User hit 'API error 402: ... You requested up to 131072 tokens, but can
only afford 4511' on the assist alias: no provider ever sent max_tokens,
so OpenRouter's credit pre-check billed the routed model's full
worst-case output; user also asked to bound session history to the last
5 requests/responses.
Architect decisions:
- AI_MAX_TOKENS (num, default 2048): sent as max_tokens on OpenRouter
and generationConfig.maxOutputTokens on Gemini — a real per-request
cost ceiling. llamacpp unchanged (local/free, no pre-check).
- AI_SESSION_TURNS (num, default 40 kept back-compat; messages, 2 per
exchange — 10 = last 5 conversations): resolved lazily in session_push
because config loads after the hardcoded line-25 default.
- Both registered in the bin/pos-ai POS_CONFIG @General section, so they
appear in 'pos config ai' with num: validation.
Reviewer hardening (CHANGES_REQUIRED -> fixed): unguarded env input could
reach jq tonumber (0/-5/010/abc all savable via config-ui's ^-?[0-9]+$)
and abort the CLI; both providers and session_push now guard with
^[1-9][0-9]*$ and fall back to the default.
Verified: fake-curl shim smoke (16 provider-body + 12 session-window
checks incl. the 010-regression proof), make gen idempotent, make check
OK, make lint 0 FAIL/0 WARN, make test 17 files / 299 checks / 0 fail
(~49s), bash -n clean, git diff --check clean. Reviewer ACCEPT (twice).
Tester regression round (permanent provider-body + session-pruning
coverage) intentionally not run this cycle — user's call; remains a
documented follow-up.
User report after the llamacpp app install: 'installed llama.cpp unknown',
valid flags rejected (randomly per run), 'Model not found' for the HF
downloader's own layout, and a systemd user-bus failure over SSH. Detective
(real b10822 binary, FACT) found four independent causes:
- version: llama-server --version prints to STDERR; detect_llama_version's
2>/dev/null swallowed it -> always 'unknown'. Now captures 2>&1 + accepts
semver/build tokens (incl. build 1.2.3 edge)
- validation: printf|grep -q under pipefail -> SIGPIPE rc=141 race randomly
rejected flags present in the 59 KB --help. Now pipe-less grep (no race);
20x determinism regression test
- model resolution: resolve_model accepted files only, but the HF downloader
creates <models>/<repo>/file.gguf dirs. Now expands a dir with exactly one
*.gguf (never silently picks; multi-gguf lists + errs)
- port: llama.cpp default 8080 vs tool/adapter 8088; validation reliability
means --port is now always pinned in the unit
- user bus: headless/SSH sessions lack XDG_RUNTIME_DIR -> ensure_user_bus in
lib/common.sh pre-flights all three systemctl --user tools with remediation
text; pos ai server --no-unit direct-run escape hatch (pidfile) for boxes
with no bus
- find_llamacpp narrowed to llama-server/llama-server-cuda (bare 'server'
fallback hazard); installer post-install sanity (version+help execute,
symlink targets resolve)
Architect decisions DQ1-DQ6 recorded. Tester: 4 new regression files
(version-from-stderr, 25x flag-validation determinism, model dir expansion,
bus pre-flight + E2E) + 3 fixture updates; suite 16 files / 269 checks.
Verified: make gen idempotent; make check OK; make lint 0 FAIL, 0 WARN;
make test 269/269 (~49s); bash -n clean; git diff --check clean.
Service manager (start/stop/status/models/logs) with systemd user
service generation, GPU auto-detection, model selection from pos ai hf
downloads. Provider adapter integrates with pos ai ask as --provider
llamacpp. Config extends existing ai scope with LLAMACPP_* keys.
87 test cases / 0 failed. make gen/check/lint 0 FAIL / 0 WARN.
Bash-native tool using curl/jq to download AI models from HF Hub.
Subcommands: download (single file/repo/gguf filter), search, list, remove.
Auth via HF_TOKEN in ai.env, resume support, disk space pre-flight,
rate limit handling, .hf-meta metadata tracking.
46 test cases / 0 failed. make gen/check/lint 0 FAIL / 0 WARN.
New tool that classifies URLs by domain and delegates to pos media mp3
(audio) or pos media mp4 (video). Listener gains URL detection step
between prefix map and AI bridge — bare URLs auto-download.
