v2.6.1

Reliability pass on the permission + diff lifecycle: no more dead permission cards after a CLI crash, no spurious toasts on every Allow/Reject, tool rows finish in the right state, and inline edit won't clobber a file you edited mid-run. All eight fixes came out of a detailed adversarial audit of the extension.

Fixed

• Permission cards no longer get stuck after a CLI crash. If the CLI exited
  while a permission card was pending, the card stayed on screen but clicking
  Allow did nothing (the request id no longer existed on the reconnected
  process). The disconnect path now drops the pending resolvers, closes any
  orphaned diff tabs, and dismisses the cards.
• No more "no pending permission for this diff" toast on every Allow/Reject.
  The tab-close handler fired an implicit reject for the extension's own
  programmatic diff close after a permission resolved, popping a confusing
  toast on every normal approval. It now only reacts when a permission is
  genuinely still pending for the closed diff.
• Cancel / New chat no longer leaves orphaned diff tabs behind. A proposed-
  change diff opened for an in-flight write/edit stayed open — showing
  live-looking Accept / Reject lenses — after you cancelled or started a new
  session. Those tabs are now closed and their tracking dropped.
• Tool rows reach their final state. A tool that streamed multiple updates
  (in-progress → completed/failed) could get dropped from tracking on the first
  update, so it never dimmed for completed or turned red for failed. The row
  now settles correctly and only stops being tracked once it's terminal.
• Idle-timeout watchdog no longer fires against the wrong run. After a
  cancel-and-resend or a new session, a watchdog armed for the previous prompt
  could fire a spurious cancel. Cancelling now disarms it.
• Inline edit won't clobber a file you edited while it was running. If the
  document changed during the (possibly minutes-long) model call, the edit is
  now skipped with a warning instead of silently overwriting the wrong region.
• Internal: control-plane request timers are now cleared when the response
  arrives, and per-turn tool-call element references are released — small
  resource-hygiene fixes with no user-visible behaviour change.

v2.6.0

Chat-surface polish: the conversation survives a window reload, light themes are readable, the keybindings no longer fight VS Code's defaults, Esc stops a run, "New chat" keeps your CLI (and MCP servers) warm, and the permission card shows every option the agent offers.

Added

• Esc stops an in-flight run from the chat input — mirrors the Stop button.
• Reject-always in the permission card. The card now renders the actual
  option set the CLI offers (Allow once / Allow for this session / Reject /
  Reject for this session) instead of a hardcoded three, so a future CLI
  option shows up with no extension change.

Changed

• "New chat" no longer respawns the CLI. It now opens a fresh session on
  the running process, so warm MCP servers stay up and there's no relaunch
  lag — the CLI already supports multiple sessions per process.

Fixed

• Your conversation comes back after a window reload. The CLI re-attaches
  the previous session server-side, but the webview rendered blank — so it
  looked like the chat was lost while the agent still had the full context.
  The transcript is now replayed on reconnect ("Restored previous chat").
• Readable on light themes. ~30 hardcoded translucent-white surfaces and a
  dark code-block background painted white-on-white (and dark-on-light code) on
  light themes. They now use VS Code theme tokens (--vscode-*), so code,
  diffs, borders, and buttons adapt to whatever theme you run.
• Keybindings stop shadowing VS Code defaults. ⌘⇧C (open chat) collided
  with "Open New External Terminal" and ⌘⇧X (send selection) collided with
  the Extensions view; both moved to ⌘⌥-based shortcuts (⌘⌥C / ⌘⌥X).

v2.5.1

Security: the manual-mode permission gate now fails closed. If the extension can't put the agent into manual mode on connect, it refuses to send prompts (with a clear message) instead of letting the agent run tools unguarded.

Security

• Manual-mode gate fails closed. On connect the extension asks the CLI to
  enter manual mode (so dangerous tools require approval). If that request fails
  — e.g. a transient error or a CLI too old to support it — the session would
  otherwise stay in unguarded auto mode. The extension now retries once and,
  if it still can't arm the gate, refuses to send prompts and surfaces an
  actionable error ("update the Codeep CLI and reload"), rather than silently
  running tools without confirmation. Pairs with the CLI 2.7.0 ACP hardening.

