Fridica

Fridica is a local persona that connects your Slack identity to Claude Code or Codex. It listens in channels you choose, decides when to participate, works in configured project directories, and replies in Slack threads as you. Replies use your first-person voice, without a Fridica introduction or visible signature. Machine-readable metadata remains for loop protection; legacy signed messages are still recognized. Your own messages never trigger your agent.

This first release runs one owner per daemon and one Slack app per owner. Anyone in an allowed channel can trigger workspace actions. It includes both CLI backends, SQLite context and task storage, clarification conversations, and loop limits. It does not include a shared relay, browser OAuth onboarding, MCP server, or automation of the Claude/Codex desktop UI.

Install

Use macOS or Linux with Python 3.11 or newer. Install and authenticate either Claude Code or Codex CLI before you install fridica.

Sandbox dependencies on Linux

Both backends run agent commands inside a bubblewrap sandbox on Linux; macOS uses the built-in sandbox-exec and needs nothing extra.

Backend	bubblewrap	socat	Notes
Claude	required (system package)	required	Fridica starts Claude with sandbox.failIfUnavailable, so Claude refuses to run at all when either is missing
Codex	bundled, but a system bwrap on PATH is preferred	not needed	Installing the system package lets one AppArmor profile cover both backends

1. Install the packages.

# Debian / Ubuntu
sudo apt install bubblewrap socat

# Fedora
sudo dnf install bubblewrap socat

2. Allow bubblewrap to create user namespaces (Ubuntu 24.04 and later). Ubuntu's default AppArmor policy blocks unprivileged user namespaces, so every sandboxed command fails with bwrap: loopback: Failed RTM_NEWADDR: Operation not permitted even though the packages are installed. Check the setting:

sysctl kernel.apparmor_restrict_unprivileged_userns

If it prints 1, install an AppArmor profile that grants bwrap the capability (the profile applies to bwrap only, not to the commands it runs inside the sandbox), then reload AppArmor:

sudo tee /etc/apparmor.d/bwrap > /dev/null <<'EOF'
abi <abi/4.0>,
include <tunables/global>

profile bwrap /usr/bin/bwrap flags=(unconfined) {
  userns,
  include if exists <local/bwrap>
}
EOF
sudo systemctl reload apparmor

If it prints 0 or No such file or directory, skip this step. This follows Claude Code's sandboxing guide; Codex uses the same system bwrap, so the profile fixes both backends.

3. Verify.

fridica doctor   # expect: PASS AI sandbox

doctor checks that the packages are present and then runs /bin/true inside a bubblewrap user namespace, so it fails with the exact bwrap: error when either step above is incomplete. start runs the same check and refuses to launch on failure. Without it, a running daemon would return the generic "I couldn't complete this request" reply within seconds for any request that needs a command. The local log records the agent's exit status and a bounded tail of its stderr; that diagnostic text is never sent to Slack.

init creates ~/.config/fridica/config.toml, without overwriting existing configuration, and copies two editable files beside it: manifest.yaml for the Slack app and contract.md, the rulebook every agent run reads (see Agent contract). The repository list is shared and ships with the package (see Repository list). For a different location, use fridica init --config /path/config.toml.

Install via pypi

pip install fridica
fridica init

Install locally to an existing python virtual environment

git clone https://github.com/chengcli/fridica
pip install -e .
fridica init

Configure Slack

Use one Slack app per person, with a user token for that person's account. Do not share tokens. Other channel members need no installation to interact with your running Fridica; anyone in an allowed channel can request workspace actions. Separate owners must not share an app: multiple Socket Mode connections divide events rather than broadcasting them to every connection. See Slack's Socket Mode documentation.

1. Create the app and set permissions

Open Slack app management, choose Create New App → From an app manifest, select your workspace, and paste slack/manifest.yaml. The manifest sets up public channels. Verify these settings before installation:

Slack settings page	Setting	Required value
Socket Mode	Enable Socket Mode	On
Basic Information → App-Level Tokens	Generate token and scopes	connections:write; save the xapp- token
Event Subscriptions	Enable Events	On
Event Subscriptions → Subscribe to events on behalf of users	Public-channel event	message.channels
OAuth & Permissions → User Token Scopes	Public-channel messages	channels:history
OAuth & Permissions → User Token Scopes	Channel information and membership checks	channels:read
OAuth & Permissions → User Token Scopes	Send replies as your account	chat:write
OAuth & Permissions → User Token Scopes	User-profile access included in the manifest	users:read

users:read is included for user-profile access, but current mention rendering does not require a profile lookup. Bot Token Scopes and bot event subscriptions are not used. You do not need a public Request URL with Socket Mode.

For private channels, also add these before installation:

Slack settings page	Additional values
OAuth & Permissions → User Token Scopes	groups:history, groups:read
Event Subscriptions → Subscribe to events on behalf of users	message.groups

Keep the public-channel settings if monitoring both types. DMs and group DMs are not supported. The authorized person must belong to every configured channel. Event subscriptions and their corresponding scopes are both necessary; see Slack's Events API and private-channel events.

Do not add unrelated scopes:

• links:read and links:write are for shared-link events and custom unfurls; ordinary replies containing URLs need only chat:write. Fridica disables unfurls.
• File upload/download, reactions, channel administration, email, and bot mention scopes are not needed for current functionality. Local workspace file access is controlled by the agent, not Slack scopes.
• metadata.message:read was listed in older manifests, but Slack documents it as a bot/legacy-bot scope, not a user-token scope. Do not add a bot token just for this scope. Fridica still attempts to attach metadata to outgoing replies; metadata availability and acceptance are separate from ordinary message access. See Slack's metadata scope reference and link permissions.