Domain rules: music.youtube.com/soundcloud/bandcamp → mp3,
youtube/vimeo/twitch → mp4 --best, unknown → configurable default.
28 test cases / 70 assertions / 0 failed.
make gen && make check green, make lint 0 FAIL / 0 WARN.
Generalizes the Telegram listener with a configurable text-prefix map
(telegram_prefixes.env): any non-command message '<word> <text>' runs
the mapped command with <text> appended as ONE quoted argument — e.g.
opencode=opencode turns 'opencode check cpu' into opencode "check cpu".
Routing order per message: text-prefix map → built-in Gemini ai bridge →
/command map → Unknown command. A mapped word shadows the Gemini bridge.
The prefix verb is reworked: bare = list map + bridge word; 'prefix
<word> <cmd...>' = map; 'prefix <word>' = show; 'prefix -r <word>' = remove.
The Gemini trigger word itself is now set via 'pos config telegram'
(TELEGRAM_AI_PREFIX).
Also extracted run_and_reply() to share the /command-map (60s) and
prefix (120s) execution semantics; fixed a latent set -e abort on
invalid templates in prefix_map_set's check_syntax call.
Verified: 27/27 routing-harness assertions, full CLI verb suite,
dispatch smoke, pos config telegram render, bash -n, make gen && make check,
make lint 0 FAIL / 0 WARN, shellcheck -S style (0 new findings).
New bin/pos-docker-stack: docker ps output grouped by Docker Compose
project (stack). Each stack is a sorted section (name, status, ports);
containers without a compose project land in a Standalone group at the
end. Running only by default, -a|--all includes stopped/exited. Status
colored on a terminal; summary line 'Stacks: N containers: N
standalone: N'; exit 0 when empty.
Data via docker ps --format with \x1f delimiters (project label
com.docker.compose.project from compose v2); parsed with awk -F'\x1f'
+ IFS=$'\x1f' read — tab/pipe delimiters are IFS whitespace or appear
in values (DEV.md:213). Dash padding via sed, not tr (multi-byte).
Deps guard (docker) before --help; no stdin.
Docs: POS.md docker row + detail, howto/docker.md table + section,
bin/pos usage EXAMPLES, AGENT_Context Common Tasks row. Verified:
stub suite 23/23, live daemon runs, dispatch, make gen && make check,
make lint 0 FAIL / 0 WARN.
x-systemd.automount in a unit file's Options= is only honored when read
from /etc/fstab, so persisted mounts never armed an automount; the unit was
also static (no [Install]) so enable was a no-op. Result: after a reboot the
share sat inactive and the disk never mounted on access.
persist now writes both <name>.mount and <name>.automount units (the
automount with WantedBy=multi-user.target) and enables/starts the automount;
unpersist stops/disables/removes both units.
persist uses x-systemd.automount, so the CIFS share is never actually
mounted until first access and findmnt -t cifs stays empty — list reported
'No SMB mounts' despite configured persistent units. list now also scans
$UNIT_DIR/*.mount for Type=cifs and reports them under 'Persistent
(automount)' (where, what). Empty output still prints 'No SMB mounts'.
Once the archive verifies, USB detection runs (so a stick plugged in while
the backup ran is found): mounted removable storage is auto-detected via
lsblk -J + a recursive jq filter (rm, mounted, type part|disk — JSON makes
spacey mountpoints safe), or BACKUP_USB_ROOT pins a fixed stick and skips
detection. None mounted → one re-scan prompt ('s' skips, EOF from cron
skips silently, rc stays 0); one stick → y/N confirm; several → numbered
pick (0 = skip). The copy lands in <usb>/backups/ (mkdir -p, chmod 600
best-effort — a vfat chmod failure warns, never fails the copy) and the
transfer is proven 100% by sha256 source-vs-copy before any success is
announced; a mismatch warns with both hashes, notifies 'USB copy FAILED',
and exits 1. The ERR trap is re-armed before the USB phase so a copy
failure no longer notifies 'Backup FAILED'. Docs: usage() Environment,
POS.md backup row, howto/system.md (USB section + env table + mismatch
troubleshooting), DEV.md system.env list. Stub suite
(/tmp/opencode/backup-test, HOME-isolated, sudo/gpg/lsblk/sender stubs,
corrupting-cp + vfat-chmod overrides, per-test lsblk JSON fixtures):
40/40 green.