Changed

• Pin the codeep-review CI workflow to the immutable codeep-action@v1.0.2.

v2.5.0

Blockquotes render in chat again, and file edits containing a $ now apply literally instead of corrupting. Also hardens link rendering against an attribute-injection edge case.

Fixed

• Blockquotes render again. The chat markdown renderer escaped > to >
  before checking for blockquote lines, so the blockquote branch was dead code
  and > quote showed as plain text. Fixed.
• Edits containing $ apply literally. Synthesizing the "after" text for an
  edit_file diff used String.replace(old, new), which interprets $&, $$,
  $1 in the replacement — corrupting any edit whose new text contained $
  (template literals, shell vars, regex). Now inserted verbatim.
• Link rendering hardened. The markdown escaper didn't escape ", so a
  model-supplied link URL containing a quote could break out of the href
  attribute. Quotes are now entity-encoded.

v2.4.1

Push your profile to the dashboard from the editor.

Added

• Codeep: Sync Profile to Dashboard — runs /me sync through the CLI to push
  your ~/.codeep/profile.md to codeep.dev (and pull it on a fresh machine).

v2.4.0

Codeep adapts to you: edit your user profile from the editor, toggle opt-in auto-learn, and a "Personalize Codeep" walkthrough step. Pairs with CLI 2.2.0's /me profile.

Added

• Codeep: Edit Profile / Codeep: Edit Project Profile. Open (and scaffold
  on first use) ~/.codeep/profile.md and the workspace .codeep/profile.md —
  a short description of how you like to work (reply language, style, stack,
  "always / never"). The CLI/ACP agent injects these into its context on every
  run, so @codeep and the chat adapt to you.
• Codeep: Toggle Profile Auto-Learn + the codeep.autoLearnProfile setting.
  Lets Codeep quietly learn durable preferences from sessions into a separate
  profile.learned.md. Pushed live to the running CLI and applied on every
  connect. Off until you turn it on; review with the CLI /me, clear with
  /me forget.
• "Personalize Codeep" walkthrough step in the Get Started walkthrough.

v2.3.0

Deeper editor integration: a @codeep participant in the native Chat view, a #codeepSkills agent tool, generate commit messages from the Source Control panel, a native Sessions tree, JSON validation for MCP config, and Workspace Trust support.

Also ships the editor features from the (unreleased) 2.2: Code Actions lightbulb (Explain / Improve / Add tests / Add doc comment / Fix), a status-bar model picker, a codeep.baseUrl setting for self-hosted OpenAI-compatible endpoints, and a Get Started walkthrough. See the 2.2.0 entry below for details.

Added

• @codeep chat participant. Invoke Codeep from the native VS Code Chat
  view — type @codeep and ask, or use @codeep /explain and @codeep /review
  with a selection. Answers come from your configured Codeep provider/model (via
  the CLI), not VS Code's model picker. Runs on its own session, independent of
  the sidebar chat.
• #codeepSkills language-model tool. Exposes the workspace's Codeep skill
  bundles (.codeep/skills/*/SKILL.md) to VS Code agent mode and #-references,
  so the native agent can discover and follow your project's own workflows.
• Generate Commit Message. A sparkle button in the Source Control title
  (and Codeep: Generate Commit Message in the palette) reads your staged
  diff — falling back to the working-tree diff — and writes a Conventional
  Commits message into the commit box. It asks before replacing a message
  you've already typed.
• Sessions tree view. A native Sessions view in the Codeep sidebar
  lists your saved conversations (title + age). Click to load one into the
  chat; use the inline trash to delete; the title bar has New Session +
  Refresh. Stays in sync with the chat panel.
• MCP config validation. .codeep/mcp_servers.json (project and global)
  now gets JSON schema validation + autocomplete — catches a mistyped
  command/args/env before you start a session.

