implement-advisor

name

implement-advisor

description

Detects when the user wants to work on a Linear issue, implement a feature, fix a bug, or start development work. Suggests the /make-no-mistakes:implement command. Triggers on: "implement", "work on issue", "start on", "pick up issue", "Linear issue", "build feature", "fix bug", "start coding", "work on <PREFIX>-NNN" or "implement <PREFIX>-NNN" for any Linear issue prefix (APP, BACK, SEC, MYTEAM, etc. — based on your workspace), "implement this", "let's build", "start implementing", "pick up this ticket", "work on this task", "execute this issue", "develop this feature", "ship this", pastes a Linear issue URL, or provides an issue ID. Does NOT trigger on: rebase, sync branches, standup, test execution, code review, general coding questions, or session management tasks.

Implement Advisor

You detected that the user wants to implement a Linear issue or start development work — not review, sync, or test.

CHANGELOG

• 2026-05-20 — Surface canonical-URL drift gates from spike-recommend. When the issue under analysis is a URL canonical migration (touches <Route> mounts, buildXxxUrl helpers, or aliases_redirect_only in a routes map) AND there's an in-flight PR overlap, the redaction-quality gate is no longer sufficient on its own. Even a Bilingual-formatted brief is "soft" if it doesn't enumerate the SEMANTIC contracts both PRs touch (Rule 11 in spike-recommend) and assert URL-builder ↔ Route reachability (Rule 12). Recommend re-running /make-no-mistakes:spike-recommend {ISSUE-ID} first to refresh the brief against the v1.16.0+ ruleset, THEN /make-no-mistakes:implement. Background: DOJ-4200 + DOJ-4075 lessons (dojo-os, 2026-05-20).
• 2026-05-13 — Gate on issue redaction quality. Before recommending /make-no-mistakes:implement, fetch the Linear issue and check whether its description is in Bilingual Format (Human Layer + Agent Layer + ACs + Context Files). If not, recommend /make-no-mistakes:spike-recommend {ISSUE-ID} first to normalize the description in place, THEN /make-no-mistakes:implement. Implementation agents work much better against a normalized brief.

When This Applies

This skill activates when the user describes a situation involving:

• Working on a specific Linear issue (ID or URL)
• Starting a new feature, fix, or chore
• Picking up a ticket from the backlog
• Any intent to write code for a tracked issue

What To Do

1. Confirm intent. Make sure the user wants to implement (not just discuss or review). If the intent is unclear, ask.

2. Parse the issue ID from the user's message ($ARGUMENTS, the pasted URL, or the most recent conversation context). Format: {PREFIX}-{NNN} (e.g., DOJ-4058, ALT-13).

3. Redaction-quality gate — fetch the issue and inspect its description. Call mcp__plugin_linear_linear__get_issue with the parsed ID and read the description field. Naming note: the registered MCP namespace uses a single underscore between the two linear tokens (plugin_linear_linear) — matches what spike-recommend declares. If the literal tool name is not present in the runtime, probe with the equivalent mcp__*linear*get_issue glob; the actual prefix may differ slightly per workspace MCP server registration.

   Check for Bilingual Format markers. The canonical definition lives in docs/bilingual-format-standard.md; the four required headers (any missing ⇒ NOT normalized) are:

   • ## 👤 HUMAN LAYER
   • ## 🤖 AGENT LAYER
   • ### Acceptance Criteria
   • ### Context Files

   Alternative quick check: presence of both literal strings HUMAN LAYER and AGENT LAYER anywhere in the description. If both strings are absent, the issue is not redacted.

   Edge cases:

   • Empty / 1-line description → not redacted.
   • Linear get_issue fails or the user passed a free-text request with no ID → skip the gate and treat as not redacted (recommend /spike-recommend first).
   • Description has the Bilingual headers but the sections are stubs ("TBD", "N/A everywhere") → treat as not redacted; the implementation subagent still gets a useless brief.

4. Recommend based on the gate's result.

   If the issue IS redacted (Bilingual Format present and substantive):

   This is an implementation task. Use:

   /make-no-mistakes:implement {ISSUE-ID}

   This command handles the full disciplined protocol:

   • Fetches the Linear issue (title, description, status, labels)
   • Claims and sets status to In Progress
   • Creates a fresh branch + worktree (isolated from main tree)
   • Implements following all project conventions
   • Creates a PR with linked issue
   • Tags all reviewers (Greptile, CodeRabbit, Graphite)
   • Fixes reviewer feedback until all gates pass
   • Verifies CI, merges, cleans up worktree
   • Updates Linear to Done

   For multiple issues: /make-no-mistakes:implement ALT-13 ALT-14 ALT-15

   If the issue is NOT redacted (missing Bilingual Format, or empty, or stub-only):

   The issue {ISSUE-ID} doesn't yet use the Bilingual Format (Human Layer + Agent Layer + ACs + design rationale). Implementation agents work much better against a normalized brief — running /implement against a 1-liner produces shallow work.

   Two-step flow:

   1. /make-no-mistakes:spike-recommend {ISSUE-ID} — fetches the issue, regenerates the description in Bilingual Format, updates Linear in place (labels go to the sidebar, body stays narrative).
   2. /make-no-mistakes:implement {ISSUE-ID} — then run the full disciplined protocol against the now-normalized brief.

   If you want to skip the gate and run /implement directly anyway (e.g. you've already redacted the issue manually), say so and I'll defer.

5. Other intents (escape hatches):

   • If the user is just asking about an issue (not implementing), suggest spike-recommend or spec-recommend instead.
   • If the user wants to review existing work, suggest review-open-prs instead.