@pmelab/gtd 47.0.0 → 49.0.0

package/src/steering/review.ts

@@ -1338,3 +1338,13 @@ export const reviewDecisions = (before: string, after: string): readonly ReviewD
     return old?.choice === decision.choice && old.text === decision.text ? [] : [decision]
   })
 }
+
+/** Each `##` chunk in document order: its title, 0-based heading line and how many hunk pointers it holds. */
+export const reviewChunks = (
+  content: string,
+): readonly { title: string; line: number; pointers: number }[] =>
+  parseReviewDoc(content).changesets.map((c) => ({
+    title: c.title,
+    line: c.headingLine,
+    pointers: c.files.length,
+  }))

package/src/testkit/Layers.ts

@@ -205,7 +205,6 @@ const missingTool = (name: string) => () =>
 const missingToolCommands: ToolCommands["Type"] = {
   ui: missingTool("ui"),
   lsp: missingTool("lsp"),
-  loop: missingTool("loop"),
   contextHook: missingTool("context-hook"),
   judge: () =>
     Promise.reject(

package/src/wire/ContextLimit.ts

@@ -1,4 +1,4 @@
-// The context-limit decision every driver shares: the loop's `gtd context-hook`
+// The context-limit decision every driver shares: `gtd context-hook`
 // and the Claude Code mod both answer a hook event through `contextHook`, so
 // the two cannot drift on thresholds, wording or which tokens count.
 

package/src/workflows/feature/coverage.feature

@@ -61,13 +61,13 @@ Feature: a feature is done only when every acceptance criterion has a passing te
       [{"id":"AC-1","answer":"yes","p":0.9},{"id":"AC-2","answer":"yes","p":0.9}]
       """
     When the process is driven
-    Then the process rests at "build.review.composing"
+    Then the process rests at "build.review.reviewing"
     And the commit subjects end with:
       """
       gtd(check): implement.applying → implement.health.check
       gtd(check): implement.health.check → implement.coverage
       gtd(judge): implement.coverage → implement.closing
-      gtd(check): implement.closing → build.review.composing
+      gtd(check): implement.closing → build.review.reviewing
       """
     And the commit "gtd(judge): implement.coverage → implement.closing" has a body containing "Gtd-Mark: covers=AC-1"
     And the commit "gtd(judge): implement.coverage → implement.closing" has a body containing "Gtd-Mark: covers=AC-2"
@@ -144,7 +144,7 @@ Feature: a feature is done only when every acceptance criterion has a passing te
       """
     And step "implement.uncovered" changes nothing
     When the process is driven
-    Then the process rests at "build.review.composing"
+    Then the process rests at "build.review.reviewing"
     And the commit "gtd(human): implement.uncovered → implement.closing" has a body containing "Gtd-Mark: covers=AC-1"
     And the commit "gtd(human): implement.uncovered → implement.closing" has a body containing "Gtd-Mark: uncovered=AC-2"
 

package/src/workflows/feature/fast-path.feature

@@ -64,7 +64,7 @@ Feature: a trivial change takes the same few beats at either risk tier
       """
     And step "implement.applying" runs its script
     When the process is driven
-    Then the process rests at "build.review.composing"
+    Then the process rests at "build.review.reviewing"
     And "src/greeting.ts" contains "Hello, world"
     And the commit subjects end with:
       """
@@ -80,7 +80,7 @@ Feature: a trivial change takes the same few beats at either risk tier
       gtd(check): implement.health.check → implement.coverage
       gtd(judge): implement.coverage → implement.closing
       gtd(check): implement.closing → build.quality.correctness.reviewing
-      gtd(agent): build.quality.correctness.reviewing → build.review.composing
+      gtd(agent): build.quality.correctness.reviewing → build.review.reviewing
       """
     And the output of gtd "summary" <summary> "needs a human reviewer"
 

package/src/workflows/feature/feature.ts

@@ -38,7 +38,8 @@ import { recordRisk, riskTierOn, type RiskTier } from "./tier.js"
 // judged with `flagMode` on: it lands as the implementation, or (refactoring
 // first when the judge asks) is built fresh from → the health loop → the
 // triaged quality lap, whose leftovers are carried on → after the last slice,
-// human review of the whole change and every slice's findings, which
+// a reviewer's walkthrough of the whole change, every slice's findings
+// among its chunks, for the human, which
 // signs off (the episode ends back at `idle`) or sends a design → implement
 // lap.
 