Changed

• Workspace Trust — the extension now declares limited support for
  untrusted workspaces. Codeep runs a local agent that can edit files and run
  commands, so in untrusted folders you'll be reminded to review permission
  prompts carefully.
• Minimum VS Code raised to 1.95 — required for the stable Chat Participant
  and Language Model Tools APIs.

Notes

• No new CLI requirement — 2.3.0 builds on ACP methods already in the shipped
  CLI. Pure additive UI; safe upgrade with zero migration.

v2.1.1

Custom OpenAI-compatible endpoints now work through the extension. Point Codeep at a self-hosted vLLM / LiteLLM / LM Studio server and use it in the VS Code chat — no commercial provider required.

Added

• Custom (OpenAI-compatible) endpoint support (via CLI 2.1.1). Run any
  OpenAI-compatible server — vLLM, LiteLLM, LM Studio, text-generation-webui —
  and the extension talks to it through the same codeep agent. Configure it
  once in the CLI: set provider custom + customBaseUrl
  (e.g. http://host:8000/v1) in ~/.codeep/config.json or via
  /settings → Custom Base URL, then pick your model with /model. The
  openai provider also honors the OPENAI_BASE_URL env var.

Requires

• CLI 2.1.1+ on your PATH. The extension is a thin client that spawns
  codeep acp, so endpoint resolution happens in the CLI — there's no separate
  extension setting and nothing to configure inside VS Code beyond the shared
  ~/.codeep config. Run npm i -g codeep@latest or brew upgrade codeep.

Notes

• No extension code changes — this is a version-parity release so the
  marketplace listing reflects that CLI 2.1.1 unlocked self-hosted / custom
  endpoints for editor users too. Safe upgrade with zero migration.

v2.1.0

Surfaces the new CLI 2.1.0 /recall command in the chat autocomplete and the Settings panel. Search across every saved session — not just the current one — with --summarize for an LLM recap of what you actually did.

Added

• /recall chip in chat autocomplete + Settings → Commands. Type
  /recall <query> to search across ALL your saved sessions (vs
  /search, which scans only the current conversation). Append
  --summarize for an LLM-written recap of what you accomplished across
  the matching sessions, or --resume to jump straight into the
  top-matching one. The extension just inserts the command — the CLI
  does the cross-session search, ranking, and summary.

Requires

• CLI 2.1.0+ on your PATH. The extension is a thin client over the
  local codeep binary, so /recall only works once codeep --version
  reports 2.1.0 or newer. Run npm i -g codeep@latest or
  brew upgrade codeep if your shell still shows 2.0.x.

Notes

• Pure additive command-list entry — no webview, activation, or settings
  schema changes. Safe upgrade with zero migration. The CLI 2.1.0 release
  also adds AI-generated session titles (so /recall and /sessions
  read like "OAuth2 migration" instead of raw IDs) and portable
  personalities + custom-command sync via codeep account sync.

v2.0.3

Surfaces the new CLI 2.0.3 commands as one-click chips in the Settings panel: /personality (six built-in agent tones plus your own from .codeep/personalities/*.md) and /insights (activity summary over the last N days).

Added

• /personality chip in Settings → Commands → Personalization.
  Click to scaffold a /personality prompt in the chat input. Type a
  preset name (concise, verbose, security, senior-reviewer,
  junior-mentor, ship-it) or off to clear. Custom personalities
  from .codeep/personalities/<name>.md (project) and
  ~/.codeep/personalities/<name>.md (global) work the same way —
  CLI handles the lookup, extension is just a launcher.
• /insights chip in Settings → Commands → Agent flow. Click to
  insert /insights — append --days N for a custom window
  (default 7 days). Output renders in the chat as a Markdown table:
  runs, tool actions, active time, by-project / top-tools /
  most-touched files / recent runs.

Requires

• CLI 2.0.3+ on your PATH (the extension is a thin client over
  the local codeep binary, so the chips just insert text — the CLI
  does the actual work). Run npm i -g codeep@latest or
  brew upgrade codeep if your shell still shows 2.0.2 or older
  from codeep --version.

Notes

• No webview / activation / settings schema changes — pure additive
  chips in the existing Settings panel. Safe upgrade with zero
  migration.