| name | cubrid-grill-and-implement |
| description | Iteratively implement CUBRID C/C++ code changes by looping a writer subagent against a relentless CUBRID PR-style reviewer subagent, with a build gate every round, until the reviewer explicitly approves or a round cap is hit. Accepts a JIRA issue, a GitHub PR URL, a spec path, or an inline spec - refuses to start without one. Trigger on phrases like 'implement and grill', 'grill the implementation', 'CUBRID grill-and-revise', 'implement CBRD-XXXXX with grill', 'implement and adversarially review', or 'grill until it compiles and passes review'. |
| argument-hint | <CBRD-XXXXX | pr-url | spec-path | inline-spec> |
| allowed-tools | Bash(bash *), Bash(gh *), Bash(git *), Bash(jq *), Bash(just *), Bash(tee *), Bash(tail *), Bash(cat *), Bash(test *), Bash(grep *), Bash(realpath *), Read, Write, Edit, Glob, Grep, mcp__plugin_oh-my-claudecode_t__lsp_diagnostics, mcp__plugin_oh-my-claudecode_t__lsp_diagnostics_directory, mcp__plugin_oh-my-claudecode_t__lsp_hover, mcp__plugin_oh-my-claudecode_t__lsp_goto_definition, mcp__plugin_oh-my-claudecode_t__lsp_find_references, mcp__plugin_oh-my-claudecode_t__lsp_document_symbols, Agent |
CUBRID Grill-and-Implement
Iteratively implement CUBRID code changes by looping a writer subagent (plans, then writes code) against a relentless reviewer subagent (CUBRID PR-style review with LSP/clangd) until the reviewer approves or a round cap is hit. The loop guarantees the final tree compiles, every reviewer concern is addressed in code (not in argument), and the writer never short-circuits with self-approval.
This is the code-implementation counterpart to /grill-with-docs. Same two-agent loop and VERDICT contract; the artifact is a working CUBRID code change in the user's tree, not a document.
NOTE for installers: The agent-spawning tool in this Claude Code harness is named Agent. If your runtime exposes it as Task instead, substitute Task for Agent in the allowed-tools frontmatter and in every spawn step below. Do not list both.
Cost Characteristic
Each round runs an opus writer plus a critic-class reviewer over a possibly-large CUBRID diff with LSP, plus a full build (15-min timeout). Roughly 10x more expensive per round than /grill-with-docs. Default max_rounds=3. If at any reviewer round the cumulative line count from git diff --stat <baseline_ref> -- exceeds 5000, print a warning and require an explicit yes from the user before continuing past round 2.
When to Use
- User says "implement and grill", "grill the implementation", "CUBRID grill-and-revise", "implement CBRD-XXXXX with grill", "implement and adversarially review", or "grill until it compiles and passes review".
- User wants an adversarially-reviewed CUBRID code change driven by a JIRA ticket, a PR to extend, or a written spec.
- User is willing to trade tokens for rigor.
Arguments
/cubrid-grill-and-implement CBRD-26583 - Implement the JIRA ticket./cubrid-grill-and-implement https://github.com/CUBRID/cubrid/pull/6950 - Extend or fix the existing PR./cubrid-grill-and-implement /path/to/spec.md - Implement the spec file./cubrid-grill-and-implement "<inline multi-line spec>" - Implement the inline spec./cubrid-grill-and-implement - Refuse with the four input options. Do not assume, do not infer from git status, do not pick the most recent JIRA mention from the conversation.
Required Input
Exactly one of:
- JIRA issue:
CBRD-XXXXX - fetched via /cubrid-jira CBRD-XXXXX.
- PR URL: existing GitHub PR to extend or fix - fetched via
gh pr view and gh pr diff.
- Spec file path: a markdown spec describing what to build.
- Inline spec: multi-line description passed as the argument.
State to Track Across Rounds
input_kind - one of jira / pr / spec-path / inline-spec. Used as the trigger for PR-mode-only rules; do not key those rules off how baseline_ref was resolved.input_payload - the resolved spec content (JIRA description+comments, PR diff+body, file contents, or inline text).baseline_ref - captured before round 1 writer pass per Step 3; never reassigned.round - starts at 1; increments after every reviewer or build-gate failure round.max_rounds - default 3; user can override at invocation.last_critique - "" on round 1; otherwise either the reviewer's most recent numbered critique or the synthesized build-failure critique.last_reviewer_critique - "" until the first reviewer round runs; thereafter the most recent reviewer-issued critique. Independent of last_critique. Used for the empty-diff prepend in Step 5 and the cap-reached print in Step 7.previous_round_diff_summary - the diff stat captured at the end of the most recent build-passing, non-empty round; "" on round 1; may be stale if subsequent rounds reverted.build_status - pass / fail / skipped for the most recent round.cubrid_grill_log - the per-round build-log path, recomputed at the start of every Step 5 invocation: /tmp/cubrid-grill-build-${baseline_ref:0:12}-r${round}.log. Pass this literal string through state. Do not use $$.last_writer_summary - the line beginning with Summary: from the writer subagent's response, captured at Step 4 exit; the Summary: prefix is stripped, and the captured value is the trimmed remainder. "" until Step 4 has run at least once. Used by Step 6 to seed the reviewer's "Justified rename/move/reformat" check.
