`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
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
