init

name	init
description	One-time platform setup for an operator workstation — km configure, km init, multi-instance resource_prefix isolation, Slack/Lambda bootstrap, and rollout sequences after sidecar/Lambda changes

Platform Initialization

This skill covers operator-workstation tasks that set up or upgrade a Klanker Maker install. Most are run once per environment; a few (km init --sidecars, make build) repeat whenever you edit km source.

Audience: the operator running km on their workstation. For agent-side workflows inside a sandbox, see klanker:sandbox, klanker:email, or klanker:slack.

Cross-references

• klanker:user — broad operator CLI tour (creating sandboxes, agent runs, lifecycle)
• klanker:vscode — VS Code Remote-SSH onboarding (its own one-time km init --sidecars requirement)
• klanker:cluster — k8s cross-account IRSA (its own provisioning flow)
• klanker:slack — agent-side Slack posting (after this skill bootstraps the bridge)

Step 1: Configure the install

km configure

Prompts for:

• resource_prefix (default km) — controls every per-install resource name ({prefix}-sandboxes table, {prefix}-sandbox-email SES rule set, etc.). Set this once; it cannot be safely changed later.
• email_subdomain (default sandboxes) — DNS subdomain for sandbox email addresses.

The chosen values are persisted to km-config.yaml and exported into every terragrunt invocation via KM_RESOURCE_PREFIX / KM_EMAIL_SUBDOMAIN. To reset the prefix back to km, use km configure --reset-prefix (this is the only way — re-running km configure preserves the existing value).

Step 2: Build the binary

make build

Always make build, never bare go build. The Makefile passes ldflags for the version string; missing them makes km --version lie and breaks the km init --sidecars upload path (S3 key embeds the version).

Step 3: Bootstrap or upgrade regional infrastructure

km init --plan             # Phase 84.2: real terragrunt plan per module + destroy-class safety gate (NEVER applies)
km init --plan --i-accept-destroys   # same plan, but clear exit code if only protected-type destroys are blocking (per-invocation; no auto-apply)
km init --dry-run=true     # static info dump — module ordering + skip annotations (does NOT run terragrunt plan)
km init --dry-run=false    # actually apply

Use --plan to see what terragrunt apply would actually change — it runs a real plan per module and trips on any destroy/replace of a curated protected resource type (Phase 84 incident protection). --dry-run=true is the older info-dump that lists module ordering and env-var skip annotations; it does NOT run terragrunt plan.

km init defaults to --dry-run=true. Forgetting --dry-run=false produces a no-op that looks like a deploy ran. After a successful apply, verify Lambda config picked up:

aws lambda get-function-configuration --function-name km-slack-bridge \
  --query '{MemorySize:MemorySize, Timeout:Timeout, Vars:Environment.Variables}'

Expect MemorySize=1024, Timeout=60, and Vars containing KM_ARTIFACTS_BUCKET. A Vars map with only TOKEN_ROTATION_TS means km slack rotate-token blew away the env vars and Terraform hasn't re-applied — re-run km init --dry-run=false.

Fast-path variants

km init --sidecars     # rebuild + upload sidecar binaries (km-slack, km-presence, audit-log, etc.) to S3
                       # AND refresh the management Lambda's bundled km binary so new schema fields are recognized
km init --lambdas      # build Lambda zips locally + bump create-handler env stamp
                       # NOTE: --lambdas does NOT actually deploy bridge Lambdas via terragrunt
                       # Use plain `km init --dry-run=false` for full bridge deploy

When to run km init --sidecars:

• After editing any code under sidecars/ (km-slack, km-presence, audit-log, http-proxy, dns-proxy, tracing)
• After adding a profile schema field that the management Lambda must understand (otherwise remote km create fails on unknown field)
• After every km CLI version bump that ships new sidecar binaries

Existing sandboxes do NOT get new sidecars retroactively. km destroy && km create to roll forward.

Step 4: Bootstrap Slack (optional)

If sandboxes will post to Slack:

km slack init                    # interactive bootstrap
# Or non-interactive:
km slack init --bot-token xoxb-... --invite-email [e-mail]

km slack init validates the bot token (calls auth.test), writes SSM parameters, creates the shared channel, sends a Slack Connect invite to --invite-email, and deploys the km-slack-bridge Lambda.

Add inbound event webhook

km slack init --force --signing-secret <signing-secret-from-slack-app>
# Then paste the printed Events URL into:
#   Slack App → Event Subscriptions → Request URL

The signing secret enables HMAC-SHA256 verification of inbound /events webhooks (Slack message events → SQS FIFO → sandbox poller). Required for bidirectional chat in per-sandbox channels.

Slack rotations

km slack rotate-token --bot-token <new-token>             # validate, persist to SSM, force bridge cold-start, smoke test
km slack rotate-signing-secret --signing-secret <new>     # rotate signing secret

After rotate-token, re-run km init --dry-run=false to restore Lambda env vars (the rotation force-cold-starts but does not re-apply Terraform env, so KM_ARTIFACTS_BUCKET etc. can be lost).

