fast-implement
Implement a piece of work in a single session, without the coordinator pipeline's approval gates.
SKILL.mdskills/first-party/pvmalove/fast-implement/SKILL.md
GitHubFast implement
Objective: Implement the work described by the user in the spec or tickets, in this session, with no architect step, no independent QA role, and no coordinator approval gates.
This is the short path. Use it for a ticket whose direction is not in question and whose blast radius
is small. For anything else use /implement, which runs the gated architect → developer →
code-review → QA pipeline.
Execution in Three Phases
A strict pipeline, resolved in order: Pre-flight (confirm the ticket is actually startable, and by this agent) → Coding (TDD, tests, an explicitly approved review, commit, and push) → PR & Wrap-up (offer the separate /to-pull-requests command). Only the developer can select /to-pull-requests.
Phase 1: Pre-flight
On GitLab, every glab command names the project with -R <project-url>; <project-url> and the other GitLab placeholders (<host>, <project-id>) are defined in docs/agents/issue-tracker.md → GitLab → Conventions.
- Resolve the ticket.
- A specific ticket is named (an issue number, URL, or a
.scratch/<feature>/issues/NN-*.mdpath): use it.- Fail fast on
hitl: if it carrieshitl(seedocs/agents/triage-labels.md), stop immediately and tell the user to run/to-guideinstead — don't check anything else on this ticket, this agent only implementsafkwork. - Fail fast without
pipeline::fast: for anafkticket, require the explicitpipeline::fastlabel (GitHub/GitLab) or**Pipeline:** pipeline::fastfield (local tracker) — seedocs/agents/triage-labels.md. If the label/field is absent entirely or ispipeline::full, stop immediately and tell the user to run/implementinstead.
- Fail fast on
- An epic is named (carries
status::specs, or is otherwise the parent of a decomposition) rather than a specific ticket: pick the next ticket yourself instead of asking. Never auto-pick ahitlticket — that execution mode always routes through/to-guide, not this agent.- GitHub/GitLab: run the same frontier query
/wayfinderuses (docs/agents/issue-tracker.md#wayfinding-operations), scoped to the epic's sub-issues and filtered topipeline::fast+afk— open, unblocked, unclaimed, first in decomposition order. Claim the chosen ticket (gh issue edit <n> --add-assignee @meon GitHub,glab issue update <n> -R <project-url> --assignee @meon GitLab) before any other write, the same way/wayfinderclaims a ticket, so a concurrent session (e.g. a parallel worktree) doesn't pick the same one. If the filtered frontier is empty, stop and tell the user why: if open, unblocked, unclaimed tickets remain but all arehitl, say so and point at/to-guide; if unblockedafktickets remain but none carrypipeline::fast(label missing orpipeline::full), say so and point at/implement; otherwise explain that nothing is unblocked yet, or everything is already claimed. - Local tracker: read each
.scratch/<feature>/issues/NN-*.mdfile in filename order — a purely linear chain — and take the first one that is**Workflow:** status::ready,**Execution:** afk, and**Pipeline:** pipeline::fast. If tickets remain but everystatus::readyone ishitl, say so — naming them — and point at/to-guideinstead of picking one; ifafkones remain but none carry**Pipeline:** pipeline::fast, say so — naming them — and point at/implementinstead. - Once chosen this way, treat the ticket exactly like one named explicitly for the rest of this process.
- GitHub/GitLab: run the same frontier query
- Nothing is named: when this repo defines a git workflow doc (e.g.
docs/agents/git-workflow.md) with an "Issue First" rule, stop and ask the user to name an existing ticket or run/to-spec//to-ticketsfirst — don't start the work. Otherwise, skip this check entirely.
- A specific ticket is named (an issue number, URL, or a
- Check blockers, on the resolved ticket, whatever its current
status::*label. Read its blockers from this repo's tracker (native GitHub/GitLab dependency links, or theBlocked by:/**Blocked by:**field — seedocs/agents/issue-tracker.md).- Any blocker still open → stop and tell the user which ones. Don't start the work. If the ticket isn't already
status::blocked, set it (same label swap as step 3). - All blockers closed/resolved (or none) → continue. A
status::blockedticket passes throughstatus::readyin step 3.