2. Install and export tokens

Under OAuth & Permissions, select Install to Workspace and authorize as the person Fridica will represent. An administrator may need to approve the app and requested scopes. Copy the User OAuth Token starting with xoxp-, not a bot token (xoxb-). Both tokens must belong to the same app and intended workspace.

For SLACK_APP_TOKEN, open Basic Information → App-Level Tokens and copy the xapp- token already generated in step 1 with connections:write. Token generation should be complete by this point; reuse that token rather than creating another. For SLACK_USER_TOKEN, use the xoxp- User OAuth Token from OAuth & Permissions after installation.

Export them in the terminal where Fridica will run:

export SLACK_APP_TOKEN='xapp-your-token'
export SLACK_USER_TOKEN='xoxp-your-token'

After changing scopes, reinstall the app, update the exported user token if Slack replaces it, and restart Fridica. Save event-subscription changes as well. Never commit tokens to Git or put them in an agent-accessible workspace.

3. Configure your local identity and channels

After fridica init and exporting your user token, detect your identity and choose channels from a numbered list:

fridica configure --detect

This gets owner_id and workspace_id from Slack's auth.test and discovers joined, non-archived channels using conversations.list. Select channel numbers separated by commas; Fridica saves their IDs automatically. It never enables all discovered channels without your selection. Blank input cancels without changing the file. Private-channel discovery requires groups:read; if a channel-type read scope is missing, a warning explains which scope to add. Receiving private messages still requires groups:history and message.groups. Detection does not require AI setup or the app-level token and posts no messages.

For noninteractive setup, select by channel name, not ID:

fridica configure --detect --channel-name general --channel-name my-project

Names must uniquely match discovered channels. Repeat --channel-name to select multiple channels. Unknown names or failed discovery leave the file unchanged.

Manual ID options remain available:

fridica configure --owner-id U123ABC --workspace-id T123ABC --channel-id C123ABC
fridica configure --channel-id C123ABC --channel-id G456DEF

Each option is optional, but supply at least one. Repeated --channel-id options replace the full channel list; omitted settings and TOML comments are preserved. Use --config /path/config.toml for a nondefault file. The command validates ID formats locally; it does not discover IDs, contact Slack, or change Slack permissions. Restart Fridica afterward. Owner/workspace IDs must match the user token, and a different identity needs a separate state_path rather than reusing old state.

Edit ~/.config/fridica/config.toml. Set owner_id to the member ID of the person who authorized the user token, workspace_id to the Slack workspace ID, and channels to the exact channel IDs to monitor. Use an existing local project directory. Choose a backend and authenticate its CLI separately from Slack.

Example configuration (replace the IDs and directory):

owner_id = "U123ABC"
workspace_id = "T123ABC"
channels = ["C123ABC"]
workspace = "~/projects/my-project"
additional_workspaces = []
backend = "codex"
profile = "I maintain the simulation package and help diagnose test failures."
general_messages = true
context_limit = 50
timeout = 600
cooldown = 60
max_turns = 6
resume_sessions = true
session_timeout = 1209600
allowed_domains = ["*"]

Remote working folder over SSH

The working folder can live on another machine. Write the roots as host:/absolute/path, where host is an alias from ~/.ssh/config (or user@host):

workspace = "dart9:/mnt/data1/projects/my-project"
additional_workspaces = ["dart9:/mnt/data1/projects/shared-lib"]

Every per-turn run then happens on that host: the reply, the tool-less classification, summaries and debriefs, and the environment checks. Roots on other hosts, such as additional_workspaces = ["dart9:/mnt/data1/projects"] next to a local workspace, do not take part in replies at all: they make that host available to the heavy tasks described below, each confined to its own roots (see Heavy tasks). Fridica starts each one as ssh -T host 'cd /path && codex exec ...' (or claude -p ...) with no PTY, feeds the prompt on stdin, and reads the structured result from stdout, so SSH itself is the transport and the security boundary; nothing listens on a port and no filesystem is mounted. Slack, the state database, the dashboard, and the configuration stay on the machine that runs fridica start.

Requirements on the remote host:

• ssh host true must succeed from this machine without any prompt (key authentication; Fridica uses BatchMode=yes). Put the alias, user, and identity file in ~/.ssh/config. Fridica multiplexes all of its connections through one ControlMaster socket (in $XDG_RUNTIME_DIR/fridica, or /tmp/fridica-ssh-<uid>), so a burst of runs does not trip the server's connection limits.
• The backend CLI must be installed and signed in for the remote user's login shell: ssh host 'command -v codex && codex login status' (or claude auth status) is what fridica doctor runs.
• For Claude, bwrap and socat on the remote host (see the sandbox dependencies above); for Codex, a system bwrap is used when present.
• With a remote workspace, every other root carries a host: prefix too. read_only_workspaces stay on the workspace's host. file_access (scoped file access) covers local roots only: it is rejected with a remote workspace, while remote roots in additional_workspaces simply become heavy-task hosts next to it.

fridica doctor adds an SSH connection check that connects, confirms the roots are directories, and checks the remote OS, then runs the executable, capability, sandbox, and sign-in checks on the host. Each further heavy-task host gets its own Heavy-task host check covering the connection, its roots, the backend and its sign-in, and bwrap when it declares GPUs. A run that cannot reach the host is logged as an SSH failure (status 255) rather than an agent error. The remote agent is bounded by timeout when the host has it, so a dropped connection cannot leave it running.

