E1 promoted-skill execution rebaseline — 2026-08-27

Scope and evidence boundary

This is the E1-specific rebaseline of the promoted skill surface. It was inspected at repository HEAD 2f87aceb; the manifest and selected canonical skill files had no local diff at inspection time. The wider execution-channel workshop and its retained probe reports remain useful historical evidence, but their project-venv activation forcing case is not the current Commonplace command model.

ADR 064 now makes the uv tool executable directory the command authority. One user-level uv tool installation supplies the active commonplace-* command version for an OS user. Projects may have a .venv for their own dependencies, but that environment, activation, and direnv are not prerequisites for Commonplace commands. After uv tool update-shell, the consuming shell, IDE, desktop agent, or service must be fully restarted before bare-name lookup is a valid test.

E1 therefore starts from two separate concerns:

1. The health preflight must establish uv ownership and bare-name resolution in the launch class being diagnosed.
2. After a package command resolves, promoted procedures must not depend on POSIX shell programs for their load-bearing behavior.

This run inspected source text from a POSIX environment. It did not execute any procedure in native Windows PowerShell. The workshop's current evidence directory contains only POSIX reports. ADR 064 records an earlier Windows installation experiment, but that is not a native-Windows E1 result for the current promoted skill set. E1's Windows runtime gate remains open.

Manifest-derived selection

The selection authority is MANIFEST.promoted_skills in src/commonplace/scaffold_manifest.py. A repeat audit must read that tuple, preserve its order, map each value to kb/instructions/<skill-name>/SKILL.md, and fail if a selected source is missing or duplicated. It must not select a remembered list or stop after a remembered count. Generated .agents/skills/ and .claude/skills/ projections are drift-check surfaces, not additional inventory members.

The tuple currently resolves to the following dated observation:

Manifest value	Canonical audit source
cp-skill-write	kb/instructions/cp-skill-write/SKILL.md
cp-skill-validate	kb/instructions/cp-skill-validate/SKILL.md
cp-skill-connect	kb/instructions/cp-skill-connect/SKILL.md
cp-skill-convert	kb/instructions/cp-skill-convert/SKILL.md
cp-skill-health-check	kb/instructions/cp-skill-health-check/SKILL.md
cp-skill-ingest	kb/instructions/cp-skill-ingest/SKILL.md
cp-skill-snapshot-web	kb/instructions/cp-skill-snapshot-web/SKILL.md
cp-skill-revise-autoreason	kb/instructions/cp-skill-revise-autoreason/SKILL.md
cp-skill-write-multistage	kb/instructions/cp-skill-write-multistage/SKILL.md
cp-skill-ground	kb/instructions/cp-skill-ground/SKILL.md

Ten entries were observed in this run. That number describes this snapshot; it is not the selection mechanism or a future completion condition.

Classification rule

• Shell-neutral now means one executable is invoked with ordinary arguments and no shell control flow, expansion, pipeline, redirection, environment assignment, or POSIX-only utility semantics. The skill may still require the executable as a declared prerequisite.
• Neutralize behind package/runtime operations means shared, load-bearing behavior should use a tested commonplace-* command, a shared Python helper, or a runtime-native filesystem/process operation. Duplicating the whole procedure in PowerShell would create two implementations of the same semantics.
• Paired preflight is reserved for checks that must work before the package command can be trusted, or before an optional external capture tool is used. These checks need explicit POSIX and PowerShell forms with the same reported fields.
• Dependency means the skill should not receive an interim translation. Its portable form follows another packet that owns the missing semantics.

