Files
archie/docs/prds/skill-curator.md
T
sam e832b8ef0c docs(prds): normalise statuses and remove writing dated to its own day
Three defects the first sweep left behind, each caught by a rule the guard
did not yet have.

Statuses. RULES.md gives three; the tree had 25 values, and the free-text ones
were carrying progress. "Implemented (wiring landed)" and "Ratified (rev. 2e).
Rev. 2e adds the OpenTaskByPR lookup. Rev. 2d revised section 10..." were both
changelogs living in the one field the other rules exempt. Every PRD now
declares Draft, Approved or Finalised and nothing else; five that declared
nothing now declare one. A revision number moves to **Revision:**, a pointer
to the authoritative document to **Authority:**, and what a past revision
changed to git. CLAUDE.md's citation of state-store-contract rev. 2e still
resolves.

Stale cross-references. "Supersedes the 2026-09-03 open question", "this
bullet previously said", "an earlier draft of this document claimed" all point
at text the reader cannot see, because it was replaced. A supersede marker
survives only when it names a file, a section or a symbol that can be found.

Temporal hedges. 100 uses of today, currently and right now across 31 files.
Each dates its sentence to the day it was written: "[notify] today is one
webhook URL, read in one place" stops being true the moment this design lands,
while still reading as the premise it rests on.

The header-field exemption was also too wide. "**Date:** 2026-09-03, Channel
resolved 2026-09-22" smuggled a dated status into the one line the rules did
not read; the exemption is now a Date holding only a date and a Status holding
only one of the three.
2026-09-22 00:38:35 +10:00