Give a dead single-file HTTP/FTP download a fresh URL: same dir + file name
(--continue=true resumes the partial), forgets the old source from
download.retry, verifies the new link (a dead replacement is diagnosed and
marked permanent instead of silently looping). status flags downloads needing
this with a 'needs fresh link' line. Torrents/active/multi-file rejected with
hints; --dir/--split/--tmux supported.
restart <gid>: re-queue from history — torrents via rebuilt magnet
(urn:btih: + &tr= trackers), HTTP via original URIs keeping dir/out;
--continue=true resumes partials, complete files verify instantly.
retry <gid|all>: waits out internet outages (NET_PROBE seam,
--interval/--max-wait), re-queues and re-verifies; aria2 error 3 = real
problem → diagnosed + marked permanent (url:/bt: ids in download.retry,
skipped by retry all, manual restart overrides); --once/--quiet for the
healer timer.
Healer: pos-aria2-retry.{service,timer} user units — arms on download
start (add/torrent/metalink/restart), disables when nothing left.
watch <gid> auto-restarts after an outage.
Fixes from stub-suite review: ensure_healer missing from submit paths;
RESTART_NAME lost across do_restart subshell (download_name helper);
restart exited 1 (tmux test as last statement).
Stub harness (/tmp/opencode/dl-test) 119/119 green; make gen && make check green.
Docs: POS.md rows, howto/network.md outage recipe, SYSTEMD.md user units.
Persistent aria2c as a systemd user service (pos-aria2.service, enable --now,
linger warning) on localhost:6800 with a generated RPC secret in
~/.config/linux_post_install/download.env (chmod 600, env override). 18
commands: start/stop/status, add/torrent/metalink (auto-start, --tmux live
view), list/info/files/peers, pause/resume/remove/purge/move, limit/set,
watch (2s live repoll). JSON built via jq -nc --arg, never string
interpolation. Deps: aria2 in preinstall PACKAGES, aria2c/jq/curl guards
before --help. Docs: POS.md + howto/network.md + indices + Common Tasks row.
pos communication matrix sender: send (plain/--markdown/--room), login
(password->access token via m.login.password), test. Implements the
lib/notify.sh sender contract, so NOTIFY_PLATFORM=telegram,matrix now
fans out for real; matrix.env config scope registered for pos config.
pos communication matrix listener: systemd user daemon long-polling
/sync (since token, compact m.room.message filter); reacts to own user's
messages (/ and ! both resolve), threaded m.in_reply_to replies, @quiet
marker, ai bridge with per-room session, interactive editor. Added to
INTERACTIVE_CMDS.
Docs: POS.md, howto/communication.md, HOWTO.md, usage EXAMPLES.
Verified against a mock homeserver (send shape, login, owner filter,
replies, exit codes, editor). make gen && make check green.
Each line of event.env is an independent rule: ["msg" if ] <check> <op> <thr>.
Check runs on every pass; first numeric output compared float-safe; op is the
rightmost 'op threshold' pair so checks with their own >/< parse fine. Alerts
once on false->true + one recovery message on true->false (no repeats while
the condition holds); per-rule state keyed by rule-line hash in
~/.local/share/linux_post_install/eventer/state/.
Subcommands: run (timer entrypoint), config (interactive add/remove/edit with
check-validation), list (rules + live values), enable [interval] (systemd
user timer pos-event-trigger.timer; 5m..weekly or OnCalendar; graceful without
a user manager, loginctl enable-linger attempt), disable, status. --dry-run
honors the DEV.md convention. Alerts via lib/notify.sh (Telegram default).
New: bin/pos-system-event-trigger, lib/eventer-lib.sh, config/event.env
template (no-clobber via postinstall), install.sh lib install, INTERACTIVE_CMDS
entry. Docs: POS.md system row, HOWTO.md index, howto/event-trigger.md.
make gen && make check green; functional tests cover trigger/recovery/no-repeat,
float+unit parsing, editor add/remove/edit + validation + dry-run, timer
enable/disable/status, dispatcher routing.
- bin/pos-ai-gemini: one-shot ask (stdout-only, pipe-friendly), interactive
multi-turn chat REPL (q/Ctrl+C, /reset), models list; --model override,
default gemini-2.5-flash; key via x-goog-api-key header, never printed.
- Config scope 'ai' (AI_GEMINI_API_KEY secret, AI_GEMINI_MODEL) in
~/.config/linux_post_install/ai.env via 'pos config ai'; config/ai.env
template installed no-clobber by postinstall.