@@ -152,11 +153,15 @@ const buildSlice = async (sliceBase: string, build: SliceBuild): Promise<void> =
   })
 }
 
-/** The quality lap of slice `last.slice`, then the human review since `base`, listing what earlier slices' `laps` left too. */
+/**
+ * The quality lap of slice `last.slice`, then a reviewer's walkthrough since
+ * `base` for the human, carrying what earlier slices' `laps` left too.
+ */
 const reviewedTail = (
   base: string,
   lensBase: string,
   last: { readonly slice: string; readonly laps: SliceBuild["laps"] } = { slice: "", laps: [] },
+  lap: "first" | "feedback" = "first",
 ): Promise<ReviewOutcome> =>
   tail(
     (escalations, report) =>
@@ -165,6 +170,8 @@ const reviewedTail = (
           last.laps.length === 0
             ? report
             : joinReports([...last.laps, { slice: last.slice, report }]),
+        walkthrough: true,
+        lap,
       }),
     { green: true, lensBase },
   )
@@ -189,7 +196,7 @@ const featureFeedback = async (
     })
     const tier = riskTierOn() ? recordRisk() : undefined
     closed = await implementing({ tier, amend: true })
-    outcome = await reviewedTail(outcome.base, lensBase)
+    outcome = await reviewedTail(outcome.base, lensBase, undefined, "feedback")
   }
 }
 

package/src/workflows/feature/flag-mode.feature

@@ -67,7 +67,7 @@ Feature: flag mode off lands the naive pass with no release flag, seam judge or
       """
     And step "build.quality.correctness.reviewing" changes nothing
     When the process is driven
-    Then the process rests at "build.review.composing"
+    Then the process rests at "build.review.reviewing"
     And "src/calc.ts" contains "subtract"
     And "src/calc.test.ts" contains "[AC-1] subtracts"
     And the commit subjects end with:
@@ -85,5 +85,5 @@ Feature: flag mode off lands the naive pass with no release flag, seam judge or
       gtd(check): implement.health.check → implement.coverage
       gtd(judge): implement.coverage → implement.closing
       gtd(check): implement.closing → build.quality.correctness.reviewing
-      gtd(agent): build.quality.correctness.reviewing → build.review.composing
+      gtd(agent): build.quality.correctness.reviewing → build.review.reviewing
       """

package/src/workflows/feature/lens-review.feature

@@ -1,12 +1,11 @@
 @inmem
-Feature: the quality lenses are the only agentic review
+Feature: the quality lap's findings are triaged before the review
 
   A judge classifies each `## ` finding the lenses wrote to `.gtd/QUALITY.md`:
   a mechanical one is fixed by the fixer, noise is dropped with its `p` kept in
   a trailer, and a design one becomes a note in `.gtd/REVIEW.md`, with answer
-  boxes. gtd writes that file itself — no reviewer turn runs — and lists,
-  without tick boxes, what was fixed automatically and what was dropped as
-  noise.
+  boxes, in the chunk the reviewer gave it. gtd lists, without tick boxes,
+  what was fixed automatically and what was dropped as noise.
 
   Background:
     Given the shipped workflows, as gtd install scaffolds them
@@ -75,7 +74,7 @@ Feature: the quality lenses are the only agentic review
       """
     And step "build.quality.sorting" runs its script
     And step "build.quality.closing" runs its script
-    And step "build.review.composing" runs its script
+    And step "build.review.carrying" runs its script
 