Agent contract

~/.config/fridica/contract.md holds the rules that every agent run must read and obey. init copies the packaged default there; edit it freely. The daemon reloads the file for each run, so changes apply to the next reply without a restart. Two headings are required:

Section	Sent to	Purpose
## Participation	the tool-less classification call	when to join a conversation that did not @mention you
## Replies	the call that does the work	voice, scope, what may be changed, how to end a thread
any other ## heading	the call that does the work, after ## Replies	project or team rules, such as the default's ## Repo rules

Prose before the first heading is for people and is never sent to the model. Fridica appends only the conversation data (owner, profile, task, bounded thread history, and the new message) plus a one-line note when a thread's session is being resumed. Rules that the code enforces regardless of the contract: replies are limited to 3500 characters, the status must be complete, waiting, or blocked, the sandbox and workspace roots come from config.toml, and Slack tokens never reach the agent. A contract that is missing either heading, has an empty section, or exceeds 64 KiB fails fridica doctor, and until it is fixed replies are refused with the generic "I couldn't complete this request" notice while the reason is logged locally.

To keep the contract elsewhere, set contract = "path/to/rules.md" in config.toml; relative paths resolve from the config file's directory. Without that setting Fridica uses contract.md beside config.toml when it exists, otherwise the packaged default. The default's text is the previous built-in rule set, so upgrading without editing changes nothing.

Repository list

Fridica ships one repository list for the whole team: src/fridica/repos.toml in this repository, installed as fridica/repos.toml and read by every agent run. It is the same for everyone, so changes go through a pull request to main; the test suite validates the file on every pull request and refuses unknown fields, missing names or URLs, non-https URLs, and duplicate names. Upgrading Fridica delivers the new list; init does not copy it.

[[repos]]
name = "snapy-cli"
url = "https://github.com/chengcli/snapy"
collaborators = ["Cheng Li", "Tianhao Le", "Xi Zhang"]   # first entry is the owner
notes = "Hydrodynamic core"

name, an https url, and at least one collaborator are required; notes is optional, and names must be unique. The first collaborator is the repository owner. The agent treats the owner's word as final for that repository: requests from other people get the analysis or preparation they ask for, but merging, releasing, or changing conventions is stated as the owner's decision, and when a thread carries conflicting instructions the owner's are followed. The owner is spelled out as an owner field in the data the agent receives. Entries deliberately carry no local path: each person keeps checkouts wherever they like, and the agent finds the checkout for a chosen repository under its workspace roots by matching the git remote URL. Listing a repository grants no access; reads and writes still follow the workspace roots in config.toml.

The list is sent to the agent as data inside every classification and reply call, never as instructions. The contract's ## Replies rule tells the agent to resolve which repository a request means by matching names and URLs and the requester against the collaborators, to name the chosen repository in its reply, and, when more than one entry could match or none does, to ask with status waiting listing the candidates instead of guessing.

To try a change before opening the pull request, set repos = "path/to/list.toml" in config.toml (relative paths resolve from the config file's directory). doctor then reports the list as a local override. An invalid file fails doctor and blocks replies with the generic notice until fixed.

4. Verify reception, then replies

1. Run fridica doctor. It checks token format and local AI setup, not granted Slack scopes. start checks Slack identity and channel membership.
2. Run fridica start --observe-only, then send a new test message in a configured channel. Expect INFO Observed event ... in ...; no agent or delivery.
3. Stop with Ctrl-C and run fridica start. Have another person @mention you in a new thread. Your own messages never trigger your agent.

Symptom	Check
Listening as ..., but no observed event	Enable Events, user event subscriptions and matching history scopes, saved changes/reinstallation, channel ID and membership, matching tokens, and no competing Socket Mode daemon
Events observed, no reply	Stop observe-only mode; use another person's explicit @mention; check AI setup and local failure logs
Reply says "I couldn't complete this request" within seconds	The agent process exited before doing work. Read the WARNING Agent response unavailable log line, which includes the agent's stderr tail, and run fridica doctor. See Sandbox dependencies
missing_scope when sending	Confirm chat:write is a User Token Scope, reinstall, refresh the token if changed, and restart
Metadata-related rejection	Inspect the exact error; adding link or unrelated scopes does not fix metadata restrictions
Old event reported as failed/interrupted/ambiguous	Inspect locally; restarting does not replay agent actions or uncertain replies

Observe-only never invokes AI or sends replies. Messages sent before startup are not fetched. Scopes, subscriptions, membership, and workspace policy all affect delivery; a successful connection alone does not verify reception.

Run

fridica doctor
fridica start --observe-only
fridica start

doctor prints a PASS or FAIL for each local check: operating system, configuration, the agent contract, the repository list, each Slack token's format, AI executable availability, required CLI flags, sandbox dependencies, and AI sign-in. It runs claude auth status or codex login status for the configured backend without invoking a model or printing account details. The sandbox check confirms that the sandbox dependencies are on PATH and that bubblewrap can create a user namespace; start performs the same check and refuses to run when it fails. Independent checks continue after failures; checks blocked by invalid configuration or a missing executable show SKIP. The command exits nonzero if any check fails or is skipped. Sign-in status does not guarantee that a later model request will succeed or that credits are available. start verifies the Slack user and workspace identity and channel membership. --observe-only records messages without invoking either model or posting replies. Stop with Ctrl-C or SIGTERM.

