All notable changes to this project are documented here. The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.
0.3.0 - 2026-06-01
v1 candidate — documentation, security, and stability hygiene (no behaviour
changes). Still 0.x / Beta until a live smoke test on a real instance.
SECURITY.md— private vulnerability reporting, scope, and response policy.- README Stability section: the SemVer public contract from
1.0.0(CLI flags,.envkeys, Talk commands, status JSON, forward-compatible migrations). CONTRIBUTING.md: the live smoke test is now a required gate before a stable release, plus a pre-1.0 checklist (incl. flipping the classifier).- Issue forms (
bug_report,feature_request), an issue-templateconfigrouting security reports to private advisories, and a PR template.
__version__is derived from package metadata — single source of truth inpyproject.toml(no second bump in__init__.py).- Clarified OpenCode compatibility: tested against 1.15.11, expected compatible with the 1.15+ HTTP API.
- Refreshed
docs/publishing.mdfor the live (post-first-publish) state.
0.2.8 - 2026-06-01
- Republish so the PyPI project description reflects the corrected README
(the
uv tool install opencode-talk-bridgeinstructions). The 0.2.7 page was built from the README commit just before the PyPI-primary install flip.
0.2.7 - 2026-06-01
First release published to PyPI — uv tool install opencode-talk-bridge
(or pipx install …) now works without a git URL.
- PyPI Trusted-Publishing workflow (
.github/workflows/publish.yml): on eachvX.Y.Ztag it guards the tag↔version match, builds,twine checks, and publishes to PyPI via OIDC (no token). Package classifiers + adocs/publishing.mdsetup guide.
0.2.6 - 2026-06-01
nextcloud-talk-coreis now a normal PyPI dependency (>=1.0.2,<2) instead of a pinned git URL — it has been published to PyPI. Simpler installs and resolution; droppedallow-direct-referencesand the git-dep notes in the README/CI. Verified our used API (TalkClient/OCSClient methods) is unchanged in 1.0.2 and the full suite passes against it.
0.2.5 - 2026-06-01
- No server text leaks into chat. The prompt worker caught generic
exceptions and posted
Fehler: {exc}; anOpenCodeErrorcarries up to ~300 chars of the OpenCode HTTP response body, which could reach Talk. The worker (both session-setup and prompt paths) now catchesOpenCodeErrorand posts a genericoc_errormessage, logging the details only — matching the command handlers. Remainingerror: {exc}paths only fire for local validation errors (e.g. a bad model string), which carry no server body.
0.2.4 - 2026-05-31
Hardening pass from a code review.
- Error isolation. Command handlers run inline on the poll thread and only
caught
OpenCodeDownError; an OpenCode HTTP 4xx/5xx (OpenCodeError) could escape and kill a conversation's poll thread. The poll loop now isolates every message (logs + notifies + keeps polling), and command dispatch catchesOpenCodeErrorwith a clean message. /taskvalidation. Reject0-minute schedules and empty prompts (minutes >= 1and a non-empty prompt are now required).
- Added tests for
__main__(CLI wiring:--check/--init/run, 0→90%) and the Talk gateway / message parsing (57→95%); overall coverage 76→82%.
0.2.3 - 2026-05-31
/sessions,/commands, and/skillsare project-scoped in OpenCode, but the bridge listed them against the globalOPENCODE_DIRECTORYand ignored the per-conversation project chosen via/projects. They now use the conversation's directory, so/sessionsshows the same sessions as the OpenCode desktop/TUI for that project (and follows/projectsswitches)./projectsitself is global (all known projects), matching desktop.
0.2.2 - 2026-05-31
opencode-talk-bridge --init— interactive.envwizard. Prompts for the essentials (Talk credentials, allowlist, OpenCode URL), hides the app password, refuses to overwrite without confirmation, and writes the filechmod 600.
0.2.1 - 2026-05-31
/skills— browse & run OpenCode skills (viaGET /skill).
run_commandnow always sends the requiredargumentsfield — without it the command endpoint returned HTTP 400, so/commands(and any command run) was broken./commandsno longer lists skills (entries withsource == "skill"); those live under/skills.
