Changelog

All notable changes to c9watch are documented here. The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

0.11.1 — 2026-10-15

Added

Hidden Claude Code sessions are counted. A Claude Code session that runs in an app another Claude Code session opened (for example open -a Terminal from an agent) inherits CLAUDE_CODE_CHILD_SESSION=1, so Claude Code writes no registry entry and no transcript for it. c9watch cannot show these sessions. The Monitor tab now shows one line with how many there are and how to fix it (quit that app and open it again from the Dock); click it for each pid and folder. c9watch list reports them in diagnostics.hiddenClaudeSessions. See docs/accuracy-matrix.md (#160).

Fixed

The web dashboard loads sessions again. The getSessions WebSocket reply sent the session list and the detection diagnostics as one pair, so the web page received [[...], {...}] and failed on projectPath. The reply is now the plain session array. The bug was there since v0.5.0; the desktop app was not affected. Thanks to @pax for the report and the fix (#161, #162).

0.11.0 — 2026-10-15

Added

c9watch wait. Blocks until a live session needs you (--until needs-input), finishes its turn (done), stops (ended), or changes at all (any-change), then prints that event. --timeout exits 124 like timeout(1); an unknown session exits 2. Scripts and orchestrator agents can use it in place of reading a terminal pane in a loop. See docs/cli-events.md (#155).
Stable watch event format. Every c9watch watch line now has "schema": 1, a top-level status, and previousStatus. Existing fields are unchanged. --json is accepted as a no-op for scripts (#155).
“While you were away” summary on waiting cards. A card that waits for you now shows what it waits for (for example “Approval: Bash” or “Question”; the tool name only, never its arguments), the last line the agent said, and what it did since you last opened it, pressed FOCUS, or left the c9watch window: files edited, commands run, subagents done, and tool calls. The counts come from the local transcript (no AI summary, no network) through a bounded index that reads only new bytes each poll. Claude Code only; other agents show ”—” (#156).
Codex asks for you. A Codex session that asks a question or requests permissions now shows as Needs Attention at once. A command that needs approval outside the sandbox shows as Needs Attention after 35 s; Codex does not record approval prompts, so this is a heuristic and can be late or wrong (#157).
Status accuracy matrix. Each agent source × situation (working, waiting, idle, killed, pid reuse, subagent) is now a test. Known gaps are listed in docs/accuracy-matrix.md (#157).

Fixed

Killed Codex CLI and Cursor Agent CLI sessions disappear on the next poll. Before, a killed session could stay on the dashboard for 30 minutes to 4 hours. c9watch now checks whether an agent process still holds the transcript open (#157).
A reused pid no longer revives a dead session. When a new process gets the pid of a dead session, c9watch compares the process start time with the session start time (#157).
The Codex quota reader reports the running c9watch version instead of a hard-coded 0.10.0 (#153).

0.10.1 — 2026-10-14

Fixed

Killed sessions no longer stay active. A session killed outside c9watch (for example a claude -p stopped by a timeout wrapper) could keep showing as Needs Attention or Working when Auto mode had fallen back to the legacy detector. The legacy detector no longer hands a session that another process names (a registry row or --resume <id>) to a different live process in the same folder (#148, #131).
Auto mode returns to the CLI detector on its own. After falling back to the legacy detector, Auto mode now retries claude agents --json every 5 minutes instead of only when the window gains focus. The CLI detector and forced modes still never probe on a poll (#148).

0.10.0 — 2026-10-08

Added

Cursor Agent monitoring (monitor-only). Detect Cursor Agent root and subagent sessions, show their lifecycle and provider metadata, and include them in Monitor, History, conversation search, and the provider filter (#120).
Provider-scoped session identity. Claude Code, Codex, and Cursor session IDs no longer collide across detection, enrichment, CLI output, notifications, and frontend selection (#120).
Pi Agent monitoring (read-only). Monitor, History, conversation, search, Cost, and CLI support for Pi, with provider-reported cost data. Pi sessions cannot be opened, stopped, or renamed from c9watch (#121).
OpenCode monitoring (preview). Connect c9watch to one OpenCode HTTP server you start yourself (opencode --port 4096) in Settings → Integration. Adds session cards, an OpenCode filter and badge, parent/child grouping, and on-demand conversation reads. It does not discover servers on its own. See docs/opencode-preview.md (#127).
Attention inbox and source health. A deduplicated inbox on the Monitor page for sessions that wait for you, a per-session health badge (fresh, stale, partial, unavailable, unknown) so an old observation no longer looks live, and a declared return target for attention notifications (#141).
Subscription usage indicators. Per-provider quota meters in the toolbar and tray for Claude Code, Codex, and Cursor, with reset times and a Settings → Usage page. Claude Code needs an optional status-line bridge (#126).
Elapsed-window marker under usage meters. A small triangle shows how far through each quota window you are, and the reset line shows “N% elapsed” (#140).
Native notification settings. Choose which events notify you, brief or detailed content, sound, and a per-session cooldown. Reply notifications show a short excerpt of the assistant’s reply (#125).
Full-screen chat layout in Monitor and History, and a resizable Monitor sidebar (#126).
Session kind and entrypoint badges. Background-pinned Claude Code sessions show a BG badge. Sessions not launched from the plain cli entrypoint show that entrypoint (#134).
Claude Code hook bridge (opt-in). c9watch hooks --install registers async hooks so Claude Code reports permission prompts directly. Hooks never delay Claude Code or affect permission decisions. --uninstall removes only c9watch’s entries. Not available on Windows. See docs/claude-hooks.md (#136).