- Any blocker still open → stop and tell the user which ones. Don't start the work. If the ticket isn't already
- Mark it in progress before creating the issue branch or editing any file, in every run mode, cloud and single-session runs included. A ticket carries exactly one
status::*label (seedocs/agents/triage-labels.md), so replace, don't add:- GitHub:
gh issue edit <n> --remove-label status::ready --remove-label status::blocked --add-label status::in-progress, then confirm withgh issue view <n> --json labels --jq '[.labels[].name]'thatstatus::in-progressis the onlystatus::*label. - GitLab:
glab issue update <n> -R <project-url> --unlabel status::ready,status::blocked --label status::in-progress, then confirm withglab issue view <n> -R <project-url> -F jsonthatstatus::in-progressis the onlystatus::*label in itslabels. - Local tracker: set the file's
**Workflow:**line tostatus::in-progress. - If the label write fails, stop and report it; don't start coding on an unmarked ticket.
- GitHub:
- Git pre-flight, before editing files:
- Resolve the exact integration branch from the ticket's
## Integration Branchsection or, for a child ticket that omits it, from its parent epic. An absent value is a blocker; do not infer a branch from memory, the current checkout, or a service name. - The current branch must match
branch_patternand be an issue branch. If it does not, fetch the integration branch and createfeature/issue-<ID>-<slug>from it before coding. Never commit or push directly to the project base branch or anintegration/*branch.
- Resolve the exact integration branch from the ticket's
Phase 2: Coding
- Agree the seams before the first test. Name the requirement and the public seam where its behavior is observed, with the public signatures and existing fixtures. When the ticket or spec already names them, restate them and ask the developer to confirm; otherwise propose them. Stop for the developer's answer. Never invent a seam and never write a test before the developer has agreed it.
- Use
/tddat those seams. Run.claude/hooks/test-author-preflight.shonce before the first test. Exactly{"status":"ok"}selects the test-author cycle below. Any other answer (disabled, a missing or failing hook, unavailable isolation, notest_pathsortest_runner) selects the ordinary/tddroute in this session: tell the developer the sanitizedreason(or that the hook failed) in one line, do not start an unprotected author, and do not re-run preflight hoping for another answer. Neither route adds a coordinator, architect or QA step. - Test-author cycle. It is the same cycle as the full route (
Red-Green-Refactor with the optional test authorin the orchestrationdeveloper.md, and the tdd skill), and you play the developer. For each implementation step, in order:- Handoff: call the
test-authorwith the requirement, the agreed seam, the public signatures, the existing fixtures and one scenario. Never pass a proposed algorithm or implementation snippet as the basis of an expectation. - Reply: it answers in compact English with
status(red,needs_skeletonorcontract_gap), the requirement, the inputs, the source of the expectation, the test id and, forred, the sanitized runner JSON plus the failed assertion, the expectation and the actual result from the run'slog.txt. - Verify the expectation against the requirement and its stated source before any implementation. An expectation derived from the implementation, or one the requirement does not support, goes back to the author as a failed call. An accepted expectation is never weakened to reach Green; only a justified contract change (the requirement or the agreed seam changed with the developer) allows correcting it, with a new author call and a new verification.
- Statuses:
redcontinues.needs_skeleton: write only types and signatures with no logic, then rerun the author on the same scenario.contract_gap: clarify the requirement or seam with the developer, never guess it. - Not a behavioral Red: a compile or import error, a fixture error, and an environment or runner refusal are never Red, whatever
statussays. A test that is Green immediately does not count as a cycle. Each is a failed author call. - Green, then Refactor: after a confirmed behavioral Red, write the minimal implementation that passes the unchanged test and run it by its test id; refactor only on Green, re-run the test, then request the next test.
- Attempt limit: after three consecutive failed author calls on one test, stop the local retries and report the blocker to the developer. A guard denial counts as one failed call, and the denied call is never rephrased. Local retries create no dispatch and no ledger state.
- Handoff: call the
- Run typechecking regularly, single test files regularly, and the full test suite once at the end.
- Ask the developer: “Провести code review?” Stop for their answer.
- Yes: run
/code-review. It launches the Standards and Spec subagents through the coding application's manually configured mechanism, waits for both reports, and returns its separate## Standardsand## Specreport to this primary session. Address any requested changes, then repeat the relevant tests before continuing. - No: record that the developer declined review and continue.
- Yes: run
- Ask the developer for explicit permission to commit and push. Stop for their answer.
- After approval, verify that the current branch still matches
branch_patternand is neitherbase_branchnorintegration/*; commit the completed work with a Semantic Commit Message and push it to the current issue branch. Report the commit and push result to the developer.
Phase 3: PR & Wrap-up
After a successful push, offer /to-pull-requests <ticket> as the next command. Do not invoke it automatically, open a PR, run qa-gate, or close the ticket in this skill. Leave status::in-progress on the ticket: this skill never removes it. Closing the ticket and moving its unblocked dependents to status::ready belong to /to-pull-requests after the merge.