v0.29.2 - prompt cache awareness

Highlights

• Improves API cost visibility by parsing provider prompt-cache usage, including DeepSeek cache hit/miss fields.
• Adds a stable OpenAI prompt_cache_key and cache-friendly prompt layout while preserving current-turn recency.
• Adds Target Context to every model evidence pack so the copilot model sees its read-only supervisor role, active scope, target session(s), and no-tools/no-injection boundary.

Verification

• Local: uv run --extra tui python -m unittest discover (728 tests OK)
• GitHub Actions: main tests, tag tests, and publish all passed

cc-copilot 0.29.1

Fix cockpit error recovery and watch isolation ahead of wider rollout.\n\n- Keep automatic /watch updates in the watch dock/monitor instead of the main chat.\n- Preserve pre-stream model failures as rewindable in-memory error turns without persisting them as successful answers.\n- Restore transient TUI error history when returning to the originating cockpit session.\n- Align REPL error turns with rewind behavior.\n- Bump release metadata and update README/CHANGELOG.

v0.28.1

Patch release for the watch UX loop.\n\n- Watch Now now runs as a flow update with stable baseline context, previous Now, current step, and current-step evidence.\n- Step boundary decisions are made by the watch flow agent instead of only raw evidence thresholds.\n- Auto Digest is now posterior: it closes the previous semantic step when the flow agent starts the next one, instead of firing from event/time cadence alone.\n- Watch monitor cards avoid empty command-completion steps and show recent evidence instead of placeholder digest text.

cc-copilot 0.28.0

Highlights

• Adds @ scope mentions and saved multi-session scope groups.
• Adds multi-session activity tabs: all + per-session views with stable Tab navigation.
• Promotes /watch into an automatic observer loop with in-place monitor, semantic steps, multi-session tabs, ephemeral process history, and light presets.
• Raises the default cc-copilot evidence context budget to 256k tokens.
• Adds Ollama Cloud as a first-class backend.

Published to PyPI as cc-copilot 0.28.0.

cc-copilot 0.27.0

• Add researched watch-mode design doc covering opt-in long-task monitoring, micro updates, rolling digests, and scope safety.
• Add copilot-written watch process updates plus the new /watch digest periodic summary path.
• Buffer watch evidence with cadence control and pause watch when the evidence scope changes.
• Show session titles before session ids in the attached-session HUD.

cc-copilot 0.18.1

Patch release for the prompt navigation update.\n\n- Fix Sticky Prompt positioning for short conversations, long conversations, and exact prompt-boundary scroll positions.\n- Keep the sticky prompt aligned with the selected prompt after keyboard or mouse navigation.\n- Keep slash-style provider model IDs on router backends unless the user explicitly prefixes another backend.\n\nPyPI: https://pypi.org/project/cc-copilot/0.18.1/

cc-copilot 0.18.0

Codex-aware /here, expanded /model provider catalog, and faster cockpit prompt browsing.

Highlights:

• /here can attach to current live Codex sessions as well as Claude.
• /model adds more OpenAI-compatible providers and provider refs like /model openai/gpt-5.5 or /model google/gemini-3.1-flash-lite.
• Composer prompt history uses Up/Down; Esc clears drafts; double-Esc opens rewind.
• The chat pane has a sticky prompt header; empty-input Left/Right jumps between prior prompts, and the header is clickable.

v0.17.1 — launch split 2:1 + click-to-focus

launch split-screen fixes, both user-reported:

• 2:1 split — the cockpit pane is now a third of the window (-l 33%); the agent is the main act.
• Click-to-focus works — sessions created by launch get mouse on. Stock tmux ships with the mouse off, where clicking a pane does nothing — anyone not fluent in tmux prefix keys couldn't reach the cockpit pane at all, and every keystroke landed in the agent. When launch splits inside your tmux, your options are left untouched.

Install/upgrade: uv tool install "cc-copilot[tui]" · uv tool upgrade cc-copilot

v0.17.0 — one-command startup

One command to start it all.

• cc-copilot launch (alias up) starts your agent and the cockpit side by side in a tmux split. Inside tmux: the cockpit takes a new pane and the agent execs in your pane — cc-copilot never becomes the agent's parent. Outside tmux: a cc-copilot session is created with both panes and attached. launch codex and launch -- claude --resume work; without tmux it says so and opens the cockpit alone.
• cockpit/chat --next waits for the project's transcripts to change — a new session appearing, or an existing one growing (claude --resume appends in place) — and pins to it. No more latching onto yesterday's transcript or dying in a fresh project.
• Bare cc-copilot now opens the cockpit on a TTY; scripts, hooks, and pipes still get the usage error.
• Hardened against the pane-dies-with-the-diagnostic family: [tui] preflight before any pane opens, agent binary absolutized (tmux panes resolve commands with the server's PATH), env-prefixed cockpit command (fish/tcsh default-shells), exact-match tmux session names, no leaked or wrongly-killed sessions on setup failure.
• Also in this release: README repositioned — a copilot for AI coding agents ("Cockpit" is now the internal name for the TUI).

26 new tests (411 total). Install: uv tool install "cc-copilot[tui]"

v0.16.0 — model catalog + picker + 6 new providers

Pick your model, not just your provider — plus 6 new providers.

• Curated model catalog per API provider (every id verified against live provider docs): the /model picker is now two-level — backend, then model — with notes, deprecation warnings, and a ✓ on the current choice. Typed fast-paths: /model deepseek-v4-pro (model-only), /model deepseek:deepseek-v4-pro (both), any free-form id accepted everywhere.
• DeepSeek default → deepseek-v4-flash (deepseek-v4-pro selectable). deepseek-chat/deepseek-reasoner are deprecated upstream on 2026-07-24 and marked as such in the picker. OpenAI default gpt-4o → gpt-5.4.
• New providers (OpenAI-compatible, zero new deps): moonshot (Kimi k2.6), zai (GLM 5.1), qwen (DashScope; mainland endpoint via DASHSCOPE_API_BASE), groq, xai (Grok 4.3), gemini-api (Google's OpenAI-compat endpoint). All get the inline key-prompt on switch.
• cc-copilot backends --models lists the whole catalog; cc-copilot init gains a numbered model menu.
• Fix: a model picked for any provider no longer leaks into ollama's default via CC_COPILOT_MODEL.

16 new tests (385 total). Upgrade: uv tool upgrade cc-copilot / pipx upgrade cc-copilot.