When to restart

The daemon loads its code and configuration once at startup and re-reads only a few files for each agent run, so what changed decides whether a restart is needed:

Change	Restart needed?
contract.md (agent rules)	No; re-read for every run, applies to the next reply
repos.toml (shared repository list, including an upgrade that delivers a new one)	No; re-read for every run
Settings the dashboard can edit (model, reasoning_effort, max_turns, max_wait_replies, directory access)	No; the listener reloads them between requests
Any other config.toml key: channels, identity, tokens, backend, allowed_domains, resume_sessions, session_timeout	Yes
Fridica's own code: a pip install --upgrade, a git pull on an editable install, or any edited .py, .js, or .html file	Yes
Sandbox packages or the AppArmor profile	No; the sandbox is set up for each run. If start had refused to launch because of them, simply start it again

A merged and pulled branch therefore needs a restart even when the daemon is already running the same feature from your working tree: the process still holds the old modules in memory. Restart with Ctrl-C in the daemon's terminal or tmux pane followed by fridica start; no message is lost, because incoming events are acknowledged and stored before processing, and undelivered replies resume. The startup log lists earlier events that ended blocked or uncertain so you can inspect them; it never replays them. All subcommands accept --config PATH; python -m fridica is also supported.

Fridica responds to mentions of the owner and follow-ups while a task is waiting for clarification. Other messages pass through a separate classification call with tools disabled, governed by the ## Participation section of the agent contract. Classification failure means silence. Set general_messages = false to disable unsolicited participation. The owner’s own messages supply context but never directly trigger their agent.

An explicit human @mention always requests a threaded reply, even with general participation disabled or a cooldown active. If the thread has exhausted its action budget or an earlier task needs inspection, Fridica replies with a brief explanation without running more actions or retrying old work. This does not override observe-only mode, channel restrictions, duplicate suppression, or automated-message loop protection. Slack rejection or uncertain delivery can still prevent a reply; inspect the local logs rather than automatically resending.

Only the structured final answer is delivered to Slack. The default contract's ## Replies rules exclude internal commentary, unsolicited summaries, and tool transcripts from that answer; CLI progress and stderr are never used as reply text. Sanitized failure categories stay in local logs, while Slack receives a short actionable notice. This output boundary does not guarantee that a model will never put unwanted prose in its final answer. Known participant IDs in reply prose become Slack mentions, displayed as people's names (and potentially notifying them); code and URLs are preserved. No additional scopes are required for mention rendering. Closing answers (complete or blocked) are instructed not to @mention anyone; they should omit direct address or use plain names. Mentions are reserved for waiting replies that need someone's response. Built-in blocker notices do not mention anyone.

Replies stay in their original thread. Each thread has a persistent six-turn default budget (max_turns). When the last allowed reply has been delivered, Fridica wraps the thread up: it posts a short stop notice in the thread, asks the agent for a summary of the whole discussion (a tool-less call governed by the contract's ## Thread summaries section), posts that summary as a new top-level message in the channel, and prepares the new thread with a fresh turn budget and the old thread's session so the work continues with full context. The summary post @mentions nobody, so the new thread only continues when a person replies to it; two agents cannot chain threads indefinitely. The exhausted thread stays paused (visible in the dashboard with the reason and the new thread's timestamp) and later mentions there are recorded but not answered. If the summary cannot be produced or posted, the stop notice says so, the failure is logged, and nothing is retried automatically.

Every reply also carries the agent's judgement of whether the discussion is finished: the original request resolved, every action item raised in the thread done or explicitly handed off, and nobody waiting on anyone. When a delivered reply says finished (only possible with status complete), Fridica asks the agent for a debrief (a tool-less call governed by the contract's ## Debriefs section) and posts it as a new top-level channel message headed "Debrief: this discussion is finished." It names people plainly and @mentions nobody. A thread is debriefed once per finish; if the conversation continues afterwards, a later finished reply produces a fresh debrief. A finished thread that is also at its turn limit gets the debrief instead of the continuation summary. A failed debrief is logged and not retried. Generated messages initiate responses only when explicitly addressed or following an active task. A per-channel cooldown limits unsolicited replies. Other agents' metadata is a loop-control hint, not an authorization credential.

Messages from other people's Fridica instances arrive as that person's own messages. If their Slack app also has a bot user, Slack adds a bot_id to the user-token post; Fridica still accepts it because the user field names the sender. Only messages without a user (true bot posts) and edits, joins, and other subtypes are ignored. When an ignored event @mentions you, the daemon logs a warning naming the event and the fields that caused the rejection so a missing reply can be traced without reading Slack.

Continuity between turns

With resume_sessions = true (the default), each Slack thread maps to one backend session. The first reply in a thread starts a session (claude --session-id or a persisted codex exec thread) and Fridica stores its identifier with the thread's task. Later turns resume it with claude --resume or codex exec resume, so the agent keeps its own reasoning, tool results, and file knowledge instead of re-reading the workspace from scratch. The bounded Slack history is still sent as reference, with the new message marked as the only new input. Classification calls remain stateless.