Subagent Choices
- Writer:
subagent_type=executor, model=opus. CUBRID is C/C++ with concurrency, page-buffer, WAL, and MVCC rules - sonnet under-thinks too often.
- Reviewer: prefer
subagent_type=critic. Fallback chain in this exact order: critic -> code-reviewer -> executor. Operational definition of "unavailable": the spawn tool returns an error mentioning unknown/unregistered subagent type, or the subagent type is absent from the runtime agent catalog. Try the next entry only on unavailability. A substantive review failure (the agent ran but returned no parseable verdict) is handled by Step 6's re-prompt-and-stop rule, not by the chain. Use the bare names listed here. If your harness exposes them as namespaced (e.g., oh-my-claudecode:critic), the skill must attempt the bare name first and, on unknown subagent type errors, attempt the namespaced variant before falling through to the next chain entry.
- Never have the orchestrator self-review. If every chain entry fails to spawn, surface the error and stop the loop.
Execution Steps
Step 1: Parse argument and refuse-on-empty
If the argument is empty, print the four input options and stop. Otherwise classify the argument:
- matches
^CBRD-\d+$ -> jira
- matches
^https?://github.com/.+/pull/\d+ -> pr
- existing file path ->
spec-path
- otherwise ->
inline-spec
Step 2: Verify working tree is a CUBRID checkout
Resolve and source the shared helper from the sibling cubrid-common skill, then require a CUBRID source checkout:
common="<this-skill-dir>/../cubrid-common/scripts/cubrid-common.sh"
source "$common"
cubrid_require_source_tree "$PWD"
If this fails, refuse to start with the helper's message. Do not proceed past this step. The helper owns the project(CUBRID) detection regex so this rule stays consistent with the other CUBRID skills.
Step 3: Fetch context and capture baseline
Per input_kind:
jira: invoke /cubrid-jira CBRD-XXXXX.pr: gh pr view <url> --json title,body,headRefName,baseRefName and gh pr diff <url>. PR-mode preconditions: require the user to already be on the PR's headRefName (validate via gh pr view --json headRefName --jq .headRefName vs git branch --show-current); refuse with a clear error if not. The PR diff and body are treated as already-applied work and as the spec for extending or fixing the PR. If the PR is closed/merged and gh pr diff returns empty, refuse with: "PR diff is empty (closed or merged). Pass a JIRA, spec path, or inline spec instead." If the PR title contains a CBRD ticket, also fetch JIRA context.spec-path: Read the file.inline-spec: store the argument verbatim.
Always Read the resolved reference.md path (see Reference Files below) and any CLAUDE.md / AGENTS.md in the working tree's relevant directories.
Capture baseline:
- For
jira / spec-path / inline-spec: baseline_ref = $(git rev-parse HEAD).
- For
pr: detect the upstream CUBRID remote (do NOT hardcode origin; CUBRID checkouts commonly use cub, vk, hg, etc.). Use UPSTREAM_REMOTE=$(cubrid_detect_remote) from the shared cubrid-common helper. Run git fetch "$UPSTREAM_REMOTE" <baseRefName> to refresh the remote-tracking ref. Then validate with git rev-parse --verify "$UPSTREAM_REMOTE/<baseRefName>"; if that fails, refuse with the message the helper emits and stop. On success: baseline_ref = $(git merge-base HEAD "$UPSTREAM_REMOTE/<baseRefName>"). The reviewer must judge the PR's full delta, not only the writer's new edits on top of the PR head.
Print baseline_ref to the user. Initialize round=1, max_rounds=3 (or user override), last_critique="", last_reviewer_critique="", previous_round_diff_summary="", build_status="skipped".
Step 4: Writer pass
Spawn the writer subagent. Pass:
input_kind, input_payload, baseline_ref, round, last_critique, previous_round_diff_summary.- The resolved absolute path to
reference.md.
- The contents of
references/writer-prompt.md verbatim.
references/writer-prompt.md is the single source of truth for writer rules - do not duplicate any rules inline in the spawn prompt. Wait for the writer to finish before continuing.
Capture last_writer_summary as follows: scan the writer's response from the bottom up for the first line matching ^Summary:[[:space:]]*(.*)$; the captured group (trimmed) is last_writer_summary. If no such line is found (writer protocol violation against rule 3), set last_writer_summary = "", surface a warning at Step 6 input, and let the reviewer flag the missing summary as a finding.
Step 5: Build gate
At the start of this step, compute cubrid_grill_log = /tmp/cubrid-grill-build-${baseline_ref:0:12}-r${round}.log and use the literal string in every command below.
Detect the build command and validate the build environment:
On non-zero status:
build_status=fail.- Synthesize
last_critique as: line 1 = Build failed at round <round>; address the build error first, then re-address any prior reviewer items (the reviewer will re-raise unfixed items in the next reviewer round). line 2+ = Tail of <cubrid_grill_log>: followed by the output of tail -n 40 "$cubrid_grill_log". Treat the build log as opaque text; never interpolate it through a shell context. Pass it to the writer as a single string argument.
- Do not modify
last_reviewer_critique. Build-failure synthesis replaces only last_critique for the round; carry-forward across build failures is by re-review, not by concatenation.
- Increment
round. If round > max_rounds go to Step 7 (cap-reached); else go back to Step 4.
On zero status:
build_status=pass.- Capture
previous_round_diff_summary = $(git diff --stat <baseline_ref> --).
- If the writer produced an empty diff (
git diff --quiet <baseline_ref> -- returns 0):
- If
round == 1: print the writer's response, the spec, baseline_ref, and stop with: "Writer made no changes on round 1; clarify the spec and re-invoke."
- If
round >= 2: synthesize last_critique. If last_reviewer_critique is non-empty, set last_critique = "Writer reverted prior changes; re-address the previous reviewer critique below." prepended to last_reviewer_critique. If last_reviewer_critique is empty (no reviewer round has run yet), set last_critique = "Writer reverted or made no edits; the prior critique was a build failure - re-read it and produce a real change." Increment round and go back to Step 4.
- If the cumulative line count in
previous_round_diff_summary exceeds 5000 and round >= 2, print the warning per Cost Characteristic and require explicit yes before continuing.
- Continue to Step 6.
Step 6: Reviewer pass
Spawn the reviewer subagent per the fallback chain. Pass:
- The contents of
references/reviewer-prompt.md verbatim.
- The resolved absolute path to
reference.md.
baseline_ref.input_kind (so the reviewer can gate PR-mode-only categories).- The output of
git diff --stat <baseline_ref> -- (size summary only).
- The output of
git diff --name-only <baseline_ref> -- (changed-file paths only).
input_payload so the reviewer can judge fit-to-spec.previous_round_diff_summary.last_writer_summary (captured at Step 4 exit per the State definition above; used to verify any rename/move/reformat justification). If last_writer_summary == "", prepend a one-line warning to the reviewer prompt: "Writer protocol violation: no Summary: line in the writer's response. Flag the missing summary as a finding."- The verdict contract: end the response with exactly one of
VERDICT: APPROVED or VERDICT: REVISE on its own line, no markdown, no surrounding punctuation.
Do not pass the full diff body. Instruct the reviewer to fetch each file's diff itself via git diff <baseline_ref> -- <path> and to Read each changed file for full-function context.
Parse the verdict:
APPROVED -> go to Step 7 with success.REVISE -> save everything before the verdict line as last_critique and as last_reviewer_critique; increment round; if round > max_rounds go to Step 7 with cap-reached, else go back to Step 4.- No parseable verdict -> re-prompt the same reviewer once with the contract reminder. If the re-prompted response also lacks a parseable verdict (i.e., two consecutive responses with no verdict), surface the issue and stop the loop. Malformed-verdict failures do not trigger the fallback chain - the chain is reserved for unavailability, and verdict-format failures of a registered reviewer are surfaced to the user instead.
Step 7: Terminate
-
Success: print the labeled template below verbatim. The TODO-marker heading prints even when the marker list is empty (use (none) so the absence-of-markers signal is explicit). Do not commit.
Approved at round <N>.
baseline_ref: <ref>
Final diff stat:
<output of `git diff --stat <baseline_ref> --`>
Changed files:
<output of `git diff --name-only <baseline_ref> --`, one path per line>
TODO markers added (review or resolve before invoking cubrid-pr-create):
<output of `git diff <baseline_ref> -- | grep -nE '^\+.*\b(TODO|FIXME|XXX)\b'`, or "(none)" if empty>
Compute the marker list with git diff <baseline_ref> -- | grep -nE '^\+.*\b(TODO|FIXME|XXX)\b' (per-file equivalent: iterate git diff --name-only <baseline_ref> -- and grep each). The heading must always print so the user always sees the absence-of-markers signal.