Per-skill inventory

Selected skill	Current executable or channel-sensitive loci	Disposition	Dependency
cp-skill-write	Targeted rg; direct commonplace-validate <path>; commonplace-relocate-note; exact snapshot-hash check on the conditional source route	The three commands are shell-neutral single-process calls. Use the shared snapshot verification operation for the source route rather than inventing a shell hash command.	None for the ordinary write path. Source-grounded writes share the checksum work described below.
cp-skill-validate	One Bash if/for program using glob expansion, test, basename, dirname, command substitution, and fail-fast exits	Remove the shell program after V1. The skill becomes one shell-neutral commonplace-validate <target> invocation and passes all, types, landings, redirects, a collection, or a path unchanged.	V1 owns recursive collection discovery, complete accumulation, and repository phases for all.
cp-skill-connect	Standalone title, description, and body rg calls; a by-tag rg -l | xargs -r rg pipeline whose no-match guard is load-bearing	Keep standalone rg calls as shell-neutral process invocations. Do not produce a PowerShell copy of the tag pipeline. Replace it with the exact shared tag/path resolver and retain a no-match test that proves search does not widen.	The tag pipeline depends on the exact resolver owned by tag-contract convergence, tracked from T1.
cp-skill-convert	Backlink rg -l; git mv; backlink edits through runtime tools	rg and git mv are shell-neutral single-process calls. Git remains an explicit prerequisite. No paired shell program is needed.	None.
cp-skill-health-check	POSIX-only layout, projection, and legacy-residue blocks; paired but non-identical uv ownership blocks; direct validator, pytest, and repair commands	Add field-equivalent POSIX/PowerShell preflights for project/layout presence, canonical and projected skill presence, uv command ownership, and legacy .envrc/.venv inspection. Keep validation, uv run pytest, uv tool install, and uv tool update-shell as shell-neutral calls. The old .envrc body is a historical residue signature, not an active dependency.	Must remain independently executable when commonplace-* lookup fails; it cannot depend on V1 for its preflight.
cp-skill-ingest	Skill dispatch and direct validation; exact-byte SHA-256 checks; platform temporary backup; byte-copy, restore, and verification	Dispatch and validation are shell-neutral. Put checksum and handled-failure backup/restore semantics behind shared package/runtime byte operations. The current cp/Copy-Item examples acknowledge both channels but do not yet specify one parity-tested operation.	Mandatory connection discovery inherits cp-skill-connect's portable completion. No direct V1 dependency.
cp-skill-snapshot-web	Shell-neutral GitHub/X package calls; POSIX command -v checks; mktemp, redirection, awk, cp, cat, wc, arithmetic, and test in PDF/HTML capture and assembly	Pair only optional-tool discovery (command -v / Get-Command). Move temporary-directory creation, subprocess execution, metadata stripping, byte assembly, checksum, verification, and cleanup behind a package capture operation or shared helpers. Temporary paths must come from the platform temporary-directory API.	Independent of V1 and the tag resolver.
cp-skill-revise-autoreason	POSIX pwd/basename/date, command substitution, variables, mkdir, cp, printf, redirection, and repeated blind-packet copies; later git mv	Replace run setup, path construction, copying, packet materialization, result writing, cleanup, and restoration with runtime-native or package operations. Resolve concrete paths before actor dispatch instead of sending shell-variable placeholders. The final git mv remains a shell-neutral process call.	Independent of V1 and the tag resolver.
cp-skill-write-multistage	Direct commonplace-validate <path>; abstract runtime file/workshop operations; exact snapshot-hash check on the conditional source route	Validation is already shell-neutral. Keep workshop orchestration runtime-native and use the shared snapshot verification operation for the source route. No paired shell blocks are needed.	No direct V1 dependency. Its conditional grounding path inherits cp-skill-ground and ingest portability.
cp-skill-ground	No literal shell program; exact-byte checksum, byte-preserving Quotes splice, one-write mutation, restoration, and direct validation	Validation is shell-neutral. Expose or reuse package/runtime operations for checksum, byte-preserving splice, and verified restore so an executor does not improvise platform commands for the load-bearing preservation contract.	A URL without an ingest inherits ingest and snapshot portability. No direct V1 or tag-resolver dependency.

Cross-cutting conclusions

The user-level tool model removes the old activation problem

No promoted procedure should look for a project .venv to find Commonplace. Only the health skill inspects .venv and .envrc, and it does so to identify legacy residue or a project-owned environment. The operative command probe is bare-name resolution plus ownership relative to uv tool dir --bin, followed by a full restart test in the failing launch class.

Pair only preflight, not workflow semantics

The required paired surface is small:

• health: layout, projection, uv ownership, and legacy-residue inspection;
• snapshot-web: discovery of optional capture executables.

Everything else either already has stable argv semantics or performs shared file, checksum, capture, or orchestration behavior that should have one implementation. Large paired Bash/PowerShell workflow blocks would duplicate failure handling and byte-preservation rules.

Two upstream packets own two exact removals

• V1 removes the complete Bash program from cp-skill-validate; a PowerShell translation would preserve the current incomplete all semantics.
• The tag resolver removes connect's xargs -r pipeline; a PowerShell pipeline would create a second tag membership implementation and still need an independently tested empty-set guard.

Direct validation of one path in write, ingest, multistage write, and ground does not wait for V1. Only full-product all semantics do.

Byte-level work is a shared portability surface

Snapshot SHA-256 verification appears in write, ingest, multistage write, and ground. Snapshot-web creates and hashes the bytes; ingest and ground also require exact copy, splice, restoration, and post-write verification. These procedures should reuse one byte-oriented package layer rather than select sha256sum, cp, or PowerShell equivalents independently. Existing internal snapshot hashing can be the starting point, but this audit does not select a new public command or implement one.

allowed-tools remains an unresolved execution-interface contract

Every currently selected skill includes Bash in allowed-tools; none names PowerShell. Static source inspection cannot establish whether each consuming runtime treats Bash as a generic process tool, ignores the field, or denies a native PowerShell interface. Do not count this declaration as either Windows support or a demonstrated failure. E1 needs a documented projection/runtime rule: normalize the capability, generate channel-specific allowlists, or prove that the field is non-binding on the native-Windows consumer. The manifest-derived static audit should flag a new selected skill until its execution-interface declaration has a disposition.

Static and runtime closure checks

The eventual static check should derive the selected set from MANIFEST.promoted_skills and require exact set equality with its audit records. For every selected source and required linked execution reference, it should flag at least:

• unpaired Bash conditionals, loops, pipelines, redirections, substitutions, environment assignments, and glob expansion;
• xargs, mktemp, awk, sed, wc, cat, and cp in load-bearing command positions;
• hardcoded /tmp, .venv/bin, .venv\\Scripts, or path separators when they are operative rather than historical signatures;
• a Bash-only execution-interface or allowed-tools declaration without a recorded platform disposition;
• selected skills missing from the audit and stale audit rows no longer in the manifest.

Static coverage does not establish runtime behavior. E1 still needs a real native-Windows PowerShell run, recorded as a new evidence report, covering the health preflight, arbitrary-path validation, validate all, connect's empty-tag path, and snapshot temporary-file handling after their owning changes land. Linux and Windows CI should separately exercise user-level uv-tool install, pristine init, the same bare-name commands, and the focused promoted-skill paths. No native-Windows result was collected by this audit.