A stored session is resumed only while the thread stays active. After session_timeout seconds without a reply in that thread (default 1209600, two weeks), the next turn starts a fresh session and stores its identifier; set 0 to never resume. If the backend reports that a stored session no longer exists (for example after the provider's session files were removed), Fridica logs it and starts a fresh session for that thread. Other failures are not retried. Resumed turns use the same sandbox, permission, and workspace settings as new ones; for Codex, the resume subcommand receives the equivalent -c settings because it lacks the --sandbox and --add-dir flags. Set resume_sessions = false to return to one ephemeral session per reply.

Workspace authority

By default (file_access = true) the agent works on the local roots through scoped file operations (see Scoped file access); heavy tasks then run only on remote hosts. With file_access = false, or with a remote workspace, the selected agent instead reads, edits, and runs commands using its provider-supported sandbox in the configured workspace roots. Task-command network access follows allowed_domains, which defaults to every host (see Network access). Provider API access is still needed to run the model. Claude requires its sandbox dependencies. Fridica does not enable bypass-permission flags or automatically approve broader access. Claude enables native Edit and Write tools in acceptEdits mode for the workspace and additional_workspaces, alongside sandboxed Bash for file operations such as renaming or deleting files. Classification still has no tools. Codex continues to use its workspace-write sandbox. No blanket permission-bypass flag is enabled. OS file permissions, managed policies, and provider-protected paths still apply; this does not grant administrator access or unrestricted writes outside the roots. Blocked actions require local intervention; there is no remote approval UI. Inside a run, Claude may attempt calls the policy forbids, such as network access or a write outside the roots; the CLI denies each one, tells the model, and the model continues. The reply is still delivered, the contract requires it to say what it could not do, and the local log lists every denied call with its tool and target (command or path, never file contents) so you can widen access deliberately. Codex enforces the same policy inside its own sandbox.

Only grant access to project directories you intend Slack participants to use. The provider sandboxes may permit reads beyond writable project directories and use temporary files; Fridica does not claim complete filesystem read isolation. Managed provider settings and project instructions remain part of the execution environment. Slack tokens are removed from agent subprocess environments, but do not store credentials in project files accessible to the agent.

Fridica supplies its own bounded conversation history for each invocation and, with resume_sessions, resumes only sessions it created; existing desktop conversations are not imported. The optional model setting is passed to the selected provider. No model name or paid API key is required by Fridica itself; each CLI uses its own authentication and billing.

Heavy tasks

Per-turn replies are one short CLI run each. With heavy_tasks = true, the reply agent may decide that a request needs long-running or hardware-heavy work (a full build, a long test suite, a GPU or multi-core job) and hand it to a persistent worker instead of doing it in the turn: it returns a self-contained brief in the structured reply's escalate field, names the host in escalate_host, and tells the requester that the job has started. Fridica then runs the brief on that host, locally or over SSH, and posts the worker's report as a follow-up message in the same Slack thread when it finishes.

Hosts. Every host that owns a root is a candidate: the workspace's own host (local, or its SSH alias) and every other alias among additional_workspaces. With file_access = true the local roots stay under scoped file access and are not a candidate, so at least one remote root is required and the reply agent hands work off with the planner's escalate operation (host name in path, brief in content). The agent sees each host's roots and [resources.<host>] hardware as data and picks the one the job needs; a worker on a host can only write inside that host's roots, and it starts in the first root listed for that host. With a local workspace and additional_workspaces = ["dart9:/mnt/data1/projects"], replies run on this machine and a GPU job runs on dart9 inside /mnt/data1/projects. A host the agent names that is not configured falls back to the first candidate: the workspace's own host, or the first remote host under file_access.

• With Codex the worker is codex app-server, a long-lived process speaking JSON-RPC over JSONL on stdin/stdout (OpenAI marks the command experimental). Fridica sends initialize, thread/start (or thread/resume for a thread it created earlier), and one turn/start per job, and reads the final agent message when the turn completes. With Claude it is claude -p --input-format stream-json --output-format stream-json with the same sandbox, permission, and tool flags as task runs.
• The worker keeps the per-turn policy: workspace-write sandbox, network access only when allowed_domains lists hosts, and no approvals (a host that declares GPUs is confined differently; see GPU access below). Fridica has no approval UI, so any approval request from the agent is declined and logged. One difference for Codex: codex app-server has no --ignore-user-config, so the owner's ~/.codex/config.toml on the working host, including any MCP servers it defines, applies to heavy jobs (the feature switches Fridica passes still turn off apps, plugins, hooks, browser and computer use). Keep that file minimal on a host that runs heavy tasks.
• One worker per Slack thread. The process stays alive between jobs and exits after heavy_task_idle seconds without work (default 1800); the backend thread it created is remembered in the state database, so the next job in that Slack thread resumes it in a fresh process. A job is bounded by heavy_task_timeout seconds (default 14400, four hours). While a job runs, the reply agent sees worker.state = "running" in its data and answers progress questions itself; a second brief is ignored until the first finishes.
• A failed or timed-out job posts a short notice in the thread and is not retried. A job cut off by restarting Fridica is reported as interrupted on the next start and is not resumed automatically.
• Anyone in an allowed channel can trigger hours of compute this way, which is why the setting is off by default. fridica doctor checks that the backend provides codex app-server or --input-format stream-json.

Declare the hardware heavy tasks may use per host, [resources.local] for this machine and [resources.<alias>] for each SSH host among the roots (a plain [resources] table describes the workspace's host):

[resources.dart9]
cpus = 8                          # sets OMP_NUM_THREADS for the worker
gpus = [0, 1]                     # device indices; sets CUDA_VISIBLE_DEVICES ([] = no GPU)
gpu_type = "NVIDIA A100 80GB"
memory_gb = 128
notes = "Jobs longer than 10 minutes go through Slurm: srun --gres=gpu:1."