-
Cap reached with REVISE open: print last_critique verbatim. If last_reviewer_critique differs from last_critique (e.g., the cap was hit on a build-failure round), print last_reviewer_critique verbatim under the heading "Most recent reviewer critique (still unresolved):". Print the final diff stat, baseline_ref, and ask the user to choose (a) accept as-is, (b) extend cap by N rounds, or (c) abandon. Cap-extension protocol: if the user picks (b), prompt for N in a follow-up turn ("How many additional rounds? Enter a positive integer."), parse the integer reply, set max_rounds += N, then resume from Step 4. Do NOT reset round. Any reply that is not (a) or a valid positive integer for (b) is treated as (c) - leave the tree as-is, do not loop further unless the user re-invokes the skill. Never silently loop past the cap.
Worked example. Suppose round=4, max_rounds=3, the cap-reached branch has fired, and the user picks (b) with N=2. The orchestrator sets max_rounds=5 and does not touch round. The next loop iteration's check round > max_rounds evaluates 4 > 5 → false → continue from Step 4 at round 4. The loop terminates either when the reviewer approves or when round reaches 6 after two more REVISE rounds. Resetting round would silently re-do prior work and burn the extension; not resetting it preserves the "extend by N more rounds" semantics the user asked for.
-
Reviewer infrastructure failure: print the error, baseline_ref, and the current diff stat so the user can resume manually.
Anti-patterns
- Don't sanitize the critique before passing it to the writer. Harsh stays harsh.
- Don't be the reviewer yourself. Spawn a real subagent every round, even when the diff looks small.
- Don't auto-commit, auto-stage, or auto-stash. The user's git state is theirs.
- Don't re-baseline mid-loop.
baseline_ref is captured once.
- Don't run the reviewer on a broken build. Build gate first, every round.
- Don't manufacture findings to look thorough.
- Don't pre-read the changed files or the full diff body in the orchestrator. The reviewer fetches per-file diffs and Reads files itself.
- Don't fall back to a different reviewer subagent on a malformed verdict. Surface and stop.
- Don't leave unflagged deferrals. Every stub, deferred edge case, known limitation, missing test, or assumption-to-revisit must carry a marker at the exact site in the source -
// TODO(CBRD-XXXXX): <what and why> (or /* TODO(CBRD-XXXXX): */) when the JIRA ticket is known, otherwise // TODO: <what and why>. For a missing test, the production code path that lacks coverage is the "exact site." The reviewer treats unflagged deferrals as a finding.
- Don't introduce unjustified non-functional churn in PR mode (
input_kind == 'pr'). Reformatting, renames, block moves, or drive-by fixes that do not serve the spec are findings - except when the writer explicitly justifies them in the one-line summary using the form "Justified rename: -> ; reason: " (or analogous "Justified move:" / "Justified reformat:"). Any unsummarized rename/move/reformat is unjustified.
Out of Scope
- No auto-PR creation.
cubrid-pr-create handles that as a separate user-invoked step.
- No worktree management. Assume the user is already in the working tree they want changed.
- No CUBRID test-case generation. The writer focuses on production code; the user runs
create-testcases separately after approval.
- No commit/squash/rebase logic. The user owns git history.
Shared snippets
Shared CUBRID Helpers
Use cubrid-common/scripts/cubrid-common.sh for common preflight helpers. It provides:
cubrid_require_source_tree "$PWD" for CUBRID checkout validation.cubrid_require_just_build_recipe "$PWD" and cubrid_require_valid_preset_mode "$PWD" for the build gate.cubrid_detect_remote for resolving UPSTREAM_REMOTE.
cubrid_detect_remote fails closed when no git remote points at CUBRID/cubrid, rather than guessing. The variable is named UPSTREAM_REMOTE (not ORIGIN_REMOTE) so the name reflects its semantics — "the remote that points at the canonical upstream repository" — regardless of whether it resolves to origin, cub, or some other local name.
Reference Files
references/reviewer-prompt.md - relentless reviewer persona, code-review variant. Loaded and passed verbatim to every reviewer subagent.references/writer-prompt.md - single source of truth for writer rules. Loaded and passed verbatim to every writer subagent.~/.claude/skills/cubrid-pr-review/reference.md - CUBRID knowledge base (six-place error-code rule, lock/page-buffer/WAL/MVCC protocols, build-mode guards, key data structures, false-positive guidance). Resolve ~ to $HOME at runtime. If the file does not exist, refuse with: "cubrid-pr-review/reference.md not found at . Run just install in your skills repo (or equivalent) to populate it." Skill development against an uninstalled local copy is a developer-workflow concern (e.g., just install symlinks during dev, or a one-off symlink from $HOME/.claude/skills/cubrid-pr-review/reference.md into your dev tree), not a skill-spec concern.
