- A2AWatchRegistry.Shared was initialised before DefaultPollDelays (static fields run in source order), so the bot's
registry had null poll delays and every poll watch crashed on its first wait, then sat in the registry doing
nothing until a restart. The delays now come first; a test reads them through Shared.
- TaskNotFound from the remote ends the watch at once with a "(lost)" notice instead of retrying for a day.
- A crashed watch loop restarts by polling after a pause; the third crash in a row ends it with a "(stopped)" notice.
Hour-long remote tasks were never delivered: a subscribe watch held one SSE stream for up to a day, the remote
sent nothing while its agent worked without text, and once the quiet connection died on the way (no FIN/RST) the
watch waited on it with no error and no log line.
- A subscription lasts at most 5 minutes (SubscriptionSlice), then GetTask and a new subscription: a pause, not a
failure. A slice ended mid-read counts as ours even when it surfaces as an I/O error.
- The shared handler connects with TCP keepalive (60 s, every 15 s, 4 probes) and HTTP/2 pings.
- Log lines when a watch starts or is restored, when a slice ends on a running task, when a result is delivered or
kept; the Telegram host's console log carries a UTC timestamp.
- action=status on a settled task ends its watch, so the result isn't delivered a second time.
- A "context busy" refusal of send says nothing was sent, names the watched task in that context and that its
result will arrive by itself; refusals never touch a watch.
A skill's SKILL.md frontmatter may declare session-start (output into
<skill_session_context>, at session start, /new, resume and after compaction),
turn-end (after every run, in the background; output or a failure reaches the
next run as a hidden notice, repeats suppressed until a clean run) and
protected-paths (write/edit refused inside them). Hooks run from the skill
directory with VOLCANIC_WORKSPACE, VOLCANIC_SKILL_DIR and the chat's TZ; waits
are bounded by each hook's timeout and failures never break a run. /skills
shows declared hooks and broken fields.
ADMIN_USER_IDS (Telegram:AdminUserIds) takes a comma separated list; the older
ADMIN_USER_ID / Telegram:AdminUserId still works, also takes a list and is merged.
Empty entries and duplicates are ignored, non-numeric entries fail validation.
Impersonation is now per admin: one admin per target, an impersonating admin
can't be impersonated, and every admin's chat menu lists /impersonate.
ElevenLabs speaks for the voice tool when ELEVENLABS_API_KEY and ELEVENLABS_VOICE_ID
are set (TTS_PROVIDER=elevenlabs|openai picks, the other is the fallback). Opus output
(opus_48000_64) goes to Telegram as is; voice names resolve through /v2/voices;
instructions become audio tags on eleven_v3 and are ignored on other models. The key
is redacted from tool output and never reaches agent processes.
Skill directories may now declare prompt or watcher jobs as files (cron or one-off at).
The file is the source of truth: the scheduler re-reads every chat's jobs/ every minute,
around agent runs and before /jobs and /skills, and schedules, re-plans or removes them.
Ids are <skill>/<job>; run state lives in .volcanic/scheduler/skill-jobs-state.json, never
in jobs.json. Watchers run from the skill directory with VOLCANIC_WORKSPACE and
VOLCANIC_SKILL_DIR. Broken files are shown with the reason in /jobs, /skills, schedule list
and the new <skill_jobs> prompt section; cancel answers with the file to edit. Every part of
a skill directory is optional; file reading and frontmatter parsing move to Abstractions.
The 1.0 SDK ignores unknown part fields, so a file part shaped {kind:file, file:{bytes|uri,
name, mimeType}} arrived as an empty part and vanished. A2AWireCompat rewrites such parts
to {raw|url, filename, mediaType} in JSON and SSE responses before the SDK reads them, and
marks any part it still can't read so the result lists it with its field names.
File parts were only taken from artifacts as inline bytes up to 10 MB; parts in the
status message or the history were dropped silently, URL parts were never downloaded,
and a plain message reply lost its files. Files now go to inbox/a2a/<taskId>/ from
artifacts, status message and recent agent messages (deduplicated), URLs are streamed
to disk (agent key only on the agent's origin, else https without it), the cap is the
new MaxReceivedFileBytes (200 MB), and results list path, bytes and media type or why
a file was not saved. Re-saving a task overwrites, so action=status refetches files.
GOOGLE_CLIENT_SECRET is no longer denied by name in the core defaults (the
*_SECRET glob still removes it); the Telegram host allows it by default so
workspace skills (gmail, gdrive) can use the Google OAuth client. Its value
stays in SecretValues: redacted from tool output, and aliases with the same
value are still removed.
A skill directory may hold INSTRUCTIONS.md, with or without SKILL.md. Its body
(frontmatter cut) goes into every system prompt in <skill_instructions>, re-read
on every run: 16 KB per file, 64 KB in total, truncation marked with the file's
path. /skills marks skills with instructions and lists instruction-only dirs.
Skills live in <cwd>/.claude/skills/<name>/SKILL.md (frontmatter name + description).
SkillCatalog re-reads them on every run and the prompt lists name, description and
path; the model reads the full SKILL.md with the read tool when a task matches.
Streamable HTTP/SSE servers from the workspace's .volcanic/mcp/servers.json.
Headers are sealed with AES-GCM; the chat and the visible fields are the
associated data, so only /mcp_add|remove|tools change servers. Tools
<server>__<tool> go live in the session. An unreachable server is skipped
and reported in the system prompt. Keys are redacted from tool results, and
Telegram deletes messages that carry one.
OpenAI gpt-4o-mini-tts (female voice nova by default; TTS_MODEL/TTS_VOICE) renders OGG/Opus,
sent with sendVoice in order with rendered output. Up to 4500 characters per call, split
into 1500-character parts at paragraph/sentence/word boundaries; all parts are synthesized
before any is sent, and API errors come back to the agent as text with the key hidden.
