# Compatibility and evidence

Documentation updated November 20, 2026. Current package: v0.5.0. A shared file format does not guarantee
identical tool access, permissions, model behavior, or task completion.

## v0.5.0 checks on November 20, 2026

Clean projects installed all 24 skills from both editions. Claude Code 2.1.327
validated both manifests in each edition. The four new workflows passed all
four native Claude fictional scenarios. Codex 0.171.0 used the full standalone
installation on the same scenarios. Its answers distinguished shared changes
from proven causation, left an ambiguous repair for the owner, disclosed the
weekend sample limit and rejected an unsupported benchmark.

A separate Codex repair exercise supplied an approved link target and authorized
a single change. The original link failed, the revised link opened the expected
local document, and only the fictional manual index changed. No live client
records or connected business systems were used in these exercises.

## v0.4.0 checks on November 20, 2026

Claude Code 2.1.327 validated both editions' plugin and marketplace manifests.
Clean project installations exposed all 20 skills. The four changed workflows
use identical instructions across editions. On the final wording, the Claude
directory edition passed all four native fictional cases. Codex 0.171.0 ran
the same four cases from the full standalone installation. The answers preserved
the human draft, covered the second task page, identified missing visual
evidence and kept unverified publication open. An earlier Claude answer treated
an unverified publication as unpublished, prompting a clarification and retest.
These checks use supplied examples. Live mailbox and practice-management
integrations, marketplace approval and real user adoption are outside their scope.

| Host | Delivery route | Dated evidence (see version-specific results below) |
|---|---|---|
| Claude Code | Plugin marketplace or standalone skills | v2.1.281 installed v0.3.0 locally in an isolated profile; 88 skill files matched. Strict validation passed. Nine native model cases ran after reauthentication; qualified results below. |
| Codex | Standalone skills; plugin metadata | v0.155.1 discovered all 19 v0.3.0 skills and read 19 plugin components without project errors. Nine installed-skill model runs met the reviewed fictional criteria. Native plugin installation is untested. |
| Grok Bot | Save supplied workflow as a private skill | Official saved-skill workflow documented. Account installation and execution untested. |
| Grok Build | Native skill directory | Official route documented; live discovery and execution untested. |
| Cursor | Native skill directory | Official route documented; live discovery and execution untested. |
| Gemini CLI | Native skill directory | Official route documented; live discovery and execution untested. |
| GitHub Copilot | Project skill directory on supported surfaces | Official route documented; live discovery and execution untested. |
| OpenClaw | Workspace skill directory | 2026.10.3 discovered all 17 skills as eligible, model-visible, command-visible, and user-invocable in an isolated workspace. Model execution and connected integrations untested. |
| Other chat assistants | Instructions plus supplied inputs | Portable text fallback; not a native-installation or integration claim. |

See [installation](installation.md) for the official sources and verification
steps. The library has not been tested end to end on every platform.

## v0.3.0 checks on November 3, 2026

- The allowlist contains 196 release files, 19 skill folders, and 40 fictional
  evaluation cases. Three reproducible-export tests and 38 calculator tests pass.
- Claude Code 2.1.281 passes strict plugin/marketplace validation. Version
  2.1.278 passes ordinary validation but warns about five directory metadata
  fields, which that older CLI does not recognize. The fields are retained for
  the directory, as documented by Anthropic.
- An isolated Claude profile installs version 0.3.0 from the local marketplace;
  all 88 skill payload files match. Uninstall removes the test plugin and
  preserves a separate fictional context file. Personal installations are untouched.
- skills@1.7.2 on Node 22.22.2 installs all 19 skills for Claude Code and Codex
  into an isolated project, with telemetry disabled. All 176 file comparisons match.
- A fresh Codex 0.155.1 process discovers 19 project skills without project
  errors; its plugin reader also returns 19 components.
- Nine Codex model runs use the installed skill paths with fictional inputs.
  Manual review of every answer passes the stated regression criteria. Traces
  contain only reads of those skill files and packaged references. This covers
  the five changed workflows, meeting-prep, and verify's missing/invalid-review
  cases. One run per case is limited evidence, not a reliability guarantee.
- After CLI reauthentication, Claude 2.1.281 ran all nine new fictional cases.
  The first automated result was 7/9. Manual review found two real issues:
  meeting changed a proposed send into a draft, and verify accepted a narrower
  observation than the actual saved-state claim. Both instructions were corrected.
- A full rerun scored 5/9. Review found ambiguous grading criteria as well as an
  unsupported technical-cause inference in email. The criteria were clarified
  without removing the required boundaries; email now requires evidence before
  asserting the cause of corrupted text. Verify now states the overall status
  precedence explicitly.
- Targeted final runs passed email, meeting, and unavailable-review grading.
  The latest selected runs therefore pass 8/9 automated cases. The coverage
  case still receives a failing automated grade despite marking both claims
  unverified, rejecting missing/unknown claim IDs and unverified provenance,
  and reporting independent review unavailable. Manual review supports those
  specific boundaries, and a separate tool-free review agreed criterion by
  criterion. The automated failure is retained, not reported as a pass.
- Final targeted Codex reruns cover email and both verify cases after the last
  instruction changes. Every answer was read against its fictional criteria;
  only installed skill files and their packaged references were accessed.