Fixed

Needs Attention accuracy. Running sub-agents no longer flag their parent session. c9watch now uses the waiting status that claude agents --json reports, and merges allow rules from user and project settings.json (#136).
Claude Code not found when launched from Finder, the Dock, or a login item. c9watch adds the usual install directories to PATH at startup. The fallback scanner no longer matches Claude Desktop helper processes, so STOP can no longer kill one (#139).
Dead sessions in claude agents --json. Entries whose process is gone are dropped before status inference. Rows with a missing PID no longer cause a parse error (#132).
Pi sessions end when the process ends. A killed Pi session no longer shows LIVE for up to 4 hours. When the process evidence is unclear, the old freshness rule stays (#141).
Cursor detection no longer blocks on the composer database. WAL changes invalidate the cache, failed composer loads retry, and stale overlays clear (#141).
Cursor transcript cache. Appends verify the cached prefix, so a rewrite falls back to a full parse. An incomplete final line stays pending until it is complete (#120).
Codex archive memory and cache. Bounded archive memory, released consumed buffers, and a persistent cache that detects same-length and middle rewrites (#122, #123).
Codex conversation content is kept across filtering and polling (#124).
Rename safety across providers. Explicit Codex and Cursor rename targets are rejected. A request without a provider is rejected before writing when the raw ID exists under another provider (#120).
OpenCode conversation loading is single-flight and bounded, and shows an explicit Retry error. Window transitions complete at once in the native app, so a hidden window no longer stalls animations (#127).

Improved

Lower energy use. c9watch no longer runs claude agents --json on every poll, holds updates while a window is hidden, and animates the status bar with transform and opacity. The contributor measured with powermetrics, against an earlier commit in the same PR (not against v0.9.0): energy use fell 97% with the dashboard open and 93% with it closed. The status bar no longer has a per-block glow or a left-to-right ripple (#138).
Less work per poll. Subagent detection and message counts are cached per file and read incrementally, so old sessions are not re-parsed (#133).
Codex archive reads reuse a cache, and conversation tools load on demand (#122).
History reads the first prompt with a bound and scans history off the main thread (#141).

Known limits

OpenCode is a preview. It needs a manual server connection.
Pi is read-only. It cannot be opened, stopped, or renamed from c9watch.
Cursor is monitor-only.
Codex shows working and idle only. It has no approval or question state.
Pi: one process keeps at most one transcript. If pi writes in its own project and then /resumes into another project, the resumed card shows ended until that pi exits.
Pi: an idle pi in another working directory, or a process whose start time cannot be read, can keep a killed card alive.
Pi: a killed card can reappear for one poll if a node or bun process cannot be read.

0.9.0 — 2026-08-16

Added

Codex App and CLI monitoring. c9watch now detects both Codex surfaces, displays provider-aware CODEX badges, and groups Codex subagents under their parent sessions.
Shared provider filtering. The All | Claude Code | Codex filter now applies to Monitor, History, Cost, and Memory, including provider-specific memory files and Codex model cost estimates.
Claude agent metadata backend. Use claude agents --json when available, with a legacy fallback for older Claude Code installations.
User-controlled updates. The new SETTINGS tab and update banner let users decide when an available app update is installed.
TODO side panel. Conversation previews can show the session’s TODO list without leaving the preview overlay.
PM bg backend. PM workers can use claude --bg instead of claude --print, with automatic detection for newer Claude Code versions and C9WATCH_WORKER_BACKEND=bg|print|auto override support.
c9watch spawn --prompt <text> is available when the bg backend is active.
Added EventStatus::Awaiting and InboxEvent::awaiting() for workers that are waiting for user input.

