fix: healthcheck.sh exits prematurely under set -e (#302) by nathanschram · Pull Request #303 · littlebearapps/untether

Nathan Schram (nathanschram) · 2026-04-14T10:39:55Z

Summary

pass()/fail() used ((var++)) which returns the pre-increment value — on first call (var=0) that trips set -e and the script exits after the first check. Switched to explicit $((var + 1)) assignment.
Error-log count piped journalctl through grep -c ., counting -- No entries -- as a match (false positive on clean systems). Now uses grep -vc '^-- ' to drop meta lines.

Fixes #302.

Test plan

Ran scripts/healthcheck.sh --version 0.35.1rc5 against live staging (post rc5 install) — 5 passed, 0 failed, exit 0.
Ran the no-errors path (clean logs, expected version match).
Confirm no regression for the FAIL paths — if staging ever emits an ERROR log within 60s of restart, the check should still flag it (logic unchanged apart from meta-line filtering).

🤖 Generated with Claude Code

…#158, #159) * ci: add CODEOWNERS, update action SHA pins, add permission comments - Create .github/CODEOWNERS requiring @littlebearapps/core review - Pin setup-uv to v7.4.0 (6ee6290f), download-artifact to v8.0.1 (3e5f45b2) - Add precise version comments on all action SHAs (codeql v3.32.6, pypi-publish v1.13.0, action-gh-release v2.5.0, fetch-metadata v2.5.0) - Document write permissions with why-comments (OIDC, releases, auto-merge) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * feat: add release guard hooks and document protection in CLAUDE.md Defence-in-depth hooks prevent Claude Code from pushing to master, merging PRs, creating tags, or triggering releases. Feature branch pushes and PR creation remain allowed. - release-guard.sh: Bash hook blocking master push, tags, releases, PR merge - release-guard-protect.sh: Edit/Write hook protecting guard files and hooks.json - release-guard-mcp.sh: GitHub MCP hook blocking merge and master writes - hooks.json: register all three hooks - CLAUDE.md: document release guard, update workflow roles, CI pipeline notes Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix: clarify /config default labels and remove redundant "Works with" lines Default labels now explain what "default" means for each setting: - Diff preview: "default (off)" — matches actual behaviour (was "default (on)") - Model/Reasoning: "default (engine decides)" - API cost: "default (on)", Subscription usage: "default (off)" - Plan mode home hint: "agent decides" - Diff preview home hint: "buttons only" Added info lines to plan mode and reasoning sub-pages explaining the default behaviour in more detail. Removed all 9 "Works with: ..." lines from sub-pages — they're redundant because engine visibility guards already hide settings from unsupported engines. Fixes #119 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix: suppress redundant cost footer on error runs When a run fails (e.g. subscription limit hit), the diagnostic context line from _extract_error() already shows cost, turns, and API time. The 💰 cost footer was duplicating this same data in a different format. Now the cost footer only appears on successful runs where it's the sole source of cost information. Error runs still show cost in the diagnostic line, and budget alerts still fire regardless. Also adds usage field to mock Return dataclass (matching ErrorReturn) so tests can verify cost footer behaviour on success runs. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * feat: suppress stall notifications when CPU-active + heartbeat re-render When cpu_active=True (extended thinking, background agents), suppress Telegram stall warning notifications and instead trigger a heartbeat re-render so the elapsed time counter keeps ticking. Notifications still fire when cpu_active=False or None (no baseline). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * chore: staging 0.34.5rc2 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix: CI release-validation tomllib bytes/str mismatch tomllib.loads() expects str but was receiving bytes from sys.stdin.buffer.read() and open(...,'rb').read(). First triggered when PR #122 changed the version (rc1 → rc2). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * docs: integrate screenshots into docs with correct JPG references - Add 44 screenshots to docs/assets/screenshots/ - Fix all image refs from .png to .jpg across 25 doc files - README uses absolute raw.githubusercontent.com URLs for PyPI rendering - Fix 5 filename mismatches (session-auto-resume→chat-auto-resume, etc.) - Comment out 11 missing screenshots with TODO markers - Add CAPTURES.md checklist tracking capture status Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: convert markdown images to HTML img tags for GitHub compatibility Switch from MkDocs `![alt](src){ loading=lazy }` syntax to HTML `<img>` tags with width="360" and loading="lazy". Fixes two GitHub rendering issues: `{ loading=lazy }` appearing as visible text, and oversized images with no width constraint. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: fix 3 screenshot mismatches and replace 3 screenshots - first-run.md: rewrite resume line text to match footer screenshot - interactive-control.md: update planmode show admonition to match screenshot (auto not on) - switch-engines.md: swap engine-footer.jpg for multi-engine-switch.jpg - Replace startup-message.jpg with clean v0.34.4 capture (was rc/6-projects) - Replace cooldown-auto-deny.jpg with post-outline approve/deny buttons - Replace file-put.jpg with photo save confirmation Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: add iOS caption limitation note to file transfer guide Telegram iOS doesn't show a caption field when sending documents via the File picker, so /file put <path> captions aren't easily accessible. Added a note with workarounds (use Desktop, send as photo, or let auto-save handle it). Updated screenshot alt text to match actual screenshot content. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: temp swap README image URLs to feature branch for preview Will revert to master before merging. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: lay out all 3 README screenshots in a single row Reduce from 360px to 270px each and combine into one block so all three hero screenshots sit side by side. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: swap 3rd hero screenshot for config-menu for visual variety Replace plan-outline-approve (too similar to approval-diff-preview) with config-menu showing the /config settings grid. The three hero images now tell: voice input → approve changes → configure everything. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: add captions under README hero screenshots Small captions: "Send tasks by voice (Whisper transcription)", "Approve changes remotely", "Configure from Telegram". Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: use table layout for README hero screenshots with captions Fixes stacking issue — in a broke inline flow. A table keeps images side by side with captions underneath each one. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: replace table layout with single hero collage image Composite image scales proportionally on mobile instead of requiring horizontal scroll. Captions baked into the image via ImageMagick. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: swap middle hero screenshot for full 3-button approval view Replace approval-diff-preview with approval-buttons-howto showing Approve / Deny / Pause & Outline Plan — more visually impressive. Caption now reads "Approve changes remotely (Claude Code)". Added footnote linking to engine compatibility table. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: swap config-menu for parallel-projects in hero collage Third hero screenshot now shows 10+ projects running simultaneously across different repos — much more compelling than a settings menu. New caption: "Run agents across projects in parallel". Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: revert README image URL to master for merge Swap hero-collage URL back from feature/github-hardening to master. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: staging 0.34.5rc3 - fix: preserve all EngineOverrides fields when setting model/planmode/reasoning (was silently wiping ask_questions, diff_preview, show_api_cost, etc.) - fix: /config home page resolves "default" to effective values - feat: file upload auto-deduplication (append _1, _2 instead of requiring --force) - feat: media groups without captions now auto-save instead of showing usage text - feat: resume line visual separation (blank line + ↩️ prefix) - fix: claude auto-approve echoes updatedInput in control response Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: expand permission policies for Codex CLI and Gemini CLI in /config Codex gets a new "Approval policy" page (full auto / safe) that passes --ask-for-approval untrusted when safe mode is selected. Gemini's approval mode expands from 2 to 3 tiers (read-only / edit files / full access) with --approval-mode auto_edit for the middle tier. Both engines now show an "Agent controls" section on the /config home page. Engine-specific model default hints replace the generic "from CLI settings" text. Also adds staging.sh helper, context-guard-stop hook, and docs updates. Closes #131 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: staging 0.34.5rc4 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: /config UX overhaul, resume line toggle, cost budget settings, model metadata /config UX cleanup: - Convert all binary toggles from 3-column (on/off/clear) to 2-column (toggle + clear) for better mobile tap targets - Merge Engine + Model into combined "Engine & model" page - Reorganise home page to max 2 buttons per row across all engines - Split plan mode 3-option rows (off/on/auto) into 2+1 layout - Add _toggle_row() helper for consistent toggle button rendering New features: - #128: Resume line /config toggle — per-chat show_resume_line override via EngineOverrides with On/Off/Clear buttons, wired into executor - #129: Cost budget /config settings — per-chat budget_enabled and budget_auto_cancel overrides on the Cost & Usage page, wired into _check_cost_budget() in runner_bridge.py Model metadata improvements: - Show Claude Code [1m] context window suffix: "opus 4.6 (1M)" - Strip Gemini CLI "auto-" prefix: "auto-gemini-3" → "gemini-3" - Future-proof: unknown suffixes default to .upper() (e.g. [500k] → 500K) Bug fixes: - #124: Standalone override commands (/planmode, /model, /reasoning) now preserve all EngineOverrides fields including new ones - Error handling: control_response.write_failed catch-all in claude.py, ask_question.extraction_failed warning, model.override.failed logging Hardening: - Plan outline sent as separate ephemeral message (avoids 4096 char truncation) - Added show_resume_line, budget_enabled, budget_auto_cancel to EngineOverrides, EngineRunOptions, normalize/merge, and all constructors Tests: 1610 passed, 80.56% coverage, ruff clean. Integration tested on @untether_dev_bot across all 6 engine chats. Closes #128, closes #129, fixes #124 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: trigger CI for PR #132 * fix: address 11 CodeRabbit review comments on PR #132 Bug fixes: - claude.py: fix UnboundLocalError when factory.resume is falsy in ask_question.extraction_failed logging path - ask_question.py: reject malformed option callbacks instead of silently falling back to option 0 - files.py: raise FileExistsError when deduplicate_target exhausts 999 suffixes instead of returning the original (overwrite risk) - config.py: disambiguate Codex "Full auto" (fa) vs Gemini "Full access" (ya) callback IDs and toast labels Hardening: - codex.py: add --ask-for-approval to _EXEC_ONLY_FLAGS guard - model.py: add try/except to clear path (matching set path) - reasoning.py: add try/except to clear path (matching set path) - loop.py: notify user when media group upload fails instead of silently dropping - export.py: log session count instead of identifiers at info level - config.py: resolve resume-line default from config instead of hardcoding True - staging.sh: pin PyPI index in rollback/reset with --pip-args Skipped (not applicable): - CHANGELOG.md: RC versions don't get changelog entries per release discipline - docs/tutorials TODO screenshot: pre-existing, not introduced by PR - .claude/hooks/context-guard-stop.sh: ContextDocs plugin hook, not Untether source Tests: 1611 passed, 80.48% coverage, ruff clean. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: replace bare pass with debug log to satisfy bandit B110 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: setup wizard security + UX improvements - Auto-set allowed_user_ids from captured Telegram user ID during onboarding (security: restricts bot to the setup user's account) - Add "next steps" panel after wizard completion with pointers to /config, voice notes, projects, and account lock confirmation - Update install.md: Python 3.12+ (not just 3.14), dynamic version string, /config mention for post-setup changes - Update first-run.md: /config → Engine & model for default engine Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: plan outline UX — markdown rendering, buttons, cleanup (#139, #140, #141) - Render outline messages as formatted text via render_markdown() + split_markdown_body() instead of raw markdown (#139) - Add approve/deny buttons to last outline message so users don't have to scroll up past long outlines (#140) - Delete outline messages on approve/deny via module-level _OUTLINE_REGISTRY callable from callback handler; suppress stale keyboard on progress message (#141) - 8 new tests for outline rendering, keyboard placement, and cleanup - Bump version to 0.35.0rc5 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: /continue command — cross-environment resume for all engines (#135) New `/continue` command resumes the most recent CLI session in the project directory from Telegram. Enables starting a session in your terminal and picking it up from your phone. Engine support: Claude (--continue), Codex (resume --last), OpenCode (--continue), Pi (--continue), Gemini (--resume latest). AMP not supported (requires explicit thread ID). Includes ResumeToken.is_continue flag, build_args for all 6 runners, reserved command registration, resume emoji prefix stripping for reply-to-continue, docs (how-to guide, README, commands ref, routing explanation, conversation modes tutorial), and 99 new test assertions. Integration tested against @untether_dev_bot — all 5 supported engines passed secret-recall verification via Telegram MCP. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: agent outbox file delivery + fix cross-chat ask stealing (#143, #144) Outbox delivery (#143): agents write files to .untether-outbox/ during a run; Untether sends them as Telegram documents on completion with 📎 captions. Config: outbox_enabled, outbox_dir, outbox_max_files, outbox_cleanup. Deny-glob security, size limits, auto-cleanup. Preamble updated for all 6 engines. Integration tested across Claude, Codex, OpenCode, Pi, and Gemini. AskUserQuestion fix (#144): _PENDING_ASK_REQUESTS and _ASK_QUESTION_FLOWS were global dicts with no chat_id scoping — a pending ask in one chat would steal the next message from any other chat. Added channel_id contextvar and scoped all ask lookups by it. Session cleanup now also clears stale pending asks. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: v0.35.0 changelog completion + fix #123 updatedInput - Complete v0.35.0 changelog: add missing entries for /continue (#135), /config UX overhaul (#132), resume line toggle (#128), cost budget (#129), model metadata, resume line formatting (#127), override preservation (#124), and updatedInput fix (#123) - Fix #123: register input for system-level auto-approved control requests so updatedInput is included in the response - Add parameterised test for all 5 auto-approve types input registration - Remove unused OutboxResult import (ruff fix) Issues closed: #115, #118, #123, #124, #126, #127, #134 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: staging 0.35.0rc6 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: rc6 integration test fixes (#145, #146, #147, #148, #149) - Reduce Telegram API timeout from 120s to 30s (#145) - OpenCode error runs show error text instead of empty body (#146) - Pi /continue captures session ID via allow_id_promotion (#147) - Post-outline approval uses skip_reply to avoid "not found" (#148) - Orphan progress message cleanup on restart (#149) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: post-outline notification reply + OpenCode empty body (#148, #150) - #148: skip_reply callback results now bypass the executor's default reply_to fallback, sending directly via the transport with no reply_to_message_id. Previously, the executor treated reply_to=None as "use default" which pointed to the (deleted) outline message. - #150: OpenCode normal completion with no Text events now falls back to last_tool_error. Added state.last_tool_error field populated on ToolUse error status. Covers both translate() and stream_end_events(). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: suppress post-outline notification to avoid "message not found" (#148) After outline approval/denial, the progress loop's _send_notify was firing for the next tool approval, but the notification's reply_to anchor could reference deleted state. Added _outline_just_resolved flag to skip one notification cycle after outline cleanup. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: note OpenCode lacks auto-compaction — long sessions degrade (#150) Added known limitation to OpenCode runner docs and integration testing playbook. OpenCode sessions accumulate unbounded context (no compaction events unlike Pi). Workaround: use /new before isolated tests. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: skip_reply on regular approve path when outline was deleted (#148) The "Approve Plan" button on outline messages uses the real ExitPlanMode request_id, routing through the regular approve path (not the da: synthetic path). When outline messages exist, set skip_reply=True on the CommandResult to avoid replying to the just-deleted outline message. Also added reply_to_message_id and text_preview to transport.send.failed warning for easier debugging. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: update changelog for rc6 integration test fixes (#145-#150) Updated fix descriptions for #146/#150 (OpenCode last_tool_error fallback) and #148 (regular approve path skip_reply). Added docs section for OpenCode compaction limitation. Updated test counts. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * style: fix formatting after merge resolution Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: address CodeRabbit review comments on PR #151 - bridge.py: replace text_preview with text_len in send failure warning to avoid logging raw message content (security) - runner_bridge.py: move unregister_progress() after send_result_message() to avoid orphan window between ephemeral cleanup and final message send - cross-environment-resume.md: add language spec to code block Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: resolve /config "default" labels to effective on/off values (#152) Sub-pages showed "Current: default" or "default (on/off)" while buttons already showed the resolved value. Now all boolean-toggle settings show the effective on/off value in both text and buttons. Affected: verbose, ask mode, diff preview, API cost, subscription usage, budget enabled/auto-cancel, resume line. Home page cost & resume labels also resolved. Plan mode, model, and reasoning keep "default" since they depend on CLI settings and aren't simple on/off booleans. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: update changelog for rc7 config default labels fix (#152) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: update documentation for v0.35.0 - fix missing nav entries in zensical.toml (cross-env resume, Gemini/Amp runners) - rewrite inline-settings.md for /config UX overhaul (2-column toggles, budget/resume toggles) - update plan-mode.md with outline rendering, buttons-on-last-chunk, ephemeral cleanup - update interactive-control tutorial with outline UX improvements - add orphan progress cleanup section to operations.md - add engine-specific approval policies to interactive-approval.md - add per-chat budget overrides to cost-budgets.md - update module-map.md with Gemini/Amp and new modules (outbox, progress persistence, proc_diag) - update architecture.md mermaid diagrams with all 6 engines - bump specification.md to v0.35.0, add progress persistence and outbox sections - add v0.35.0 screenshot entries to CAPTURES.md Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: broaden frozen ring buffer stall escalation beyond MCP tools (#155) Frozen ring buffer escalation was gated on `mcp_server is not None`, so general stalls with cpu_active=True and no MCP tool running were silently suppressed indefinitely. Broadened to fire for all stalls after 3+ checks with no new JSONL events regardless of tool type. New notification: "CPU active, no new events" for non-MCP frozen stalls. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: tool approval buttons no longer suppressed after outline approval (#156) After "Approve Plan" on an outline, the stale discuss_approve action remained in ProgressTracker with completed=False. The renderer picked up its stale "Approve Plan"/"Deny" buttons first, then the suppression logic at line 994 stripped ALL buttons — including new Write/Edit/Bash approval buttons. Claude blocked indefinitely waiting for approval. Fix: after suppressing stale buttons, complete the discuss_approve action(s) in the tracker, reset _outline_sent, and trigger a re-render so subsequent tool requests get their own Approve/Deny buttons. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: add workflow mode indicator to startup message, fix startup crash on topics validation failure (#158, #159) Features: - Startup message now shows mode: assistant/workspace/handoff - Derived from session_mode + topics.enabled config values - _resolve_mode_label() helper in backend.py Bug fixes: - Fix UnboundLocalError crash when topics validation fails on startup (#158) - Moved import signal and shutdown imports before try block in loop.py - Downgrade can_manage_topics check from fatal error to warning (#159) - Bot can now start without manage_topics admin right - Existing topics work fine; only topic creation/editing affected Tests: - 17 new unit tests for stateless/handoff mode (test_stateless_mode.py) - _should_show_resume_line, _chat_session_key, ResumeResolver, ResumeLineProxy - Integration-level: stateless shows resume lines, no auto-resume, chat hides lines - 3 new tests for mode indicator in startup message (test_telegram_backend.py) Docs: - New docs/reference/modes.md — comprehensive reference for all 3 workflow modes - Updated docs/reference/index.md and zensical.toml nav with modes page * docs: comprehensive three-mode coverage across all documentation New: - docs/how-to/choose-a-mode.md — decision tree, mode comparison, mermaid sequence diagrams, configuration examples, switching guide, workspace prerequisites Updated: - README.md — improved three-mode description in features list - docs/tutorials/install.md — added mode selection step (section 10) - docs/tutorials/first-run.md — added 'What mode am I in?' tip - docs/reference/config.md — cross-linked session_mode/show_resume_line to modes.md - docs/reference/transports/telegram.md — added mode requirement callouts for forum topics and chat sessions sections - docs/how-to/chat-sessions.md — added session persistence explanation (state files, auto-resume mechanics, handoff note) - docs/how-to/topics.md — expanded prerequisites checklist with group privacy, can_manage_topics, and re-add steps - docs/how-to/cross-environment-resume.md — added handoff mode terminal workflow with mermaid sequence diagram - docs/how-to/index.md — added 'Getting started' section with choose-a-mode - zensical.toml — added choose-a-mode to nav * docs: add three-mode summary table to README Quick Start section * feat: migrate to dev branch workflow — dev→TestPyPI, master→PyPI Branch model: - feature/* → PR → dev (TestPyPI auto-publish) → PR → master (PyPI) - master always matches latest PyPI release - dev is the integration/staging branch CI changes: - ci.yml: TestPyPI publish triggers on dev push (was master) - ci.yml, codeql.yml: CI runs on both master and dev pushes - dependabot.yml: PRs target dev branch Hook changes: - release-guard.sh: updated messages to mention dev branch - release-guard-mcp.sh: updated messages to mention dev branch - Both hooks already allow dev pushes (only block master/main) Documentation: - CLAUDE.md: updated 3-phase workflow, CI table, release guard docs - dev-workflow.md: added branch model section - release-discipline.md: added dev branch staging notes * ci: retrigger CI for PR #160 * feat: allow Claude Code to merge PRs targeting dev branch only Release guard hooks now check the PR's base branch: - dev → allowed (TestPyPI/staging) - master/main → blocked (PyPI releases remain Nathan-only) Both Bash hook (gh pr merge) and MCP hook (merge_pull_request) updated with base branch checking via gh pr view. --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