- Telegram listener: non-command text starting with 'ai ' (case-insensitive)
is answered by Gemini via 'pos ai gemini ask'; owner-chat only, errors
reply with the pos config ai hint. Future intents slot in as case arms.
- ai-gemini added to INTERACTIVE_CMDS (chat reads stdin).
- Docs: POS.md ai section + listener bridge, howto/ai.md, HOWTO/README
index rows, bin/pos usage example; make gen refreshed context/completions.
pos tree derives the hierarchy live from bin/pos-* filenames + # POS:
/# POS_SUBCMDS: headers, so it always matches what the dispatcher can
run. Category-less tool (like pos-config); --depth N limits nesting.
Also sharpen the agent-facing docs that the work exposed:
- AGENTS.md: drop the stale "one category-less exception" claim; document
the category-less pattern (pos-config, pos-tree) and that gen only reads
the text after the first em-dash in # POS: headers
- DEV.md: category-less vs categorized rule + header extraction note
- AGENT_Context "Adding a New Tool": new step 0 — pin the exact CLI verb
and runtime context (dev/repo-only vs installed) before writing code
- POS.md: new config + tree sections; category-less tools get their own
sections
- README: pos tree bullet following the pos X convention
- Drop sender 'config'/'config set' subcommand — redundant with 'pos config
telegram' (same POS_CONFIG registry: masked token input/display, chat-id
validation, chmod 600); sender/listener error hints now point there.
- Listener: load_map now uses a \x1f delimiter, so mapped /commands
containing '|' are no longer truncated at the first pipe.
- entertainment send: forward extra plugin args (previously dropped); pass
'--' before the message so leading '-' plugin output isn't parsed as an
option.
- write_config_key/cfg_write: replace unescaped sed writes with grep-v+append
so values with &, |, \ round-trip safely.
- sync_systemd: daemon-reload after removing timer units.
- config-ui: 'digits' validation accepts negative group/supergroup chat ids.
Map entries may now carry a menu description (/cmd::desc=bash) shown in
the bot's "/" menu; it falls back to the bash command (~40 chars) when
omitted. The command list is pushed automatically after every map edit,
on --enable, and at daemon start, plus a manual --sync-commands flag
(wired into POS_FLAGS completion). 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; an empty map clears the menu.
Fixes found by unit-testing the sync path: map_has always returned false
(awk END{exit 1} overrode the match), map_set had a local cross-reference
that broke under set -u, and warn() leaked into the generated JSON (now
stderr). Docs + AGENT_TODO updated; make gen && make check pass.
The --send flag duplicated the send subcommand — completion suggested both.
Now 'pos communication telegram <TAB>' completes to just 'sender' and
'listener'. send is the single canonical action.
- git mv bin/pos-communication-telegram bin/pos-communication-telegram-sender
- remove --send branch + POS_FLAGS entry; add --markdown to completion
- lib/notify.sh: notify_sender_name() maps platform telegram -> telegram-sender
(notify_send otherwise looks for bin/pos-communication-telegram)
- pos-system-health --send 'sent:' check + pos-entertainment-send use the new
name and the send subcommand
- completion: keys with nested tools but no direct tool complete to the group
suffixes (sender/listener); --type/--parse-mode value completion moved to
the new word positions
- docs updated (POS.md, AGENT_Context, DEV.md, SCRIPTS.md, notify.env,
HOWTO.md, postinstall.sh); removed phantom webhook/log/broadcast subcommands
from howto/communication.md
- lib/notify.sh: route notify_send to every platform in NOTIFY_PLATFORM
(notify.env, default telegram; comma-separated = send to all). New
platforms need only a bin/pos-communication-<p> sender implementing
'send <value> [--markdown]' (Matrix/Synapse ready)
- pos-communication-telegram: add --markdown as alias for --parse-mode
markdown to match the sender contract
- lib/common.sh: load_system_env() — shared ~/.config/linux_post_install/
system.env for pos-system-* tools (env exported > file > default)
- pos-system-health/backup: load system.env and show effective dynamic
values (NOTIFY_PLATFORM, HEALTH_BACKUP_MAX_AGE_DAYS, BACKUP_SERVICE_ROOTS)
in --help
- config/system.env + config/notify.env templates copied by postinstall
- systemd/pos-health.service: EnvironmentFile for both configs so the
daily digest honors them
Your Name