Linux_post_install/AGENT_TODO.md

AGENT_TODO — Worklist & Idea Backlog

Living list of what we are doing, what is next, and what we might do later. Deep history lives in git: git log --follow AGENT_TODO.md, git blame, and the individual feature commits — the Done section below is just a readable summary (newest last).

Conventions

• Now — items actively being worked on this session (only a few).
• Next — queued, well-scoped items.
• Later — idea backlog. Ideas marked NOT NOW were evaluated and rejected for the stated reason; revisit only if circumstances change.
• When a task is completed: move it from Now/Next into Done (dated one-line) in the same commit that finishes the work.

Done

• 2026-08-09 — pos media mp3/mp4 hardened + smart format selection. Both tools: yt-dlp calls go through spawn (honor $DRY_RUN; --dry-run prints the exact command and skips dep checks), -o/--output, --no-playlist, --cookies (file existence check), clean ffmpeg/yt-dlp guards, # POS_FLAGS: for completion, full metadata (--embed-metadata --embed-chapters --embed-thumbnail --no-overwrites, mp3 also --convert-thumbnails jpg + --parse-metadata "%(artist,uploader)s:%(artist)s" so the uploader fills the artist tag). mp3 gains --by-artist (~/Music/<artist>/<title>.mp3). mp4 gains -f <id> / --best / --worst (no prompt), conflict validation, and an interactive picker that shows a curated -F table ([audio]/[video]/[combo] grouping, raw clutter dropped) on stderr — stdout carries only the chosen id (ui_pick lesson) — with id validation against the real table and empty/best default. Docs: howto/media.md rewritten (flags, metadata, by-artist, troubleshooting); make gen && make check green.

• 2026-08-09 — Telegram ai … now answers about a message you reply to: the listener extracts reply_to_message.text (falls back to caption) from each update and passes it to handle_message; the AI bridge prefixes the prompt with [Reply context — the message you are replying to]. So replying to a /status output and asking ai check this details gives the model the actual output. Applies only to the AI bridge (mapped /commands untouched); reply context rides in the user turn so the session records what was analyzed. Docs: howto/ai.md bridge section.

• 2026-08-09 — pos ai gemini sessions + Telegram-friendly replies. --session <name> gives ask/chat persistent memory (~/.local/share/linux_post_install/ai/<name>.json, capped at 40 turns, pruning keeps the first user turn as scene); new sessions subcommand (list / reset <name>). Telegram listener now keeps one session per chat (telegram-<chat_id>) with ai /reset to clear. New --system "<text>" flag injects a Gemini systemInstruction (via jq merge) sent every turn but never stored in the session file; the listener passes a Telegram-voice prompt ("reply like a friendly Telegram chat, use emojis") and strips markdown (**, *, backticks, #, links, lists, blockquotes) from replies before sendMessage, since messages go out as plain text. Docs: howto/ai.md (flags, sessions, bridge memory/formatting), make gen && make check green.

• 2026-08-09 — Fixed pos config secret-value corruption: cfg_read_secret's cursor-advance echo went to stdout and, since the function is called via $(...), a leading \n ended up inside every secret value → the env file got AI_GEMINI_API_KEY="\n<key>", unreadable by cfg_value/load_config (menu showed (not set), pos ai gemini demanded a key). The newline now goes to the terminal (echo >&2). Defense in depth: cfg_write/write_config_key strip CR and truncate multi-line pastes (warn), cfg_value and the ai/telegram load_configs strip CR on read. Verified on a real PTY (piped tests couldn't reproduce — non-TTY stdin skips the echo path).

• 2026-08-09 — ai category — pos ai gemini (ask/chat/models) via Google Gemini REST API. ask prints only the answer (pipe/script/Telegram-friendly), chat is a multi-turn REPL (q/quit/Ctrl+C, /reset, empty input re-prompts), models lists generateContent-capable ids and flags the default; --model override; default gemini-2.5-flash. Config scope ai (AI_GEMINI_API_KEY secret + AI_GEMINI_MODEL) in ~/.config/linux_post_install/ai.env, edited via pos config ai; config/ai.env template installed no-clobber by postinstall; ai-gemini added to INTERACTIVE_CMDS. Telegram listener now answers non-command messages starting with ai via pos ai gemini ask (owner chat only; error replies carry the pos config ai hint) — future intents (reminders) slot in as more case arms in handle_message. Docs: POS.md ai section + listener bridge, howto/ai.md, HOWTO/README index rows, bin/pos usage example.

