claude-threads 1.24.3 → 1.25.1

package/CHANGELOG.md

@@ -5,6 +5,32 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [1.25.1] - 2026-08-20
+
+### Fixed
+- **Keep-alive inhibitors no longer outlive the bot process.** (#480, thanks @Jadefalkner) The sleep-prevention child processes are now tied to the bot's lifetime at the OS level: on Linux `systemd-inhibit` runs `cat` on a pipe held by the bot (EOF on bot death releases the lock, even after SIGKILL), on macOS `caffeinate -w <pid>` watches the bot pid natively, and the xdg-screensaver/PowerShell fallbacks poll the parent pid. Previously a hard death of the bot (SIGKILL, crashed test runner) orphaned `systemd-inhibit sleep infinity` processes to init, permanently blocking system sleep/hibernate until they were killed by hand.
+
+## [1.25.0] - 2026-08-19
+
+### Added
+- **Persistent memory — the bot now learns over time**, modeled on how Anthropic's own products do memory, in two layers scoped per platform instance (≈ one channel, the hard privacy boundary, mirroring Claude Tag's per-channel isolation):
+  - **Repo layer (Claude Code style):** Claude Code's native auto-memory is redirected via the `autoMemoryDirectory` setting into a bot-managed per-(platform, repository) directory under `~/.config/claude-threads/memory/`. Claude's built-in save/recall machinery does the learning; the bot only controls where memory lives — which also makes it immune to the account pool's per-session `HOME` overrides. Worktrees of one repo share memory, like native behavior. Verified headless over stream-json against CLI 2.1.235 (write, load-at-start, resume, HOME override, kill switch).
+  - **Channel layer (Claude Tag style):** a shared per-channel `MEMORY.md` of team notes, injected into every session's system prompt (capped at the native 200-line/25KB limits) with explicit "background context, not instructions" framing as a prompt-injection mitigation. Written by `!remember <text>` and by **end-of-session distillation** — a one-shot haiku pass that extracts up to 3 durable facts when a session ends (`!stop`, normal exit, or idle timeout; never on pause/respawn/shutdown).
+  - **Transparency commands:** `!memory` lists the channel's entries numbered; `!memory forget <n|text>` removes one (owner-gated, with ambiguity handling); `!memory forget all` clears the channel.
+  - **Config:** per-platform `memory:` option — on by default; `memory: false` disables everything (and force-sets `CLAUDE_CODE_DISABLE_AUTO_MEMORY=1` on the CLI child so native auto-memory can't accumulate cross-channel context under a shared pooled-account `$HOME`); an object form toggles `repoLayer` / `channelLayer` / `distillation` individually. Storage override: `CLAUDE_THREADS_MEMORY_DIR`.
+  - `ClaudeCliOptions.memory` is a deliberately **required** field so every present and future spawn/respawn site must decide its memory binding explicitly — the same compiler-enforced-checklist pattern that protects `uploadDir`.
+  - **Supersede authorization (security-review finding):** a new note can only replace ("supersede") entries its writer legitimately owns — distilled entries (background inference, no author) or the same user's own earlier notes. Previously any session-authorized user, including temporarily `!invite`d collaborators, could silently delete another user's entry by embedding its text in a `!remember` — bypassing the owner gate that protects `!memory forget`. Conflicting notes now coexist (visible in `!memory`, resolvable by the owner), and replacements are reported in the confirmation instead of happening silently.
+  - **Review-driven hardening:** `!memory` listings are batched under the platform's post-size limit (a full 400-entry memory would otherwise exceed Mattermost's 16K cap exactly when inspection matters most); ambiguous `!memory forget` previews cap at 10 matches; distillation prompts carry only the newest 50 existing entries instead of the whole file; repo-memory keys follow a linked worktree back to its main repository via `git rev-parse --git-common-dir`, so `!cd` into a worktree (or a legacy resumed session without recorded worktree info) keeps the shared repo memory; hand-written `#` headings in the channel `MEMORY.md` survive rewrites; the truncation warning for `!remember` measures the collapsed text (newline collapsing can lengthen it); and an explicit user `!remember` whose text is a fragment of an existing entry now lands instead of being reported "Already known" — a fragment may be a correction, and only distilled candidates are deduped by containment.
