Deepsky Sustain CLI

Use deepsky sustain as an observe -> decide -> act toolset. Keep planning in the agent, not in the CLI.

CLI Preflight

Before using any sustain command:

Check whether deepsky sustain is already available.
If the command is missing or unusable, install the published CLI globally:

npm install -g @superise/deepsky-cli
deepsky sustain --help

If the install fails, report the exact npm error and stop instead of pretending the CLI is available.

Do not assume a local workspace checkout is already wired into PATH. The supported fallback is the published npm package.

Boundary

Use only the sustain commands this package actually implements.
Do not assume legacy commands such as set-model or mcp-server exist.
The only sustain setup flow restored here is deepsky sustain setup openclaw.
OpenClaw provider bootstrap is separate: use deepsky setup openclaw when OpenClaw still needs the Deepsky provider, wallet prerequisite handling, or repository skill installation.
Prefer explicit sustain commands over ad hoc wallet transfers for recharge.
Report wallet-side auth dependencies clearly instead of inventing local workarounds.

Units And Money Semantics

Do not mix up account balance and recharge input units.

health-check.balance is the platform-side account balance. Treat it as the platform's USD-like balance or credits, not as CKB.
criticalBalance and lowBalance are thresholds on that platform-side balance, not on wallet CKB.
top-up <amount> always takes a CKB amount as input.
exchangeAmount from order creation or top-up results is the platform-side credited amount after exchange-rate conversion, not the requested CKB amount.
newBalance is again the platform-side balance after the credited amount is applied.

When reporting outcomes:

State the requested top-up amount in CKB.
State the credited or exchanged amount separately when exchangeAmount is available.
Do not describe exchangeAmount as "only X CKB arrived". It is not a CKB field.
If the wallet rejects or limits a transfer, quote the exact wallet error instead of inferring a hidden local limit.

Autonomous Delegation

If the user says anything like:

"keep yourself alive"
"self-supervise"
"manage your own balance"
"watch your account"
"你自己维持余额"
"你自己监控并充值"

treat that as delegation of ongoing sustain responsibility.

When that happens, do not stop at explaining commands. Start acting:

Run the CLI preflight first.
Gather the current state with health-check, forecast, and config show.
Make an immediate decision based on balance, runway, and the user's risk tolerance.
If ongoing supervision is wanted, register recurring checks with deepsky sustain setup openclaw unless the user explicitly prefers app automation.
Report what you decided and why.

The CLI does not do the planning for you. The agent remains responsible for observe -> decide -> act behavior.

Natural Language Triggers

Map intent to commands like this:

"Check your balance" -> deepsky sustain health-check --json
"How long can you keep running?" -> deepsky sustain forecast --json
"Show models" -> deepsky sustain list-models --json
"Show sustain settings" -> deepsky sustain config show
"Top up 3000 CKB" -> deepsky sustain top-up 3000
"Dry-run a recharge" -> deepsky sustain top-up <amount> --dry-run --json
"Retry failed top-up orders" -> deepsky sustain retry-orders --json
"Configure OpenClaw for Deepsky" -> deepsky setup openclaw
"Bootstrap OpenClaw provider" -> deepsky setup openclaw
"Set up recurring sustain checks" -> deepsky sustain setup openclaw
"Install sustain cron jobs" -> deepsky sustain setup openclaw
"Clean OpenClaw setup" -> deepsky clean openclaw
"Remove the Deepsky provider only" -> deepsky clean openclaw --provider-only
"Remove sustain cron jobs only" -> deepsky clean openclaw --jobs-only
"Clear your login" -> deepsky sustain logout

Observe

Start any survival task by gathering state:

deepsky sustain --help
deepsky sustain health-check --json
deepsky sustain forecast --json
deepsky sustain list-models --json
deepsky sustain config show

Use --json whenever the result will drive a follow-up decision.

Act

Use these commands to change state:

deepsky setup openclaw
deepsky setup openclaw --defaults
deepsky setup openclaw --api-key <key>
deepsky sustain top-up <amount>
deepsky sustain top-up <amount> --json
deepsky sustain retry-orders --json
deepsky sustain setup openclaw
deepsky sustain setup openclaw --json
deepsky sustain setup openclaw --tick-every 20m --retry-every 10m --session isolated
deepsky clean openclaw
deepsky clean openclaw --provider-only
deepsky clean openclaw --jobs-only
deepsky sustain config set <key> <value>
deepsky sustain config unset <key>
deepsky sustain config reset
deepsky sustain logout

Follow these rules:

Use deepsky setup openclaw when the user needs the Deepsky OpenClaw provider configured before sustain automation can work.
That top-level setup now treats the wallet as a prerequisite: with the default local wallet MCP URL it auto-installs or auto-starts the SupeRISE Agent Wallet before configuring OpenClaw.
For non-default wallet MCP URLs, deepsky setup openclaw expects the configured wallet health endpoint to already be reachable and fails early when it is not.
deepsky setup openclaw --defaults runs non-interactively with defaults, keeps the current primary model unchanged, and reminds the user to switch the OpenClaw primary model to Deepsky manually.
Top-level setup also runs Install skills, which silently installs all skills from both https://github.com/appfi5/deepsky-tools.git and https://github.com/appfi5/superise-for-agent in global copy mode.
Choose a top-up amount based on the user's intent and current balance posture.
Let wallet-side policy enforce transfer limits. If the wallet rejects the amount, report the wallet error directly.
Use top-up for recharge because it creates the order, transfers CKB, and submits the tx hash as one flow.
Use retry-orders when transfer likely succeeded but platform-side submission or finalization did not complete cleanly.
Use deepsky clean openclaw when the user wants to remove the Deepsky provider config, Deepsky sustain jobs, or both.
Use logout to clear stale local auth state.
Keep requested CKB, credited platform amount, and resulting balance separate in your reasoning and reporting.

