Two defects from the original review of the seed.
Prerelease ordering was worse than first recorded. parse_semver returned
(major, minor, patch, raw_string), so 1.0.0-rc1 and 1.0.0 tied on the numeric
fields and then compared as strings — "1.0.0-rc1" > "1.0.0". A release
candidate therefore shadowed its own release for `newest` and for `>=`, not
just for the no-version case.
parse_semver now returns a SemVer section 11 precedence key: numeric fields, a
release/prerelease rank, then dot-separated prerelease identifiers with
numeric ones compared numerically. Build metadata is ignored. Beyond ordering,
prereleases are excluded from `any`, `newest` and `>=` entirely; only an exact
pin selects one, so publishing a release candidate never changes what existing
consumers resolve to. A package holding only prereleases now says so rather
than reporting a bare not-found.
Packaging copied the whole source directory, so a stray .git, virtualenv or
scratch file landed in the catalog and registry. Section 2 already required
otherwise — tools MUST ignore unknown non-reserved files unless a manifest
field references them — so this is conformance rather than a new rule. What is
new is that omissions are reported instead of silent:
not packaged (not a reserved path, not referenced by the manifest): .git/, .venv/, notes.txt
LICENSE joins the reserved paths. Strict packaging would otherwise drop a
package's license text while faithfully copying its `license` field, which
contradicts section 14's instruction to surface licensing on publish and
install.
Spec: 2 (LICENSE, packaging obligation, reporting), 17.1 new, 10.3 note.
Reference CLI: parse_semver rewritten with is_prerelease; select_version and
pick_version updated; copy_package and report_skipped replace copy_immutable.
Tests 65 -> 78.
examples/pqrst-estimate carries a LICENSE and a license field, exercising the
new reserved path.
Also fixes a leaked loop variable in package_members that would have reported
a bad `template` path as an `evals` error.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bjefh8NUiEiahN4JLwoSKM
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 388925@bnt-lap001
Assistant-Session: 3507023f-e0fd-4a1e-9d90-a0d4217d1502
The answer to the capability question was conditional: keep both fields if
they carry the required/observed distinction, fix the terminology if they do
not. They did not.
Section 9 opened with "records known requirements or observations", mixing
both in one field — `models` was observational ("known to be compatible or
evaluated") while `capabilities` was prescriptive ("expected from the
execution environment"). Section 10 then described dependencies as what a
prompt "expects". Both fields said expected, so the overlap was real
ambiguity rather than redundancy, and the fix is terminology.
`dependencies` now means **required**; `compatibility` means **observed**. A
consumer must not refuse to run a package because its environment is absent
from a compatibility list. The same capability name may legitimately appear in
both: required to run at all, and separately observed to work well on
particular models. `compatibility.aliases` records the same capability under
other names, so a consumer can recognize a requirement its environment labels
differently.
Dependencies now have three kinds, separated by what the format can do about
them: `prompts` it resolves by id and version; `context` names what it does
not package at all; `capabilities` are what the environment must be able to
do. Context entries use `name` rather than `id`, because nothing can look them
up, and `description` is required because nothing else can explain an
unpackaged dependency. A capability takes no version and no
`requirement: generate` — it is not an artifact and cannot be fetched, pinned
or generated. Capability names are free-form kebab-case, validated for shape
and not membership, exactly as tags are.
Section 10.1 also draws the line the format had never stated: an input is
content the caller passes for one use; a context dependency is a standing fact
about the environment.
Spec: 9 rewritten, 9.1 and 10.1 and 10.2 new, 10 reframed, 18 (rules 19-20),
4 updated. Former 10.1/10.2 renumbered to 10.3/10.4 with cross-references.
Reference CLI: validate_capabilities, validate_context_dependencies, and a
`resolve` section listing required capabilities and context under "this tool
cannot verify these" rather than implying it checked. Tests 51 -> 65.
Also drops an invented `session-review` capability from the example package in
favour of an honest `long-context` observation.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bjefh8NUiEiahN4JLwoSKM
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 388925@bnt-lap001
Assistant-Session: 3507023f-e0fd-4a1e-9d90-a0d4217d1502
`evals/` was a reserved path holding unvalidated blobs: section 12 named the
directory and gave an illustrative snippet, but nothing was specified, so no
tool could act on an eval file.