   Scenario: a mechanical finding is fixed and listed, noise dropped and listed, a design one noted
     Given judge step "build.quality.triage" answers:
@@ -89,6 +88,20 @@ Feature: the quality lenses are the only agentic review
       """
     And step "build.fix.quality.fixing" deletes ".gtd/QUALITY.md"
     And step "build.health.check" changes nothing
+    And step "build.review.reviewing" writes ".gtd/REVIEW.md" with:
+      """
+      # Review: abc1234
+
+      <!-- base: 0000000 -->
+
+      ## Subtract
+
+      - [ ] ./src/calc.ts#2-2 — subtract(a, b)
+
+      ## AC-1 ignores floating point
+
+      - [ ] ./src/calc.ts#2-2 — where it applies
+      """
     When the process is driven
     Then the process rests at "build.review.await-review"
     And ".gtd/QUALITY.md" does not exist
@@ -99,6 +112,10 @@ Feature: the quality lenses are the only agentic review
 
       Should subtract round? The criterion does not say.
 
+      In the code:
+
+      - [ ] ./src/calc.ts#2-2 — where it applies
+
       Tick exactly one:
 
       - [ ] fix as suggested
@@ -129,8 +146,9 @@ Feature: the quality lenses are the only agentic review
       gtd(judge): build.quality.triage → build.quality.sorting
       gtd(check): build.quality.sorting → build.fix.quality.fixing
       gtd(agent): build.fix.quality.fixing → build.health.check
-      gtd(check): build.health.check → build.review.composing
-      gtd(check): build.review.composing → build.review.await-review
+      gtd(check): build.health.check → build.review.reviewing
+      gtd(agent): build.review.reviewing → build.review.carrying
+      gtd(check): build.review.carrying → build.review.await-review
       """
 
   Scenario: an unsure verdict makes a finding a note, never a fix or a drop
@@ -138,6 +156,28 @@ Feature: the quality lenses are the only agentic review
       """
       [{"id":"Q1","answer":"mechanical","p":0.5},{"id":"Q2","answer":"noise","p":0.6}]
       """
+    And step "build.review.reviewing" writes ".gtd/REVIEW.md" with:
+      """
+      # Review: abc1234
+
+      <!-- base: 0000000 -->
+
+      ## Subtract
+
+      - [ ] ./src/calc.ts#2-2 — subtract(a, b)
+
+      ## subtract has no negative test
+
+      - [ ] ./src/calc.ts#2-2 — where it applies
+
+      ## add could be a method
+
+      - [ ] ./src/calc.ts#2-2 — where it applies
+
+      ## AC-1 ignores floating point
+
+      - [ ] ./src/calc.ts#2-2 — where it applies
+      """
     When the process is driven
     Then the process rests at "build.review.await-review"
     And ".gtd/REVIEW.md" contains "## subtract has no negative test"
@@ -148,8 +188,9 @@ Feature: the quality lenses are the only agentic review
     And the commit subjects end with:
       """
       gtd(judge): build.quality.triage → build.quality.sorting
-      gtd(check): build.quality.sorting → build.review.composing
-      gtd(check): build.review.composing → build.review.await-review
+      gtd(check): build.quality.sorting → build.review.reviewing
+      gtd(agent): build.review.reviewing → build.review.carrying
+      gtd(check): build.review.carrying → build.review.await-review
       """
 
   Scenario: a mechanical finding the fix cap leaves becomes a note
@@ -166,6 +207,20 @@ Feature: the quality lenses are the only agentic review
       A class would read better.
       """
     And step "build.health.check" changes nothing
+    And step "build.review.reviewing" writes ".gtd/REVIEW.md" with:
+      """
+      # Review: abc1234
+
+      <!-- base: 0000000 -->
+
+      ## Subtract
+
+      - [ ] ./src/calc.ts#2-2 — subtract(a, b)
+
+      ## add could be a method
+
+      - [ ] ./src/calc.ts#2-2 — where it applies
+      """
     When the process is driven
     Then the process rests at "build.review.await-review"
     And ".gtd/QUALITY.md" does not exist
@@ -186,8 +241,9 @@ Feature: the quality lenses are the only agentic review
       gtd(agent): build.fix.quality.fixing
       gtd(agent): build.fix.quality.fixing → build.health.check
       gtd(check): build.health.check → build.quality.closing
-      gtd(check): build.quality.closing → build.review.composing
-      gtd(check): build.review.composing → build.review.await-review
+      gtd(check): build.quality.closing → build.review.reviewing
+      gtd(agent): build.review.reviewing → build.review.carrying
+      gtd(check): build.review.carrying → build.review.await-review
       """
 