• 2026-08-09 — Entertainment plugins gold + weather now emit emoji-visualized Telegram messages. Gold: headline is USD/gram (XAU/oz ÷ 31.1034768), ounce as reference, bid/ask, cleaned timestamp (+00:00/fractional seconds stripped). Weather: per-WMO-code emoji (☀️/🌙 day-night aware for clear sky), °C + feels-like, humidity, wind with unit spacing. Both verified live; emojis are safe in the default plain send mode.

• 2026-08-09 — pos tree: prints the live pos command tree (categories → commands → subcommands) by deriving the hierarchy from bin/pos-* filenames + # POS: / # POS_SUBCMDS: headers, so it always matches what the dispatcher can run. Category-less like pos-config; --depth N limit; pos help tree works. Docs: POS.md tree section, bin/pos usage example, make gen regenerated the AGENT_Context tree/dispatch/filetable + _pos_flags[tree].

• 2026-08-09 — Telegram sender config / config set removed — redundant with pos config telegram (same # POS_CONFIG: registry, masked token display + input, chat-id validation, chmod 600); sender/listener error hints now point there. Deep-review bugfixes in the same commit: mapped /command values containing | are no longer truncated (load_map switched from a | to a \x1f delimiter — previously /up=echo hi | head silently ran echo hi ); pos entertainment send <plugin> [args…] actually forwards the extra args (every arg was shifted in the flag loop, so $@ was empty) and passes -- before the message so leading-- plugin output isn't parsed as an option; write_config_key (entertainment-lib) and cfg_write (config-ui) replaced unescaped sed -i "s|^K=.*|K=\"$v\"|" with grep-v+append so values with &/|/\ no longer mangle (also the path all telegram config now flows through); sync_systemd daemon-reloads after removing timer units; digits config validation accepts negative group/supergroup chat ids (-100…).

• 2026-08-09 — Fixed telegram listener editor crash on remove/edit/test: ui_pick printed its menu listing to stdout, so idx="$(ui_pick)" captured the menu and the number, and MAP_CMDS[$idx] (arithmetic array subscript) blew up with "syntax error in expression". Menu decoration now goes to stderr; only the picked index is emitted on stdout. Pre-existing bug (before the ::desc work), exposed by the description column.

• 2026-08-09 — Telegram listener pushes its mapped /commands to the bot's / menu via setMyCommands (auto after every map edit, on --enable, and at daemon start; manual --sync-commands flag). Map lines may carry a menu description: /cmd::short description=bash command (falls back to the bash command, ~40 chars). Names are validated against Telegram's lowercase [a-z0-9_] rule — invalid ones are skipped from the menu with a warning but still resolve when typed; empty map clears the menu. Fixed latent bugs found by the sync work: map_has (awk END{exit 1} overrode the match), and warn() went to stdout so it leaked into the generated JSON (now stderr).

• 2026-08-09 — pos config <TAB> scope completion is now cached at gen time (_pos_config_scopes array emitted by make gen from the # POS_CONFIG: registry) instead of scanning ~40 tools per TAB — a per-keypress subshell storm that wedged interactive shells for minutes on the loaded homelab box. Two stuck -bash sessions (69%/38% CPU) killed. plugin_marker/plugin_keys hardened with || true so config_keys no longer aborts mid-scan under set -euo pipefail on mixed lib/plugin dirs (installed layout) — fixes missing plugin keys in pos config entertainment.

• 2026-08-09 — pos config <scope> interactive config editor: reads the # POS_CONFIG: registry across tools into a single runtime config (~/.config/linux_post_install/*.env, one file per scope, chmod 600); secret masking with show/hide toggle, digits:/num:/url: validation, - to clear, blank keeps; *plugins marker expands plugin vars (entertainment) from entertainment-lib.sh; desc::example value-format hints shown in the editor; gen-docs now handles category-less tools (pos-config), fixed a set -e+pipefail bug that truncated the header registry.

• 2026-08-09 — pos-communication-telegram → pos-communication-telegram-sender: one canonical send (dropped the legacy --send flag, which duplicated the send subcommand in completion). pos communication telegram <TAB> now completes to just sender listener. lib/notify.sh maps platform telegram → telegram-sender via notify_sender_name(); entertainment-send + health --send check updated. Removed phantom subcommands from howto/communication.md (webhook/logs/broadcast/file never existed).

• 2026-08-09 — Structure/convention audit fix: --dry-run now truly dry (spawn() honors DRY_RUN, install.sh exports it to child phases, postinstall mutations run-wrapped); gen-docs.sh no longer chmods regenerated files to 0600; make check now syntax-checks apps/entertainment/features/templates; .gitignore protects config/authorized_keys + config/rclone.conf; honest --send confirmation; docs refreshed (notify.sh in lib lists, pos-health systemd units, tsui, scripts/, INTERACTIVE_CMDS).