* ci: add CODEOWNERS, update action SHA pins, add permission comments - Create .github/CODEOWNERS requiring @littlebearapps/core review - Pin setup-uv to v7.4.0 (6ee6290f), download-artifact to v8.0.1 (3e5f45b2) - Add precise version comments on all action SHAs (codeql v3.32.6, pypi-publish v1.13.0, action-gh-release v2.5.0, fetch-metadata v2.5.0) - Document write permissions with why-comments (OIDC, releases, auto-merge) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * feat: add release guard hooks and document protection in CLAUDE.md Defence-in-depth hooks prevent Claude Code from pushing to master, merging PRs, creating tags, or triggering releases. Feature branch pushes and PR creation remain allowed. - release-guard.sh: Bash hook blocking master push, tags, releases, PR merge - release-guard-protect.sh: Edit/Write hook protecting guard files and hooks.json - release-guard-mcp.sh: GitHub MCP hook blocking merge and master writes - hooks.json: register all three hooks - CLAUDE.md: document release guard, update workflow roles, CI pipeline notes Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix: clarify /config default labels and remove redundant "Works with" lines Default labels now explain what "default" means for each setting: - Diff preview: "default (off)" — matches actual behaviour (was "default (on)") - Model/Reasoning: "default (engine decides)" - API cost: "default (on)", Subscription usage: "default (off)" - Plan mode home hint: "agent decides" - Diff preview home hint: "buttons only" Added info lines to plan mode and reasoning sub-pages explaining the default behaviour in more detail. Removed all 9 "Works with: ..." lines from sub-pages — they're redundant because engine visibility guards already hide settings from unsupported engines. Fixes #119 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix: suppress redundant cost footer on error runs When a run fails (e.g. subscription limit hit), the diagnostic context line from _extract_error() already shows cost, turns, and API time. The 💰 cost footer was duplicating this same data in a different format. Now the cost footer only appears on successful runs where it's the sole source of cost information. Error runs still show cost in the diagnostic line, and budget alerts still fire regardless. Also adds usage field to mock Return dataclass (matching ErrorReturn) so tests can verify cost footer behaviour on success runs. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * feat: suppress stall notifications when CPU-active + heartbeat re-render When cpu_active=True (extended thinking, background agents), suppress Telegram stall warning notifications and instead trigger a heartbeat re-render so the elapsed time counter keeps ticking. Notifications still fire when cpu_active=False or None (no baseline). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * chore: staging 0.34.5rc2 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix: CI release-validation tomllib bytes/str mismatch tomllib.loads() expects str but was receiving bytes from sys.stdin.buffer.read() and open(...,'rb').read(). First triggered when PR #122 changed the version (rc1 → rc2). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * docs: integrate screenshots into docs with correct JPG references - Add 44 screenshots to docs/assets/screenshots/ - Fix all image refs from .png to .jpg across 25 doc files - README uses absolute raw.githubusercontent.com URLs for PyPI rendering - Fix 5 filename mismatches (session-auto-resume→chat-auto-resume, etc.) - Comment out 11 missing screenshots with TODO markers - Add CAPTURES.md checklist tracking capture status Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: convert markdown images to HTML img tags for GitHub compatibility Switch from MkDocs `![alt](src){ loading=lazy }` syntax to HTML `<img>` tags with width="360" and loading="lazy". Fixes two GitHub rendering issues: `{ loading=lazy }` appearing as visible text, and oversized images with no width constraint. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: fix 3 screenshot mismatches and replace 3 screenshots - first-run.md: rewrite resume line text to match footer screenshot - interactive-control.md: update planmode show admonition to match screenshot (auto not on) - switch-engines.md: swap engine-footer.jpg for multi-engine-switch.jpg - Replace startup-message.jpg with clean v0.34.4 capture (was rc/6-projects) - Replace cooldown-auto-deny.jpg with post-outline approve/deny buttons - Replace file-put.jpg with photo save confirmation Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: add iOS caption limitation note to file transfer guide Telegram iOS doesn't show a caption field when sending documents via the File picker, so /file put <path> captions aren't easily accessible. Added a note with workarounds (use Desktop, send as photo, or let auto-save handle it). Updated screenshot alt text to match actual screenshot content. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: temp swap README image URLs to feature branch for preview Will revert to master before merging. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: lay out all 3 README screenshots in a single row Reduce from 360px to 270px each and combine into one block so all three hero screenshots sit side by side. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: swap 3rd hero screenshot for config-menu for visual variety Replace plan-outline-approve (too similar to approval-diff-preview) with config-menu showing the /config settings grid. The three hero images now tell: voice input → approve changes → configure everything. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: add captions under README hero screenshots Small captions: "Send tasks by voice (Whisper transcription)", "Approve changes remotely", "Configure from Telegram". Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: use table layout for README hero screenshots with captions Fixes stacking issue — in a broke inline flow. A table keeps images side by side with captions underneath each one. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: replace table layout with single hero collage image Composite image scales proportionally on mobile instead of requiring horizontal scroll. Captions baked into the image via ImageMagick. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: swap middle hero screenshot for full 3-button approval view Replace approval-diff-preview with approval-buttons-howto showing Approve / Deny / Pause & Outline Plan — more visually impressive. Caption now reads "Approve changes remotely (Claude Code)". Added footnote linking to engine compatibility table. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: swap config-menu for parallel-projects in hero collage Third hero screenshot now shows 10+ projects running simultaneously across different repos — much more compelling than a settings menu. New caption: "Run agents across projects in parallel". Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: revert README image URL to master for merge Swap hero-collage URL back from feature/github-hardening to master. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: staging 0.34.5rc3 - fix: preserve all EngineOverrides fields when setting model/planmode/reasoning (was silently wiping ask_questions, diff_preview, show_api_cost, etc.) - fix: /config home page resolves "default" to effective values - feat: file upload auto-deduplication (append _1, _2 instead of requiring --force) - feat: media groups without captions now auto-save instead of showing usage text - feat: resume line visual separation (blank line + ↩️ prefix) - fix: claude auto-approve echoes updatedInput in control response Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: expand permission policies for Codex CLI and Gemini CLI in /config Codex gets a new "Approval policy" page (full auto / safe) that passes --ask-for-approval untrusted when safe mode is selected. Gemini's approval mode expands from 2 to 3 tiers (read-only / edit files / full access) with --approval-mode auto_edit for the middle tier. Both engines now show an "Agent controls" section on the /config home page. Engine-specific model default hints replace the generic "from CLI settings" text. Also adds staging.sh helper, context-guard-stop hook, and docs updates. Closes #131 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: staging 0.34.5rc4 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: /config UX overhaul, resume line toggle, cost budget settings, model metadata /config UX cleanup: - Convert all binary toggles from 3-column (on/off/clear) to 2-column (toggle + clear) for better mobile tap targets - Merge Engine + Model into combined "Engine & model" page - Reorganise home page to max 2 buttons per row across all engines - Split plan mode 3-option rows (off/on/auto) into 2+1 layout - Add _toggle_row() helper for consistent toggle button rendering New features: - #128: Resume line /config toggle — per-chat show_resume_line override via EngineOverrides with On/Off/Clear buttons, wired into executor - #129: Cost budget /config settings — per-chat budget_enabled and budget_auto_cancel overrides on the Cost & Usage page, wired into _check_cost_budget() in runner_bridge.py Model metadata improvements: - Show Claude Code [1m] context window suffix: "opus 4.6 (1M)" - Strip Gemini CLI "auto-" prefix: "auto-gemini-3" → "gemini-3" - Future-proof: unknown suffixes default to .upper() (e.g. [500k] → 500K) Bug fixes: - #124: Standalone override commands (/planmode, /model, /reasoning) now preserve all EngineOverrides fields including new ones - Error handling: control_response.write_failed catch-all in claude.py, ask_question.extraction_failed warning, model.override.failed logging Hardening: - Plan outline sent as separate ephemeral message (avoids 4096 char truncation) - Added show_resume_line, budget_enabled, budget_auto_cancel to EngineOverrides, EngineRunOptions, normalize/merge, and all constructors Tests: 1610 passed, 80.56% coverage, ruff clean. Integration tested on @untether_dev_bot across all 6 engine chats. Closes #128, closes #129, fixes #124 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: trigger CI for PR #132 * fix: address 11 CodeRabbit review comments on PR #132 Bug fixes: - claude.py: fix UnboundLocalError when factory.resume is falsy in ask_question.extraction_failed logging path - ask_question.py: reject malformed option callbacks instead of silently falling back to option 0 - files.py: raise FileExistsError when deduplicate_target exhausts 999 suffixes instead of returning the original (overwrite risk) - config.py: disambiguate Codex "Full auto" (fa) vs Gemini "Full access" (ya) callback IDs and toast labels Hardening: - codex.py: add --ask-for-approval to _EXEC_ONLY_FLAGS guard - model.py: add try/except to clear path (matching set path) - reasoning.py: add try/except to clear path (matching set path) - loop.py: notify user when media group upload fails instead of silently dropping - export.py: log session count instead of identifiers at info level - config.py: resolve resume-line default from config instead of hardcoding True - staging.sh: pin PyPI index in rollback/reset with --pip-args Skipped (not applicable): - CHANGELOG.md: RC versions don't get changelog entries per release discipline - docs/tutorials TODO screenshot: pre-existing, not introduced by PR - .claude/hooks/context-guard-stop.sh: ContextDocs plugin hook, not Untether source Tests: 1611 passed, 80.48% coverage, ruff clean. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: replace bare pass with debug log to satisfy bandit B110 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: setup wizard security + UX improvements - Auto-set allowed_user_ids from captured Telegram user ID during onboarding (security: restricts bot to the setup user's account) - Add "next steps" panel after wizard completion with pointers to /config, voice notes, projects, and account lock confirmation - Update install.md: Python 3.12+ (not just 3.14), dynamic version string, /config mention for post-setup changes - Update first-run.md: /config → Engine & model for default engine Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: plan outline UX — markdown rendering, buttons, cleanup (#139, #140, #141) - Render outline messages as formatted text via render_markdown() + split_markdown_body() instead of raw markdown (#139) - Add approve/deny buttons to last outline message so users don't have to scroll up past long outlines (#140) - Delete outline messages on approve/deny via module-level _OUTLINE_REGISTRY callable from callback handler; suppress stale keyboard on progress message (#141) - 8 new tests for outline rendering, keyboard placement, and cleanup - Bump version to 0.35.0rc5 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: /continue command — cross-environment resume for all engines (#135) New `/continue` command resumes the most recent CLI session in the project directory from Telegram. Enables starting a session in your terminal and picking it up from your phone. Engine support: Claude (--continue), Codex (resume --last), OpenCode (--continue), Pi (--continue), Gemini (--resume latest). AMP not supported (requires explicit thread ID). Includes ResumeToken.is_continue flag, build_args for all 6 runners, reserved command registration, resume emoji prefix stripping for reply-to-continue, docs (how-to guide, README, commands ref, routing explanation, conversation modes tutorial), and 99 new test assertions. Integration tested against @untether_dev_bot — all 5 supported engines passed secret-recall verification via Telegram MCP. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: agent outbox file delivery + fix cross-chat ask stealing (#143, #144) Outbox delivery (#143): agents write files to .untether-outbox/ during a run; Untether sends them as Telegram documents on completion with 📎 captions. Config: outbox_enabled, outbox_dir, outbox_max_files, outbox_cleanup. Deny-glob security, size limits, auto-cleanup. Preamble updated for all 6 engines. Integration tested across Claude, Codex, OpenCode, Pi, and Gemini. AskUserQuestion fix (#144): _PENDING_ASK_REQUESTS and _ASK_QUESTION_FLOWS were global dicts with no chat_id scoping — a pending ask in one chat would steal the next message from any other chat. Added channel_id contextvar and scoped all ask lookups by it. Session cleanup now also clears stale pending asks. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: v0.35.0 changelog completion + fix #123 updatedInput - Complete v0.35.0 changelog: add missing entries for /continue (#135), /config UX overhaul (#132), resume line toggle (#128), cost budget (#129), model metadata, resume line formatting (#127), override preservation (#124), and updatedInput fix (#123) - Fix #123: register input for system-level auto-approved control requests so updatedInput is included in the response - Add parameterised test for all 5 auto-approve types input registration - Remove unused OutboxResult import (ruff fix) Issues closed: #115, #118, #123, #124, #126, #127, #134 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: staging 0.35.0rc6 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: rc6 integration test fixes (#145, #146, #147, #148, #149) - Reduce Telegram API timeout from 120s to 30s (#145) - OpenCode error runs show error text instead of empty body (#146) - Pi /continue captures session ID via allow_id_promotion (#147) - Post-outline approval uses skip_reply to avoid "not found" (#148) - Orphan progress message cleanup on restart (#149) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: post-outline notification reply + OpenCode empty body (#148, #150) - #148: skip_reply callback results now bypass the executor's default reply_to fallback, sending directly via the transport with no reply_to_message_id. Previously, the executor treated reply_to=None as "use default" which pointed to the (deleted) outline message. - #150: OpenCode normal completion with no Text events now falls back to last_tool_error. Added state.last_tool_error field populated on ToolUse error status. Covers both translate() and stream_end_events(). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: suppress post-outline notification to avoid "message not found" (#148) After outline approval/denial, the progress loop's _send_notify was firing for the next tool approval, but the notification's reply_to anchor could reference deleted state. Added _outline_just_resolved flag to skip one notification cycle after outline cleanup. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: note OpenCode lacks auto-compaction — long sessions degrade (#150) Added known limitation to OpenCode runner docs and integration testing playbook. OpenCode sessions accumulate unbounded context (no compaction events unlike Pi). Workaround: use /new before isolated tests. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: skip_reply on regular approve path when outline was deleted (#148) The "Approve Plan" button on outline messages uses the real ExitPlanMode request_id, routing through the regular approve path (not the da: synthetic path). When outline messages exist, set skip_reply=True on the CommandResult to avoid replying to the just-deleted outline message. Also added reply_to_message_id and text_preview to transport.send.failed warning for easier debugging. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: update changelog for rc6 integration test fixes (#145-#150) Updated fix descriptions for #146/#150 (OpenCode last_tool_error fallback) and #148 (regular approve path skip_reply). Added docs section for OpenCode compaction limitation. Updated test counts. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * style: fix formatting after merge resolution Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: address CodeRabbit review comments on PR #151 - bridge.py: replace text_preview with text_len in send failure warning to avoid logging raw message content (security) - runner_bridge.py: move unregister_progress() after send_result_message() to avoid orphan window between ephemeral cleanup and final message send - cross-environment-resume.md: add language spec to code block Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: resolve /config "default" labels to effective on/off values (#152) Sub-pages showed "Current: default" or "default (on/off)" while buttons already showed the resolved value. Now all boolean-toggle settings show the effective on/off value in both text and buttons. Affected: verbose, ask mode, diff preview, API cost, subscription usage, budget enabled/auto-cancel, resume line. Home page cost & resume labels also resolved. Plan mode, model, and reasoning keep "default" since they depend on CLI settings and aren't simple on/off booleans. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: update changelog for rc7 config default labels fix (#152) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: update documentation for v0.35.0 - fix missing nav entries in zensical.toml (cross-env resume, Gemini/Amp runners) - rewrite inline-settings.md for /config UX overhaul (2-column toggles, budget/resume toggles) - update plan-mode.md with outline rendering, buttons-on-last-chunk, ephemeral cleanup - update interactive-control tutorial with outline UX improvements - add orphan progress cleanup section to operations.md - add engine-specific approval policies to interactive-approval.md - add per-chat budget overrides to cost-budgets.md - update module-map.md with Gemini/Amp and new modules (outbox, progress persistence, proc_diag) - update architecture.md mermaid diagrams with all 6 engines - bump specification.md to v0.35.0, add progress persistence and outbox sections - add v0.35.0 screenshot entries to CAPTURES.md Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: broaden frozen ring buffer stall escalation beyond MCP tools (#155) Frozen ring buffer escalation was gated on `mcp_server is not None`, so general stalls with cpu_active=True and no MCP tool running were silently suppressed indefinitely. Broadened to fire for all stalls after 3+ checks with no new JSONL events regardless of tool type. New notification: "CPU active, no new events" for non-MCP frozen stalls. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: tool approval buttons no longer suppressed after outline approval (#156) After "Approve Plan" on an outline, the stale discuss_approve action remained in ProgressTracker with completed=False. The renderer picked up its stale "Approve Plan"/"Deny" buttons first, then the suppression logic at line 994 stripped ALL buttons — including new Write/Edit/Bash approval buttons. Claude blocked indefinitely waiting for approval. Fix: after suppressing stale buttons, complete the discuss_approve action(s) in the tracker, reset _outline_sent, and trigger a re-render so subsequent tool requests get their own Approve/Deny buttons. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: add workflow mode indicator to startup message, fix startup crash on topics validation failure (#158, #159) Features: - Startup message now shows mode: assistant/workspace/handoff - Derived from session_mode + topics.enabled config values - _resolve_mode_label() helper in backend.py Bug fixes: - Fix UnboundLocalError crash when topics validation fails on startup (#158) - Moved import signal and shutdown imports before try block in loop.py - Downgrade can_manage_topics check from fatal error to warning (#159) - Bot can now start without manage_topics admin right - Existing topics work fine; only topic creation/editing affected Tests: - 17 new unit tests for stateless/handoff mode (test_stateless_mode.py) - _should_show_resume_line, _chat_session_key, ResumeResolver, ResumeLineProxy - Integration-level: stateless shows resume lines, no auto-resume, chat hides lines - 3 new tests for mode indicator in startup message (test_telegram_backend.py) Docs: - New docs/reference/modes.md — comprehensive reference for all 3 workflow modes - Updated docs/reference/index.md and zensical.toml nav with modes page * docs: comprehensive three-mode coverage across all documentation New: - docs/how-to/choose-a-mode.md — decision tree, mode comparison, mermaid sequence diagrams, configuration examples, switching guide, workspace prerequisites Updated: - README.md — improved three-mode description in features list - docs/tutorials/install.md — added mode selection step (section 10) - docs/tutorials/first-run.md — added 'What mode am I in?' tip - docs/reference/config.md — cross-linked session_mode/show_resume_line to modes.md - docs/reference/transports/telegram.md — added mode requirement callouts for forum topics and chat sessions sections - docs/how-to/chat-sessions.md — added session persistence explanation (state files, auto-resume mechanics, handoff note) - docs/how-to/topics.md — expanded prerequisites checklist with group privacy, can_manage_topics, and re-add steps - docs/how-to/cross-environment-resume.md — added handoff mode terminal workflow with mermaid sequence diagram - docs/how-to/index.md — added 'Getting started' section with choose-a-mode - zensical.toml — added choose-a-mode to nav * docs: add three-mode summary table to README Quick Start section * feat: migrate to dev branch workflow — dev→TestPyPI, master→PyPI Branch model: - feature/* → PR → dev (TestPyPI auto-publish) → PR → master (PyPI) - master always matches latest PyPI release - dev is the integration/staging branch CI changes: - ci.yml: TestPyPI publish triggers on dev push (was master) - ci.yml, codeql.yml: CI runs on both master and dev pushes - dependabot.yml: PRs target dev branch Hook changes: - release-guard.sh: updated messages to mention dev branch - release-guard-mcp.sh: updated messages to mention dev branch - Both hooks already allow dev pushes (only block master/main) Documentation: - CLAUDE.md: updated 3-phase workflow, CI table, release guard docs - dev-workflow.md: added branch model section - release-discipline.md: added dev branch staging notes * ci: retrigger CI for PR #160 * feat: allow Claude Code to merge PRs targeting dev branch only Release guard hooks now check the PR's base branch: - dev → allowed (TestPyPI/staging) - master/main → blocked (PyPI releases remain Nathan-only) Both Bash hook (gh pr merge) and MCP hook (merge_pull_request) updated with base branch checking via gh pr view. * docs: add workflow mode indicator and modes.md to CLAUDE.md --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