On whenever OPENAI_API_KEY is set.
Add docs/a2a.md (concepts, file formats, endpoints, sessions, task
lifecycle, configuration, security notes, curl examples) describing
what src/Volcanic.AgenticCore.A2A.Server actually does. Add an A2A
project paragraph to architecture.md, an "A2A" section to
telegram-bot.md (/a2a, notify_owner, isolated sessions, port 8080),
A2A_ENABLED/A2A_PUBLIC_URL and the aspnet image to docker.md, and
project-map/test-command updates to CLAUDE.md.
Mark the A2A server design spec implemented and record the deviations
found while building it (A2AHostOptions, synchronous
IA2ACallerDirectory.Find, generic WorkspaceCallerDirectory, per-status
push payload, split file-part error handling, the outbound guarded
handler, the per-caller server/router architecture, the restart grace
period, the owner= log field, the aspnet image, and the actual push
retry schedule). Note the aspnet image in the docker-deployment spec
too, without rewriting its historical decisions.
While on, the admin's updates are routed as the target user's (allow-list,
chat, workspace, session, settings, jobs) and the target chat's output goes
to the admin's chat. The target's own updates are dropped with a notice that
an admin is working with their account. In memory; /impersonate 0 stops.
Claude Code 2.1.287 loads cc-plugin-* built-ins regardless of --setting-sources,
which tripped the fail-closed init check. Disable every known built-in in
--settings, and when init lists only built-ins that slipped through, disable them
process-wide and replay the request once. Other plugins still fail closed.
token.json is sealed with AES-256-GCM (key derived from the OAuth client secret,
bound to the chat), so the agent's processes can read the file but not the
refresh token. Plain files, including VolcanicSecretary's refresh_token format,
are rewritten sealed on first read.
The scheduler gains IScriptEnvironment: extra per-run variables set after the
process environment filter and redacted from the run's output. The Telegram
host uses it to pass GOOGLE_ACCESS_TOKEN, refreshed to cover the job's timeout;
the refresh token and client secret never reach the script.
Each chat's jobs live in <workspace>/.volcanic/scheduler/jobs.json. On start the
scheduler loads every chat's file, moves the retired shared
<agentDir>/scheduler/jobs.json into them and renames it .migrated-<ts>.
A VolcanicSecretary jobs.json (bare array) in a chat's folder is converted
in place for that chat; the original is kept as .secretary-<ts>.
Also adds the VolcanicSecretary data migration plan.
Jobs are prompts or watchers. A reminder is a prompt the agent passes on, so
it lands in the transcript. Stored reminder jobs load as prompts
("Remind me: <text>"). Also records the Coolify deploy target in CLAUDE.md.
Tokens now live in <cwd>/.volcanic/google-calendar/token.json, as in the
VolcanicSecretary reference, instead of <agentDir>/google-calendar/<chat
key>/. GoogleCalendarOptions.TokenDirectory, when set, keeps the old
per-chat layout under that root. Spec, docs and CLAUDE.md updated.
Entrypoint drops the capability bounding set and sets no_new_privs. Security docs
and spec no longer claim the agent can't reach the secrets; docker exec fallbacks
use a clean environment. PLAYWRIGHT_ENABLED drops the nested compose default.
Tests pin the Dockerfile's Playwright install step and the compose model fallback.
Status implemented, verification intro rewritten, and a verification record of the final checks (dotnet build/test, both image variants, offline probes of uid, /proc environ and the diagnostics socket, playwright-mcp --version offline, headless Chromium as app, docker compose config, compose start with a fake token -> 401 -> exit 1). Open question 2 (Playwright compose default) marked resolved.
Image contents, build arguments, the environment table with what the agent's processes can't see, volumes, Coolify steps, first Claude login, updating (the npm layer is cached: build without cache or pin CLAUDE_CODE_VERSION), extensions, troubleshooting and security.
Two-stage build, agent tools, Node 24, claude CLI, data dirs owned by app, DOTNET_EnableDiagnostics=0. Adds the MCP spec's deferred test that the Dockerfile's PLAYWRIGHT_MCP_VERSION equals PlaywrightPreset.DefaultPackage.
Spec change: the image keeps root as its user and the entrypoint drops to app with setpriv. With USER app, compose's docker-init (PID 1) runs as app and its /proc/1/environ exposes every secret to the agent's bash. The spec also lists the csprojs the restore layer copies (GoogleCalendar, Mcp, Scheduler included) and sets HOME=/home/app in the image.
- JobScheduler: a stored job whose zone lookup throws no longer fails StartAsync; timer re-armed under the lock;
a delivery cut off by shutdown isn't recorded (one-offs fire again after restart); CancelJobAsync in one step
- Schedules.ParseAt: ISO 8601 with a date only (no machine-clock date for a bare 09:00)
- core: a FollowUp prompt that loses the race to a starting run joins it as a follow-up, not steering
- Telegram: TimeZoneFor tolerates non-numeric targets; RunPromptAsync waits only when queued; one agent dir
- ScriptRunner: bad workspace paths are Failed, not exceptions; zone id helpers share TimeZoneIds.ToIana's rule
- tests: no wall-clock sleeps in scheduler tests; transcript-entry, disabled-scheduler and failing-prompt coverage
- docs: spec notes for the above; README row for the scheduler
Spec: a skipped one-off fires after the next start; the late note's placement and no-exit-code failure wording are pinned down; IsDeliverable is part of the watcher invariant and must not create a session; saves snapshot inside a serialising gate.
Spec: on Windows the command goes through Arguments like the bash tool, since cmd.exe doesn't parse ArgumentList quoting; the workspace is created if missing, and a timeout vs cancellation outcome is spelled out.