• 2026-08-07 — pos system health --send notification-only; listener @quiet prefix (run mapped command without replying, for commands that self-notify). /status=@quiet pos system health --send = exactly one digest.

Now

• (none — Tier 1 shipped: pos system health, lib/notify.sh, digest timer)

Next

• Wire alerting into more tools as they are added (default: source lib/notify.sh, call notify_send on success/failure).

Later — idea backlog

• Tier 2: watch plugins — pos system watch <event>: poll conditions and alert on change (public IP changed, disk > 90%, backup skipped, fail2ban spike). Reuses notify_send + a systemd timer per watch.
• Tier 2: pos health extras — temperature/fan/load average thresholds, ss -tln port checks for known services, SMART status for disks.
• Tier 3: backup rotation + remote target — keep-N rotations, upload to rclone remote after verify, --remote flag, digest reports rotation age.
• Tier 3: pos secret vault — gpg/age-encrypted key-value store; backend for future tools that need stored tokens.
• Tier 3: pos inventory — machine manifest (OS, packages, services, mounted disks, USB devices) exportable as markdown/JSON.
• Tier 4: pos self update — pull repo, make gen && make check, re-run install.sh to refresh /usr/local/bin.
• Tier 4: pos new — scaffold a new tool from templates/pos-tool.sh (category, name, POS header, exec bit, doc stubs).
• NOT NOW: per-category bin/ subdirectories — flat bin/ + filename dispatch scales fine; revisit only if bin/ passes ~40 files.
• NOT NOW: split lib/entertainment-lib.sh — fine under 600 lines; revisit if it grows.

Done (summary, newest last)

• 2026-08-06: Fix entertainment timer 1h not firing — interval_to_oncalendar emitted invalid OnCalendar=*-*-* */N:00:00 (systemd rejects */N in the hour field); now *-*-* 00/N:00:00. Dropped the cron fallback entirely: scheduling is systemd user timers only (sync_cron/interval_to_cron/cron_block removed), status simplified, Nd intervals rejected with a clear error.
• 2026-08-06: Nested pos subcommands — # POS_SUBCMDS: header annotation (telegram, docker-compose, docker-vbox) + make gen emits a _pos_subcmds completion map; nested tools (telegram listener) auto-list under their parent instead of as a flat sibling (telegram-listener) in pos <category> and tab-completion; generic tool-level completion (subcommands + flags + --help).
• 2026-08-06: Telegram listener — pos communication telegram listener: interactive /command → bash map editor + owner-only polling daemon as a systemd user service (map in ~/.config/linux_post_install/telegram_commands.env, re-read per message; /help, unknown-command reply, 60s timeout, stdout reply).
• 2026-08-06: NFS in pos system — pos system nfs-server (status/share/ unshare/list/reload/enable/disable, idempotent /etc/exports edits, generic default with Tailscale/WireGuard/LAN examples) + pos system nfs-client (mount/unmount/list + persistent mounts as systemd .mount units ordered after network-online.target, no fstab); nfs-kernel-server + nfs-common added to preinstall PACKAGES.
• 2026-08-06: pos HOW-TO guide set — DOC/HOWTO.md index + per-category DOC/howto/*.md (network, docker, media, system, ssh, usb, communication, entertainment) with flags, recipes, config, and troubleshooting; wired into DOC/README, root README, AGENTS.md.
• 2026-08-06: Multi-platform alerting — lib/notify.sh routes via NOTIFY_PLATFORM (notify.env, default telegram; sender contract for Matrix/Synapse later), system.env shared config for health/backup, dynamic effective values in --help, telegram --markdown alias.
• 2026-08-06: Tier 1 — pos system health (dashboard + --send), lib/notify.sh (wired into backup + firewall), daily digest timer via postinstall.
• 2026-08-06: Document Map index + Entertainment section in AGENT_Context (cf36780).
• 2026-08-06: Entertainment module — plugins (weather/joke/gold), pos entertainment config/enable/disable/send/status, auto-trigger + Telegram send.
• 2026-08-05: pos communication telegram — --parse-mode (plain/markdown/html).
• 2026-08-05: doc/code sync gate — make gen + make check + pre-commit hook.
• 2026-08-05: pos usb server — USB Redirector control tool (494eae2).
• 2026-08-05: pos <category> --help auto-discovery in the dispatcher.
• 2026-08-05: AGENTS.md with lazy-loaded DOC references.