Company OS Workspace And AgentChat Launch

This is the shortest path to a working Company OS workspace with AgentChat.

For a customer install, use the published installer. It installs the stack and, on macOS arm64, the desktop only after its verified release manifest is published:

curl -fsSL https://raw.githubusercontent.com/HSM-II/company-os-install/main/install.sh | bash

The same installer accepts --upgrade, --rollback, --uninstall, and --purge. Uninstall preserves Company OS data; purge explicitly removes it.

For a source checkout, the complete normal path is one command:

cd /path/to/hyper-stigmergic-morphogenesisII
./company-os

It starts or reuses the local stack, waits for the API and AgentChat runtime, opens the Workspace start page, and prints exactly one clickable URL. Source startup that builds the Rust runtime allows up to 10 minutes for readiness; prebuilt startup allows 30 seconds. HSM_AGENTCHAT_WAIT_SECS overrides either window.

company-os: http://127.0.0.1:3050/workspace/start

scripts/company-os-up.sh is the internal low-level supervisor, not the documented user command. ./company-os owns its up, status, and down lifecycle; its installer verbs delegate to the same contract used by the published installer.

The checkout launcher is a development path: it still needs local Rust, Node/npm, and reachable Postgres (Docker can supply it), plus model access for executing provider-backed turns. It keeps those details inside the launcher; if one is unavailable, it emits the single company-os failed and log block instead of asking for a port, URL, or copied token. A bundled public CLI has not yet replaced this source-development dependency boundary.

What Starts

When defaults apply:

Surface	URL
Company Console / Workspace	http://127.0.0.1:3050/workspace/start
Company OS API (hsm_console)	http://127.0.0.1:3847
API health	http://127.0.0.1:3847/api/company/health
AgentChat runtime health	http://127.0.0.1:3050/api/agent-chat-runtime

Ports can be changed with HSM_COMPANY_CONSOLE_PORT and HSM_CONSOLE_PORT.

Desktop from a source checkout

./company-os desktop

This command starts or reuses the stack through the canonical Zeron desktop, waits up to 30 seconds for the API, AgentChat health route, and Workspace page, then opens Company OS.app and the ready Workspace URL in your browser. It prints that same URL so it is easy to reopen. If the Workspace is unavailable, the command reports the failed readiness check and log path instead of sending you to an API health page. A source checkout silently builds and packages a local preview app when one is not installed; the signed public desktop release remains gated until its release manifest is published. The source-preview bundler currently supports macOS on Apple silicon only.

Desktop AgentChat uses the same Company OS skill-routing and governed turn path as the Company Console. Open JEV ranks trusted Company Skills and can abstain; AgentChat performs the work. Recommendation outcomes are recorded as preliminary evidence for later verified learning, not treated as proof that a skill caused success. Skill selection remains advisory and does not grant tool authority.

Terminal chat with OMP

From this checkout, ./company-os chat opens the installed OMP TUI in the current directory after the host admits a signed, company-bound run. The machine-local companyos shortcut can point to this command for use in Tern. Use ./company-os login --company ID when the CLI session lapses; the browser is only the authorization handshake. ./company-os chat --prompt TEXT runs one turn and exits. The terminal run records model, tool, approval, denial, and settlement evidence in the host ledger.

./company-os connect omp installs a read-only MCP profile for manual OMP sessions. A bare omp session using that profile is not a governed Company OS run. ./company-os tui remains the Zeron terminal projection, and ./company-os desktop opens Zeron itself.

For Tern's native command palette, key bindings, Workspace link, and explicit OMP setup/diagnostics, see the Company OS Tern plugin.

Lifecycle and advanced flags

./company-os

Useful flags:

Command	What it does
./company-os status	Prints the ready workspace URL when API, AgentChat, and Workspace are healthy
./company-os down	Stops checkout-owned Company OS listeners
./company-os up --no-open	Starts or reuses the stack without opening the browser
./company-os up --force-restart	Stops checkout-owned stale listeners before launching
./company-os up --smoke	Starts or reuses the stack, then runs the live AgentChat smoke

The launcher refuses to kill non-repo processes that happen to be using the same ports. If that happens, stop that process yourself or choose different ports.

Normal Product Path

After launch, use /workspace/start:

1. Create or select a company.
2. Choose a model lane.
3. Run the proof demo.
4. Use AgentChat from the Workspace.

Use Lab Mode only when you want advanced runtime, harness, and evaluation surfaces.

Remote host recovery

If a remote agent needs to recover a Mac-hosted Company OS while the Console proxy is returning 502, install the independently supervised host-recovery boundary described in services/company-os-host-recovery/README.md. It exposes one fixed, bearer-protected POST /recover action on port 3070; it is not an arbitrary shell bridge. The action starts the canonical AgentChat launcher (which delegates to scripts/company-os-up.sh) and verifies the local Console and AgentChat health routes before reporting success.

Optional checks

curl -sfS http://127.0.0.1:3847/api/company/health
curl -sfS http://127.0.0.1:3050/api/agent-chat-runtime
bash scripts/company-os-harness.sh smoke

/api/company/health answers 503 (body still populated, probe_ok:false) when Postgres is unreachable or a registered supervised worker is stale or stopped, so curl -f, the compose healthcheck, and the fly probe all fail together instead of reporting a process that cannot reach its ledger as alive. degraded supervision (a worker starting or one failed tick) and a deployment with no optional workers keep 200.

More context:

• Operator guide: docs/company-os/HSM_II_COMPANY_OS_OPERATOR_GUIDE.md
• Harness guide: config/COMPANY_OS_HARNESS.md
• Eval/meta-harness: docs/EVAL_AND_META_HARNESS.md
• Product truth map: config/company-os-truth.json

Company Skill selection

AgentChat asks Open JEV to suggest one trusted Company Skill before a responding turn by default (recommend). The composer exposes Off, Suggest, and Use when confident; the last option (auto_load) adds only the selected, version- and content-hash-verified skill guidance to that turn's prompt. Configure the local Open JEV provider as described in JEV_JUDGE_PROVIDER.md.

This selector is advisory: it does not add tools or authority, and if it is unavailable AgentChat still handles the request normally.

Block To Paste Into An LLM

Task: Start HSM-II Company OS so Workspace and AgentChat are usable.
Repository: hyper-stigmergic-morphogenesisII. Clone path = REPO_ROOT.

1. cd REPO_ROOT
2. Run: ./company-os
3. Use the one Workspace URL it prints. Do not start another terminal or ask
   for ports, tokens, or copied URLs. If it emits a failure/log block, diagnose
   only the named missing dependency or service.

Use ./company-os status for a ready URL and ./company-os down to stop
checkout-owned services. Do not invent alternate entrypoints;
scripts/company-os-up.sh is an internal implementation detail.

Small Local Models

Small local models are supported, but measured performance matters. A previous 8-task meta-harness run on qwen2.5:1.5b through Ollama finalized 6/8 tasks correctly. The failures were vague planning prompts where the small model produced narrative text instead of structured tool calls; the verifier rejected optimistic completion.

Example local-model launch:

HSM_LLM_PROVIDER_ORDER=ollama \
DEFAULT_LLM_MODEL=qwen2.5:1.5b \
HSM_AGENT_CHAT_PROVIDER=ollama \
OLLAMA_MODEL=qwen2.5:1.5b \
./company-os