Changed

PM orchestration is now disabled by default. Normal builds omit the PM CLI subcommands and hide the HUMANS/WORKERS toggle, worker badges, worker title prefix, and Workers panel. The implementation remains opt-in through the pm-orchestration Cargo feature and PM_ORCHESTRATION_ENABLED UI flag.
History search now defaults to phrase matching and exposes case-sensitive and whole-word toggles.
The MEMORY tab uses an accordion layout with improved reading size and animation.
Added Cmd+1 through Cmd+4 shortcuts for tab navigation.

Fixed

Preserved Codex CLI search metadata and prefixes across search results.
Dropped history entries whose JSONL files were cleared instead of showing unviewable stale cards.

Improved

Lower memory usage for Codex session monitoring. Incremental rollout parsing and bounded monitor representations reduce peak RSS by 65.8% on a controlled workload while preserving full conversation and tool details. (#116)
Opening a session now focuses the matching macOS Terminal/iTerm2 target by tty, and can focus the exact Supacode tab and surface when its coordinates are available.
Added a dry-run-by-default utility for cleaning stale Cargo targets from other registered worktrees.

0.8.1 — 2026-04-20

Fixed

c9watch inbox, cost, adopt launched the GUI instead of the CLI — src-tauri/src/main.rs kept a hardcoded allowlist of subcommand names for dispatching to the CLI handler. The three subcommands added in v0.8.0 (inbox, adopt) and v0.7.0 extension (cost) weren’t added to the list, so running them on a GUI-inclusive build fell through and launched the Tauri desktop app. Added all three to the allowlist.

0.8.0 — 2026-04-20

Added

PM orchestration CLI — spawn, send messages to, and manage child Claude Code worker sessions from a parent “PM” session. New commands: c9watch spawn, send, workers, stop, adopt, inbox, tasks. Workers show a “WORKER” badge on session cards; PMs show a “PM” badge with a Workers panel in the expanded overlay. (#84)
Subagent visibility — detect and display Task-tool subagents spawned inside any session. 2-row card layout with click-to-preview transcript and async completion resolution. (#92)
Entry animations — three-tier cascade pacing across MONITOR, HISTORY, MEMORY, and COST tabs plus overlay panel fly-ins. History preloaded at startup to avoid tab-switch lag. (#93)
Cost per session — new c9watch cost --session <id> and --session-prefix <any> CLI flags for per-session breakdown in JSON. (#94)
USD/TOKENS toggle in cost tab plus per-card cost pill on SessionCard and overlay charts. (#85)
CLI auto-symlink — installing the desktop app now symlinks the c9watch CLI into ~/.local/bin/ automatically. (#83)

Fixed

App icon padding — added ~100px inset to icon.png master so macOS Sequoia no longer oversizes it in the Dock. Master restored to 1024×1024 after Tauri CLI downscale. (#97)
debug_log test races — serialized ring-buffer tests to prevent flaky CI. (#82)

Improved

Pink agent accent + de-duped nav header — agent sessions now use pink for visual distinction; overlay navigation header consolidated. (#90)
Shared UI utilities — extracted time-utils.ts, status-utils.ts, unified preview state, and usage-stats helper from PR #92 for reuse across cards and overlay. (#96)
Removed leaked internal planning docs — repo now ignores docs/superpowers/ by default. (#89)

0.7.0 — 2026-04-06

Added

CLI for scriptable session management — c9watch list, view, history, search, stop, watch, self, status, tasks commands for agent-to-agent monitoring (#75)
One-line CLI installer for macOS and Linux (install-cli.sh)
Cost records split by date for accurate daily totals — sessions spanning midnight now attribute costs to the correct day (#78)
Session names in cost tab — display custom title or first user message alongside session ID (#79)
Conversation preview in cost tab — click any session row to open the conversation overlay (#79)
DATE/COST sort toggles in cost tab with ascending/descending order (#79)
History tab shows latest prompt text and native custom titles from JSONL files (#80)

Fixed

Session detection on macOS now uses cmd args instead of binary path for more reliable process matching (#77)
PID-to-session mapping after /clear now uses session metadata for accuracy (#73)
Notification title now uses renamed session title instead of generic text (#74)

Improved

History tab layout redesigned with CSS grid for better alignment (#80)
Session count shown per project in cost tab (#79)

0.6.0 — 2026-03-23

Added

Session metadata improvements — richer session info display (#65)
NeedsPermission renamed to NeedsAttention with user question detection (#66)
Draggable title bar and mobile responsive styling improvements (#58)
5 new token distance milestones: Angel Falls, Mt. Vesuvius, Krubera Cave, Mt. Olympus, Mt. Etna (#63)