Survival Loop

When the user delegates self-supervision:

Run the CLI preflight.
Run deepsky sustain health-check --json.
Run deepsky sustain forecast --json.
If balance is low, run deepsky sustain list-models --json.
If OpenClaw is not yet configured for Deepsky, run deepsky setup openclaw first.
If recurring supervision is expected, run deepsky sustain setup openclaw.
If balance is critical or runway is too short, choose an amount and run deepsky sustain top-up <amount>.
If the wallet rejects the amount, report the wallet-side limit or policy instead of inventing local rules.
If a recharge looked partial, run deepsky sustain retry-orders --json.
Report the decision and the reason.

Threshold Strategy

Unless the user overrides policy, use the configured thresholds like this:

healthy: balance is above lowBalance. Stay quiet. Do not top up unless there is some unusual context the user explicitly cares about.
low: balance is at or below lowBalance but still above criticalBalance. Treat this as a preventive-action zone. Check forecast, inspect pricing context if useful, and usually decide for yourself whether a preventive top-up reduces interruption risk. Prefer self-handling over asking the user.
critical: balance is at or below criticalBalance. Treat this as immediate-action territory. Choose a top-up amount and execute it by default so the account recovers before work is interrupted.

Extra rules:

Compare thresholds against platform balance only, never against wallet CKB.
A recharge request is expressed in CKB, but the success criterion is whether account balance recovers after exchange-rate conversion.
If the wallet rejects the requested CKB amount, surface the wallet rejection exactly and adjust from there instead of inventing a local max.
If forecast shows the runway is too short, treat that as justification to act more aggressively even if the current status is only low.
Only interrupt the user when autonomy is blocked: wallet failures, platform failures, authentication failures, or manual-review situations.

Sustain Config For Decisions

The agent should not invent hidden sustain policy. Inspect the configured thresholds first:

deepsky sustain config show
deepsky sustain config get criticalBalance
deepsky sustain config get lowBalance
deepsky sustain config get requestTimeoutMs

If the user gives an explicit policy like "treat anything below 20 as critical", update config:

deepsky sustain config set criticalBalance 20
deepsky sustain config set lowBalance 100

Do not invent local recharge bounds. Wallet-side policy is authoritative for transfer limits.

How To Choose A Top-Up Amount

The amount is still the agent's decision unless the user gave one explicitly.

Use:

current balance from deepsky sustain health-check --json
runway from deepsky sustain forecast --json
the user's current activity level and tolerance for interruption
any wallet-side limit or rejection message returned during recharge

Guidelines:

Do not always choose the smallest possible amount.
Prefer enough runway for the user's near-term usage.
If the user is actively relying on the agent, bias toward fewer future interruptions.
If the wallet rejects the amount, surface the exact rejection and adjust from there.
Remember that the top-up input is CKB while the target outcome is platform-side balance recovery after conversion.
When in doubt, choose the amount that reduces the chance of another near-term interruption rather than the amount that minimizes immediate spend.

Scheduled Runs

When the user asks for recurring monitoring, automatic keepalive, or a timed sustain loop:

Prefer deepsky sustain setup openclaw for one-click recurring sustain setup when OpenClaw is available.
Before sustain scheduling, use deepsky setup openclaw if the Deepsky OpenClaw provider has not been configured yet.
The default sustain setup installs the keepalive review loop. It starts at 20m, then retunes to 2h when healthy, 1h when low, and 20m when critical.
The retry-orders loop is scheduled only when a top-up enters pending-retry state, and is removed once pending orders are cleared or escalated to manual review.
Use --tick-every for a different initial health-check cadence, and --retry-every or --session if the user asks for a different retry cadence or OpenClaw target when retry scheduling is needed.
Keep the scheduled loop focused on observe -> decide -> act, not on ad hoc wallet transfers.
Only rely on automatic top-up inside the scheduled loop when the user explicitly delegated autonomous recharge.
If OpenClaw is unavailable or the user explicitly wants app-managed scheduling, fall back to app automation instead.

Default OpenClaw setup:

deepsky sustain setup openclaw

Customized OpenClaw setup:

deepsky sustain setup openclaw --retry-every 30m --session isolated

If the user explicitly asks for --session main, still use setup openclaw, but remember that current OpenClaw requires main-session jobs to be registered as system events instead of announced chat turns.

Reporting Language

When you describe sustain work to the user:

Call it a sustain check or survival review.
Report the decision, balance, requested CKB amount, credited amount, and next action clearly.
Prefer concrete outcomes over internal deliberation.
Default to quiet success. Only send an update when you took a material action, hit an exception, or need human help.

Auth Assumption

marketPublicKey is configured explicitly, or
wallet MCP exposes Nervos public identity through nervos.identity, including address and publicKey

If wallet MCP does not provide that contract yet, stop and report the wallet-side dependency.

deepsky-sustain