Every eval file must now declare a `schema`, and CPF defines exactly one —
`canned-prompts/eval-rubric/v0.1`. Unrecognized schemas stay legal and are
skipped rather than rejected, so the format gains something actionable without
becoming an evaluation language, which remains a non-goal.
The schema splits along the same seam as T01 and T03. Render checks
(`contains`, `not_contains`, `resolves_all`) assert properties of the rendered
prompt text, need no model, and are therefore run by the reference CLI. Output
criteria describe a good result and are declared but not run, because judging
them requires a model. That division is now the format's consistent answer to
"deterministic locally, or not".
An eval references a fixture already declared in the manifest's `examples`
rather than carrying its own copy, so an example that is also an eval fixture
stays honest — both break together.
An eval declares assessment and must not record outcomes. Results are run
evidence and live outside the immutable package, per INTENT.md and section 17.
Spec: 12 rewritten with 12.1, 18 (rules 17-18), 21 (`eval` verb).
Reference CLI: read_eval, validate_eval, load_example_values,
run_render_checks, cmd_eval; a failed render check exits non-zero.
Tests 42 -> 51.
examples/pqrst-estimate/evals/quality.yaml is a real eval with four render
checks and four output criteria, and it passes.
Fixes a latent bug reaching a fixture exposed: coerce_value assumed every
value was a command-line string, so a YAML fixture carrying a real type
(include_rationale: true) crashed on .lower(). Typed values are now validated
but not re-parsed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bjefh8NUiEiahN4JLwoSKM
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 388925@bnt-lap001
Assistant-Session: 3507023f-e0fd-4a1e-9d90-a0d4217d1502
T01 had already delivered half of composition without naming it: a derived
default binds an input to a prompt dependency, which is transclusion — run
package B, use its output. What was missing was the deterministic half.
`include` inlines another package's rendered template as text. No model is
involved, so the reference CLI can actually perform it, and a shared preamble,
rubric or style block becomes a versioned package instead of copied text. This
is the concrete way to honor INTENT principle 9 without any runtime. `derive`
stays as it was. Both are input defaults, so composition reuses the resolution
machinery rather than adding a second one.
No template inheritance. Four of this repo's own documents argue against it:
INTENT principle 3 (hidden context defeats reuse), section 19's "make package
contents visible before execution", section 17's requirement that behavior
changes produce a new version, and the non-goal on range resolution.
Version selectors: an exact pin is the expected form, with `any`, `newest` and
`>= X.Y.Z` as explicit opt-ins so looseness is written rather than implied by
absence. Selectors are evaluated per dependency against what is available —
no solver, no cross-dependency constraint satisfaction — which is what keeps
them outside the range-resolution non-goal, and the spec says so.
Also defines `type` (template | fragment), which appeared once in the section 4
manifest surface and was specified nowhere.
Spec: 3.2 (type), 5.1 (inclusion resolution rule, renumbered), 6.1 (included
default), 10.1 and 10.2 (new), 18 (rules 14-16), 21.
Reference CLI: validate_version_selector, select_version,
prompt_dependencies replacing prompt_dependency_ids,
check_composition_reference, CatalogComposer with cycle detection, and
resolve_inputs gaining composer= and inherited=. Tests 21 -> 42.
Examples: house-style is a real fragment package; pqrst-estimate composes it
and is bumped 0.1.0 -> 0.2.0 per section 17.
Fixes an ordering bug found while testing: inputs resolved before parameters,
so an included package could not see the including package's parameters and
silently fell back to its own defaults — the fragment rendered tone=neutral
where the including package said blunt. Parameters now resolve first; the
report still lists inputs first.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bjefh8NUiEiahN4JLwoSKM
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 388925@bnt-lap001
Assistant-Session: 3507023f-e0fd-4a1e-9d90-a0d4217d1502
Section 17 already defined what a registry does when the same id@version is
republished; nothing defined what a *consumer* does. That is where namespace
conflict actually bites — in the catalog, after installing from two
registries.
Identity is now registry-scoped: an id names a package within a registry, the
way a path names a file within a repository. A local-first format with no
signing, no federation and no central authority cannot enforce global
uniqueness, and an unenforceable guarantee is worse than none — it invites
consumers to conflate two packages that merely share a name. A qualified
`<registry>:<id>` reference distinguishes them, and `:` is now barred from ids
so the separator stays available.