Fixed

Cost pricing updated — Opus 4.5/4.6 corrected to $5/$25 (standard) and $30/$150 (fast), Haiku 4.5 to $1/$5 (#64)
Session titles no longer forced to uppercase with pixel font (#67)
History “newest” sort now uses last activity time instead of creation time (#68)
JetBrains IDE “Open” action now focuses existing window instead of opening a new one (#69)

0.5.0 — 2026-03-14

Added

Memory tab with two-panel viewer for browsing Claude Code memory files and Claude command integration (#41)
Token distance visualizer — animated rice stack overlay with 17 real-world landmarks, native share sheet, and Instagram-ready PNG export (#62)
FDA permission banner — heuristic detection when Full Disk Access is missing, with deep-link to System Settings (#48)
Debug console (Cmd+Shift+D) — hidden panel showing real-time diagnostic logs for troubleshooting session detection (#48)
Custom title and ACTIVE badge display in history tab (#52)
Multi-word AND search in history — search terms are combined with AND logic for more precise results (#51)
List item numbers in history session rows (#50)
Restore minimized terminal windows when clicking Open on a session (#49)
Thinking toggle restored in conversation preview (#45)
Product website at c9watch.mclee.dev (#42)
Website migrated to Starlight documentation framework (#54)

Fixed

Path encoding mismatch — dots in directory names now correctly encoded as dashes for session matching (#57)
Path encoding aligned with Claude Code’s algorithm — all non-alphanumeric characters replaced with dashes (#48)
Sliding window rendering for large conversations — prevents DOM overload (#53)
Cloudflare Workers deploy configuration for website (#43)

0.4.0 — 2026-03-01

Added

Session history search tab — browse and search all past Claude Code sessions with instant metadata filter + debounced deep content search (#33)
Full conversation viewer overlay for history sessions with message rendering, tool toggle, message nav sidebar, and copyable RESUME command chip (#33)
Collapsible project groups in history BY PROJECT view with collapse/expand all (#33)
Search result snippets with keyword highlighting (#33)
Click a deep search result to scroll to and highlight the matching message in the conversation viewer (#36)
Inline image rendering for screenshots pasted in user messages (#38)
Cost tracker dashboard tab with daily, by-project, and by-model spending views (#34)
Rust cost backend with per-model pricing tables (Sonnet, Opus, Haiku) and mtime-based caching (#34)
Tab bar in native macOS title bar area with drag region and grip dots (#33)

Improved

Drag dots handle shows hover brightness effect for better UX feedback (#33)
Removed non-functional thinking toggle — JSONL files never contain thinking blocks (#38)

Fixed

Search highlight blink after animation fade, wrong message highlighted on deep search, and NavMap scroll targeting wrong element (#37)

0.3.0 — 2026-02-27

Added

Native tray popover with session overview — click the menu bar icon to see all sessions at a glance (#25)
Pixel grid status bar with sweep animation on state changes (#25)
Fullscreen space support — popover uses NSPanel to appear above fullscreen apps (#25)
JetBrains IDE support: 15 IDEs (PhpStorm, IntelliJ IDEA, WebStorm, PyCharm, GoLand, CLion, Rider, RubyMine, DataGrip, Android Studio, Aqua, Fleet, RustRover) with 3-tier path resolution via Toolbox scripts dir, user Applications, and system Applications (#26)

Improved

Test coverage increased from 53% to 65% (#31)
Clippy warnings resolved and rustfmt applied throughout Rust codebase (#31)

Fixed

Popover not appearing above fullscreen app Spaces (#25)
App quitting when main window is closed — tray icon now keeps app alive (#25)
“Open Dashboard” button not working after main window was closed (#25)

0.2.1 — 2026-02-16

Fixed

Strip ‘v’ prefix from version in latest.json for updater compatibility (#23)

0.2.0 — 2026-02-16

Added

WebSocket server for mobile/remote access — view sessions from any device on the same network (#6)
QR code connection for instant mobile browser pairing
Custom session titles with inline editing (#9)
Linux support via AppImage (#2)

Improved

~60% CPU reduction — optimized polling and status detection, from ~15% to ~5-9% (#14, #19)
Simplified notifications — removed custom permission banner, macOS handles prompts natively (#20)
Better iTerm2 click-to-focus using tty matching instead of window title matching (#5)

Fixed

Status flickering when sessions are actively working (#19)
Duplicate notification firing (#19)
Register missing get_terminal_title command (#21)

0.1.0 — 2026-02-08

Initial release

Automatic session discovery — detects Claude Code sessions by scanning running processes at the OS level
Real-time dashboard with status indicators (Working, Needs Permission, Idle)
Status view (grouped by state) and project view (grouped by directory)
Session control — expand to read full message history, approve permissions, manage agents
Auto-updater for future releases
Built with Tauri, Rust, and Svelte for minimal resource usage