….0rc8 * ci: add CODEOWNERS, update action SHA pins, add permission comments - Create .github/CODEOWNERS requiring @littlebearapps/core review - Pin setup-uv to v7.4.0 (6ee6290f), download-artifact to v8.0.1 (3e5f45b2) - Add precise version comments on all action SHAs (codeql v3.32.6, pypi-publish v1.13.0, action-gh-release v2.5.0, fetch-metadata v2.5.0) - Document write permissions with why-comments (OIDC, releases, auto-merge) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * feat: add release guard hooks and document protection in CLAUDE.md Defence-in-depth hooks prevent Claude Code from pushing to master, merging PRs, creating tags, or triggering releases. Feature branch pushes and PR creation remain allowed. - release-guard.sh: Bash hook blocking master push, tags, releases, PR merge - release-guard-protect.sh: Edit/Write hook protecting guard files and hooks.json - release-guard-mcp.sh: GitHub MCP hook blocking merge and master writes - hooks.json: register all three hooks - CLAUDE.md: document release guard, update workflow roles, CI pipeline notes Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix: clarify /config default labels and remove redundant "Works with" lines Default labels now explain what "default" means for each setting: - Diff preview: "default (off)" — matches actual behaviour (was "default (on)") - Model/Reasoning: "default (engine decides)" - API cost: "default (on)", Subscription usage: "default (off)" - Plan mode home hint: "agent decides" - Diff preview home hint: "buttons only" Added info lines to plan mode and reasoning sub-pages explaining the default behaviour in more detail. Removed all 9 "Works with: ..." lines from sub-pages — they're redundant because engine visibility guards already hide settings from unsupported engines. Fixes #119 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix: suppress redundant cost footer on error runs When a run fails (e.g. subscription limit hit), the diagnostic context line from _extract_error() already shows cost, turns, and API time. The 💰 cost footer was duplicating this same data in a different format. Now the cost footer only appears on successful runs where it's the sole source of cost information. Error runs still show cost in the diagnostic line, and budget alerts still fire regardless. Also adds usage field to mock Return dataclass (matching ErrorReturn) so tests can verify cost footer behaviour on success runs. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * feat: suppress stall notifications when CPU-active + heartbeat re-render When cpu_active=True (extended thinking, background agents), suppress Telegram stall warning notifications and instead trigger a heartbeat re-render so the elapsed time counter keeps ticking. Notifications still fire when cpu_active=False or None (no baseline). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * chore: staging 0.34.5rc2 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix: CI release-validation tomllib bytes/str mismatch tomllib.loads() expects str but was receiving bytes from sys.stdin.buffer.read() and open(...,'rb').read(). First triggered when PR #122 changed the version (rc1 → rc2). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * docs: integrate screenshots into docs with correct JPG references - Add 44 screenshots to docs/assets/screenshots/ - Fix all image refs from .png to .jpg across 25 doc files - README uses absolute raw.githubusercontent.com URLs for PyPI rendering - Fix 5 filename mismatches (session-auto-resume→chat-auto-resume, etc.) - Comment out 11 missing screenshots with TODO markers - Add CAPTURES.md checklist tracking capture status Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: convert markdown images to HTML img tags for GitHub compatibility Switch from MkDocs `![alt](src){ loading=lazy }` syntax to HTML `<img>` tags with width="360" and loading="lazy". Fixes two GitHub rendering issues: `{ loading=lazy }` appearing as visible text, and oversized images with no width constraint. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: fix 3 screenshot mismatches and replace 3 screenshots - first-run.md: rewrite resume line text to match footer screenshot - interactive-control.md: update planmode show admonition to match screenshot (auto not on) - switch-engines.md: swap engine-footer.jpg for multi-engine-switch.jpg - Replace startup-message.jpg with clean v0.34.4 capture (was rc/6-projects) - Replace cooldown-auto-deny.jpg with post-outline approve/deny buttons - Replace file-put.jpg with photo save confirmation Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: add iOS caption limitation note to file transfer guide Telegram iOS doesn't show a caption field when sending documents via the File picker, so /file put <path> captions aren't easily accessible. Added a note with workarounds (use Desktop, send as photo, or let auto-save handle it). Updated screenshot alt text to match actual screenshot content. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: temp swap README image URLs to feature branch for preview Will revert to master before merging. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: lay out all 3 README screenshots in a single row Reduce from 360px to 270px each and combine into one block so all three hero screenshots sit side by side. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: swap 3rd hero screenshot for config-menu for visual variety Replace plan-outline-approve (too similar to approval-diff-preview) with config-menu showing the /config settings grid. The three hero images now tell: voice input → approve changes → configure everything. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: add captions under README hero screenshots Small captions: "Send tasks by voice (Whisper transcription)", "Approve changes remotely", "Configure from Telegram". Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: use table layout for README hero screenshots with captions Fixes stacking issue — in a broke inline flow. A table keeps images side by side with captions underneath each one. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: replace table layout with single hero collage image Composite image scales proportionally on mobile instead of requiring horizontal scroll. Captions baked into the image via ImageMagick. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: swap middle hero screenshot for full 3-button approval view Replace approval-diff-preview with approval-buttons-howto showing Approve / Deny / Pause & Outline Plan — more visually impressive. Caption now reads "Approve changes remotely (Claude Code)". Added footnote linking to engine compatibility table. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: swap config-menu for parallel-projects in hero collage Third hero screenshot now shows 10+ projects running simultaneously across different repos — much more compelling than a settings menu. New caption: "Run agents across projects in parallel". Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: revert README image URL to master for merge Swap hero-collage URL back from feature/github-hardening to master. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: staging 0.34.5rc3 - fix: preserve all EngineOverrides fields when setting model/planmode/reasoning (was silently wiping ask_questions, diff_preview, show_api_cost, etc.) - fix: /config home page resolves "default" to effective values - feat: file upload auto-deduplication (append _1, _2 instead of requiring --force) - feat: media groups without captions now auto-save instead of showing usage text - feat: resume line visual separation (blank line + ↩️ prefix) - fix: claude auto-approve echoes updatedInput in control response Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: expand permission policies for Codex CLI and Gemini CLI in /config Codex gets a new "Approval policy" page (full auto / safe) that passes --ask-for-approval untrusted when safe mode is selected. Gemini's approval mode expands from 2 to 3 tiers (read-only / edit files / full access) with --approval-mode auto_edit for the middle tier. Both engines now show an "Agent controls" section on the /config home page. Engine-specific model default hints replace the generic "from CLI settings" text. Also adds staging.sh helper, context-guard-stop hook, and docs updates. Closes #131 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: staging 0.34.5rc4 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: /config UX overhaul, resume line toggle, cost budget settings, model metadata /config UX cleanup: - Convert all binary toggles from 3-column (on/off/clear) to 2-column (toggle + clear) for better mobile tap targets - Merge Engine + Model into combined "Engine & model" page - Reorganise home page to max 2 buttons per row across all engines - Split plan mode 3-option rows (off/on/auto) into 2+1 layout - Add _toggle_row() helper for consistent toggle button rendering New features: - #128: Resume line /config toggle — per-chat show_resume_line override via EngineOverrides with On/Off/Clear buttons, wired into executor - #129: Cost budget /config settings — per-chat budget_enabled and budget_auto_cancel overrides on the Cost & Usage page, wired into _check_cost_budget() in runner_bridge.py Model metadata improvements: - Show Claude Code [1m] context window suffix: "opus 4.6 (1M)" - Strip Gemini CLI "auto-" prefix: "auto-gemini-3" → "gemini-3" - Future-proof: unknown suffixes default to .upper() (e.g. [500k] → 500K) Bug fixes: - #124: Standalone override commands (/planmode, /model, /reasoning) now preserve all EngineOverrides fields including new ones - Error handling: control_response.write_failed catch-all in claude.py, ask_question.extraction_failed warning, model.override.failed logging Hardening: - Plan outline sent as separate ephemeral message (avoids 4096 char truncation) - Added show_resume_line, budget_enabled, budget_auto_cancel to EngineOverrides, EngineRunOptions, normalize/merge, and all constructors Tests: 1610 passed, 80.56% coverage, ruff clean. Integration tested on @untether_dev_bot across all 6 engine chats. Closes #128, closes #129, fixes #124 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: trigger CI for PR #132 * fix: address 11 CodeRabbit review comments on PR #132 Bug fixes: - claude.py: fix UnboundLocalError when factory.resume is falsy in ask_question.extraction_failed logging path - ask_question.py: reject malformed option callbacks instead of silently falling back to option 0 - files.py: raise FileExistsError when deduplicate_target exhausts 999 suffixes instead of returning the original (overwrite risk) - config.py: disambiguate Codex "Full auto" (fa) vs Gemini "Full access" (ya) callback IDs and toast labels Hardening: - codex.py: add --ask-for-approval to _EXEC_ONLY_FLAGS guard - model.py: add try/except to clear path (matching set path) - reasoning.py: add try/except to clear path (matching set path) - loop.py: notify user when media group upload fails instead of silently dropping - export.py: log session count instead of identifiers at info level - config.py: resolve resume-line default from config instead of hardcoding True - staging.sh: pin PyPI index in rollback/reset with --pip-args Skipped (not applicable): - CHANGELOG.md: RC versions don't get changelog entries per release discipline - docs/tutorials TODO screenshot: pre-existing, not introduced by PR - .claude/hooks/context-guard-stop.sh: ContextDocs plugin hook, not Untether source Tests: 1611 passed, 80.48% coverage, ruff clean. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: replace bare pass with debug log to satisfy bandit B110 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: setup wizard security + UX improvements - Auto-set allowed_user_ids from captured Telegram user ID during onboarding (security: restricts bot to the setup user's account) - Add "next steps" panel after wizard completion with pointers to /config, voice notes, projects, and account lock confirmation - Update install.md: Python 3.12+ (not just 3.14), dynamic version string, /config mention for post-setup changes - Update first-run.md: /config → Engine & model for default engine Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: plan outline UX — markdown rendering, buttons, cleanup (#139, #140, #141) - Render outline messages as formatted text via render_markdown() + split_markdown_body() instead of raw markdown (#139) - Add approve/deny buttons to last outline message so users don't have to scroll up past long outlines (#140) - Delete outline messages on approve/deny via module-level _OUTLINE_REGISTRY callable from callback handler; suppress stale keyboard on progress message (#141) - 8 new tests for outline rendering, keyboard placement, and cleanup - Bump version to 0.35.0rc5 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: /continue command — cross-environment resume for all engines (#135) New `/continue` command resumes the most recent CLI session in the project directory from Telegram. Enables starting a session in your terminal and picking it up from your phone. Engine support: Claude (--continue), Codex (resume --last), OpenCode (--continue), Pi (--continue), Gemini (--resume latest). AMP not supported (requires explicit thread ID). Includes ResumeToken.is_continue flag, build_args for all 6 runners, reserved command registration, resume emoji prefix stripping for reply-to-continue, docs (how-to guide, README, commands ref, routing explanation, conversation modes tutorial), and 99 new test assertions. Integration tested against @untether_dev_bot — all 5 supported engines passed secret-recall verification via Telegram MCP. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: agent outbox file delivery + fix cross-chat ask stealing (#143, #144) Outbox delivery (#143): agents write files to .untether-outbox/ during a run; Untether sends them as Telegram documents on completion with 📎 captions. Config: outbox_enabled, outbox_dir, outbox_max_files, outbox_cleanup. Deny-glob security, size limits, auto-cleanup. Preamble updated for all 6 engines. Integration tested across Claude, Codex, OpenCode, Pi, and Gemini. AskUserQuestion fix (#144): _PENDING_ASK_REQUESTS and _ASK_QUESTION_FLOWS were global dicts with no chat_id scoping — a pending ask in one chat would steal the next message from any other chat. Added channel_id contextvar and scoped all ask lookups by it. Session cleanup now also clears stale pending asks. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: v0.35.0 changelog completion + fix #123 updatedInput - Complete v0.35.0 changelog: add missing entries for /continue (#135), /config UX overhaul (#132), resume line toggle (#128), cost budget (#129), model metadata, resume line formatting (#127), override preservation (#124), and updatedInput fix (#123) - Fix #123: register input for system-level auto-approved control requests so updatedInput is included in the response - Add parameterised test for all 5 auto-approve types input registration - Remove unused OutboxResult import (ruff fix) Issues closed: #115, #118, #123, #124, #126, #127, #134 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: staging 0.35.0rc6 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: rc6 integration test fixes (#145, #146, #147, #148, #149) - Reduce Telegram API timeout from 120s to 30s (#145) - OpenCode error runs show error text instead of empty body (#146) - Pi /continue captures session ID via allow_id_promotion (#147) - Post-outline approval uses skip_reply to avoid "not found" (#148) - Orphan progress message cleanup on restart (#149) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: post-outline notification reply + OpenCode empty body (#148, #150) - #148: skip_reply callback results now bypass the executor's default reply_to fallback, sending directly via the transport with no reply_to_message_id. Previously, the executor treated reply_to=None as "use default" which pointed to the (deleted) outline message. - #150: OpenCode normal completion with no Text events now falls back to last_tool_error. Added state.last_tool_error field populated on ToolUse error status. Covers both translate() and stream_end_events(). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: suppress post-outline notification to avoid "message not found" (#148) After outline approval/denial, the progress loop's _send_notify was firing for the next tool approval, but the notification's reply_to anchor could reference deleted state. Added _outline_just_resolved flag to skip one notification cycle after outline cleanup. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: note OpenCode lacks auto-compaction — long sessions degrade (#150) Added known limitation to OpenCode runner docs and integration testing playbook. OpenCode sessions accumulate unbounded context (no compaction events unlike Pi). Workaround: use /new before isolated tests. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: skip_reply on regular approve path when outline was deleted (#148) The "Approve Plan" button on outline messages uses the real ExitPlanMode request_id, routing through the regular approve path (not the da: synthetic path). When outline messages exist, set skip_reply=True on the CommandResult to avoid replying to the just-deleted outline message. Also added reply_to_message_id and text_preview to transport.send.failed warning for easier debugging. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: update changelog for rc6 integration test fixes (#145-#150) Updated fix descriptions for #146/#150 (OpenCode last_tool_error fallback) and #148 (regular approve path skip_reply). Added docs section for OpenCode compaction limitation. Updated test counts. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * style: fix formatting after merge resolution Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: address CodeRabbit review comments on PR #151 - bridge.py: replace text_preview with text_len in send failure warning to avoid logging raw message content (security) - runner_bridge.py: move unregister_progress() after send_result_message() to avoid orphan window between ephemeral cleanup and final message send - cross-environment-resume.md: add language spec to code block Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: resolve /config "default" labels to effective on/off values (#152) Sub-pages showed "Current: default" or "default (on/off)" while buttons already showed the resolved value. Now all boolean-toggle settings show the effective on/off value in both text and buttons. Affected: verbose, ask mode, diff preview, API cost, subscription usage, budget enabled/auto-cancel, resume line. Home page cost & resume labels also resolved. Plan mode, model, and reasoning keep "default" since they depend on CLI settings and aren't simple on/off booleans. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: update changelog for rc7 config default labels fix (#152) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: update documentation for v0.35.0 - fix missing nav entries in zensical.toml (cross-env resume, Gemini/Amp runners) - rewrite inline-settings.md for /config UX overhaul (2-column toggles, budget/resume toggles) - update plan-mode.md with outline rendering, buttons-on-last-chunk, ephemeral cleanup - update interactive-control tutorial with outline UX improvements - add orphan progress cleanup section to operations.md - add engine-specific approval policies to interactive-approval.md - add per-chat budget overrides to cost-budgets.md - update module-map.md with Gemini/Amp and new modules (outbox, progress persistence, proc_diag) - update architecture.md mermaid diagrams with all 6 engines - bump specification.md to v0.35.0, add progress persistence and outbox sections - add v0.35.0 screenshot entries to CAPTURES.md Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: broaden frozen ring buffer stall escalation beyond MCP tools (#155) Frozen ring buffer escalation was gated on `mcp_server is not None`, so general stalls with cpu_active=True and no MCP tool running were silently suppressed indefinitely. Broadened to fire for all stalls after 3+ checks with no new JSONL events regardless of tool type. New notification: "CPU active, no new events" for non-MCP frozen stalls. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: tool approval buttons no longer suppressed after outline approval (#156) After "Approve Plan" on an outline, the stale discuss_approve action remained in ProgressTracker with completed=False. The renderer picked up its stale "Approve Plan"/"Deny" buttons first, then the suppression logic at line 994 stripped ALL buttons — including new Write/Edit/Bash approval buttons. Claude blocked indefinitely waiting for approval. Fix: after suppressing stale buttons, complete the discuss_approve action(s) in the tracker, reset _outline_sent, and trigger a re-render so subsequent tool requests get their own Approve/Deny buttons. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: add workflow mode indicator to startup message, fix startup crash on topics validation failure (#158, #159) Features: - Startup message now shows mode: assistant/workspace/handoff - Derived from session_mode + topics.enabled config values - _resolve_mode_label() helper in backend.py Bug fixes: - Fix UnboundLocalError crash when topics validation fails on startup (#158) - Moved import signal and shutdown imports before try block in loop.py - Downgrade can_manage_topics check from fatal error to warning (#159) - Bot can now start without manage_topics admin right - Existing topics work fine; only topic creation/editing affected Tests: - 17 new unit tests for stateless/handoff mode (test_stateless_mode.py) - _should_show_resume_line, _chat_session_key, ResumeResolver, ResumeLineProxy - Integration-level: stateless shows resume lines, no auto-resume, chat hides lines - 3 new tests for mode indicator in startup message (test_telegram_backend.py) Docs: - New docs/reference/modes.md — comprehensive reference for all 3 workflow modes - Updated docs/reference/index.md and zensical.toml nav with modes page * docs: comprehensive three-mode coverage across all documentation New: - docs/how-to/choose-a-mode.md — decision tree, mode comparison, mermaid sequence diagrams, configuration examples, switching guide, workspace prerequisites Updated: - README.md — improved three-mode description in features list - docs/tutorials/install.md — added mode selection step (section 10) - docs/tutorials/first-run.md — added 'What mode am I in?' tip - docs/reference/config.md — cross-linked session_mode/show_resume_line to modes.md - docs/reference/transports/telegram.md — added mode requirement callouts for forum topics and chat sessions sections - docs/how-to/chat-sessions.md — added session persistence explanation (state files, auto-resume mechanics, handoff note) - docs/how-to/topics.md — expanded prerequisites checklist with group privacy, can_manage_topics, and re-add steps - docs/how-to/cross-environment-resume.md — added handoff mode terminal workflow with mermaid sequence diagram - docs/how-to/index.md — added 'Getting started' section with choose-a-mode - zensical.toml — added choose-a-mode to nav * docs: add three-mode summary table to README Quick Start section * feat: migrate to dev branch workflow — dev→TestPyPI, master→PyPI Branch model: - feature/* → PR → dev (TestPyPI auto-publish) → PR → master (PyPI) - master always matches latest PyPI release - dev is the integration/staging branch CI changes: - ci.yml: TestPyPI publish triggers on dev push (was master) - ci.yml, codeql.yml: CI runs on both master and dev pushes - dependabot.yml: PRs target dev branch Hook changes: - release-guard.sh: updated messages to mention dev branch - release-guard-mcp.sh: updated messages to mention dev branch - Both hooks already allow dev pushes (only block master/main) Documentation: - CLAUDE.md: updated 3-phase workflow, CI table, release guard docs - dev-workflow.md: added branch model section - release-discipline.md: added dev branch staging notes * ci: retrigger CI for PR #160 * feat: allow Claude Code to merge PRs targeting dev branch only Release guard hooks now check the PR's base branch: - dev → allowed (TestPyPI/staging) - master/main → blocked (PyPI releases remain Nathan-only) Both Bash hook (gh pr merge) and MCP hook (merge_pull_request) updated with base branch checking via gh pr view. * docs: add workflow mode indicator and modes.md to CLAUDE.md * fix: dual outline buttons (#163), entity URL sanitisation (#157), changelog migration - Strip approval buttons from progress message when outline is visible — only outline message shows Approve/Deny/Cancel (#163) - Reset outline state via source_has_approval tracking so future ExitPlanMode requests work correctly (#163) - Sanitise text_link entities with invalid URLs (localhost, loopback, file paths, bare hostnames) by converting to code entities — prevents silent 400 errors that drop the entire final message (#157) - Merge v0.34.5 changelog into v0.35.0 — v0.34.5 was never released (latest PyPI is v0.34.4), all rc1-rc7 work is v0.35.0 17 new tests (2 for #163, 15 for #157). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: staging 0.35.0rc8 fix: restore frozen ring buffer stall escalation (#155) The #163 fix (6f43e5b) accidentally removed all frozen ring buffer code from runner_bridge.py. Restored from 8fcad32: - _frozen_ring_count tracking and ring buffer snapshot comparison - frozen_escalate gating (fires notification after 3+ frozen checks despite cpu_active=True) - _has_running_mcp_tool() for MCP server name extraction - _STALL_THRESHOLD_MCP_TOOL (15 min, configurable via watchdog) - MCP-aware notification text ("MCP tool may be hung", "CPU active, no new events", "MCP tool running") - 8 new tests + 2 updated existing tests - mcp_tool_timeout watchdog setting docs: integration testing S1 MCP threshold, tutorials index, glossary, outbox screenshot, CAPTURES checklist Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: CI lint — unused import in test, bandit nosec for loopback blocklist - Remove unused ActionEvent import in test_has_running_mcp_tool_returns_server_name - Add # nosec B104 to _LOOPBACK_HOSTS — it's a URL blocklist, not a bind address Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