   Scenario: a fixer that renames a chunk it leaves counts nothing fixed
@@ -204,6 +260,32 @@ Feature: the quality lenses are the only agentic review
       Still open.
       """
     And step "build.health.check" changes nothing
+    And step "build.review.reviewing" writes ".gtd/REVIEW.md" with:
+      """
+      # Review: abc1234
+
+      <!-- base: 0000000 -->
+
+      ## Subtract
+
+      - [ ] ./src/calc.ts#2-2 — subtract(a, b)
+
+      ## subtract has no negative test
+
+      - [ ] ./src/calc.ts#2-2 — where it applies
+
+      ## add could be a method
+
+      - [ ] ./src/calc.ts#2-2 — where it applies
+
+      ## AC-1 ignores floating point
+
+      - [ ] ./src/calc.ts#2-2 — where it applies
+
+      ## null check in parse()
+
+      - [ ] ./src/calc.ts#2-2 — where it applies
+      """
     When the process is driven
     Then the process rests at "build.review.await-review"
     And ".gtd/REVIEW.md" contains "## null check in parse()"
@@ -223,17 +305,37 @@ Feature: the quality lenses are the only agentic review
       All fixed.
       """
     And step "build.health.check" changes nothing
+    And step "build.review.reviewing" writes ".gtd/REVIEW.md" with:
+      """
+      # Review: abc1234
+
+      <!-- base: 0000000 -->
+
+      ## Subtract
+
+      - [ ] ./src/calc.ts#2-2 — subtract(a, b)
+      """
     When the process is driven
     Then the process rests at "build.review.await-review"
     And ".gtd/REVIEW.md" contains "- subtract has no negative test ("
     And ".gtd/REVIEW.md" does not contain "check whether it was fixed"
 
-  Scenario: signing off a composed review ends the process
+  Scenario: signing off a review with no notes ends the process
     Given judge step "build.quality.triage" answers:
       """
       [{"id":"Q1","answer":"noise","p":0.9},{"id":"Q2","answer":"noise","p":0.9},{"id":"Q3","answer":"noise","p":0.9}]
       """
     And step "build.review.closing" runs its script
+    And step "build.review.reviewing" writes ".gtd/REVIEW.md" with:
+      """
+      # Review: abc1234
+
+      <!-- base: 0000000 -->
+
+      ## Subtract
+
+      - [ ] ./src/calc.ts#2-2 — subtract(a, b)
+      """
     When the process is driven
     Then the process rests at "build.review.await-review"
     When gtd lands
@@ -242,7 +344,8 @@ Feature: the quality lenses are the only agentic review
     And ".gtd/REVIEW.md" does not exist
     And the commit subjects end with:
       """
-      gtd(check): build.review.composing → build.review.await-review
+      gtd(agent): build.review.reviewing → build.review.carrying
+      gtd(check): build.review.carrying → build.review.await-review
       gtd(human): build.review.await-review → build.review.closing
       gtd(check): build.review.closing → idle
       """

package/src/workflows/feature/one-step-land.feature

@@ -64,7 +64,7 @@ Feature: a naive pass the judge rules land is itself the implementation
 
   Scenario: a land judgment applies the naive pass and runs from idle to the review gate with no rebuild
     When the process is driven
-    Then the process rests at "build.review.composing"
+    Then the process rests at "build.review.reviewing"
     And "src/greeting.ts" contains "Hello, world"
     And "src/acceptance.test.ts" contains "[AC-1] holds"
     And the commit subjects end with:
@@ -81,7 +81,7 @@ Feature: a naive pass the judge rules land is itself the implementation
       gtd(check): implement.health.check → implement.coverage
       gtd(judge): implement.coverage → implement.closing
       gtd(check): implement.closing → build.quality.correctness.reviewing
-      gtd(agent): build.quality.correctness.reviewing → build.review.composing
+      gtd(agent): build.quality.correctness.reviewing → build.review.reviewing
       """
     And the commit "gtd(human): idle → unwind" has a body containing "Gtd-Var: flag_risk_tier=0"
     And the commit "gtd(agent): seam.naive → seam.check" has a body containing "Gtd-Kept:"

package/src/workflows/feature/refactor-absorbs.feature

@@ -113,7 +113,7 @@ Feature: a refactor that carried the whole feature still gets a quality lap over
 
