Public: anyone can read this; the crew can change it.

"How is my forge?" has one answer everywhere: the same needs-you stack as the… #12183

shipped production is serving it opened this

Landed in pull #12186 · run

opened this

"How is my forge?" has one answer everywhere: the same needs-you stack as the home page, each with a recommended choice, and one call that makes the recommended choices

The owner, 2026-11-13: "the top of my homescreen, there is a red component with things i should do next. the current one is the app template. this stack of items should easily accessible via mcp so any agent can get helping right away. like if just ask how is my forge? i should see these items and i should be able to say just make the recommended choices and it gets them done".

Today: the home page's needs-you line has its own source; the asks tool returns what waits on a person (37 rows now, some 11 days old) with options like ["Ship","Say what to change","Close it"] but NO recommendation, and it is not among the 13 tools an agent sees by default (it must be found through find_tool); health returns ok:false with operator chores ("no drill has ever run… bun scripts/drills/run.ts …") that are a different list again; now shows runs. An agent asked "how is my forge?" has to call three tools, gets three lists, and can act on none of them in one step.

What should be true:

  • ONE stack. The home page's needs-you component, the asks page, the TUI and MCP read the same function: an ordered list of what needs the owner, most consequential first, each item: a plain sentence, why it matters (what waits behind it), the choices that exist, and the RECOMMENDED choice with a one-line reason and how reversible it is. The count on home is the length of this list (a binding: declare it so the bindings harness checks it).
  • One default tool answers the question. Either now grows it or a tool named for the question (forge / how: pick the plainest; it must be in the default tool list and the first thing contribute/llms.txt tells an agent to call): { says: "one or two sentences: is it serving, is the crew moving, what is unusual", needs: [...the stack...], recommended: "<id of the single call that applies them>" }. Under ~1.5k tokens for a typical forge: titles and choices, links for depth.
  • "Make the recommended choices": one call (decide {items: "recommended"} or per item {item, choice:"recommended"}) applies the recommended choice for every item whose recommendation is safe to take without a person: reversible and inside what this forge's owner has already allowed (land a green, approved, non-conflicting pull; close a pull whose change already landed; retry a parked run; dismiss a stale notice; wake a stopped app). Items that are not safe (anything that deletes, spends above a stated amount, changes access, touches another person's forge, or is a product decision) are NOT applied: they come back listed with the question to ask the owner. Dry run by default for agents, with the plan row by row; the receipt names what was done and gives the undo where one exists (the fleet bulk actions built for the home table already preview, apply and keep receipts: reuse them).
  • Recommendations are honest: derived from state (checks, approvals, conflicts, age, what waits behind), never a default "Ship"; when the forge cannot tell, the recommendation is "ask the owner" with the question.
  • The same works per app (repo) and for a hosted forge's owner on their own forge.

Done when: an agent with the default tools calls one tool and gets the stack the home page shows (a test renders home and calls the tool on the same state and compares ids and order); "apply recommended" on a seeded forge lands the green approved pull, closes the already-landed one, retries the parked run, leaves the destructive and the product-decision items untouched and says so; the receipt's undo restores the reversible ones; asks, health's needs and home no longer disagree (health keeps "serving or not" only: #12071).

commented

Parked by the crew: the check stayed red through 2 automatic tries (pull #12186 closed; its branch scout/12183-how-is-my-forge-has-one-answer-everywher at 925109fae38e keeps the work so far). Planner: split this smaller, or rewrite what done looks like; a person can also say what to try.

try 1: docs/MCP.md, docs/features/forge-needs.md, src/actions/needs.ts, src/http/mcp/catalog.ts, src/http/mcp/tools/forge.ts, src/http/mcp/tools/index.ts, test… try 2: src/actions/needs.ts, test/forge-needs-one-stack.test.ts try 3: docs/MCP.md, src/actions/needs.ts, src/apps/egress.ts, src/core/capacity.ts, src/spawn/origin_broker.ts, test/boundary-outbound.test.ts, test/conversati…

Sign in to join this conversation.

↑↓ navigate ↵ open ⌘↵ its app esc close

Keyboard shortcuts

Docs