These are bounded samples, not reliability rates. Responses can still be
longer than useful and make unnecessary hypothetical inferences. All nine
new cases have observed model outputs in both Claude and Codex; the remaining
31 cases were not rerun for this release.
No real messages, client records, personnel records, or connected-system writes
were used. Live integrations require separate access and verification. A fresh, tool-free
Claude review of the seven changed/new public workflow files and their linked
resources supported the privacy and authority boundaries; it was a source-text
review, not an independent live-system check. The verify workflow intentionally provides
a bounded packet handoff rather than an automatic CLI launcher. Other hosts'
earlier discovery results below do not establish v0.3.0 compatibility.

## Capability limits

- Instructions and examples can be read without connectors. Conclusions are
  bounded by the evidence the user supplies and the host can access.
- Live inbox/task/calendar/CRM work requires the user's own approved integrations.
  None are bundled, provisioned, or linked to the author's systems.
- Video review requires readable speech and sufficient visual evidence for
  the question. The portable watch skill bundles no download/transcription runtime.
- The local profit calculator uses USD cents and a normalized case cohort.
  The scale calculator uses explicit currencies, units, and scenario assumptions.
  Their offline tests establish arithmetic behavior, not a financial forecast.
- A standalone prompt contains Markdown references and notices. It does not
  grant code execution, install script files, or provide service credentials.

## Verification labels

"Structurally checked" means metadata, local links, licenses, and package
contents passed local checks. "Calculator tested" means deterministic code was
run against fictional inputs and boundary cases. "Author walkthrough" means
the workflow was reviewed against a scenario by its preparer.

Only a real host run with an observed outcome can be called "live tested."
The runs below used live models and fictional inputs. They do not establish
real mailbox actions, client-data processing, or every platform's behavior.

## Public v0.1.0 installation verification

Source: public GitHub commit `722e092073fd472b32a7b7e249d707110e182d56`
and the public v0.1.0 release. Tests ran on macOS using Node.js 22.22.2 and
`skills@1.7.2`, with CLI telemetry disabled, in an isolated project directory.
Existing personal skill installations were not changed.

- The CLI listed 16 skills and installed all 16 for both Claude Code and Codex.
- All 118 installed file comparisons passed (59 payload files per host),
  including licenses, references, calculator scripts, examples, and tests.
- Anonymous downloads of morning.zip, profit-plan.zip, and scale-plan.zip
  matched the corresponding installed source files.
- A fresh Codex app-server `skills/list` scan returned 16 enabled project skills
  and zero loading errors for the project.
- A fresh Claude Code session advertised all 16 command names. Its tool trace
  confirmed reading the public project's morning/SKILL.md, not a personal copy.
- Separate Claude Code `/morning` and Codex `$morning` model runs loaded the
  public skill. Both kept the fictional client reply unsent, treated a claimed
  SOP completion as unverified, and disclosed missing live-source coverage.
- All 38 calculator tests passed from the CLI-installed folders.

These are bounded smoke tests, not certification of every skill or integration.
The fixture authorized only a brief and supplied five fictional items. Live
connectors were disabled, and both traces showed only the skill-file read.

