Read AGENTS.md and, when it exists, docs/llm/INDEX.md for conventions and affected wiki pages; a missing wiki must not block work. If AGENTS.md contains Read CONVENTIONS.md before editing., read the root CONVENTIONS.md too. Apply only rules whose scopes match the changed files (most specific wins); route same-scope conflicts through user clarification.
If application source exists, call the available Skill entry whose local name is read-codebase-context to trace the relevant code path; understand callers, data flow, and error paths before editing. For a source-less project, read the design, tasks, and delta without creating source, and state that callers and error paths do not exist yet. When the active plan has an ## Impact map, follow read-codebase-context's ## Refresh an impact map (an unusable baseline means tracing from scratch; planned source that does not exist yet uses the source-less map rule). Read current source for every existing file to edit; for a planned new file, follow the approved design and read the source after creating it. The map is navigation, not evidence.
Follow the approved plan-feature output when one exists. For a change, read the change files that exist (design.md, tasks.md, delta.md), and read tasks.md's ## Approval Gate before editing application code when it exists; a Bounded or bug-fix change may have only delta.md.
- A tasks.md proceeds only with Status: approved, which covers design.md, tasks.md, and any delta.md together. If it is missing or pending, stop; plan-feature step 9 asks for and records the approval.
- A legacy plan or an older tasks.md with a Required gate follows the gate rules in ../using-devkit/references/legacy.md; read it when you meet one.
Apply using-devkit's ## Change lifecycle: an archived change and a completed legacy plan are historical, so route new work through brainstorm-feature; an open change or plan supplies its remaining tasks.

Without a plan, require an approved bounded design from brainstorm-feature, or call it before editing. That approval counts only from the current conversation, a delta.md Approved: line, or a handoff's Approval: line; a resumed session with none of these re-presents the design through brainstorm-feature. A clear-small-change notice from brainstorm-feature satisfies this gate. If implementation uncovers impact outside the user's named scope, stop and return to the approval gate.
Read ## Decision sources below; conversation memory is not one. Persist a newer user answer through the clarification flow below before using it.
After tracing the real flow, apply ## Implementation ladder below.
Follow using-devkit's baseline rule.

Implementation ladder

Apply in order and stop at the first option that satisfies the approved behavior; plan-feature, systematic-debugging, and review-and-verify use this ladder too:

Does this need to exist at all? Skip speculative work (YAGNI).
Does it already exist in this codebase? Reuse the module, helper, type, or pattern.
Does the standard library do it? Use it.
Does a native platform feature cover it? Use it.
Does an already-installed dependency solve it? Use it.
Only then, write the minimum clear new code that works.

Do not use minimalism to remove explicit requirements, trust-boundary validation, security, accessibility, or error handling that prevents data loss.

Decision sources

Read, when they exist: the change's decisions.md; tasks.md's ## Decision Log; for a legacy plan or spec, its ## Decision Log and any task-linked file under docs/agent-devkit/decisions/ listed in docs/agent-devkit/INDEX.md (legacy, read-only). Skills that must honor decisions (review-and-verify, context-handoff) read this same list.

During implementation

With a plan, work through tasks in order, starting at the first task whose Progress is not done (backfill a task with no Progress line first, per ../using-devkit/references/legacy.md). After a task's Verify check passes, set Progress: done and Result: <check> — <outcome> before the next task. A failing task stays todo; fix it within scope and rerun the check, under the stop rule in "After implementation" step 3. Continue until every task is done, unless a stop condition applies. For a source-less project, bootstrap the first planned entry point as part of its approved task, then read the created source before verifying it.
Reuse local patterns. Keep the diff focused on one logical change.
Make the smallest change that satisfies the task. Do not refactor unrelated code.
On a bug or unexpected behavior, call the available Skill entry whose local name is systematic-debugging as a repair handoff; it returns its fix/check evidence here, and implementation owns the final review and archive. Do not guess-and-check.
Run the smallest relevant check after each non-trivial change.
Do not commit (using-devkit).

Technical rulings

Resolve a choice without user input only when every viable option preserves the same observable behavior, fits the approved design, plan, decisions, and repository contract, and changes no API, schema, dependency, security boundary, scope, or data-loss risk. It must be non-destructive and reversible within the task: no migration, data rewrite, external contract change, or caller change outside it.

Choose with the implementation ladder and report:

Ruling R<n>: <choice and concise repository-grounded reason>.

Do not persist a ruling in ## Decision Log; it changes no approved requirement and can be re-derived from source. If a condition above fails, an approved artifact conflicts, or the action is destructive or irreversible, use the clarification flow below.

Clarification decisions

When implementation needs a user answer before it can continue:

Stop editing and ask one question. After the answer, restate it as Decision D<n>: <one unambiguous sentence> before taking another action.
Keep an implementation-only choice (no change to observable behavior, requirement, API, schema, security boundary, or scope) in the session; create no artifact for it.
Persist every answer that changes observable behavior or an approved requirement. For a change, append it to the change's decisions.md; when the folder does not exist yet (a Bounded or clear-small change with no delta), create it and index it per using-devkit's ## Change lifecycle. For a legacy plan or spec, append to its ## Decision Log; create no new file under docs/agent-devkit/decisions/.

Optionally memory_write only the decision's title and file path; memory is never the only copy. memory_recall (1–3 keywords, maxTokens when supported) may locate decision or handoff files; read them before relying.

Use this shape:
```
### D<n> — <short title>

Question: <what was unresolved>
Decision: <the user's answer>
Impact: <requirements, tasks, interfaces, or tests affected>
Confirmed by user: YYYY-MM-DD
```
If the answer materially changes scope, observable behavior, an interface, a schema, or risk, stop editing code. For an open tasks.md or legacy plan, call plan-feature as a route handoff to revise it; plan-feature records the decision log entry, resets the gate to pending, presents the revision, and resumes this skill after re-approval. For a change to design.md or delta.md, or when a bounded task expands beyond its approved design, call brainstorm-feature instead of silently widening scope. A non-material answer only needs the decision record from step 3.

After implementation

After code checks pass, assess documentation using approved scope, the diff, and traced callers. When docs/llm/ exists and task behavior is new, changed, missing, or stale, call document-wiki's ## Task updates (it merges any delta first and verifies the affected wiki). A missing wiki means skip, not setup; unrelated gaps stay optional findings. Resolve in-scope documentation blockers before final review, under the stop rules below.
Call the available Skill entry whose local name is review-and-verify to review the diff, run fresh verification, and confirm no stale documentation.
If the review result is Status: fail, fix the listed blockers that are inside the approved scope, run the smallest relevant check, and call the available Skill entry whose local name is review-and-verify again. Each blocker gets at most three fix attempts. Stop and report the remaining blockers, without claiming completion, when any of these holds:
- the same blocker remains after a fix attempt with no new information;
- three fix attempts for one blocker have failed (the architecture signal in systematic-debugging); end with Next: brainstorm-feature to revisit the design;
- the fix needs a change outside the approved scope; or
- the fix needs a user decision.
When this skill stops before review (a question, pending gate, or scope growth), end with the Next: block from using-devkit, naming the current task.

Red flags

Thought	Reality
"The user said implement earlier"	An instruction before the approved files existed is not approval; check the gate.
"Approval is in the chat history"	Only this conversation, an `Approved:` line, a handoff's `Approval:` line, or `tasks.md`'s `Status: approved` counts.
"One more fix attempt"	Three failed attempts on one blocker end the loop; route to `brainstorm-feature`.

implement-task

Implement Task

Before coding