[mganchas]

Description

Make the planning skill work on MetaMask Extension as well as MetaMask Mobile, and narrow when it triggers.

• planning, planning-spec and jira-brief install for MetaMask Extension with per-repo overlays. Shared body, same rules in both repos. Core stays out of scope.
• planning, planning-spec and jira-brief are maturity: experimental with no base, so a default stable sync installs none of them. planning triggers only when scope is ambiguous, the change is large, the work is security-sensitive or performance-sensitive, a contract or API is being changed, or the engineer asks for it.
• Security and performance stop for agreement before the edit, even when fully specified. A contract or API change, including a new public method, stops for a brief even when fully specified. A spec is only for a large change.
• large is tightened to a new product area, or a cross-system design with undecided scope.
• Cheap uncertainty (whether the result is kept, or whether an existing spec, plan or ticket covers the change) proceeds and ends with one confirmation line. Risk uncertainty stops for a short brief and does not recommend a spec.
• Catalog step reads every installed mms-* description on every path, and follows a matching mms-* skill over a non-mms-* one. If a later step is reached before that read, the catalog runs first. The fixed mms-performance read is dropped, since that skill does not install for Extension. On Mobile, performance-sensitive work still reads mms-performance.
• The line A Jira ticket would improve this context. is removed. jira-brief reads a ticket when one is identified; otherwise planning continues without asking.

Type of Change

• Skill improvement/update

Testing

Latest check, on 1d402d6: 19 scenarios × 3 labs. The description repair is in that commit: the accepted suggestion had split the ambiguous-scope sentence and repeated the security line. The contract trigger and the catalog recovery sentence stayed. The only run comment has that table. The table below is the first 54 runs. Labs: Claude Sonnet 5.5 (high), GPT 5.6 Terra (medium), Grok 4.7 (high). Skills were installed locally in metamask-mobile and metamask-extension checkouts with tools/install --include planning/<skill>.

Outcome per scenario. Labels: "opened, proceeded" = skill read and triggered, then the agent continued without a pause; "note, proceeded" = continued with the closing note; "brief, no spec" = short brief, no spec recommended; "short brief, no spec" = one-line pause for agreement; "spec" = planning spec recommended; "did not open" = skill not triggered.

Scenario	Claude	GPT	Grok
Typo, Mobile	opened, proceeded	did not open	did not open
Typo, Extension	did not open	opened, proceeded	did not open
Ambiguous send, Mobile	brief, no spec	brief, no spec	brief, no spec
Ambiguous swap, Mobile	brief, no spec	brief, no spec	brief, no spec
Vault, decision pending	short brief, no spec	brief, no spec	brief, no spec
Keyring, decision pending	brief, no spec	brief, no spec	brief, no spec
Explicit plan, Mobile	brief, no spec	brief, no spec	brief, no spec
Explicit plan, Extension	brief, no spec	brief, no spec	brief, no spec
Label, Mobile	did not open	opened, proceeded	did not open
Label, Extension	opened, proceeded	opened, proceeded	did not open
New area: prediction markets	spec	spec	spec
Jest (mechanical)	opened, proceeded	opened, proceeded	did not open
Vault, locked	short brief, no spec	short brief, no spec	short brief, no spec
Public method on keyring	short brief, no spec	brief, no spec	brief, no spec
New area: household, Mobile	spec	spec	spec
New area: household, Extension	spec	spec	spec
Cheap uncertainty, Mobile (send amount)	brief, stopped	note, proceeded	note, proceeded
Ticket EXT-4401, Extension	note, proceeded	note, proceeded	note, proceeded

Notes:

• Security-sensitive scenarios paused every time, without a spec.
• Over-triggering on mechanical work: the skill opened on typo, label and Jest scenarios in several runs, but then proceeded. No pause happened on those. Known gap: the Do not use this skill before every edit line does not fully stop this yet (7 of 15 mechanical runs opened the skill).
• Claude stopped for a brief on cheap uncertainty (send amount, Mobile). GPT and Grok proceeded with the note. The request was ambiguous, so the brief is defensible. This is a lab divergence, not a rule-level change.
• Claude used a short brief for the keyring public method. GPT and Grok wrote a full brief. None recommended a spec.
• Large-scope judgement differs between labs on some borderline cases. The table records what each run did.

Validation run locally:

• yarn test: 76/76 pass.
• yarn lint:changelog: pass.
• yarn audit:skills: 0 errors.

Limitations:

• The table is the first 54 runs, taken before the Jira-line removal. Later checks are summarised here: on 9d2bc38, spec and Jira paths still fired and no ticket was requested; on 7613162, the clear performance case and the borderline selector case paused, the plain UI change did not, and ticket and no-ticket paths behaved on Mobile and Extension. The only run comment is the latest, on 1d402d6: 19 scenarios times 3 labs.

Checklist

• I have read the CONTRIBUTING.md guidelines
• My skill follows the SKILL_TEMPLATE.md format
• I have tested this skill with an AI agent
• My skill does not contain any secrets, private keys, or sensitive data
• I have added appropriate documentation
• My changes don't break existing skills

[NicolasMassart]

Explanation of my above comments:

issue (blocking): Contract/API changes should explicitly trigger planning

The workflow already correctly states that contract/API changes require a brief, even when fully specified, but don't require a spec unless the change is also large.

However, contract/API changes are missing from the skill's trigger conditions, meaning mms-planning might never activate for these cases.

I'd suggest adding contract/API changes to both the frontmatter description and the When to use section.

This would make the trigger conditions consistent with the existing workflow without broadening the scope unnecessarily or requiring further changes.

[mganchas]

Re-check on 1d402d6

This is the only run comment. Earlier evidence comments were removed. 19 scenarios × 3 labs. The first 18 are the original set. The last one is new: a fully specified public method that is not security-sensitive.

Scenario	Claude	GPT	Grok
Typo, Mobile	did not open	opened, proceeded	did not open
Typo, Extension	did not open	opened, proceeded	did not open
Ambiguous send, Mobile	full brief, no spec	full brief, no spec	full brief, no spec
Ambiguous swap, Extension	full brief, no spec	full brief, no spec	full brief, no spec
Vault, shape not decided	full brief, no spec	full brief, no spec	full brief, no spec
Keyring, decision pending	full brief, no spec	full brief, no spec	full brief, no spec
Explicit plan, Mobile	full brief, no spec	full brief, no spec	full brief, no spec
Explicit plan, Extension	full brief, no spec	full brief, no spec	short brief, no spec
Label, Mobile	did not open	opened, proceeded	did not open
Label, Extension	did not open	opened, proceeded	did not open
New area: prediction markets	brief, spec	spec	spec
Jest, routine CI	opened, proceeded	opened, proceeded	did not open
Vault, re-encrypt, fully specified	short brief, no spec	short brief, no spec	short brief, no spec
Public method on keyring	full brief, no spec	full brief, no spec	full brief, no spec
New area: household, Mobile	brief, spec	spec	brief, spec
New area: household, Extension	brief, spec	brief, spec	brief, spec
Cheap uncertainty, keep the result	full brief, no spec	proceeded	cheap note, proceeded
Ticket might cover the change	cheap note, proceeded	cheap note, proceeded	cheap note, proceeded
Public method, not security, fully specified	full brief, no spec	full brief, no spec	full brief, no spec

The new contract trigger works: the non-security public method stopped for a brief with no spec in all three labs. Security still stops without a spec. Large areas still recommend a spec. Mechanical work still opens on GPT, then proceeds. That known gap is unchanged.