0.2.0 - 2026-05-31
Feature-parity push toward grinev/opencode-telegram-bot, adapted to Talk.
-
Live response streaming: the assistant reply is streamed into a single Talk message via editing (
PUT, throttled bySTREAM_THROTTLE_MS), instead of one post at the end. Toggle withRESPONSE_STREAMING. -
Tool & thinking notices from the SSE stream (
💻 bash,💭 …), each tool announced once per turn; toggle withHIDE_TOOL_MESSAGES/HIDE_THINKING. -
Agent questions (
question.asked) are surfaced into Talk and answered by a numbered option or free text. -
Numbered-text pickers (Talk has no buttons):
/sessions(switch),/model(pick),/agent(pick plan/build/…);/agent <name>sets directly. -
OpenCode client wrappers for sessions (rename/revert/unrevert/fork), projects, worktrees, models, agents, commands, MCP, and question reply/reject.
-
Breadth commands (Phase 2):
/rename,/detach,/projects,/worktree,/commands(run with streaming),/mcps(toggle), and/messages→ revert/fork. Projects/worktrees bind a per-conversation directory for new sessions. -
Background-session notifications when a non-foreground session goes idle (
TRACK_BACKGROUND_SESSIONS). -
i18n: user-facing strings localised via
messages.py(German + English), selected byBOT_LOCALE. -
Voice & files (Phase 3): incoming file/image attachments are inlined into the prompt as OpenCode file parts; voice notes are transcribed (STT, Whisper-compatible) into the prompt; optional spoken replies (TTS) shared as audio, toggled per-conversation with
/tts. Config:STT_*,TTS_*. -
Scheduled tasks (Phase 3):
/task <min> <prompt>(or/task every <min> …) and/tasklist; a daemon scheduler fires due tasks, persisted in SQLite.
- Internal refactor: unified pending-interaction registry (
pending.py), streaming substrate (streaming.py), and typed SSE dispatch (events.py). TalkGateway.sendnow returns the message id; addededit/download.
0.1.1 - 2026-05-30
- File attachments now upload to Nextcloud directly via WebDAV (
PUT, creating the target folder on demand) before sharing into the conversation. This removes the previous dependency on a desktop sync client having already uploaded the file, which could makeshare_filefail with "not found". - Attachment config simplified to a single
SHARE_WEBDAV_DIR(the server-side folder), replacingSHARE_DIR+SHARE_WEBDAV_ROOT.
webdav.py: minimal WebDAV client (reuses the app-password credentials).
0.1.0 - 2026-05-30
Initial release.
- Polling bridge between Nextcloud Talk and a local OpenCode instance — no webhooks, works on institutional Talk instances without admin access.
- OpenCode HTTP + SSE client (
opencode.py), verified against theopencode serve1.15 OpenAPI: health, session create/list, blocking prompt, abort, global permission reply, and the/global/eventstream. - Mandatory user allowlist enforced on the stable OCS
actorIdwithactorType == "users"; the bridge refuses to start with an empty allowlist. - Permission prompts for dangerous agent actions surfaced into Talk, answered
with
ja/immer/neinand replied viaPOST /permission/{id}/reply. - SQLite session store mapping each Talk conversation to an OpenCode session, model, and last-known-message cursor; persists across restarts.
- Slash-commands:
/new,/session,/model,/stop,/status,/help. - File attachments for long output / code, written to a Nextcloud share folder and shared into the conversation instead of posting a large text block.
- Atomic JSON status file as a stable contract for a supervising menubar app.
- CLI entrypoint with
--checkand graceful SIGINT/SIGTERM shutdown. - launchd user-agent example (
deploy/), README threat model, and the.env.exampletemplate. - Test suite (mocked Talk + OpenCode, no live calls) and CI matrix on Python 3.10–3.13 with ruff lint + format checks.
nextcloud-talk-core, pinned to thecore-v1.0.0git tag.httpx >= 0.27.