The table is sent to the reply agent and to the worker as data, so the agent can judge what a request needs, and the worker process is started with matching OMP_NUM_THREADS and CUDA_VISIBLE_DEVICES (exported through SSH for a remote host). Nothing is measured or enforced beyond those variables; the notes are the place for site rules such as a batch scheduler.

GPU access. Both backends sandbox commands with bubblewrap, whose minimal /dev hides the GPU device nodes: inside their sandbox nvidia-smi cannot reach the driver and CUDA finds no device. Declaring gpus for a host therefore makes its heavy worker run inside Fridica's own bubblewrap instead: /dev is bound in full so CUDA works, the whole filesystem is visible read-only, and writes are allowed only in that host's designated roots, a private /tmp, and the backend's own state directory (whose settings and hook files stay read-only, so a job cannot plant anything that would run outside the confinement later). The backend's sandbox is turned off inside (Codex threads use danger-full-access, Claude runs with its sandbox disabled and Bash allowed) because Fridica's wrapper already confines the process. The wrapper shares the host's network: the CLI itself must reach the model API, and bubblewrap cannot separate that from the commands the job runs, so allowed_domains does not restrict a GPU worker's commands. Configured roots should be real directories rather than symlinks. This needs bwrap on that host and Linux; fridica doctor checks both. The per-turn replies keep their normal sandbox, and the reply agent is told that GPU work must be escalated to a GPU host. CUDA_VISIBLE_DEVICES still limits the devices. Set gpu_access = false for a host to keep the backend's sandbox there (and lose GPU access), or leave gpus out.

Network access

By default (allowed_domains = ["*"]) commands the agent runs may reach any host. With an empty list, git fetch, pip install, and similar calls are denied inside the run, the model is told, and the reply says what it could not do. To allow only specific hosts, list them:

allowed_domains = ["github.com", "*.pypi.org"]

Entries are host names, optionally with a leading *. wildcard, and are lower-cased. A single "*" entry allows every host, which is full internet access for task commands. They apply only to task runs; classification never has network access, and the model's own API traffic is unaffected. Restart the daemon after changing the list.

Backend	Effect of a non-empty list
Claude	The sandbox proxy admits outbound requests to the listed hosts only; anything else is denied inside the run. Traffic must pass through the proxy, so HTTPS remotes are the reliable choice; SSH remotes generally do not connect from inside the sandbox.
Codex	Codex cannot filter by host, so any entry enables full network access for Codex task commands (sandbox_workspace_write.network_access=true).

Network access lets a Slack request send workspace contents to the listed hosts and fetch code from them. List only hosts you trust, keep credentials out of the workspace roots, and remember that anyone in an allowed channel can trigger a run. Leave the list empty to keep the previous behavior.

Scoped file access

file_access = true replaces native agent tools with checked file operations on the local roots. It is the default whenever workspace is local. Remote roots in additional_workspaces are unaffected: they host heavy tasks with the agent's native tools, confined to their own roots. A config that enables heavy_tasks without any remote root leaves the local roots as the only place to run them, so it either falls back to the native tools when file_access is left out, or fails when it is set explicitly.

file_access = true
workspace = "/absolute/path/project/docs"
additional_workspaces = []
read_only_workspaces = ["/absolute/path/project/data"]

workspace and additional_workspaces are the maximum writable roots. All listed roots are readable, including their descendants; read-only roots always take precedence. Unlisted paths are refused. Keep the config, contract, state database, credentials, and Fridica's installed code outside these roots. Symlinks, hard links, special files, parent traversal, and .git, .codex, .claude, .ssh, and .env components are refused. Denials also cover case and Unicode normalization aliases, conservatively on case-sensitive systems.

Level	Operation	Authorization
L0	Read a listed text file	Automatic
L1	Create or replace a text file	Local approval, or a matching active sender/channel/path grant
L2	Delete a text file	New local approval for that exact request, every time

The model proposes operations with native tools disabled. The controller checks paths and permissions, saves the exact change in SQLite, then applies it. An existing file must match the content supplied to the model and the content reviewed locally. Writes use an atomic replacement; interrupted operations are never rerun automatically. Revocation affects operations not yet claimed for execution, and does not undo completed changes.

Manage requests from a local terminal or ask your Desktop agent to run these commands after reviewing the request. No Slack message can approve or grant access. Commands work while the daemon is running and return JSON:

fridica permissions status
fridica permissions status REQUEST_ID       # includes before and proposed contents
fridica permissions approve REQUEST_ID
fridica permissions reject REQUEST_ID
fridica permissions grant --sender U123ABC --channel C123ABC --path /absolute/path/project/docs --ttl 3600
fridica permissions revoke GRANT_ID

Each command accepts --config PATH. Omit --ttl for a grant that lasts until revoked. A grant permits L1 writes only; it cannot widen the configured roots or permit deletion. Approve an already pending request separately after reviewing it. Approved changes run when the daemon next processes its local queue, and results go back to the original Slack thread. Delivery retries reuse the saved result. Root changes require a daemon restart. Desktop remains a local control client; this does not attach a CLI session to a Desktop task.

