fix(education): take the illustrate slug from the model, not the build command

[melodic-automation-lanes]

Closes #7374

Hold: a human merges this

Summary

Opened by the work-resolve-issue lane from work-resolve-issue/issue-7374-39058663258-1, which an earlier run left with commits and no report. The model did not run again, and the earlier run's outcome could not be read, so this body carries none of its text.

Fix

The diff shows the changes.

Verification

The earlier run reported no checks here.

• The earlier lane run: https://github.com/melodic-software/claude-code-plugins/actions/runs/39058663258
• This lane run: https://github.com/melodic-software/claude-code-plugins/actions/runs/39070313832

Left out: every check; a person reviews the diff before merging.

Related

#7374

[github-actions]

claude-security-review needs no new review of 1712f0e766ef8f1e8098b2eb8b29a8eb750addfd: no file of this pull request changed since the review of 37420a4.

No security issues in the incremental change since e617d36 (reviewed head 37420a4).

The change moves the illustrate model file from the shared model.json to a per-session models/${CLAUDE_SESSION_ID}.json. It touches the changelog fragment, SKILL.md and build-explainer.test.sh.

• Trust boundary: the build command still carries only fixed paths and the session ID, which the harness sets. No model-chosen or user-chosen text reaches the command line. The builder still checks the slug against ^[a-z0-9-]+$ and derives the record and page paths itself.
• Shared-directory race: per-session files remove the cross-session overwrite window. This strengthens the control, and the "read it before you write it" step and the race caveat are gone because they are no longer needed.
• Path injection: CLAUDE_SESSION_ID is a harness-generated UUID, not attacker-controlled, so it gives no traversal path.
• Test change: the test now also asserts that the build command uses a per-session model file. The edit adds a check and weakens none.
• Instruction surface: the removed text is the old shared-file read-before-write and race workaround. The per-session path replaces it, so no guardrail is left unenforced.

Only the three listed files were reviewed. I did not consult the instruction exception register, because the diff removes no standing guardrail.

[github-actions]

claude-review needs no new review of 1712f0e766ef8f1e8098b2eb8b29a8eb750addfd: no file of this pull request changed since the review of 37420a4.

No findings. I reviewed the incremental diff since e617d36 in three files: the changelog fragment, plugins/education/skills/illustrate/SKILL.md and build-explainer.test.sh. I did not run the test script or the builder.

• Skill change: SKILL.md now has the model written to illustrate/models/${CLAUDE_SESSION_ID}.json, so each session gets its own file. The old "Read it before you write it" and "run straight after the write" caveats are removed, which fits because two sessions no longer share the file. The fixed build command keeps carrying only fixed paths, and grep of plugins/education finds no leftover illustrate/model.json reference.
• Test change: Pulling buildCommand into a variable is a clean refactor. The new check pins --model "...${CLAUDE_SESSION_ID}..." in the first bash block, which is the build command. It also carries a portability-ok marker like the neighbouring checks.
• Changelog: The new sentence in the fragment matches the skill change, and the bump level (patch) is unchanged.

[melodic-automation-lanes]

pr-unblock / address-feedback finished round 1 of 3 on b0fa036.

[melodic-automation-lanes]

pr-unblock / address-feedback finished round 2 of 3 on e617d36.

[melodic-automation-lanes]

Unattended merge: risk MED at head 1712f0e766ef8f1e8098b2eb8b29a8eb750addfd. Only LOW merges without a human; this pull request waits for one.

MED evidence-unbound:claude-review: not bound: no-run
MED evidence-unbound:claude-security-review: not bound: no-run