Todos

A PERSON's outstanding items — the human half of memory, as against beans/, which is the agent's workflow management.

Stickies 2

computable adjudication and agentic test harness

Computable adjudication and agentic test harness.

Also written: caaat-harness, ca&at-harness, .c&at-harness, c@t-harness

Please be introduced to a cat who acquires things, for whatever purpose — maybe somebody knows.

Scope: the platform — skills, schemas, processes, the MCP server and the pipeline. Not what a folio IS (that is folio-assist-core), and not any folio’s own content, which lives in its own repository.

Before you know what this repository is

The graph an agent reads before it knows what this repository is.

Scope: phase one of two, and valid with NOTHING installed — no skills, no knowledge graph, no MCP, no work-plan store. One process here is the place to start. Everything else is composed on top of it, and nothing here may import from the layer above.

Feedback — 0 todo(s)

No feedback todo recorded in this graph.

Boards — 11 board(s)

idtitleshowsnotes placed
bootstrap-toolsBootstrap toolsthe whole folio0
cat-harness-toolsC@T Harness Toolsthe whole folio0
cat-harnessC@T Harnessthe whole folio0
fhir-harnessFHIR IG Harnessthe whole folio0
folio-assistant-coreFolio Assistant Corethe whole folio0
folio-assistant-scifolio-assistant-scithe whole folio0
folio-assistantFolio Assistantthe whole folio0
smart-baseSMART Basethe whole folio0
smart-immunizationsWHO SMART Guidelines — Immunizationsthe whole folio0
smart-trustWHO SMART Trustthe whole folio0
who-irisWHO IRISthe whole folio0

No board-positions.json in this graph: no note has been placed on any board.


Open notes (3)

Every note in this folio, in document order. This listing is the artefact; the board is an overlay over it.

  1. The human-todos page still says 'Not built yet' — the store now exists

    Status
    open
    Priority
    high
    Raised by
    agent
    Created
    2026-09-19
    About
    guides/beans-and-todos › human-todos

    content/docs/guides-beans-and-todos/human-todos.md opens **"Not built yet"**, and that is no longer true. todos/ is a declared graph with items on disk, a reader, a published JSON index and a board on the site.

    This todo is attached to that block, so it renders beside the sentence it contradicts. That is the per-page assignment working on a real case rather than a demonstration — the page is genuinely stale and somebody has to decide what it should say now.

    ## What the page still gets RIGHT, and must not lose

    Three of its four claims survive, and they are the valuable part:

    - **"Do not repurpose beans for it."** Still correct, and now structural: beans/ and todos/ are separate declared graphs. - **When a todo is shown.** Still undecided. This PR answers it for the navbar and the board; it does not answer whether an open item should surface unasked, and the page's objection stands — surfacing every item on every turn is noise, surfacing none is a list nobody reads. - **How one is retired.** Still undecided. TodoStatus carries resolved and wontfix, but nothing retires one, and the page's point holds: a person abandoning a reminder is ordinary and needs no ceremony, so the bean rules do not transfer.

    ## What has changed and needs writing down

    - **How a person creates one** is answered in the worst way: by an agent writing a file. There is no human path at all yet. - The 2×2 the page is missing — memory against workflow management, human against agent — with the empty quadrant named.

    ## Done when

    The page says what is true. Cheapest correct move is an edit to that block, not a rewrite: strike "Not built yet", keep the three open questions, add what now exists and who created it.

    Every todo carries a pencil to exactly this file, so the edit is one click from where the stale sentence renders.

  2. Decide which roles content-pipeline-navigator and platform-boundary-guard take on

    Status
    blocked
    Priority
    medium
    Raised by
    agent
    Created
    2026-09-19
    About
    folio-assistant-core/process/publication-workflow › agents-and-system-actors

    ci-health-watcher is settled: it is a mechanical participant in the CI process, so it takes build-pipeline and validation-pipeline, both actorKind: "system". Declared, and memoryForRoles now discriminates because of it.

    **The other two are not settled by that argument, and the reason is precise.** Both system roles are described as running a fixed program and exercising *no judgement*. That is exactly what makes them the right home for a watcher — and exactly why they are the wrong home for these two:

    - **platform-boundary-guard** decides whether a literal naming one folio belongs in platform code, and whether a new content type needs an adapter or only a profile. Those are judgements; a wrong one is not a failed check. - **content-pipeline-navigator** is closer to mechanical — it knows which stage owns which file — but it also answers "should this be a block kind", which is not.