177 lines
8.2 KiB
Markdown

# Skill curator (reference implementation) -- decision
**Status:** Approved
**Date:** 2026-09-02
**Beads issue:** `archie-core-1786637499206`
**Parent epic:** `archie-core-1786637500725` (curator engine)
## Decision
The skill curator's v1 scope is **mechanical validation and safe
normalization only** -- it never rewrites skill prose, never deletes a
skill, and never calls the model. Each pass:
1. Lists every skill under one configured root.
2. Reads each `SKILL.md` and attempts to parse its frontmatter
(`internal/skill.Parse`).
3. Records an `Action` for any structural problem: a parse failure, or a
missing `name`/`description` in the frontmatter that parsed.
4. Normalizes trivial whitespace (trailing whitespace per line, exactly
one trailing newline) and writes back **only** when normalization
actually changes the bytes, recording an `Action` for that write.
Nothing here needs judgment. "Review, prune, and improve" -- the epic's
framing -- is real for v1 but deliberately narrow: review is structural
validation, improve is whitespace normalization, and prune is *reporting*
a problem, never removing a skill. `Manifest.Skills = true` still marks
this curator agentic (existing `agentic()` rule), so it receives model
access from the registrar -- unused but nothing about the contract
or the runtime needs to change to let a later version reason about
content quality using it.
## Why this scope, not more
An autonomous process that rewrites or deletes a user's skill content
is a real risk with no undo path (no draft/approval flow for
curator writes exists yet, unlike the workflow engine's PR-based review).
The epic's own stated purpose for this issue is proving the curator
contract holds for a second, structurally different consumer than the
session-memory curator (files-as-skills vs. observations-as-memory), not
solving "what makes a skill good." Mechanical validation is real,
useful, safe to run unattended, and fully exercises
List/Read/Write -- every method `curator.SkillStore` declares except
`Delete`, which this curator's Pass never has a safe reason to call in
v1 (kept implemented on the store adapter regardless, since the
interface requires it and a future consumer may).
## Why `internal/skill` isn't `curator.SkillStore` directly
`internal/skill.Discover` parses every `SKILL.md` under a root and
returns an error for the **whole** call if any one of them fails to
parse -- the read side was built for "load the catalog to start the
daemon," where one bad skill failing the whole boot is arguably correct,
not for "review each skill and report what's wrong with it
individually." A per-skill parse failure must become one `Action`, not
abort the pass before any skill is reviewed. So
`internal/infrastructure/skillcurator` gets its own thin store adapter
that lists raw directory names and reads/writes raw `SKILL.md` bytes
per name; the skill curator itself calls `internal/skill.Parse` on each
one, isolated per skill.
## Root scope
One root only, resolved the same way `loadWorkflows` already resolves
`skillsBase` (`cfg.SkillsDir`, falling back to `cfg.WorkDir`) -- not
every root `skill.CatalogRoots`/`DefaultRoots` would read for a model
(which layers a shared dir over the per-profile one). Write and Delete
need one unambiguous target; a curator that could write into a shared,
multi-tenant skills directory is a bigger decision than this issue
covers. The model still sees the full multi-root catalog through the
existing read path -- this curator only maintains the local one.
## Shape
### `internal/infrastructure/skillcurator/store.go`
```go
// Store implements curator.SkillStore over one root directory's
// .agents/skills/*/SKILL.md files, tolerating a per-skill parse failure
// as the curator's problem to report rather than this store's problem
// to abort on.
type Store struct { root string }
func NewStore(root string) *Store
func (s *Store) List(ctx context.Context) ([]curator.SkillRef, error)
func (s *Store) Read(ctx context.Context, name string) (curator.Skill, error)
func (s *Store) Write(ctx context.Context, sk curator.Skill) error
func (s *Store) Delete(ctx context.Context, name string) error
```
`Read` best-effort parses frontmatter for `Skill.Description` (empty on
parse failure -- not this method's job to report that; `Pass` does, via
its own `skill.Parse` call on `Skill.Content`). `Write` rejects a name
that does not already exist as a directory (this store curates existing
skills; creating a new one is out of scope). `Delete` removes the whole
skill directory, matching "delete a skill," not just `SKILL.md`.
### `internal/infrastructure/skillcurator/skillcurator.go`
```go
// Curator implements domain/curator.CuratorEngine. See
// docs/prds/skill-curator.md for what a pass does and why.
type Curator struct { /* unexported: registrar view, interval */ }
func New(interval time.Duration) *Curator
func (c *Curator) Manifest() curator.Manifest {
return curator.Manifest{Interval: c.interval, Skills: true}
}
func (c *Curator) Check(ctx context.Context) (bool, error) // len(skills) > 0
func (c *Curator) Pass(ctx context.Context, in curator.PassInput) (curator.PassResult, error)
```
`Pass` walks every skill via `Bind`'s `Skills` (the filtered
`SkillStore` the registry already narrows to declared curators, see
`Registry.filter`), classifying each into exactly one of:
- **parse failure** -- `Action{Type: "skill.invalid", Reason: "frontmatter failed to parse"}`
- **missing required field** -- `Action{Type: "skill.incomplete", Reason: "name" | "description"}`
- **normalized** -- content changed by whitespace cleanup, written back,
`Action{Type: "skill.normalized", Detail: "<name>: trimmed trailing whitespace"}`
- **clean** -- no action recorded (a pass with nothing to report is not
an error, and not every skill needs an entry every time)
One pass reviews every skill on disk; there is no
per-skill "already reviewed, skip" state; this errs toward re-checking
over missing a regression, and the cost is cheap (no model calls, just
file reads).
## What we deliberately do NOT do
- **No LLM-authored rewrites of skill content.** Real "improve" in the
sense of better prose is out of scope until there is a review/approval
path for autonomous content changes to something a human wrote.
- **No automatic deletion of a skill**, however broken. `Delete` exists
on the store for a future consumer; this curator's `Pass` never calls
it.
- **No dangling-tool-reference checking** (a skill's
`metadata.archie.tools` naming a tool that no longer exists). The
mechanism to check "does this tool exist" belongs to a tool registry
this curator has no access to and shouldn't be given for this --
`curator.ToolBuilder.Build` resolves the curator's *own* declared
tools, not an arbitrary existence check against skill-referenced
names. A real fix here is a separate, better-scoped issue.
- **No multi-root awareness.** One configured root, matching
`loadWorkflows`'s existing resolution, not the full layered catalog a
chat turn sees.
- **No new curator-specific webui or chat surface.** Activity is visible
through the existing curator observability surface
(`archie-core-1786637489932`) -- `/api/curators` and
the dashboard Curators page already show any registered curator's
recent actions.
## Acceptance criteria
Restated from the issue against this shape:
1. Registers and runs through `curator.Registry` with no curator-specific
special-casing in the runtime -- proven by using the existing
`Registry.Register`/`Runtime` path, no new registry method.
2. Its changes to skills are observable and attributable -- every write
is preceded by an `Action` recording what changed and why, surfaced
through the existing curator activity tracker.
3. It operates within its declared tool set and cannot reach tools it
did not declare -- `Manifest.Tools` is empty; nothing about this
curator ever touches `Registrar.Tools`.
## Files this change adds
- `internal/infrastructure/skillcurator/store.go` -- `curator.SkillStore` adapter
- `internal/infrastructure/skillcurator/store_test.go`
- `internal/infrastructure/skillcurator/skillcurator.go` -- the `CuratorEngine`
- `internal/infrastructure/skillcurator/skillcurator_test.go`
- `internal/app/archied/bootstrap.go` -- register it in `setupCurators`, root
resolved the same way `loadWorkflows` resolves `skillsBase`