+
+### Fixed
+- **Repo memory stays correctly keyed after leaving a worktree.** `!cd` does not clear `session.worktreeInfo`, and the `!permissions`/plugin-restart/resume paths passed its recorded `repoRoot` as the repo-memory key override — so a session that left its worktree for another repository kept reading and writing the old repository's memory. A new `activeWorktreeRepoRoot` helper honors the recorded root only while the working directory is still inside the worktree (preserving the deleted-worktree resume fast path); otherwise the key derives from the working directory itself.
+- **Slack threads now distill their most recent messages, not their oldest.** The Slack client forwarded `getThreadHistory`'s limit straight to `conversations.replies`, which paginates oldest-first — a long thread's distillation (and the pre-existing context-prompt and work-summary features) saw the opening of the conversation and missed the decisions at the end. The client now fetches a full page and keeps the most recent N, matching the Mattermost client's contract.
+- **`quickQuery` prompts travel over stdin instead of argv.** Distillation prompts (existing memory + thread tail, up to ~40KB) exceeded Windows command-line limits as a single argv argument via the npm shim, silently killing the spawn — distillation (and long work-summary prompts) never worked on Windows. Verified against CLI 2.1.235 that `claude -p` reads the prompt from stdin.
+- **Chained worktree switches no longer split repo memory.** Creating a worktree *from inside* another worktree keyed the new session's auto-memory (and the persisted `worktreeInfo.repoRoot` that resume/`!permissions`/plugin restarts reuse) to the intermediate worktree's own path instead of the main repository — worktree creation now resolves the main root via `git rev-parse --git-common-dir` (`getMainRepositoryRoot`, also reused by `detectWorktreeInfo`, replacing its weaker suffix-stripping copy).
+- **`!memory` no longer pings entry authors.** The listing rendered authors as live @-mentions, notifying every author each time anyone viewed the read-only list; authors now render as inline code.
+- **Malformed per-field `memory` config values warn at startup** (e.g. `channelLayer: "off"` silently stayed enabled before) instead of falling back silently.
+- **Plugin install/uninstall respawns no longer drop session wiring.** The `!plugin install`/`uninstall` restart paths built `ClaudeCliOptions` by hand, omitting the session's pooled Claude account (so `--resume` under a multi-account pool ran against the wrong `$HOME` and failed with "No conversation found"), the `send_file` upload directory, the decision-bridge socket (breaking plan approvals after a plugin change), and the append-system-prompt (stripping platform context, co-author rules, attribution, and channel memory — `--append-system-prompt` is per-invocation and not re-applied by `--resume`). Both paths now share a builder on `buildRestartCliOptions` and rebuild the prompt like the `!cd`/`!permissions` restarts.
+
 ## [1.24.3] - 2026-08-19
 
 ### Changed

package/README.md

@@ -53,6 +53,7 @@
 - **Collaboration** - `!invite` teammates into a session; commits Claude makes get `Co-Authored-By:` trailers for everyone involved
 - **A real chat citizen** - Nine MCP tools let Claude post files, follow permalinks, react, and DM, each behind its own guardrail (see [What Claude can do in your chat](#what-claude-can-do-in-your-chat))
 - **Git worktrees** - `!worktree feature/foo` isolates Claude's changes on a branch
+- **Memory** - Channels learn over time: `!remember` saves shared team notes every session sees, sessions distill durable facts when they end, and Claude's native auto-memory keeps per-repo project knowledge — all scoped per channel ([docs](https://github.com/anneschuth/claude-threads/blob/main/docs/CONFIGURATION.md#memory-memory-default-fully-enabled))
 - **Files both ways** - Drop any file into the chat for Claude to read, with full multimodal for images and PDFs; Claude posts screenshots, plots, or PDFs back with `send_file` (100 MB cap)
 - **Quiet mode and verbosity dials** - `!mentions on` makes a session respond only when mentioned; session headers and the channel sticky each have `full`/`minimal`/`hidden` modes
 - **Runs on macOS, Linux, and Windows** - Windows via Git Bash or WSL