This first version supports UTF-8 text files up to 64 KiB, existing parent directories, and one write or deletion per request. It does not execute shell commands, tests, merges, deployments, arbitrary sends, or directory operations. The model gets up to eight planning calls per message, each stateless; this mode does not resume native workspace sessions. With heavy_tasks, a planning call may instead return escalate, which starts a persistent worker on a remote host exactly as in the native mode; the local roots are never touched by that worker. File contents passed to the model and before/after contents saved in SQLite may be private: only allow projects appropriate for the selected channel, and protect the database accordingly.

Codex planning uses a named permissions profile that grants only minimal runtime reads and reads of its temporary invocation directory, with no command network access. It requires a CLI supporting named permissions and --strict-config; unsupported configuration fails instead of falling back to workspace mode. Claude planning uses its existing empty tool list. The provider CLI and local controller remain trusted processes with their normal authentication/runtime access; this is not isolation from a compromised CLI or another process running as the same OS user. No additional model API or billing fallback is introduced.

Local dashboard

The optional monitor runs separately from the Slack listener and makes no model calls. Start it in another terminal:

fridica dashboard --config ~/.config/fridica/config.toml --port 8877

Open http://127.0.0.1:8877. Add --allow-approvals to enable local controls. The terminal prints the path to a private .dashboard-key file; enter its contents under Settings → Local approvals. The key rotates on restart. Keep the monitor local: read-only views expose Slack history and paths, and recognized-token masking is not a general secret detector.

• Inbox / Requests: review tasks, conversations and file proposals; approve or reject an exact diff in managed file-access mode. The listener applies the decision and reports to Slack. Changed files require a fresh proposal.
• Projects & access: edit managed directories and per-person write grants. Repository labels do not grant access; grants do not permit deletion, Git commands or external actions.
• Settings: edit model, effort and conversation limits. Changes apply between requests. Channel, identity, credentials and backend require a config edit and restart. Configuration editing needs the --config path used above.
• Activity: archive older entries by date or restore them. Archiving hides entries from Current; it does not delete history or reclaim disk space.

The page refreshes while visible; auto-refresh can be disabled. Closing a tab leaves services running. Stop monitor stops only the monitor.

Task context and corrections

Task & handoff shows the repository, assignee, next step and blockers. Repositories come from the shared repos.toml; changing that list requires a PR and package upgrade. Task ownership does not grant access.

Claims link to source messages and remain unverified reports. Under Correct task or conclusion, record a replacement and its evidence. Local corrections take precedence over model updates and keep an audit trail. Saving requires the local key, a current revision and no related operation in flight; it neither grants permissions nor resumes execution.

Acknowledgments and repeated replies can stay silent. Three turns without recorded progress pause the task; this is a heuristic, not automatic fact checking. Blocked or paused threads make no further model calls or replies. Use Resume for future messages or Close request to stop the thread. Resume does not replay old messages. Continuation threads share task notes and the progress counter; a no-progress pause does not create a continuation.

Task notes are bookkeeping attached to a reply, never a reason to withhold it. Before a note is recorded, each field the model produced is checked and, where it cannot be salvaged, dropped: a repository name is matched to the shared list ignoring case, an assignee may be a member ID, a <@ID> mention, or a display name that the dashboard's name cache maps to exactly one member who has posted in the channel, and a claim must be an exact excerpt of a message in the task. Every drop or correction is logged locally with the event ID and reason, the reply is delivered unchanged, and the thread stays answerable. Only the fields that pass are saved. Before this rule, a display name in the assignee field replaced the whole reply with a "task update could not be validated" notice and blocked the thread.

Cleanup

Archive a finished or closed request, then use Preview cleanup to clear its stored messages, proposals and shared task notes. Related continuations must first be closed or archived. Cleanup is irreversible; deduplication IDs and decision records remain. It does not delete project files, Slack messages, CLI transcripts or backups, and is not a secure erase of SQLite journals.

Local state and recovery

State defaults to ~/.local/state/fridica/state.sqlite3; override state_path with an absolute path outside agent workspaces. It contains message text, task results, and delivery state, so treat it as private local data. A file lock prevents two processes from opening the same state database. Context sent to the model is bounded; stored history remains until local cleanup. There is no historical Slack backfill on startup.

Incoming events are persisted before acknowledgment. Agent runs are serialized, and their results are saved before Slack delivery. Rate-limited replies retry without rerunning the agent. Interrupted executions and uncertain deliveries are not retried automatically because file changes or Slack posts may already have occurred. Startup logs their event IDs. Inspect them locally:

sqlite3 ~/.local/state/fridica/state.sqlite3 \
  "SELECT event_id,state FROM events WHERE state IN ('interrupted','ambiguous','failed','blocked');"

Check the workspace and Slack thread before requesting work again in a new thread. Restarting cannot guarantee exactly-once execution across an external agent, filesystem, and Slack. Raw subprocess output and Slack tokens are not logged. Stop the daemon and use a separate state database when changing owner.

When resume_sessions is enabled, the providers also keep their own transcripts on disk: Claude under ~/.claude/projects/ and Codex under ~/.codex/sessions/. They contain the Slack text Fridica sent and the agent's tool activity, so treat them like the state database. Removing them is safe; the next turn in an affected thread starts a new session. Setting resume_sessions = false stops new transcripts from being written.

