content gets stripped; merges resulting consecutive same-role messages to preserve Anthropic's strict role-alternation rule. Auto-heals sessions persisted before the streaming fix landed.
  - 9 unit tests covering the drop-and-merge cases.
- **Config file permissions** — now 0600 (was 0644, world-readable with API keys)
- **API key masking** — shows last 4 chars only (was showing first 8 + last 4)
- **ProviderConfig Debug** — custom impl redacts `api_key` as `[REDACTED]` in logs
- **Provider base URLs** — corrected siliconflow (.com→.cn), chutes (→llm.chutes.ai)
- **Dead models removed** — groq/llama-3.1-70b, groq/qwen-qwq-32b, groq/gemma2-9b (all decommissioned)
- **Google stream_options** — skip `include_usage` for googleapis.com (rejects it)
- **Tool→user message ordering** — insert space-content assistant between tool result and user message for strict providers
- **Finish-frame tool call dedup** — NVIDIA sends final argument chunk with finish_reason; was being dropped as "resend"
- **Model ID extraction** — health prefix (✅ 253ms) no longer embedded in model string sent to API
- **`/status` on non-Anthropic** — shows friendly message instead of crashing
- **"Calling Claude..."** — now shows actual model name in `synaps run`
- **Paste in settings** — `Event::Paste` now handled in API key, text, and custom model editors
- **Missing provider key** — `/model sambanova/llama` with no key shows clear error instead of silent Anthropic misroute
- **UTF-8 truncation panic** — `bash` and `shell` output truncation now finds valid char boundaries before truncating, preventing crash on multi-byte characters (emoji, CJK)
- **Zero-width terminal panic** — markdown word-wrap `chunks(0)` crash on rapid tmux pane resize, clamped to `.max(1)`
- **Local model connection** — friendly "is Ollama running?" instead of raw TCP error
- **`/help` updated** — mentions provider/model syntax and /settings for key management
- **`ping_print` lifecycle** — resets after all results arrive via pending counter
- **MCP tool name causes 400 errors** — Anthropic rejects `mcp_` prefixed tool names (rate limit pool misrouting). Renamed `mcp_connect` → `connect_mcp_server`, tool prefix `mcp__server__tool` → `ext__server__tool`.
- **`/saveas` on empty sessions** — `save_session()` bailed on empty `api_messages`, so the name never persisted. Now calls `session.save()` directly.
- Compaction no longer overwrites original session (loads from disk)
- Event content sanitized against prompt injection (XML tags, case-insensitive)
- Atomic inbox writes (.json.tmp → .json rename)
- inotify watcher runs on spawn_blocking (no longer starves tokio runtime)
- Events during streaming buffered and flushed after MessageHistory
- Spurious auto-triggers prevented by event_received guard
- push() calls notify_one() (was missing — events silently queued)
- Session save ordering: new session saved before old session updated
- chars().count() moved outside lock scope
- High-priority event FIFO ordering (was LIFO among Highs)
- Queue-full events left in inbox for retry (not silently dropped)
- push_priority logs evicted event ID
- **Session file corruption**: atomic writes via write-to-tmp then rename
- **Tool input parse errors surfaced to model**: malformed tool_use JSON no longer silently falls through to empty input; model sees `invalid tool input JSON: ...` and can self-correct
- **Custom theme crash**: `unreachable!()` in draw replaced with graceful fallback colors for non-Rgb themes
- ASCII logo alignment: use unicode display width for consistent centering
- 'default' theme missing from settings picker
- Context bar pinned at 100% after 2-3 turns (was using cumulative tokens / hardcoded 200K)
- Thinking blocks invisible on Opus 4.7 (display defaulted to "omitted")
- `budget_tokens: 0` sentinel leaked to non-adaptive models → 400 error
- `/settings` cycling capped at xhigh (`apply_setting` silently rejected "adaptive")
- Stale test asserting `thinking_level_for_budget(0) == "low"` (production returns "adaptive")
- Usage log world-readable at `/tmp/` → moved to `~/.cache/` with 0600

### Changed
- Compaction logic moved from chatui to `core/compaction.rs`
- `SubagentHandle`/`Registry`/`Status` moved to `runtime/subagent.rs`
- `ApiOptions` struct replaces `use_1m_context: bool` threading
- `build_auth_header` + `build_beta_header` extracted as helpers
- `clone_repo` helper deduplicates plugin installer
- All compaction prompts colocated in `core/compaction.rs`
- Tool descriptions guide model choice (subagent vs subagent_start)
- Stub tools unregistered from tool registry (model doesn't see unimplemented tools)
- SubagentState uses RwLock instead of Mutex
- **`define_settings!` macro**: settings schema + apply handler defined once in `settings/defs.rs` via declarative macro — zero drift possible (replaced manual sync + parity tests)
- **Single source of truth for commands**: removed duplicate `ALL_COMMANDS` array; `commands.rs` now sources from `skills::BUILTIN_COMMANDS`
- **`src/cmd_*.rs` → `src/cmd/` module**: subcommand handlers moved to dedicated directory, `cmd_` prefix stripped
- Binary name: `chatui`/`synaps-cli`/`synaps-agent` → `synaps` (single binary with subcommands)
- `src/chatui/main.rs` → `src/chatui/mod.rs` (module, not binary)
- watcher spawns `synaps agent` instead of standalone `synaps-agent`
- `thinking_level_for_budget()` consolidated from 4 copies into single source of truth in `core/models.rs`
- `DEFAULT_LEGACY_ADAPTIVE_FALLBACK` constant replaces magic `16384` in clamp sites
- Dead-code warnings suppressed with explanatory comments (reaper handles, settings help field)
- Auto-cache toggle removed (manual breakpoints won: 90% vs 53%)
- README rewritten from internal documentation to product landing page

### Removed
- `SPEC-WATCHER.md` — internal spec, not needed in repo
- Auto-cache config toggle (`auto_cache = true/false`)

## [0.1.0] — 2026-04-12

### Added
- Interactive TUI (`chatui`) with streaming, markdown, syntax highlighting
- Headless chat (`chat`) for scripting and piping
- Autonomous agent supervision (`watcher`) with heartbeat, cost limits, handoff
- 10 built-in tools: bash, read, write, edit, grep, find, ls, subagent, mcp_connect, load_skill
- Interactive shell sessions: shell_start, shell_send, shell_end
- 18 color themes
- MCP integration with lazy server spawning
- Skills & plugins subsystem
- OAuth + API key authentication
- WebSocket server/client transport
- `/settings` full-screen modal
- `/plugins` management UI
- Session persistence and `/resume`