* ci: add CODEOWNERS, update action SHA pins, add permission comments - Create .github/CODEOWNERS requiring @littlebearapps/core review - Pin setup-uv to v7.4.0 (6ee6290f), download-artifact to v8.0.1 (3e5f45b2) - Add precise version comments on all action SHAs (codeql v3.32.6, pypi-publish v1.13.0, action-gh-release v2.5.0, fetch-metadata v2.5.0) - Document write permissions with why-comments (OIDC, releases, auto-merge) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * feat: add release guard hooks and document protection in CLAUDE.md Defence-in-depth hooks prevent Claude Code from pushing to master, merging PRs, creating tags, or triggering releases. Feature branch pushes and PR creation remain allowed. - release-guard.sh: Bash hook blocking master push, tags, releases, PR merge - release-guard-protect.sh: Edit/Write hook protecting guard files and hooks.json - release-guard-mcp.sh: GitHub MCP hook blocking merge and master writes - hooks.json: register all three hooks - CLAUDE.md: document release guard, update workflow roles, CI pipeline notes Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix: clarify /config default labels and remove redundant "Works with" lines Default labels now explain what "default" means for each setting: - Diff preview: "default (off)" — matches actual behaviour (was "default (on)") - Model/Reasoning: "default (engine decides)" - API cost: "default (on)", Subscription usage: "default (off)" - Plan mode home hint: "agent decides" - Diff preview home hint: "buttons only" Added info lines to plan mode and reasoning sub-pages explaining the default behaviour in more detail. Removed all 9 "Works with: ..." lines from sub-pages — they're redundant because engine visibility guards already hide settings from unsupported engines. Fixes #119 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix: suppress redundant cost footer on error runs When a run fails (e.g. subscription limit hit), the diagnostic context line from _extract_error() already shows cost, turns, and API time. The 💰 cost footer was duplicating this same data in a different format. Now the cost footer only appears on successful runs where it's the sole source of cost information. Error runs still show cost in the diagnostic line, and budget alerts still fire regardless. Also adds usage field to mock Return dataclass (matching ErrorReturn) so tests can verify cost footer behaviour on success runs. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * feat: suppress stall notifications when CPU-active + heartbeat re-render When cpu_active=True (extended thinking, background agents), suppress Telegram stall warning notifications and instead trigger a heartbeat re-render so the elapsed time counter keeps ticking. Notifications still fire when cpu_active=False or None (no baseline). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * chore: staging 0.34.5rc2 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix: CI release-validation tomllib bytes/str mismatch tomllib.loads() expects str but was receiving bytes from sys.stdin.buffer.read() and open(...,'rb').read(). First triggered when PR #122 changed the version (rc1 → rc2). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * docs: integrate screenshots into docs with correct JPG references - Add 44 screenshots to docs/assets/screenshots/ - Fix all image refs from .png to .jpg across 25 doc files - README uses absolute raw.githubusercontent.com URLs for PyPI rendering - Fix 5 filename mismatches (session-auto-resume→chat-auto-resume, etc.) - Comment out 11 missing screenshots with TODO markers - Add CAPTURES.md checklist tracking capture status Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: convert markdown images to HTML img tags for GitHub compatibility Switch from MkDocs `![alt](src){ loading=lazy }` syntax to HTML `<img>` tags with width="360" and loading="lazy". Fixes two GitHub rendering issues: `{ loading=lazy }` appearing as visible text, and oversized images with no width constraint. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: fix 3 screenshot mismatches and replace 3 screenshots - first-run.md: rewrite resume line text to match footer screenshot - interactive-control.md: update planmode show admonition to match screenshot (auto not on) - switch-engines.md: swap engine-footer.jpg for multi-engine-switch.jpg - Replace startup-message.jpg with clean v0.34.4 capture (was rc/6-projects) - Replace cooldown-auto-deny.jpg with post-outline approve/deny buttons - Replace file-put.jpg with photo save confirmation Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: add iOS caption limitation note to file transfer guide Telegram iOS doesn't show a caption field when sending documents via the File picker, so /file put <path> captions aren't easily accessible. Added a note with workarounds (use Desktop, send as photo, or let auto-save handle it). Updated screenshot alt text to match actual screenshot content. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: temp swap README image URLs to feature branch for preview Will revert to master before merging. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: lay out all 3 README screenshots in a single row Reduce from 360px to 270px each and combine into one block so all three hero screenshots sit side by side. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: swap 3rd hero screenshot for config-menu for visual variety Replace plan-outline-approve (too similar to approval-diff-preview) with config-menu showing the /config settings grid. The three hero images now tell: voice input → approve changes → configure everything. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: add captions under README hero screenshots Small captions: "Send tasks by voice (Whisper transcription)", "Approve changes remotely", "Configure from Telegram". Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: use table layout for README hero screenshots with captions Fixes stacking issue — in a broke inline flow. A table keeps images side by side with captions underneath each one. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: replace table layout with single hero collage image Composite image scales proportionally on mobile instead of requiring horizontal scroll. Captions baked into the image via ImageMagick. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: swap middle hero screenshot for full 3-button approval view Replace approval-diff-preview with approval-buttons-howto showing Approve / Deny / Pause & Outline Plan — more visually impressive. Caption now reads "Approve changes remotely (Claude Code)". Added footnote linking to engine compatibility table. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: swap config-menu for parallel-projects in hero collage Third hero screenshot now shows 10+ projects running simultaneously across different repos — much more compelling than a settings menu. New caption: "Run agents across projects in parallel". Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: revert README image URL to master for merge Swap hero-collage URL back from feature/github-hardening to master. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: staging 0.34.5rc3 - fix: preserve all EngineOverrides fields when setting model/planmode/reasoning (was silently wiping ask_questions, diff_preview, show_api_cost, etc.) - fix: /config home page resolves "default" to effective values - feat: file upload auto-deduplication (append _1, _2 instead of requiring --force) - feat: media groups without captions now auto-save instead of showing usage text - feat: resume line visual separation (blank line + ↩️ prefix) - fix: claude auto-approve echoes updatedInput in control response Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: expand permission policies for Codex CLI and Gemini CLI in /config Codex gets a new "Approval policy" page (full auto / safe) that passes --ask-for-approval untrusted when safe mode is selected. Gemini's approval mode expands from 2 to 3 tiers (read-only / edit files / full access) with --approval-mode auto_edit for the middle tier. Both engines now show an "Agent controls" section on the /config home page. Engine-specific model default hints replace the generic "from CLI settings" text. Also adds staging.sh helper, context-guard-stop hook, and docs updates. Closes #131 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: staging 0.34.5rc4 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: /config UX overhaul, resume line toggle, cost budget settings, model metadata /config UX cleanup: - Convert all binary toggles from 3-column (on/off/clear) to 2-column (toggle + clear) for better mobile tap targets - Merge Engine + Model into combined "Engine & model" page - Reorganise home page to max 2 buttons per row across all engines - Split plan mode 3-option rows (off/on/auto) into 2+1 layout - Add _toggle_row() helper for consistent toggle button rendering New features: - #128: Resume line /config toggle — per-chat show_resume_line override via EngineOverrides with On/Off/Clear buttons, wired into executor - #129: Cost budget /config settings — per-chat budget_enabled and budget_auto_cancel overrides on the Cost & Usage page, wired into _check_cost_budget() in runner_bridge.py Model metadata improvements: - Show Claude Code [1m] context window suffix: "opus 4.6 (1M)" - Strip Gemini CLI "auto-" prefix: "auto-gemini-3" → "gemini-3" - Future-proof: unknown suffixes default to .upper() (e.g. [500k] → 500K) Bug fixes: - #124: Standalone override commands (/planmode, /model, /reasoning) now preserve all EngineOverrides fields including new ones - Error handling: control_response.write_failed catch-all in claude.py, ask_question.extraction_failed warning, model.override.failed logging Hardening: - Plan outline sent as separate ephemeral message (avoids 4096 char truncation) - Added show_resume_line, budget_enabled, budget_auto_cancel to EngineOverrides, EngineRunOptions, normalize/merge, and all constructors Tests: 1610 passed, 80.56% coverage, ruff clean. Integration tested on @untether_dev_bot across all 6 engine chats. Closes #128, closes #129, fixes #124 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: trigger CI for PR #132 * fix: address 11 CodeRabbit review comments on PR #132 Bug fixes: - claude.py: fix UnboundLocalError when factory.resume is falsy in ask_question.extraction_failed logging path - ask_question.py: reject malformed option callbacks instead of silently falling back to option 0 - files.py: raise FileExistsError when deduplicate_target exhausts 999 suffixes instead of returning the original (overwrite risk) - config.py: disambiguate Codex "Full auto" (fa) vs Gemini "Full access" (ya) callback IDs and toast labels Hardening: - codex.py: add --ask-for-approval to _EXEC_ONLY_FLAGS guard - model.py: add try/except to clear path (matching set path) - reasoning.py: add try/except to clear path (matching set path) - loop.py: notify user when media group upload fails instead of silently dropping - export.py: log session count instead of identifiers at info level - config.py: resolve resume-line default from config instead of hardcoding True - staging.sh: pin PyPI index in rollback/reset with --pip-args Skipped (not applicable): - CHANGELOG.md: RC versions don't get changelog entries per release discipline - docs/tutorials TODO screenshot: pre-existing, not introduced by PR - .claude/hooks/context-guard-stop.sh: ContextDocs plugin hook, not Untether source Tests: 1611 passed, 80.48% coverage, ruff clean. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: replace bare pass with debug log to satisfy bandit B110 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: setup wizard security + UX improvements - Auto-set allowed_user_ids from captured Telegram user ID during onboarding (security: restricts bot to the setup user's account) - Add "next steps" panel after wizard completion with pointers to /config, voice notes, projects, and account lock confirmation - Update install.md: Python 3.12+ (not just 3.14), dynamic version string, /config mention for post-setup changes - Update first-run.md: /config → Engine & model for default engine Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: plan outline UX — markdown rendering, buttons, cleanup (#139, #140, #141) - Render outline messages as formatted text via render_markdown() + split_markdown_body() instead of raw markdown (#139) - Add approve/deny buttons to last outline message so users don't have to scroll up past long outlines (#140) - Delete outline messages on approve/deny via module-level _OUTLINE_REGISTRY callable from callback handler; suppress stale keyboard on progress message (#141) - 8 new tests for outline rendering, keyboard placement, and cleanup - Bump version to 0.35.0rc5 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: /continue command — cross-environment resume for all engines (#135) New `/continue` command resumes the most recent CLI session in the project directory from Telegram. Enables starting a session in your terminal and picking it up from your phone. Engine support: Claude (--continue), Codex (resume --last), OpenCode (--continue), Pi (--continue), Gemini (--resume latest). AMP not supported (requires explicit thread ID). Includes ResumeToken.is_continue flag, build_args for all 6 runners, reserved command registration, resume emoji prefix stripping for reply-to-continue, docs (how-to guide, README, commands ref, routing explanation, conversation modes tutorial), and 99 new test assertions. Integration tested against @untether_dev_bot — all 5 supported engines passed secret-recall verification via Telegram MCP. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: agent outbox file delivery + fix cross-chat ask stealing (#143, #144) Outbox delivery (#143): agents write files to .untether-outbox/ during a run; Untether sends them as Telegram documents on completion with 📎 captions. Config: outbox_enabled, outbox_dir, outbox_max_files, outbox_cleanup. Deny-glob security, size limits, auto-cleanup. Preamble updated for all 6 engines. Integration tested across Claude, Codex, OpenCode, Pi, and Gemini. AskUserQuestion fix (#144): _PENDING_ASK_REQUESTS and _ASK_QUESTION_FLOWS were global dicts with no chat_id scoping — a pending ask in one chat would steal the next message from any other chat. Added channel_id contextvar and scoped all ask lookups by it. Session cleanup now also clears stale pending asks. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: v0.35.0 changelog completion + fix #123 updatedInput - Complete v0.35.0 changelog: add missing entries for /continue (#135), /config UX overhaul (#132), resume line toggle (#128), cost budget (#129), model metadata, resume line formatting (#127), override preservation (#124), and updatedInput fix (#123) - Fix #123: register input for system-level auto-approved control requests so updatedInput is included in the response - Add parameterised test for all 5 auto-approve types input registration - Remove unused OutboxResult import (ruff fix) Issues closed: #115, #118, #123, #124, #126, #127, #134 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: staging 0.35.0rc6 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: rc6 integration test fixes (#145, #146, #147, #148, #149) - Reduce Telegram API timeout from 120s to 30s (#145) - OpenCode error runs show error text instead of empty body (#146) - Pi /continue captures session ID via allow_id_promotion (#147) - Post-outline approval uses skip_reply to avoid "not found" (#148) - Orphan progress message cleanup on restart (#149) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: post-outline notification reply + OpenCode empty body (#148, #150) - #148: skip_reply callback results now bypass the executor's default reply_to fallback, sending directly via the transport with no reply_to_message_id. Previously, the executor treated reply_to=None as "use default" which pointed to the (deleted) outline message. - #150: OpenCode normal completion with no Text events now falls back to last_tool_error. Added state.last_tool_error field populated on ToolUse error status. Covers both translate() and stream_end_events(). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: suppress post-outline notification to avoid "message not found" (#148) After outline approval/denial, the progress loop's _send_notify was firing for the next tool approval, but the notification's reply_to anchor could reference deleted state. Added _outline_just_resolved flag to skip one notification cycle after outline cleanup. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: note OpenCode lacks auto-compaction — long sessions degrade (#150) Added known limitation to OpenCode runner docs and integration testing playbook. OpenCode sessions accumulate unbounded context (no compaction events unlike Pi). Workaround: use /new before isolated tests. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: skip_reply on regular approve path when outline was deleted (#148) The "Approve Plan" button on outline messages uses the real ExitPlanMode request_id, routing through the regular approve path (not the da: synthetic path). When outline messages exist, set skip_reply=True on the CommandResult to avoid replying to the just-deleted outline message. Also added reply_to_message_id and text_preview to transport.send.failed warning for easier debugging. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: update changelog for rc6 integration test fixes (#145-#150) Updated fix descriptions for #146/#150 (OpenCode last_tool_error fallback) and #148 (regular approve path skip_reply). Added docs section for OpenCode compaction limitation. Updated test counts. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * style: fix formatting after merge resolution Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: address CodeRabbit review comments on PR #151 - bridge.py: replace text_preview with text_len in send failure warning to avoid logging raw message content (security) - runner_bridge.py: move unregister_progress() after send_result_message() to avoid orphan window between ephemeral cleanup and final message send - cross-environment-resume.md: add language spec to code block Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: resolve /config "default" labels to effective on/off values (#152) Sub-pages showed "Current: default" or "default (on/off)" while buttons already showed the resolved value. Now all boolean-toggle settings show the effective on/off value in both text and buttons. Affected: verbose, ask mode, diff preview, API cost, subscription usage, budget enabled/auto-cancel, resume line. Home page cost & resume labels also resolved. Plan mode, model, and reasoning keep "default" since they depend on CLI settings and aren't simple on/off booleans. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: update changelog for rc7 config default labels fix (#152) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: update documentation for v0.35.0 - fix missing nav entries in zensical.toml (cross-env resume, Gemini/Amp runners) - rewrite inline-settings.md for /config UX overhaul (2-column toggles, budget/resume toggles) - update plan-mode.md with outline rendering, buttons-on-last-chunk, ephemeral cleanup - update interactive-control tutorial with outline UX improvements - add orphan progress cleanup section to operations.md - add engine-specific approval policies to interactive-approval.md - add per-chat budget overrides to cost-budgets.md - update module-map.md with Gemini/Amp and new modules (outbox, progress persistence, proc_diag) - update architecture.md mermaid diagrams with all 6 engines - bump specification.md to v0.35.0, add progress persistence and outbox sections - add v0.35.0 screenshot entries to CAPTURES.md Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: broaden frozen ring buffer stall escalation beyond MCP tools (#155) Frozen ring buffer escalation was gated on `mcp_server is not None`, so general stalls with cpu_active=True and no MCP tool running were silently suppressed indefinitely. Broadened to fire for all stalls after 3+ checks with no new JSONL events regardless of tool type. New notification: "CPU active, no new events" for non-MCP frozen stalls. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: tool approval buttons no longer suppressed after outline approval (#156) After "Approve Plan" on an outline, the stale discuss_approve action remained in ProgressTracker with completed=False. The renderer picked up its stale "Approve Plan"/"Deny" buttons first, then the suppression logic at line 994 stripped ALL buttons — including new Write/Edit/Bash approval buttons. Claude blocked indefinitely waiting for approval. Fix: after suppressing stale buttons, complete the discuss_approve action(s) in the tracker, reset _outline_sent, and trigger a re-render so subsequent tool requests get their own Approve/Deny buttons. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: add workflow mode indicator to startup message, fix startup crash on topics validation failure (#158, #159) Features: - Startup message now shows mode: assistant/workspace/handoff - Derived from session_mode + topics.enabled config values - _resolve_mode_label() helper in backend.py Bug fixes: - Fix UnboundLocalError crash when topics validation fails on startup (#158) - Moved import signal and shutdown imports before try block in loop.py - Downgrade can_manage_topics check from fatal error to warning (#159) - Bot can now start without manage_topics admin right - Existing topics work fine; only topic creation/editing affected Tests: - 17 new unit tests for stateless/handoff mode (test_stateless_mode.py) - _should_show_resume_line, _chat_session_key, ResumeResolver, ResumeLineProxy - Integration-level: stateless shows resume lines, no auto-resume, chat hides lines - 3 new tests for mode indicator in startup message (test_telegram_backend.py) Docs: - New docs/reference/modes.md — comprehensive reference for all 3 workflow modes - Updated docs/reference/index.md and zensical.toml nav with modes page * docs: comprehensive three-mode coverage across all documentation New: - docs/how-to/choose-a-mode.md — decision tree, mode comparison, mermaid sequence diagrams, configuration examples, switching guide, workspace prerequisites Updated: - README.md — improved three-mode description in features list - docs/tutorials/install.md — added mode selection step (section 10) - docs/tutorials/first-run.md — added 'What mode am I in?' tip - docs/reference/config.md — cross-linked session_mode/show_resume_line to modes.md - docs/reference/transports/telegram.md — added mode requirement callouts for forum topics and chat sessions sections - docs/how-to/chat-sessions.md — added session persistence explanation (state files, auto-resume mechanics, handoff note) - docs/how-to/topics.md — expanded prerequisites checklist with group privacy, can_manage_topics, and re-add steps - docs/how-to/cross-environment-resume.md — added handoff mode terminal workflow with mermaid sequence diagram - docs/how-to/index.md — added 'Getting started' section with choose-a-mode - zensical.toml — added choose-a-mode to nav * docs: add three-mode summary table to README Quick Start section * feat: migrate to dev branch workflow — dev→TestPyPI, master→PyPI Branch model: - feature/* → PR → dev (TestPyPI auto-publish) → PR → master (PyPI) - master always matches latest PyPI release - dev is the integration/staging branch CI changes: - ci.yml: TestPyPI publish triggers on dev push (was master) - ci.yml, codeql.yml: CI runs on both master and dev pushes - dependabot.yml: PRs target dev branch Hook changes: - release-guard.sh: updated messages to mention dev branch - release-guard-mcp.sh: updated messages to mention dev branch - Both hooks already allow dev pushes (only block master/main) Documentation: - CLAUDE.md: updated 3-phase workflow, CI table, release guard docs - dev-workflow.md: added branch model section - release-discipline.md: added dev branch staging notes * ci: retrigger CI for PR #160 * feat: allow Claude Code to merge PRs targeting dev branch only Release guard hooks now check the PR's base branch: - dev → allowed (TestPyPI/staging) - master/main → blocked (PyPI releases remain Nathan-only) Both Bash hook (gh pr merge) and MCP hook (merge_pull_request) updated with base branch checking via gh pr view. * docs: add workflow mode indicator and modes.md to CLAUDE.md * fix: dual outline buttons (#163), entity URL sanitisation (#157), changelog migration - Strip approval buttons from progress message when outline is visible — only outline message shows Approve/Deny/Cancel (#163) - Reset outline state via source_has_approval tracking so future ExitPlanMode requests work correctly (#163) - Sanitise text_link entities with invalid URLs (localhost, loopback, file paths, bare hostnames) by converting to code entities — prevents silent 400 errors that drop the entire final message (#157) - Merge v0.34.5 changelog into v0.35.0 — v0.34.5 was never released (latest PyPI is v0.34.4), all rc1-rc7 work is v0.35.0 17 new tests (2 for #163, 15 for #157). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: staging 0.35.0rc8 fix: restore frozen ring buffer stall escalation (#155) The #163 fix (6f43e5b) accidentally removed all frozen ring buffer code from runner_bridge.py. Restored from 8fcad32: - _frozen_ring_count tracking and ring buffer snapshot comparison - frozen_escalate gating (fires notification after 3+ frozen checks despite cpu_active=True) - _has_running_mcp_tool() for MCP server name extraction - _STALL_THRESHOLD_MCP_TOOL (15 min, configurable via watchdog) - MCP-aware notification text ("MCP tool may be hung", "CPU active, no new events", "MCP tool running") - 8 new tests + 2 updated existing tests - mcp_tool_timeout watchdog setting docs: integration testing S1 MCP threshold, tutorials index, glossary, outbox screenshot, CAPTURES checklist Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: CI lint — unused import in test, bandit nosec for loopback blocklist - Remove unused ActionEvent import in test_has_running_mcp_tool_returns_server_name - Add # nosec B104 to _LOOPBACK_HOSTS — it's a URL blocklist, not a bind address Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: update CLAUDE.md test counts for v0.35.0rc8 Total: 1578 → 1743 tests Per-file: test_exec_bridge 109→112, test_claude_control 82→89, test_callback_dispatch 25→26, test_ask_user_question 25→29, test_meta_line 43→54, test_preamble 5→6, test_config_command 195→218, test_build_args 33→39 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