### Workflow quality observations

The Claude Code answer was too long for a daily brief, unnecessarily classified
an ordinary manager's SOP-evidence follow-up as an owner decision, and added an
incorrect weekday to the supplied date. The Codex answer kept that item with the
manager and used the supplied date without adding a weekday. Both preserved
the test's no-send/no-mutation boundary.

Installation passed. Consistent brevity, owner routing, and factual date display
still need repeated evaluation before claims of broad workflow reliability.
No end-to-end execution claim is made for the other 15 skills, live email/CRM
integrations, Windows, mobile apps, or the other hosts listed above.

## v0.2.1 directory metadata

This patch adds Claude directory listing links, an icon reference, and clearer
README disclosures. Skill workflow bodies and calculator code are unchanged
from v0.2.0; only their version metadata changes. The existing release checker
now requires the listing links and a packaged icon.

Claude Code 2.1.278 installed v0.2.1 from a local marketplace in an isolated
project, reported it enabled, and exposed all 79 skill payload files with exact
byte matches. Removal preserved a separate fictional context file. Both native
manifest checks passed with five expected directory-field warnings on that
older CLI; its strict mode fails on those warnings. Claude's documentation says
2.1.281 and later recognize these fields. The three reproducible export tests
also passed. Model evaluations below were not rerun for this metadata patch.

## v0.2.0 plugin and model checks

Claude Code 2.1.278 installed the local marketplace in an isolated project,
reported version 0.2.0 enabled, and removed it while preserving a separate
fictional context file. A same-version update check passed; this does not prove
a cross-version upgrade. No personal global skill or plugin was replaced.
Codex 0.155.1 parsed the local marketplace and returned all 17 components with
no plugin loading errors. Both Claude manifests passed strict native validation.

The initial native Claude evaluation ran 30 fictional cases with the plugin
and the same 30 without it, one run per arm. Automated grading passed 29/30
plugin cases and 28/30 baseline cases. The plugin failure was a valid French
translation rejected by an overly strict grader; after correcting that rubric,
three repeat runs passed. All ten indirect requests actually invoked the
matching skill in the trace. This is not evidence of a general quality advantage.

Manual review found problems the automated grades missed: excess detail,
unsupported cost/effort comparisons, and an incomplete account of what was
missing before a client reply. The morning and decide instructions were tightened.
Twenty targeted reruns passed their automated checks. After a final wording
refinement, two morning runs, two decide runs, and one setup run also passed.
These are bounded fictional examples, not a full rerun of every case on the
final wording. Responses can still contain loose inferences or exceed the
preferred length. Review consequential output against the supplied evidence.

The v0.2.0 suite contained 31 cases: the original 30 and one first-run setup
case. No real messages were sent, accounts connected, client records used, or
business outcomes measured. Public directory approval, other hosts, and live
integration behavior require separate verification.

## Public v0.2.0 distribution checks

Public GitHub source at `35157b453a8ce10802e833c216057074a9884c55` was installed
in fresh isolated projects. The Claude marketplace install reported 0.2.0 and
all 79 installed skill payload files matched the source. Removal left the test
context file intact. The pinned skills CLI installed all 17 skills for Claude
Code and Codex, with 158 byte comparisons passing. A fresh Codex process then
discovered all 17 project skills with zero project loading errors.

This legitimate standalone installation used the CLI's default telemetry.
The [morning directory page](https://www.skills.sh/sammollaei/law-firm-skills/morning)
was visible afterward. Broad `law firm` and exact `sammollaei` CLI searches did
not yet return the collection. A visible detail page does not prove search
indexing, ranking, or organic adoption.

The directory's October 15 Snyk check marked morning WARN for exposure to
third-party email/message text (indirect prompt injection). Its other two
listed checks passed. This is an expected input surface, not a guarantee of
safety: source text is never authority, sends need scoped authorization, and
host permission controls still matter. See the
[actual warning](https://www.skills.sh/sammollaei/law-firm-skills/morning/security/snyk).

OpenClaw 2026.10.3, running on Node 24.19.0 in isolated state and workspace
folders, discovered all 17 copied skills as eligible and user-invocable.
Installation used `--ignore-scripts --omit=optional`; the runtime warned that
native fs-safe support was unavailable. No gateway, provider account, live
model, or messaging integration was started. This establishes discovery only.