Installing the same id from two registries is therefore not a conflict. The
catalog is namespaced by registry and keeps both.
Ownership is registry policy, not package data. An optional `registry.yaml`
names a registry and records namespace claims. Those claims are explicitly
descriptive — a filesystem registry cannot authenticate a publisher, and
`publish` says so rather than implying it checked. Keeping the claim out of
packages leaves artifacts free of unverifiable assertions of authority, and
means package semantics do not change when a hosted registry appears later.
Spec: 3.2 (registry-scoped identity, qualified references), 17 (immutability
scoped to a registry), 20.1 and 20.2 (new), 18 (registry-manifest
validation), 21 (qualified references, reserved `local` name).
Reference CLI: parse_reference, check_registry_name, read_registry_manifest,
registry_name, namespace_policy; registry_package_path and
catalog_package_path split; resolve_installed reports ambiguity and returns
the source registry; iter_catalog; `add --as`; closed-namespace warning on
publish. Tests 11 -> 21.
The catalog layout changed. An existing catalog is detected and reported with
instructions rather than failing as "package not found". Signing, trust
scoring and federation remain non-goals and were not touched.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bjefh8NUiEiahN4JLwoSKM
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 388925@bnt-lap001
Assistant-Session: 3507023f-e0fd-4a1e-9d90-a0d4217d1502
Closes the gap that made optional inputs unusable: rendering rule 5.1(4)
made any unresolved placeholder an error while inputs had no `default`, so
an input marked `required: false` and referenced from the template failed
every render in which the caller omitted it — including the spec's own
section 4 example.
Section 10 already carried the derivation mechanism (`requirement: generate`,
resolution deliberately undefined), so a derived default needed a binding
rather than a new concept: the input's default names a declared prompt
dependency.
Spec:
- 5.1 rewritten as "Resolution and rendering". Resolution may be
non-deterministic and must report what it derived; rendering is
deterministic and must not derive. A tool that handles only supplied
values and static defaults is stated to be conforming.
- 6.1 (new) covers both declaration forms. Reference form is preferred, with
the reason stated — an inline prompt is anonymous, so unversioned,
unprovenanced and un-evaluable — and validators should warn when a
published package derives inline.
- A derived default may declare a static fallback `value`. Without one the
input stays unresolved, which is an error; derivation never silently
yields empty content.
- 4, 10, 18 (rules 11-13), 19 (two new MUST NOTs), 21 updated accordingly.
Reference CLI:
- New `resolve` verb reporting the origin of every value.
- `Resolution` dataclass and `resolve_inputs`; `resolve_values` kept as a
wrapper so existing callers are unaffected.
- `render` refuses with a specific error naming underivable inputs rather
than substituting empty text.
- Tests 3 -> 11. Example package lifecycle re-verified end to end.
INTENT.md is unchanged: splitting resolve from render preserves success
criterion 4 (deterministic rendering) as written.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bjefh8NUiEiahN4JLwoSKM
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 388925@bnt-lap001
Assistant-Session: 3507023f-e0fd-4a1e-9d90-a0d4217d1502
Register canned-prompts under agents / practice (topic
c1d199b6-55ee-4db6-b49e-257a9f0f15ac, workplan prefix CANP-WP) via
`statehub register`, then replace the generated placeholders with
repo-specific facts.
- SCOPE.md: real boundaries drawn from INTENT.md's deliberate boundary,
current state (spec v0.1 + reference CLI, 3/3 tests pass, example
round-trips), and the developer workflow.
- AGENTS.md: drop the unresolved {CREDENTIAL_ROUTING} template token left
by the generator.
- CANP-WP-0001: bootstrap tasks closed.
- CANP-WP-0002: new workplan carrying the five § 23 open questions promoted
from "experience will decide" to "decide for v0.2" — optional-input
defaults (static or derived), registry namespaces/ownership, prompt
composition, canonical eval schemas, typed context/dependency contracts —
plus two reference-implementation conformance defects found in review
(prerelease versions sort as newest; copy_immutable packages the whole
source directory).
Also lands the previously untracked seed: INTENT.md, the CPF v0.1 spec,
the reference CLI, and examples/pqrst-estimate.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bjefh8NUiEiahN4JLwoSKM
Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 388925@bnt-lap001
Assistant-Session: 3507023f-e0fd-4a1e-9d90-a0d4217d1502