…ll (#183, #184, #167, #168) * fix: prevent Codex/OpenCode headless hangs (#184, #183) Codex (#184): always pass --ask-for-approval in headless mode. Default to "never" (auto-approve all) so Codex never blocks on terminal input. Safe permission mode still uses "untrusted". OpenCode (#183): surface unsupported JSONL event types as visible Telegram warnings instead of silently dropping them. When msgspec DecodeError occurs, _extract_event_type() tries to parse the raw JSON for the type field. If extractable, a warning ActionEvent is emitted (visible in Telegram) instead of returning []. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: auto-continue for Claude bug #34142, sleeping-process stall fix (#167, #168) Auto-continue (#167): detect when Claude Code exits after receiving tool results without processing them (last_event_type=user) and auto-resume the session. Configurable via [auto_continue] with enabled (default true) and max_retries (default 1). Sleeping-process stall (#168): CPU-active suppression now checks process_state; when main process is sleeping (state=S) but children are CPU-active (hung Bash tool), notifications fire. Stall message shows tool name ("Bash tool may be stuck") instead of generic text. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: CI lint — explicit super() for @DataClass(slots=True) compat Zero-argument super() breaks in @DataClass(slots=True) on Python <3.14 because the __class__ cell references the pre-slot class. Use explicit JsonlSubprocessRunner.decode_error_events(self, ...) instead. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: resolve 9 new ty warnings — typed test helpers, isinstance narrowing - TestShouldAutoContinue._call: replace mixed-type dict with typed keyword args to satisfy ty's union narrowing - TestDecodeErrorEvents: add isinstance(ActionEvent) checks before accessing .message and .action attributes on union type Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * ci: make ty check informational (continue-on-error) ty has 55 pre-existing warnings across the codebase. These are not regressions — the same warnings exist on dev and master. Making ty non-blocking so it doesn't prevent PR merges while still reporting warnings for visibility. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- test count: 1743 → 1765 - test_exec_bridge: 112 → 124 (auto-continue, sleeping-process stall) - test_build_args: 39 → 40 (Codex default approval) - add auto-continue feature to features list - note sleeping-process awareness in stall diagnostics - clarify ty is informational (continue-on-error) in CI table Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…189) * feat: emoji buttons + edit-in-place for outline approval (#186) Add emoji prefixes to ExitPlanMode and post-outline buttons (✅/❌/📋). Post-outline approve/deny now edits the "Asked Claude Code to outline the plan" message in-place instead of creating a second message. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: redesign startup message layout (#187) Split engine info into separate lines, add italic subheadings, rename "projects" to "directories", add bug report link. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: add 🧹 emoji to /new session clear messages Part of startup message UX improvements (#187). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: make stall warnings succinct and accurate for long-running tools (#188) Truncate Last: to 80 chars, recognise command: prefix for Bash tools, use reassuring "still running" when CPU active, drop PID diagnostics from Telegram messages, only say "may be stuck" when genuinely stuck. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: frozen ring escalation shows reassuring message for long Bash commands (#188) When a known tool is running (main sleeping, CPU active on children), frozen ring escalation now shows "Bash command still running" instead of alarming "No progress" message. Found via wpnav staging session where benchmark scripts ran for 60+ min with false warnings. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Security audit fixes (4 HIGH severity): - sanitise bot token in log URLs (#190) - cap JSONL line buffer at 10MB to prevent OOM (#191) - fix tag name injection in notify-website CI workflow (#193) - add -- separator before user prompts in gemini/amp runners (#194) Also includes: tool-active stall repeat suppression, CLAUDE.md doc updates, configurable watchdog timeouts, and 4 new tests. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat: add "Let's discuss" button to post-outline plan approval Add a 💬 Let's discuss button to the post-outline plan approval buttons (between Approve Plan/Deny and cancel). When clicked, it tells Claude Code to ask the user what they'd like to discuss about the plan before deciding to approve or deny. Implementation: - New `chat` action in claude_control.py with `_CHAT_DENY_MESSAGE` - Handles both da: prefix (synthetic) and hold-open (real request_id) paths - Clears cooldown and outline_pending state on both paths - Early toast: "Let's discuss..." - Post-outline keyboard now has 2 rows: [Approve Plan | Deny], [Let's discuss] Tests: 5 new tests, 5 updated for new button layout (1773 pass, 81% coverage) Docs: updated 9 files across how-to, tutorial, reference, and rules Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: remove unused _CHAT_DENY_MESSAGE import in test Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…tion (#220, #221) - Read OpenCode's ~/.config/opencode/opencode.json at runner construction to populate the model footer even when no untether.toml override is set. Previously the footer showed no model for the default config. (#221) - Update _ENGINE_MODEL_HINTS for opencode to show "provider/model (e.g. openai/gpt-4o)" instead of the unhelpful "from provider config", guiding users to use the required provider-prefixed format. (#220) - Gate /planmode to Claude-only; gate /usage to subscription-supported engines; add _resolve_engine helper for command-level engine checks. - Deduplicate repeated StartedEvent headers in /export markdown output for resumed sessions. - Fix Gemini CLI prompt injection: use --prompt=<value> instead of -p <value> to prevent prompts starting with - being parsed as flags. - Ensure Codex runner always includes model in meta dict. - Add 8 tests for _read_opencode_default_model and build_runner fallback, plus engine gate and export dedup tests. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: /new command now cancels running processes before clearing sessions (#222) Previously /new only cleared stored resume tokens, leaving old Claude/Codex/OpenCode processes running (~400 MB each). This leaked processes and worsened memory pressure, contributing to earlyoom kills (rc=143). Now /new cancels all running tasks for the chat before clearing sessions. Reply message shows "cancelled run and cleared" when a process was killed. - Add _cancel_chat_tasks() helper to topics.py - Add running_tasks param to both /new handlers - Pass running_tasks through all 3 call sites in loop.py - Add running_tasks field to TelegramCommandContext - 10 new tests covering cancellation behaviour - Update 7 docs to reflect new /new behaviour - Bump version to 0.35.0rc12 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: remove unused AsyncMock import (pre-existing ruff F401) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…222) (#226) When earlyoom killed Claude sessions (rc=143/SIGTERM), auto-continue detected last_event_type=user and immediately respawned all 4 killed sessions (~5 GB of new processes) into the same memory pressure. This caused a death spiral where sessions were killed and respawned repeatedly. Fix: _should_auto_continue now checks proc_returncode — signal deaths (rc>128 or rc<0) are excluded. The upstream bug #34142/#30333 exits with rc=0, so auto-continue still works for its intended purpose. - Add _is_signal_death() helper to runner_bridge.py - Add proc_returncode field to JsonlStreamState - Store returncode on stream state after process exit - Pass proc_returncode through to _should_auto_continue - 12 new tests for signal death detection and auto-continue gating Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- Auto-continue feature description: add signal death suppression note - runner_bridge.py key file: add auto-continue mention - test_exec_bridge.py: 128 → 140 tests (signal death tests) - runner-development.md: add stream state tracking section (proc_returncode) - telegram-transport.md: add /new cancellation section Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…229) (#230) * fix: improve error_during_execution hint for session archival (#228) Update error hint text from "corrupted during a restart" to "archived or expired" — better reflects the actual cause when Claude Code auto-archives a session between resume runs. Related: anthropics/claude-code#39178 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: prevent duplicate control response for already-handled requests (#229) When a user clicks Approve/Deny on a control request via Telegram, send_claude_control_response() marks it in _HANDLED_REQUESTS but can't access state.pending_control_requests. The 5-minute expiry check then sends a duplicate DENY for the same request_id, causing Claude Code to receive conflicting approve+deny responses and stall. Add reconciliation in translate() that checks _HANDLED_REQUESTS against pending_control_requests before the expiry loop: - Removes already-handled requests from pending (prevents spurious deny) - Emits action_completed to clear stale inline keyboards - Adds belt-and-suspenders guard on the expiry list comprehension The upstream Claude Code freeze is tracked in anthropics/claude-code#39666. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: Gemini JSONL parsing, ask mode toggle, diff preview buttons, doc chat IDs (#231, #232, #233, #238) - Strip non-JSON prefixes from JSONL stdout lines in decode_jsonl() — fixes Gemini CLI "MCP issues detected" warning corrupting the first event (#231) - Change ask mode toggle default from False to True to match display default, fixing inverted button state in /config (#232) - Only strip approval buttons from progress when current action is a DiscussApproval (outline flow), not for regular tool approvals — fixes diff preview buttons disappearing after plan outline (#233) - Update integration test chat IDs from stale ut-dev: to ut-dev-hf: (#238) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: backfill changelog entries, update docs and rules for v0.35.0 - Backfill changelog entries for fixes #32, #33, #59, #60, #62, #115, #134, #152, #166 and changes #36, #38 - Update contrib/untether.service KillMode from process to mixed (#166) - Update CLAUDE.md test counts (1818 tests) - Update docs (config, operations, troubleshooting, first-run) - Sync .claude/rules with current conventions Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: release v0.35.0 - Bump version from 0.35.0rc14 to 0.35.0 - Set changelog date to 2026-03-29 - Remove non-standard ### ci changelog section (internal repo infra) - Sync lockfile Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: bump requests 2.33.0 (CVE-2026-25645), ignore unfixed pygments CVE - Bump requests 2.32.5 -> 2.33.0 to fix CVE-2026-25645 - Ignore CVE-2026-4539 in pip-audit (pygments 2.19.2, no fix available) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- Fix _PENDING_ASK_REQUESTS type: dict[str, tuple[int, str]] not dict[str, str] - Fix auto-approve docs: _AUTO_APPROVE_TYPES + _TOOLS_REQUIRING_APPROVAL (not the non-existent _AUTO_APPROVE_TOOLS) - Update screenshot count 47 -> 48 - Add progress persistence to features list Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