Python interfaces

fridica.models defines Message, ConversationContext, Decision, AgentResult, AgentBackend, and Transport. fridica.replica.Replica combines configuration, storage, a backend, and a transport and owns the rules of engagement. fridica.agents holds the Claude and Codex backends; they build on fridica.runner (bounded subprocess execution with a scrubbed environment), fridica.prompts (structured-output schemas and prompt composition), and fridica.checks (the local environment checks used by doctor and start). Alternative backends implement async classify(message, context) and async respond(message, context); transports implement async send(message, result, task_id, turn) and return the confirmed message timestamp. Backend responses contain text and a status of complete, waiting, or blocked. fridica.contract.load_contract(path) parses an agent contract into its participation and replies sections; custom backends should send the matching section as their instruction.

Validate

python -m pytest
node --test tests/dashboard.test.cjs
python -m build --no-isolation
fridica --help
python -m fridica --version

Before opening a pull request, run the contributor hooks once over the whole tree:

python -m pip install pre-commit
pre-commit run --all-files   # or `pre-commit install` to run them on every commit

They check file hygiene (whitespace, file endings, merge markers, valid TOML, YAML and JSON, leftover debugger calls) and lint with ruff's default rules as configured in pyproject.toml; no formatter is applied.

Frontend regression tests use Node 22 or later and its built-in test runner, with no npm dependencies. Tests use fake Slack clients and fake agent processes and require no tokens or live model calls. For a live smoke test, select one test channel and an empty project directory, run doctor, then run start --observe-only. Have another member post a message and verify the observation log. Restart normally and ask that member to mention you with a request to create a small text file. Check the file and threaded reply. Request a file without specifying its location to exercise clarification, then try a request outside configured write roots to verify blocked behavior. Repeat with the other backend. Live tests can consume provider credits and require your Slack installation and CLI login.

CI, releases, and deployment

The workflows follow snapy's CI → automatic tag → manual PyPI publishing flow, adapted for a pure-Python package. Fridica produces one universal wheel and one source distribution, rather than platform-specific compiled wheels.

• Continuous Integration (.github/workflows/ci.yml) runs on pull requests and pushes to main. It tests Python 3.11 on Ubuntu and macOS, builds both distributions after every matrix job passes, checks package metadata, and smoke-tests the installed wheel and bundled Slack manifest. Tests use fake agents and Slack clients; no Slack/model credentials are required.
• Auto Tag on PR Merge (cd.yml) tags the exact merge commit and creates a GitHub release. The first tag is v0.1.0; subsequent merges default to a patch bump. Add one of release:major, release:minor, or release:patch to select the increment. Multiple release labels fail the job. Rerunning an already tagged merge reuses its tag and repairs a missing GitHub release. Authentication uses the automatically supplied GITHUB_TOKEN, with contents: write permission limited to the tagging job. No GitHub App, private key, or personal access token is needed. Merged fork PRs are supported through a merged-only pull_request_target event; the checkout is verified to belong to main before release code runs. Tag jobs use GitHub's concurrency queue to run serially (up to 100 pending runs).
• Publish to PyPI (release.yml) is manually dispatched with an existing stable tag, such as v0.1.0. It verifies the tag belongs to main, reruns the full CI workflow on its resolved commit, checks that both artifact versions match the requested tag, then publishes those exact artifacts. Publishing does not run on every merge or tag push.

Versions come from Git tags using hatch-vcs. fridica.__version__ and the CLI read installed package metadata. Untagged/dirty checkouts produce development versions; reinstall an editable checkout after changing tags to refresh its installed version. Full Git history is fetched in CI. Source distributions carry version metadata so they also build without Git.

Repository maintainers must configure these GitHub settings before using CD:

1. Ensure repository/organization Actions policies allow the tagging job's GITHUB_TOKEN to have Contents: write permission. The workflow requests this explicitly; do not create a token secret. If tag rules restrict v* creation, configure them to allow this workflow's tag creation. The built-in token does not bypass repository rules. Existing BUMP_BOT_APP_ID and BUMP_BOT_PRIVATE_KEY settings are unused and can be removed from Fridica.
2. Create a GitHub Actions environment named pypi. Add PYPI_API_TOKEN as an environment secret using a PyPI account authorized to publish fridica. Configure required reviewers if publication needs an approval gate. A new PyPI project may need an account-scoped token for its first upload; replace it with a project-scoped token afterward.
3. Protect main and require CI before merging. Auto-tagging reacts to a merge, so branch protection supplies its CI gate. Publishing independently reruns all tests. Require the matrix test jobs and the package job as checks.
4. After a release tag exists, open Actions → Publish to PyPI → Run workflow on main, enter the tag, and approve the pypi environment if configured. PyPI versions are immutable; use a new tag for changed artifacts rather than overwriting a published version.

Tags and releases created using GITHUB_TOKEN do not trigger downstream tag-push or release-event workflows. Fridica's publishing workflow is manually dispatched and reruns CI itself, so it does not depend on those events. See GitHub's workflow-trigger rules.

Only the package is deployed by CI. Run the daemon on each owner's machine, where their Slack tokens, agent authentication, and project directories live:

python -m pip install --upgrade 'fridica==0.1.0'
fridica doctor
fridica start

Replace 0.1.0 with the published version. Stop the running daemon before an upgrade, then restart it in the same environment. No remote daemon, Slack app, GitHub secrets, or PyPI project is provisioned by installing these workflows.

For local workflow linting with actionlint 1.7.12, use actionlint -ignore 'unexpected key "queue" for "concurrency" section' .github/workflows/*.yml. That version's schema predates GitHub's documented queue field; the exception only suppresses that schema mismatch.