phrazzld /factory-apps
Use the owned factory app before inventing local state, bespoke glue, or a
generic third-party workflow. Product repos own the concrete skills and MCP
servers; Harness Kit imports those skills under misty-* aliases and manages
MCP policy in .harness-kit/factory-mcps.yaml.
Router
| Need | App | First surface | Fallback |
|---|---|---|---|
| uptime, incidents, error timelines, health checks, service evidence, production debugging | Canary | Canary MCP from the global profile when registered |
misty-canary, /Users/phaedrus/Development/canary/bin/canary, API |
| backlog, issue cards, claims, relations, operator input requests, work status | Powder | Powder MCP from the non-adminifi-non-r90 profile when configured |
misty-powder, CLI, API |
| release intelligence, versions, changelogs, release notes, release kit, fleet adoption | Landmark | misty-landmark and landmark describe --json / dry-run CLI/action paths |
docs/agent-integration.md, docs/fleet-integration-playbook.md |
| UI/UX, Misty Step design law, tokens, static design registry, rendered design gate | Aesthetic | misty-aesthetic, @misty-step/aesthetic package, static API, law gate |
docs/ADOPTING.md, DESIGN.md |
| ad-hoc supervised dispatch, event-triggered agents, reflex loops, durable runs | Bitterblossom | misty-bitterblossom, bb CLI, read-only MCP from factory-ops profile |
product plane config |
Operating Rule
- Production debugging starts with Canary state. Query service health, incidents, checks, and recent errors before making a repo-local hypothesis.
- Backlog or issue state lives in Powder. Do not keep durable card state in chat, TODO prose, or an ad-hoc markdown list when Powder is available.
- Release questions start with Landmark. Do not hand-write release intelligence from memory when the release app can describe the repo.
- UI and artifact design starts with Aesthetic. Use its tokens, recipes, registry, and law gate before adding one-off CSS vocabulary.
- Dispatch architecture starts with Bitterblossom only when the work is Mode B: triggered, scheduled, durable, reflexive, or event-driven. Ad-hoc operator work remains Harness Kit / Mode A.
Current Audit
The live 2026-07-03 audit is in
references/capability-audit-2026-07-03.md. Load it when the question is
"are these configured?" or "what gaps remain?" before changing product repos
or system config.
Fleet Integration Standard
For active Misty Step repos, load
references/fleet-integration-standard.md before claiming a project is
factory-integrated. The standard defines the repo-level Canary receipt,
Powder backlog, and Landmark manifest/workflow evidence expected for runtime
apps, libraries, and non-release support repos.
Gotchas
- A product repo having an MCP implementation does not mean this harness has
that MCP registered. Check
.harness-kit/factory-mcps.yamland the active harness config before claiming MCP availability. - Do not add placeholder MCP servers. A broken registered tool is worse than a clear CLI/API fallback.
- Root product skills (
SKILL.md) and portable product skills under<repo>/skills/<name>/SKILL.mdare for consumers of the app. Repo-local.agents/skills/*are usually QA/deploy/dogfood runbooks for work inside that repo. Do not treat one as a substitute for the other. - Bitterblossom's MCP is read-only in the audited checkout. Mutating dispatch and run control still go through the CLI/API unless the product changes.