`/new` was only handled in `_dispatch_builtin_command()` when `topics.enabled=true`. With topics disabled (assistant mode), the command fell through to prompt dispatch and triggered an engine run. - Move `/new` out of the `topics.enabled` gate, mirroring `/ctx` pattern: topic → `handle_new_command`, chat session → `handle_chat_new_command`, stateless → cancel + reply - Add `chat_session_store` and `chat_session_key` to `TelegramCommandContext` dataclass - Remove unreachable early routing for `/new` at lines 1871-1910 (now handled by `_dispatch_builtin_command`) Verified: all 6 engine forum topic chats return "no stored sessions" instead of triggering runs. Project supergroup chats unaffected. Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

) * feat: v0.35.0 UX polish — error formatting, resume line, help links, error hints expansion (#244, #245, #246) Error messages: hints shown above raw error in code blocks, 67 error patterns (was 32) covering model, context, safety, auth, CLI, SSL, AMP/Gemini-specific. Resume line moved below cost/subscription footer for cleaner visual hierarchy. Startup message and /config menu now include help guide and bug report links. README restructured with consolidated Help Guides section. AMP -x flag fix. New docs/reference/errors.md central error reference with cross-links from all 6 engine guides and troubleshooting doc. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: startup message help/bug links point to README Help Guides section Updated URLs from old littlebearapps.com docs site to the restructured README.md anchors (#-help-guides, #-contributing) — now consistent with the /config menu links added earlier. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: ruff format fix in backend.py Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: Gemini runner defaults to yolo approval mode in headless mode (#244, #248) Gemini CLI's default (read-only) mode disables write tools entirely, causing 8-18 min stalls as the agent cascades through sub-agents. Default to --approval-mode yolo since headless mode has no interactive approval path, matching the existing Codex pattern. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat: logging audit — fill gaps in structlog coverage (#254) - Elevate settings loader failures from DEBUG to WARNING (footer, watchdog, auto-continue, preamble) so config regressions are visible in production logs and to the issue watcher - Add access control logging (message.dropped, callback.dropped) in parsing.py for unrecognised chat IDs - Add handle.engine_resolved info log in executor.py after successful runner + CWD resolution - Elevate outline cleanup failures from DEBUG to WARNING - Add credential redaction for OpenAI (sk-...) and GitHub (ghp_, ghs_, gho_, github_pat_) API key patterns in logging.py - Add file_transfer.saved and file_transfer.sent info logs - Bind session_id in structlog context vars when StartedEvent arrives - Add media_group.flush.ok, cost_budget.check, cancel.ambiguous, cancel.nothing_running debug logs - Update troubleshooting docs with key log events table and redaction note Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: expand ruff lint rules from 7 to 18, auto-fix imports (#255) Add ASYNC, LOG, I (isort), PT, RET, RUF (full), FURB, PIE, FLY, FA, ISC rule sets to ruff configuration. Tailored for Untether's async/structlog/pytest-heavy codebase. Auto-fixed: - 42 import sorts across ~35 files via isort (I) - 73 stale noqa directives cleaned by RUF100 - 3 useless if-else conditions simplified in config.py (RUF034) - 9 unused unpacked variables prefixed with _ (RUF059) - 1 endswith call merged to tuple in render.py (PIE810) - __all__ sorted in api.py (RUF022) Per-file ignores for tests: ASYNC109/110/251, PT006/012, RUF059, S110 Global ignores: FLY002, RET504/505, RUF001/005/009, PT018 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: staging 0.35.0rc16 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Upstream Claude Code bug — Bash tool children use their own session group, unreachable by Untether's process group cleanup. No TTY means no SIGHUP cascade in headless mode. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- #245: AMP CLI -x flag regression (double-dash separator broke execute mode) - #246: expanded error hints coverage (model, auth, safety, CLI, SSL categories) - #248: add issue reference alongside #244 for Gemini yolo default fix Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- Bump version from 0.35.0rc16 to 0.35.0 - Set changelog date to 2026-03-31 - Sync lockfile Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- CLAUDE.md: fix test counts (1818→1820, claude_control 94→67, exec_runner 28→22, build_args 40→42), add context-quality.md rule - architecture.md mermaid: fix Gemini CLI args (-p <prompt> → --prompt=) - gemini/runner.md: update invocation to --prompt=<value>, document --approval-mode yolo default Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…time reset (#265) * fix: reduce stall warning false positives during Agent subagent work (#264) - Add tree CPU tracking (pid + descendants) to proc_diag for accurate child process activity detection - Move diagnostic collection to every 60s cycle (ensures CPU baseline exists before first stall warning) - Add child-aware stall threshold (15 min) when child processes or elevated TCP detected - Suppress repeat stall warnings when tree CPU is active - Persist CPU baseline across stall recovery - Track total stall warnings across recovery (session.summary fix) - Improved notification messages for child process work - Add configurable subagent_timeout to WatchdogSettings Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: v0.35.1 security hardening + /ping uptime reset (#192, #193, #194, #234) Security: - Validate callback query sender against allowed_user_ids in group chats; reject unauthorised button presses with "Not authorised" toast (#192) - Escape release tag name in notify-website CI workflow using jq for proper JSON encoding instead of direct interpolation (#193) - Add sanitize_prompt() to base runner class; apply to Gemini and AMP runners to prevent flag injection from prompts starting with - (#194) Bug fix: - /ping uptime resets on service restart via reset_uptime() called from the Telegram loop startup (#234) Changelog prepared for v0.35.1 (unreleased). Closed #190 and #191 as already mitigated. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: add PyPI monthly downloads badge to README Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: validate sender on cancel callback in group chats (#192) The cancel button callback was routed directly to handle_callback_cancel() bypassing the sender validation added in _dispatch_callback(). Any group member could cancel another user's running task. Add the same allowed_user_ids check to the cancel handler. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: update changelog for cancel button sender validation (#192) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: bump version to 0.35.1 Set changelog date and sync lockfile for release preparation. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Fixes pip-audit CI failure from CVE-2026-34513 through CVE-2026-34525. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Add optional timezone field to cron triggers so schedules can be evaluated in a specific timezone rather than the server's system time. - Per-cron `timezone` field (IANA names, e.g. "Australia/Melbourne") - Global `default_timezone` in [triggers] — per-cron overrides it - DST-aware via stdlib zoneinfo (zero new dependencies) - Invalid timezone names rejected at config parse time - Backward compatible — omitting both preserves current behaviour - 12 new tests (6 settings validation + 6 cron timezone matching) - Docs updated: triggers reference, config reference, how-to guides, README Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- CLAUDE.md: add timezone-aware cron triggers to features, add triggers/cron.py and triggers/settings.py to key files, update test count 1820→1856, add trigger test file entries - testing-conventions.md: add trigger cron/settings test files to key test files table - untether-architecture SKILL.md: add trigger system section covering cron scheduler, timezone support, webhooks, and dispatch flow Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…-fetch, bug fixes (#272, #276-#281) * feat: add max effort level for Claude Code, show resolved defaults in /config - Add "max" to Claude Code reasoning levels (CLI supports --effort max) - Show actual default effort level from ~/.claude/settings.json (e.g. "default (high)") instead of generic "from CLI settings" text - Remove broken Settings guide and Troubleshooting links from /config footer, keep Help guides and Report a bug - Update docs: runner.md, model-reasoning.md, inline-settings.md Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: improve toggle button UX and add engine-specific reasoning labels - Toggle buttons now show both [On] [Off] options with ✓ on active, instead of confusing single-button that showed current state - Cost page multi-toggles use compact mode with labels to distinguish - Reasoning/effort/thinking label adapts to engine terminology: Claude → "Effort", Codex → "Reasoning", Pi → "Thinking" - Home page button and hint row use engine-specific label Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: update inline-settings guide for toggle UX and engine-specific labels - Update toggle pattern docs: single-button → two-button [On] [Off] selection - Update reasoning labels: Claude→"Effort", Codex→"Reasoning" - Remove broken Settings guide/Troubleshooting links from example - Update cost page toggle pattern description Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: add SSRF protection utility for trigger outbound requests (#276) Add shared SSRF blocker at src/untether/triggers/ssrf.py for use by upcoming webhook http_forward, external payload fetch, and cron data-fetch features. Blocks private/reserved IP ranges (loopback, RFC 1918, link-local, CGN, multicast, IPv6 equivalents, IPv4-mapped IPv6), validates URL schemes, performs DNS resolution with IP checks to prevent DNS rebinding attacks. Includes configurable allowlist, timeout/size clamping, and SSRFError exception. 73 tests covering IPv4/IPv6 blocking, URL validation, DNS resolution mocking, async validation, allowlist overrides, and edge cases. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: add SSRF protection to changelog, triggers reference, and security guide (#276) Update CHANGELOG.md with v0.35.1 SSRF entry, add ssrf.py to triggers key files table, add SSRF protection section to security how-to guide, add ssrf.py to CLAUDE.md key files. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: non-agent webhook actions — file_write, http_forward, notify_only (#277) Add `action` field to webhook config with three new action types that skip the agent run pipeline: - `file_write` — write POST body to a file path with template substitution, atomic writes, path traversal protection, deny-glob enforcement, on-conflict handling (overwrite/append_timestamp/error) - `http_forward` — forward payload to another URL with SSRF protection (#276), header template rendering with control char validation, exponential backoff on 5xx (max 3 retries), configurable method - `notify_only` — send templated Telegram message, no agent run All actions go through existing auth, rate limiting, and event filter. Optional `notify_on_success`/`notify_on_failure` flags for Telegram visibility. Default action is `agent_run` (fully backward compatible). Also adds `render_template_fields()` to templating module for non-prompt template rendering (file paths, URLs, messages) without untrusted prefix. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: add non-agent webhook actions to changelog and triggers reference (#277) Update CHANGELOG.md with v0.35.1 non-agent actions entry, add action fields to webhook config table, add non-agent actions section with examples (file_write, http_forward, notify_only), update key files. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * test: add tests for non-agent webhook actions and SSRF (#276, #277) Add test_trigger_actions.py (34 tests) covering file_write (path traversal, deny globs, size limits, conflict modes, template paths, atomic writes), http_forward (SSRF blocking, 4xx no-retry, header injection, template URLs), and notify_only (template rendering, no untrusted prefix). Extend test_trigger_settings.py (+12 tests) with action validation: required fields per action type, backward compat for existing configs. Extend test_trigger_templating.py (+4 tests) for render_template_fields (no untrusted prefix variant). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: multipart webhook uploads and cron data-fetch triggers (#278, #279) Multipart support (#278): - Webhooks accept multipart/form-data when accept_multipart=true - File parts saved with sanitised filenames, atomic writes, deny-glob and path traversal protection, configurable size limits (default 50MB) - Form fields available as template variables alongside file metadata - New config fields: accept_multipart, file_destination, max_file_size_bytes Cron data-fetch (#279): - New CronFetchConfig model with type (http_get, http_post, file_read) - HTTP fetches use SSRF protection (#276), configurable timeout, parse_as (json/text/lines), on_failure (abort/run_with_error) - File reads use path traversal protection, deny-globs, symlink rejection - Fetched data injected into prompt via store_as template variable - CronConfig now supports prompt_template alongside static prompt - Fetch failures send Telegram notification when on_failure=abort - New fetch.py module with execute_fetch() and build_fetch_prompt() Tests: 31 new tests (test_trigger_fetch.py + settings extensions) All 2012 tests pass. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: add multipart uploads and data-fetch crons to docs (#278, #279) Update CHANGELOG.md with v0.35.1 entries for multipart and data-fetch. Update triggers reference with CronFetchConfig table, multipart section, data-fetch crons section, fetch.py in key files. Update webhooks-and-cron how-to with data-fetch and non-agent action examples. Update CLAUDE.md with new trigger module files. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: staging 0.35.1rc2 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: multipart body consumption and rate limiter ineffectiveness (#280, #281) Two blocker bugs surfaced during v0.35.1rc2 integration testing: #280: multipart webhooks returned HTTP 500 with "Could not find starting boundary". Root cause — `_process_webhook` pre-reads the full request body for size/auth/rate-limit checks, which exhausts the aiohttp stream before `_parse_multipart` calls `request.multipart()`. Fix: reconstruct a `MultipartReader` from the cached raw body via a fresh `StreamReader`. Also added a short-circuit in `execute_file_write` so the raw MIME envelope isn't written to `file_path` in addition to the extracted file part at `file_destination`. #281: rate limiter never returned 429 even under heavy bursts. Root cause — `_process_webhook` awaited `dispatcher.dispatch_webhook` / `dispatch_action` before returning 202, which capped HTTP handler throughput at the Telegram outbox rate (~1/sec for private chats). The token bucket refilled faster than it drained. Fix: dispatch via `asyncio.create_task` with exception logging, tracked in a module-level set so the GC can't drop in-flight tasks. With `rate_limit = 60`, an 80-request burst now yields 60 × 202 + 20 × 429. Tests: +8 regression tests covering multipart success paths (single file, form-field + file, size limit, unsafe filename, auth rejection), rate limit under burst, and the fire-and-forget HTTP 202 contract. Full suite still passes (2020 tests, +1 skipped). The old `test_internal_error_returns_500` test was adjusted to match the new fire-and-forget semantics — dispatch errors log rather than surface as HTTP 500. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: note fire-and-forget dispatch and multipart short-circuit (#280, #281) - Update flow diagram to show HTTP 202 returns before dispatch completes. - Clarify rate_limit docs: bursts are limited at ingress, not downstream outbox. - Document that multipart + action=file_write saves only to file_destination. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: use tempfile.gettempdir() instead of hardcoded /tmp (bandit B108) The multipart upload fallback path (when file_destination is not configured) previously wrote to /tmp/untether-uploads/, which bandit flagged at Medium severity (predictable location, CWE-377) on the #278 feature branch — the CI check never went green for that PR because bandit was failing on this line. Switching to Path(tempfile.gettempdir()) / "untether-uploads" is portable and silences the warning. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- CLAUDE.md: add trigger module key files (server.py, dispatcher.py, rate_limit.py, auth.py), update test count 1856→2020, add 6 new trigger test files to key test files list - architecture.md: add actions.py, fetch.py, ssrf.py to Triggers layer diagram and edges (server→actions, cron→fetch, actions→ssrf), update summary table - testing-conventions.md: add trigger server, actions, ssrf, fetch test files to key test files table Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

