[katekruger]

De-brands the plugin, fixes two manifest defects that left v0.3.0 partly non-functional, eliminates the trigger collisions, and adds the missing governance, test, and CI layer.

CI is green on this branch: run 33023608790 — every step succeeded, including the CLI manifest validation and a full-history gitleaks scan.

1. What changed, by commit

refactor: remove Merget-specific branding
The pipeline was always brand-agnostic — it reads whatever is in the cached profile — but the manifest description, the brand-profile schema, and two skills named Merget directly, which made a general-purpose plugin read as one company's internal tool. Generalized the schema placeholders, the Figma design-system step, and the orchestrator's plan example and trigger phrasing.

fix: manifest defects that broke validation and dropped the subagent

• repository was an npm-style object where the spec requires a string — validate --strict failed outright.

• descript-editor had not loaded since v0.3.0. agents replaces its default, and every accepted form loaded zero agents:

  agents value	validate	loaded
  ["./agents"] (v0.3.0)	fail	—
  ["./agents/"], "./agents/"	fail	—
  ["./agents/descript-editor.md"]	pass	Agents (0)
  key removed	pass	Agents (1) ✓

  The third row is the trap: green validation, silently no agent. Both agents and commands overrides removed; all 21 components load from default paths.

• Corrected four doc claims that outran the code.

chore: normalize author and add marketplace.json
Author was "Kate Merget" while the other three repos and the GitHub account use "Kate Kruger". Normalized, and dropped kate@merget.ai per the GitHub-only contact policy. Added marketplace.json; claude plugin marketplace add resolves it.

fix: eliminate trigger collisions — the highest-value behavioral change. Detail in §2.

chore: governance, tests, CI, and the remaining docs
SECURITY.md, CODE_OF_CONDUCT.md (Covenant 2.1, GitHub-only contact), trademark notice in LICENSE, CLAUDE.md, .version-bump.json, scripts/, tests/ (30 checks), .github/ (CI, dependabot, 3 issue templates, PR template), docs/configuration.md, docs/development.md, README restructure.

ci: validate both manifests with one expected warning allowlisted — see §3.

2. Breaking changes

Nothing was renamed. All 21 component names are unchanged, so every existing invocation still works.

Every skill and command description was rewritten, which changes when things fire. claude plugin details proved commands and skills share one 21-item selection namespace, so each command was competing with the skill it wraps.

Component	Change	Direction
descript-studio	Dropped the "mentions Descript" catch-all plus edit/publish, which overlapped every child it routes to	narrowed
descript-edit	Dropped bare stems change/make it/fix/adjust/tweak, which reached cleanup, brandkit, and polish	narrowed
descript-clips / descript-export	Both claimed 9:16 vertical. Clips owns excerpts cut out of a longer video; export owns reframing the whole composition. Each now names the other	disambiguated
All 12 skills	Lead with the trigger instead of a capability summary; /video-* strings removed so each string has one owner	narrowed
All 9 commands	Became short action labels instead of restating their skill's trigger vocabulary	narrowed

Behavioral change in the intended direction: descript-editor now loads, so approved plans hand off to the subagent as designed instead of the orchestrator running passes inline.

3. Files moved or deleted

• docs/ARCHITECTURE.md → docs/architecture.md and docs/SKILLS.md → docs/skills.md, via two-step git mv to preserve history across a case-only rename on a case-insensitive filesystem. This was a latent bug: CLAUDE.md links docs/architecture.md, which resolved on macOS only because the FS is case-insensitive and would have broken on Linux CI.
• Nothing deleted. dist/descript-studio.plugin rebuilt in place.

4. Verification

$ scripts/validate-manifests.sh
── marketplace manifest (strict)   ✔ Validation passed
── plugin manifest                 ✔ Validation passed with warnings
OK   both manifests validate (1 expected CLAUDE.md warning)

$ python3 tests/run_tests.py       ✔ all 30 checks passed
$ scripts/bump-version.sh --check  all 3 locations agree: 0.4.0
$ scripts/build-plugin.sh --check  OK  dist/ matches source (48 entries)
$ JSON parse                       4 files parse
$ bash -n scripts/*.sh             ok

$ claude plugin details descript-studio
  Skills (21)  descript-brand-setup … video-status
  Agents (1)   descript-editor
  MCP servers (1)  descript

Trigger-collision check: 67 distinct quoted trigger phrases, zero claimed by two components. Enforced going forward by tests/test_descriptions.py::test_no_trigger_phrase_is_claimed_twice.

Token cost: always-on ~1,603 (v0.3.0) → ~1,636. The subagent now loads (it previously contributed nothing because it wasn't loading), and the tightened descriptions offset almost all of that.

One expected warning, deliberately allowlisted

The repo root is the plugin root, so CLAUDE.md sits at the plugin root and the CLI warns it will not load as plugin context. That is correct — CLAUDE.md documents repo development, and build-plugin.sh excludes it from the bundle. Rather than dropping --strict or deleting a file the standard requires, scripts/validate-manifests.sh validates the marketplace strictly, validates the plugin manifest non-strict, and fails on any error or any other warning.

Not verified

• No end-to-end run against a live Descript project. Nothing here was exercised against the real API. These are manifest, description, and documentation changes; I am not claiming the pipeline was re-tested.
• Marketplace install tested locally, not end-to-end from GitHub. claude plugin marketplace add <local path> resolves the manifest; /plugin marketplace add katekruger/descriptmcp will only work once this is on main.
• Component loading was verified by installing as a skills-directory plugin under an isolated CLAUDE_CONFIG_DIR, not through a marketplace install.

5. History findings

None. Full-history scans — mine and gitleaks in CI — found no tokens, API keys, bearer credentials, private keys, Descript project or composition IDs, or committed brand profiles. .mcp.json is Descript's documented public OAuth endpoint with no secret material. tests/test_docs.py now enforces this on every commit.

6. Follow-ups deliberately not done

• No v0.4.0 tag. Tag after merge.
• dist/descript-studio.plugin.zip — the double-wrapped original download. Kept for provenance and documented in dist/README.md, but only the .plugin is freshness-checked. Worth dropping eventually.
• descript-studio's body is 525 words, over the 500 guidance. Its skill map and pipeline table could move to references/. Left alone to keep this PR's skill-body changes to zero.
• ${CLAUDE_PLUGIN_DATA} vs ~/.descript-studio/ — the spec's data directory is arguably more correct for the brand profile, but switching orphans any existing profile. Left as-is pending your call.
• Semantic trigger overlap is not fully automatable. The test catches identical quoted phrases; two differently-worded descriptions meaning the same thing still need human review. CLAUDE.md makes that a required step when adding a skill.

🤖 Generated with Claude Code