    ## What differs depending on the answer

    Nothing breaks either way today: memory is scoped by agent and the duplication is already fixed, because one node can reach several agents. What changes is whether a fact written for one of these two reaches an agent standing in that lane, or only the agent named in the file.

    ## Options

    1. **Leave both agent-scoped, and say so in the role model.** Costs nothing, and is honest — neither is a lane anyone else acts in. The cost is that the role axis stays half-used, with one actor on it. 2. **Give them a judgement-bearing existing role** — editor and code-reviewer are the nearest. Cheap, and wrong if the fit is loose: a forced edge is one every later audit trusts as true. 3. **Add roles for them.** Honest edges. But AGENTS.md warns under skill-in-role-or-process against "inventing roles and activities to absorb tools that do not want them", and a role with one actor describes the tool rather than the work.

    No recommendation. The ci-health-watcher case was settled by a property of the work (it exercises no judgement); neither option here has an equivalent, which is why it is a person's call and not an agent's.

    ## If nothing is decided

    Option 1 is the status quo and nothing regresses.

  3. Decide what 'kick off' means mechanically for the two CI-watcher dispatch points

    Status
    in_progress
    Priority
    medium
    Raised by
    agent
    Created
    2026-09-19
    About
    folio-assistant-core/process/publication-workflow › agents-and-system-actors

    Stated 2026-09-19: a CI watcher *"should kick off during feature development once changes are made to feature branch. and when approved for final publication/merge to main."*

    The two dispatch points are clear. **What performs the dispatch is not**, and the three candidates differ enough that guessing would be wasted work.

    ## ANSWERED, 2026-09-19

    Owner: *"kick off if like task or workflow initation or bean roast"*.

    **Option 2.** A dispatch point is a **task / workflow initiation**, and it opens a **bean** — not a GitHub Actions job. That is the answer the repo's own machinery is already shaped for: an activity carrying <folio:bean op="claim"/> performs the bean operation when the step completes, so "start the watcher" and "record that it is running" are one declared step rather than two mechanisms that can disagree.

    It also settles the cost I flagged against option 2 — that a BPMN activity only fires while somebody is working. With a bean opened at initiation, the work is *durable*: a later session sees the bean whether or not the session that opened it is still alive.

    Remaining, and now narrow: which diagrams host the two steps, and whether the feature-branch one needs a process that does not yet exist.

    ## Options, with what each costs

    1. **A GitHub Actions job.** Fires without anybody present, which is the whole premise of xom7 — the failure being guarded against is a quiet stretch with nobody looking. Cost: it cannot read the session's memory, so the "dispatch an agent with a specific memory context" half does not happen. 2. **An activity in a BPMN process**, performed by a session. Gets the memory context, and puts the dispatch where workflow_next can report it. Cost: it only fires while somebody is working — precisely the stretch that is already covered. 3. **An MCP tool a workflow step calls.** Both, at the price of building the tool and deciding who calls it.

    ## Where the two dispatch points already live

    Both are steps in diagrams that exist, which is why this todo is tagged with them rather than describing them:

    - **"approved for final publication / merge to main"** is Process_Publication (draft-to-publication.bpmn), whose lane carries *Authorise the release* and *Version, tag and publish*. The dispatch point the owner named is that authorisation step. - **"once changes are made to feature branch"** is Process_CodeReview, reached from Process_Review.

    ## What is already true, so this is not started from nothing

    The two dispatch points are process events, so wherever they are performed they should be DRAWN in BPMN under processes/. The likely homes are draft-to-publication for the merge-to-main gate and a feature-branch process that may not exist yet — that needs checking rather than assuming.

    ## If nothing is decided

    Nothing fires. bun run cat check:ci-health still runs in the session-start sweep and ci-health.yml still runs weekly, so the watcher is not silent — it is just not tied to the two moments named above.

    • process: Process_CodeReview
    • process: Process_Publication
    • who: github:litlfred
    • bean: folio-assistant-29ij

folio-assistant — code under the Apache License 2.0, documentation under CC BY 3.0. Generated schema and API references are rebuilt from source on every push to main.

deployed dev