Required Slack bot scopes

Scope	Used for
chat:write	base posting
channels:history, groups:history	inbound message events
reactions:write	ACK reaction (👀) on inbound messages
files:write	transcript uploads
files:read	inbound file attachments (downloaded by the bridge)

After adding scopes to the Slack App, re-install the app and km slack rotate-token with the new token.

Slack Connect channels

Per-sandbox channels are externally shared (is_ext_shared: true) via Slack Connect (conversations.inviteShared). This requires a Pro Slack workspace (free tier returns not_allowed_token_type).

Slack Connect channels reject final transcript file uploads (files.completeUploadExternal returns silent internal_error). Per-turn chat lines + auto-thread + DDB record-mapping still work; pull the .jsonl.gz transcript from S3 instead:

aws s3 ls s3://${KM_ARTIFACTS_BUCKET}/transcripts/<sandbox-id>/

Step 5: Verify health

km doctor
km doctor --all-regions     # if you have sandboxes in multiple regions

km doctor runs ~22 checks. Critical ones after init:

Check	Means
presence_daemon_healthy	each running sandbox emitted a source:"presence" event in the last 5 minutes
slack_inbound_queue_exists	every inbound-enabled sandbox has a healthy SQS queue
slack_inbound_stale_queues	orphan SQS queues with no DDB sandbox row
slack_app_events_subscription	bot has channels:history + groups:history + reactions:write
slack_transcript_table_exists	DDB table provisioned + ACTIVE
slack_files_write_scope / slack_files_read_scope	bot has the required file scopes

Multi-instance support

A single AWS account can host multiple km installs distinguished by resource_prefix (default km, common alternates: rg, dev, staging). Every per-install resource carries a km:resource-prefix=${prefix} tag.

Second install on the same account

# 1. Configure with a distinct prefix
export KM_RESOURCE_PREFIX=rg
export KM_EMAIL_SUBDOMAIN=rg-sandboxes
km configure   # accept the prefix/subdomain prompts

# 2. CRITICAL — do not steal SES rule-set activation from the primary install
export KM_SES_ACTIVATE_RULESET=false

# 3. Build + init
make build
km init --sidecars
km init --plan            # verify plan: no aws_ses_active_receipt_rule_set in destroy list
km init --dry-run=false

AWS SES allows only ONE active receipt rule set per account/region. The second install's rule set is created but not activated — the primary install keeps inbound email. To hand off inbound email to the second install (a deliberate operator action):

aws ses list-receipt-rule-sets --query 'RuleSets[*].Name'
aws ses set-active-receipt-rule-set --rule-set-name rg-sandbox-email

Rollback (restore primary):

aws ses set-active-receipt-rule-set --rule-set-name km-sandbox-email

Tag backfill for pre-isolation installs

AWS_DEFAULT_REGION=us-east-1 AWS_PROFILE=<your-profile> \
  km doctor --backfill-tags --dry-run=true     # preview

AWS_DEFAULT_REGION=us-east-1 AWS_PROFILE=<your-profile> \
  km doctor --backfill-tags --dry-run=false    # apply

Idempotent. Second run reports Tagged: 0. Required when running outside km configure context because the tag emitter reads AWS_DEFAULT_REGION + AWS_PROFILE directly.

Rollout sequence template

When you edit km source (CLI, sidecar, userdata template, or Lambda code):

make build                                  # always first
km init --sidecars                          # if sidecars/* or userdata template changed
km init --dry-run=false                     # if Terraform modules or Lambda code changed
km doctor                                   # confirm health
# Existing sandboxes do NOT get new sidecars — km destroy && km create to roll forward

km init --lambdas alone is insufficient for bridge Lambda deploys — it builds the zip locally but does not upload it via terragrunt. Use km init --dry-run=false.

Common gotchas

• km configure no longer overwrites resource_prefix on re-run. To reset, use km configure --reset-prefix.
• systemd EnvironmentFile gotcha: systemd does NOT accept shell-style export VAR=val lines. The userdata template writes a parallel /etc/km/notify.env (no export prefix) and points systemd units at it. /etc/profile.d/km-notify-env.sh remains the source of truth for shell sessions.
• Audit pipe on resumed sandboxes: /run is tmpfs and is wiped on every boot; cloud-init does NOT re-run on second boot. /usr/lib/tmpfiles.d/km.conf uses p+ to recreate /run/km/audit-pipe at every boot, before sysinit.target. Pre-fix sandboxes that have already tripped the bug can be recovered manually:

  km shell <sandbox-id>
  sudo rm /run/km/audit-pipe && \
    sudo mkfifo /run/km/audit-pipe && \
    sudo chown km-sidecar:km-sidecar /run/km/audit-pipe && \
    sudo chmod 666 /run/km/audit-pipe && \
    sudo systemctl restart km-audit-log
  

• Stale Anthropic OAuth tokens (noBedrock: true profiles with expired ~/.claude/.credentials.json) produce api_error_status: 401 in claude -p output. Fix: km shell → claude login.

Teardown

Two-layer, must be in this order:

km uninit         # regional infra (Lambda, DynamoDB, SES, VPC)
km unbootstrap    # foundation (S3 state bucket — destructive)

Always km uninit first; km unbootstrap deletes the state bucket and any orphaned regional state becomes unrecoverable.

See OPERATOR-GUIDE.md for the full operator runbook.