When diff_preview was enabled and the user approved ExitPlanMode, every subsequent Edit/Write/Bash tool still required manual approval — defeating the plan mode UX of "review once, execute freely". Now tracks plan exit approval per-session and auto-approves diff_preview tools after the plan is approved. Also bumps to v0.35.1rc3 for staging. Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Bumps [pypa/gh-action-pypi-publish](https://github.com/pypa/gh-action-pypi-publish) from 1.13.0 to 1.14.0. - [Release notes](https://github.com/pypa/gh-action-pypi-publish/releases) - [Commits](pypa/gh-action-pypi-publish@ed0c539...cef2210) --- updated-dependencies: - dependency-name: pypa/gh-action-pypi-publish dependency-version: 1.14.0 dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

Bumps [softprops/action-gh-release](https://github.com/softprops/action-gh-release) from 2.5.0 to 3.0.0. - [Release notes](https://github.com/softprops/action-gh-release/releases) - [Changelog](https://github.com/softprops/action-gh-release/blob/master/CHANGELOG.md) - [Commits](softprops/action-gh-release@a06a81a...b430933) --- updated-dependencies: - dependency-name: softprops/action-gh-release dependency-version: 3.0.0 dependency-type: direct:production update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

Implements hot-reload for trigger configuration (crons and webhooks) via new TriggerManager class. Editing untether.toml applies changes immediately without restarting Untether or killing active runs. - New TriggerManager class in triggers/manager.py holds cron/webhook config - Cron scheduler reads manager.crons each tick; last_fired preserved across swaps - Webhook server resolves routes per-request via manager.webhook_for_path() - Config watcher handle_reload() re-reads raw TOML [triggers] section - 13 new tests in test_trigger_manager.py; 2038 existing tests still pass Closes #269. Co-authored-by: Nathan Schram <nathan@littlebearapps.com>

…t Tier 1 (#271, #286, #287, #288) * fix: stop Untether being the preferred OOM victim (#275) systemd user services inherit OOMScoreAdjust=200 + OOMPolicy=stop defaults, which made Untether's engine subprocesses preferred earlyoom/kernel OOM killer targets ahead of CLI claude (oom_score_adj=0) and orphaned grandchildren actually consuming the RAM. When lba-1 ran low on RAM, live Telegram chats died with rc=143 (SIGTERM) while the processes actually eating the RAM survived. Updates contrib/untether.service with: - OOMScoreAdjust=-100 — documents intent; kernel clamps to the parent baseline for unprivileged users (typically 100), but takes effect if the parent user@UID.service is ever overridden lower - OOMPolicy=continue — a single OOM-killed child no longer tears down the whole unit cgroup; previously every live chat died at once Also updates docs/reference/dev-instance.md with a new OOM section covering the asymmetry, the clamping caveat, and the optional sudo systemctl edit user@UID.service override for operators who want Untether's children to live longer than CLI processes. Existing installs need to copy the unit file and `systemctl --user daemon-reload`; staging picks up the change on the next `scripts/staging.sh install` cycle. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: v0.35.1rc4 — /at command, hot-reload bridge config, trigger visibility, restart Tier 1 (#271, #286, #287, #288) Bundles four rc4 features plus a CHANGELOG entry for #283 (diff_preview gate, already on dev as 8c04904). Full details in CHANGELOG.md. #269/#285 hot-reload triggers: merged separately as PR #285 (squash-merged to dev); this commit extends TriggerManager with rc4 helpers (remove_cron, crons_for_chat, webhooks_for_chat, cron_ids, webhook_ids) for Features 4b and 5 below. #288 — /at command and run_once cron flag: - new telegram/at_scheduler.py — module-level task-group + run_job holder; schedule_delayed_run(), cancel_pending_for_chat(), active_count(); per-chat cap of 20 pending delays - new telegram/commands/at.py — AtCommand backend, /at <duration> <prompt> with Ns/Nm/Nh suffixes, 60s-24h range - /cancel integration via cancel_pending_for_chat() - drain integration via active_count() in _drain_and_exit - entry-point at = untether.telegram.commands.at:BACKEND - CronConfig.run_once: bool = False; scheduler removes cron after fire if run_once=True; re-enters on reload/restart #286 — unfreeze TelegramBridgeConfig: - drop frozen=True (slots preserved); add update_from(settings) method - route_update() reads cfg.allowed_user_ids live; handle_reload() calls update_from() and refreshes state.forward_coalesce_s / media_group_debounce_s - restart-only keys still warn (bot_token, chat_id, session_mode, topics, message_overflow); others hot-reload #271 — trigger visibility Tier 1: - new triggers/describe.py — describe_cron(schedule, timezone) utility - /ping shows per-chat trigger indicator when triggers target the chat - RunContext.trigger_source field; dispatcher sets it to cron:<id>/webhook:<id>; runner_bridge seeds progress_tracker.meta['trigger'] with icon + source; ProgressTracker.note_event merges engine meta over dispatcher meta - format_meta_line() appends 'trigger' to footer parts - CommandContext gains trigger_manager, default_chat_id fields (default None); populated by telegram/commands/dispatch.py from cfg #287 — graceful restart Tier 1: - new sdnotify.py — stdlib sd_notify client (READY=1 / STOPPING=1); poll_updates sends READY=1 after _send_startup succeeds; _drain_and_exit sends STOPPING=1 at drain start - new telegram/offset_persistence.py — DebouncedOffsetWriter; loads saved update_id on startup, persists via on_offset_advanced callback in poll_incoming; flushes in poll_updates finally block - contrib/untether.service: Type=notify, NotifyAccess=main, RestartSec=2 Tests: +224 tests added across 6 new test files and 6 extended files; 2164 total tests pass with 81.55% coverage. Context files (CLAUDE.md, .claude/rules/*) and human docs (README, triggers reference, dev-instance, integration-testing, webhooks-and-cron how-to, commands-and-directives) updated. rc4 integration test scenarios R1-R10 added to integration-testing.md. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: prevent /at timers from firing after /cancel (CancelScope race) anyio.CancelScope.__exit__ swallows the Cancelled exception when the scope itself caused the cancellation. The fire/dispatch code outside the scope continued regardless. Added cancelled_caught check after the scope exits to prevent stale timers from dispatching. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: add v0.35.1rc4 integration test plan 52-test plan covering all rc4 features: /at command, run_once, hot-reload (triggers + bridge config), trigger visibility, graceful restart Tier 1, plus standard Tier 1/6/7 regression. Includes correct dev bot chat IDs (Bot API + Telethon MCP mapping), pre-test trigger config, results template, and known caveats. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: update integration test chat IDs to current dev bot The old ut-dev-hf: chat IDs (5171122044 etc.) belong to a different bot (ID 8485467124). Updated both docs to the current @untether_dev_bot chats with both Telethon and Bot API ID forms. Added note about Telegram MCP PeerUser fallback for channel/supergroup IDs. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: comprehensive v0.35.1 documentation updates HIGH priority: - config.md: add run_once to cron table, fix watch_config description to list hot-reloadable vs restart-only settings - operations.md: fix hot-reload section (transport settings ARE now partially hot-reloadable), add /ping trigger format, update_id persistence, systemd section with Type=notify/OOM notes - schedule-tasks.md: add /at command section with examples, run_once mention MEDIUM priority: - triggers.md: remove duplicate hot-reload section, keep authoritative version with watch_config requirement and last_fired note - CLAUDE.md: add diff_preview plan bypass (#283) to features list - troubleshooting.md: add entries for config hot-reload issues and /at delay not firing LOW priority: - security.md: document untrusted-payload prefix for webhooks/cron - voice-notes.md: note that voice settings hot-reload - specification.md: bump version to v0.35.1 - tutorials: update version numbers from 0.35.0 to 0.35.1 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: add 38 missing structlog calls across 13 files (logging audit) Comprehensive logging audit found gaps in security-critical paths (auth, rate limiting, SSRF), runner lifecycle (codex peer parity), state mutations (topic_state), and CLI error paths. Adds structured log statements at appropriate levels without over-logging. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: comprehensive v0.35.1 documentation updates - Expand docs/reference/changelog.md with full v0.35.1 entry (security, fixes, changes) instead of a stub pointing to GitHub - Add #190 (token redaction) and #191 (line buffer cap) to CHANGELOG.md - Add logging audit (#299) to CHANGELOG.md and docs changelog - Update CLAUDE.md test count from 2038 to 2165 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * chore: staging 0.35.1rc5 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: bump pytest 9.0.2 → 9.0.3 (CVE-2025-71176) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Two bugs in scripts/healthcheck.sh made post-deploy health checks unusable: 1. pass()/fail() used `((var++))` which returns the pre-increment value. On first call (var=0) that trips `set -e`, so only the first check ever ran and the script always exited 1. Switched to explicit assignment. 2. The error-log count piped journalctl through `grep -c .`, which counts journalctl's `-- No entries --` meta line as a match, producing a false-positive "1 ERROR-level log entries in last 60s" on clean systems. Now filters meta lines with `grep -vc '^-- '`. Verified with `scripts/healthcheck.sh --version 0.35.1rc5` on staging: all 5 checks run, exit 0. Closes #302 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

coderabbitai · 2026-04-14T10:40:06Z

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: dce32241-21df-4c06-8220-ff3bb7b8d832

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

Create PR with unit tests
Commit unit tests in branch fix/healthcheck-counter-set-e

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

Nathan Schram (nathanschram) and others added 30 commits March 20, 2026 16:03

chore: staging 0.35.0rc7

4063d42

chore: staging 0.35.0rc9

770a191

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

chore: staging 0.35.0rc10

c12c140

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

docs: add #215–#221 to v0.35.0 changelog

d6c006d

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

docs: add topics.py and test file to CLAUDE.md key files

4b135bc

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

chore: staging 0.35.0rc13

2c2fcb4

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

chore: staging 0.35.0rc14

ec5458e

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

chore: staging 0.35.0rc15

fc59ced

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

chore: release v0.35.0

f8ceeb3

- Bump version from 0.35.0rc16 to 0.35.0 - Set changelog date to 2026-03-31 - Sync lockfile Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Nathan Schram (nathanschram) and others added 14 commits March 31, 2026 18:01

chore: bump aiohttp 3.13.3 → 3.13.5 (10 CVE fixes)

55bc762

Fixes pip-audit CI failure from CVE-2026-34513 through CVE-2026-34525. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Base automatically changed from dev to master April 15, 2026 06:43

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix: healthcheck.sh exits prematurely under set -e (#302)#303

fix: healthcheck.sh exits prematurely under set -e (#302)#303
Nathan Schram (nathanschram) wants to merge 44 commits intomasterfrom
fix/healthcheck-counter-set-e

Nathan Schram (nathanschram) commented Apr 14, 2026

Uh oh!

coderabbitai bot commented Apr 14, 2026

Review skipped

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

Nathan Schram (nathanschram) commented Apr 14, 2026

Summary

Test plan

Uh oh!

coderabbitai bot commented Apr 14, 2026

Review skipped

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant