Size-adaptive red-team at /implement Phase 3 (folded into 3.1.0)

Phase 3 now sizes the red-team half like /plan sizes work: a single-file
(or trivial, no new logic) diff runs one red-team mode: both; a multi-file
or cross-cutting diff splits into parallel mode: bugs + mode: cleanup with
the cleanup run report-only, so the parallel gate stays read-only.
review-impl runs alongside in every case. No new red-team mode added.

validate.sh section 18 locks the invariant (mode literals + report-only,
matched loosely). Also: tracker-schema.md tracker-commit note reconciled
with the README, section 14 title corrected to 510, README qualifier
aligned.

Dogfooded via devloop: review-plan PASS-WITH-WARNINGS (5 findings folded
into the plan); Phase 3 split gate (dogfooding this very feature) =
review-impl PASS-WITH-WARNINGS + red-team mode:bugs (1 CONFIRMED bucket-
overlap, fixed) + red-team mode:cleanup report-only (WARNs addressed);
holistic release review SHIP-WITH-FIXES (all applied). Folded into 3.1.0
per maintainer; manifests unchanged.

Author: one-kash

CHANGELOG.md

@@ -54,6 +54,22 @@ resolves in the user's project) and gives trackers a stable home outside
   `skills/implement/project-configs/` (engineering) layout, and
   `validate.sh` now checks the examples in a single section (the former
   duplicate example-config check is removed).
+- **`/implement` Phase 3 sizes the `red-team` half by diff size**, the
+  same way `/plan` sizes work (its Trivial / Small / Medium+ / Large
+  table). A single-file change (or a trivial one with no new logic) runs
+  one `red-team` in `mode: both`, unchanged from before. A broader,
+  multi-file or cross-cutting diff (`/plan` Medium+ and Large) splits the
+  `red-team` half into parallel `mode: bugs` and `mode: cleanup` runs so
+  neither family crowds the other out. `review-impl` runs alongside in
+  every case. Because `red-team` in `mode: cleanup` can apply fixes, the
+  split invokes the `cleanup` run report-only, so all three concurrent
+  agents only report and the parallel gate stays read-only. No new
+  `red-team` mode was added; report-only is an invocation instruction
+  inside `cleanup` mode.
+- **`validate.sh` section 18** asserts the Phase 3 spawn stays
+  size-adaptive (it names the `mode: both`, `mode: bugs`, `mode: cleanup`,
+  and `report-only` markers), so a future edit can't silently revert to
+  the fixed single-agent gate.
 
 ### Migration
 - **Move in-flight trackers.** Trackers previously written under `docs/`

README.md

@@ -71,7 +71,9 @@ the same across projects and harnesses.
   overlap: `review-impl` defers correctness, robustness, standards, and
   cleanup to `red-team`. A bug that faithfully implements a flawed plan
   is caught only by `red-team`; a correct-but-off-spec change only by
-  `review-impl`.
+  `review-impl`. On a large, multi-file diff the `red-team` half fans
+  into a focused bugs pass and a focused cleanup pass in parallel; on a
+  small diff it stays a single `both` pass (see below).
 - **The bug hunter is recall-biased, then verified.** `red-team`
   surfaces candidate defects freely (conservative reviewers
   under-report), then runs a verify pass that keeps only
@@ -140,14 +142,23 @@ It takes a mode:
 | Mode | What it does |
 |---|---|
 | `bugs` | correctness angles only, then verify + sweep |
-| `cleanup` | quality angles only, the tidy pass; can apply fixes |
+| `cleanup` | quality angles only, the tidy pass; can apply fixes (report-only when run as a gate) |
 | `both` *(default)* | everything |
 
 ```
 Use the red-team agent in mode: both to review the changed files
 Use the red-team agent in mode: cleanup to tidy the changed files
 ```
 
+**Size-adaptive at the `/implement` gate.** Phase 3 sizes the diff the
+same way `/plan` sizes work: a single-file change (or a trivial one with
+no new logic) runs one `red-team` in `mode: both`; a multi-file or
+cross-cutting diff splits
+the `red-team` half into parallel `mode: bugs` and `mode: cleanup` runs so
+neither family crowds the other out. At the gate the `cleanup` run is
+invoked report-only, so the whole review stays read-only and safe to run
+alongside `review-impl`.
+
 ## Install
 
 **Claude Code.** Install from the marketplace:

skills/implement/SKILL.md

@@ -254,29 +254,53 @@ are available:
 `complete` (see Phase 1.3), is a **blocked feature**: do not treat it as
 complete or trigger this gate, stop and surface it to the user.
 
-**GATE: All chunks complete triggers parallel review.** Spawn two
-independent agents in a single message so they run concurrently. Both
-are read-only (`red-team` in `both` mode reports; it does not apply),
-so concurrent runs are safe:
+**GATE: All chunks complete triggers parallel review.** Spawn the review
+agents in a single message so they run concurrently. They are read-only
+(`review-impl` and `red-team` in `bugs`/`both`, plus `cleanup` when
+invoked report-only, all report; none apply), so concurrent runs are
+safe:
 
 1. **`review-impl` agent**, the conformance gate. Verifies plan match,
    acceptance criteria met with quoted test evidence, test quality, and
    regression against the tracker. Answers "does the implementation match
    what was planned, and does the suite prove it?" It does **not** hunt
    correctness bugs, robustness gaps, standards violations, or cleanup,
    that is `red-team`'s territory, so the two no longer overlap.
-2. **`red-team` agent** (`mode: both`), adversarial diff reviewer.
-   Hunts correctness bugs (5 angles) and flags cleanup (reuse,
-   simplification, efficiency, altitude), verifies each finding
-   (recall-biased), then sweeps for gaps. Answers "what is wrong or
-   wasteful in this diff, regardless of the plan?"
+2. **`red-team` agent**, adversarial diff reviewer. Hunts correctness
+   bugs (5 angles) and flags cleanup (reuse, simplification, efficiency,
+   altitude), verifies each finding (recall-biased), then sweeps for gaps.
+   Answers "what is wrong or wasteful in this diff, regardless of the
+   plan?"
+
+**Size the red-team half** the same way `/plan` sizes work (its Trivial /
+Small / Medium+ / Large table), by the set of files touched across the
+completed chunks:
+
+- **Small diff** (a single-file change, or a trivial one with no new
+  logic, matching `/plan`'s Small row): one `red-team` in `mode: both`.
+- **Large diff** (broader: multiple files or cross-cutting, i.e. `/plan`
+  Medium+ and Large): split the red-team half into `mode: bugs` and
+  `mode: cleanup` run in parallel, so neither family crowds the other
+  out. Invoke the `cleanup` agent **report-only** (it can otherwise apply
+  fixes), which keeps the whole gate read-only.
+
+Small diff (default):
 
 ```
 In parallel:
 - Use the review-impl agent to review implementation against [path to tracker]
 - Use the red-team agent in mode: both to review the changed files (pass the tracker path)
 ```
 
+Large diff (split the red-team half):
+
+```
+In parallel:
+- Use the review-impl agent to review implementation against [path to tracker]
+- Use the red-team agent in mode: bugs to review the changed files (pass the tracker path)
+- Use the red-team agent in mode: cleanup to review the changed files, report only (pass the tracker path)
+```
+
 Complementary by design: `review-impl` checks conformance to intent
 (conservative, PASS when the plan is met); `red-team` checks correctness
 independent of intent (recall-biased, surfaces everything, then

skills/plan/references/tracker-schema.md

@@ -1,7 +1,8 @@
 # JSON Tracker Schema
 
 The tracker file is the single source of truth for feature implementation progress.
-It is a local working file - not committed to version control.
+It is a working file; commit it for cross-machine resumability or
+gitignore it, your call.
 
 Write it **atomically** (write a temp file, then rename over the old one) so a
 crash mid-write cannot corrupt the single source of truth.

validate.sh

@@ -531,7 +531,7 @@ check_skill_frontmatter "$plan_skill" "Plan SKILL.md"
 check_skill_frontmatter "$spec_skill" "Spec SKILL.md"
 
 # ─────────────────────────────────────────────
-section "14. SKILL.md files are under 500 lines"
+section "14. SKILL.md files are under 510 lines"
 # ─────────────────────────────────────────────
 
 for skill_file in "$implement_skill" "$plan_skill" "$spec_skill"; do
@@ -639,6 +639,36 @@ for agent in agents/review-plan.md agents/review-impl.md agents/red-team.md; do
     fi
 done
 
+# ─────────────────────────────────────────────
+section "18. Phase 3 red-team spawn is size-adaptive"
+# ─────────────────────────────────────────────
+
+# /implement Phase 3 sizes the diff and either runs one red-team in
+# mode: both (small diff) or splits into parallel mode: bugs +
+# mode: cleanup (report-only) on a large diff. Guard the invariant so a
+# future edit can't silently revert to the fixed single-agent gate or
+# drop the report-only instruction that keeps the split read-only.
+# The three mode literals prove both paths are documented: mode: both
+# (small diff) plus mode: bugs and mode: cleanup (the large-diff split).
+# mode: bugs is unique to the split, so a revert to the fixed single-agent
+# gate drops it and this fires.
+for token in "mode: both" "mode: bugs" "mode: cleanup"; do
+    if grep -qi "$token" "$implement_skill"; then
+        pass "implement SKILL.md Phase 3 names the size-adaptive marker: \"$token\""
+    else
+        fail "implement SKILL.md Phase 3 is missing the size-adaptive marker: \"$token\" (the red-team spawn must document small -> mode: both vs large -> parallel mode: bugs + mode: cleanup)"
+    fi
+done
+# The split cleanup agent must be invoked report-only so the gate stays
+# read-only. Match loosely (hyphen or space) so a benign reword does not
+# spuriously fail; the invariant is the report-only instruction, not its
+# spelling.
+if grep -Eqi "report[- ]only" "$implement_skill"; then
+    pass "implement SKILL.md invokes the split cleanup agent report-only"
+else
+    fail "implement SKILL.md Phase 3 must invoke the large-diff cleanup agent report-only (keeps the parallel gate read-only)"
+fi
+
 # ─────────────────────────────────────────────
 # Summary
 # ─────────────────────────────────────────────