     Given step "build.quality.correctness.reviewing" changes nothing
     When the process is driven
-    Then the process rests at "build.review.composing"
+    Then the process rests at "build.review.reviewing"
     And "src/calc.ts" contains "subtract"
     And the commit subjects end with:
       """
@@ -123,7 +123,7 @@ Feature: a refactor that carried the whole feature still gets a quality lap over
       gtd(check): implement.health.check → implement.coverage
       gtd(judge): implement.coverage → implement.closing
       gtd(check): implement.closing → build.quality.correctness.reviewing
-      gtd(agent): build.quality.correctness.reviewing → build.review.composing
+      gtd(agent): build.quality.correctness.reviewing → build.review.reviewing
       """
     And the output lines of gtd "slices" start with:
       """

package/src/workflows/feature/refactor-first.feature

@@ -121,7 +121,7 @@ Feature: a refactor-first judgment walks one fixed path to the review gate
       [{"id":"AC-1","answer":"yes","p":0.9}]
       """
     When the process is driven
-    Then the process rests at "build.review.composing"
+    Then the process rests at "build.review.reviewing"
     And "src/calc.ts" contains "op(a, b, 1)"
     And "src/calc.ts" contains "subtract"
     And the commit subjects end with:
@@ -145,7 +145,7 @@ Feature: a refactor-first judgment walks one fixed path to the review gate
       gtd(agent): implement.building → implement.health.check
       gtd(check): implement.health.check → implement.coverage
       gtd(judge): implement.coverage → implement.closing
-      gtd(check): implement.closing → build.review.composing
+      gtd(check): implement.closing → build.review.reviewing
       """
     And the commit "gtd(human): idle → unwind" has a body containing "Gtd-Var: flag_risk_tier=0"
     And the commit "gtd(agent): seam.naive → seam.check" has a body containing "Gtd-Kept:"

package/src/workflows/feature/refactor-rebuild.feature

@@ -242,7 +242,7 @@ Feature: refactor-first writes a refactor note, refactors on its own, then rebui
       [{"id":"AC-1","answer":"yes","p":0.9}]
       """
     When the process is driven
-    Then the process rests at "build.review.composing"
+    Then the process rests at "build.review.reviewing"
     And "src/calc.ts" contains "op(a, b, -1)"
     And "src/acceptance.test.ts" contains "[AC-1]"
     And ".gtd/REFACTOR-NOTE.md" does not exist
@@ -260,7 +260,7 @@ Feature: refactor-first writes a refactor note, refactors on its own, then rebui
       gtd(agent): implement.building → implement.health.check
       gtd(check): implement.health.check → implement.coverage
       gtd(judge): implement.coverage → implement.closing
-      gtd(check): implement.closing → build.review.composing
+      gtd(check): implement.closing → build.review.reviewing
       """
     And the output lines of gtd "slices" start with:
       """

package/src/workflows/feature/refactor-review.feature

@@ -646,7 +646,7 @@ Feature: the refactor slice is reviewed on its own before the feature is impleme
         """
       When gtd lands
       And the process is driven
-      Then the process rests at "build.review.composing"
+      Then the process rests at "build.review.reviewing"
       And "src/calc.ts" contains "subtract"
 
     Scenario: a praise-only one-shot footnote signs off and the feature slice follows
@@ -688,7 +688,7 @@ Feature: the refactor slice is reviewed on its own before the feature is impleme
         """
       When gtd lands
       And the process is driven
-      Then the process rests at "build.review.composing"
+      Then the process rests at "build.review.reviewing"
       And "src/calc.ts" contains "subtract"
 
     Scenario: a one-shot footnote judged edit that collecting leaves out refuses naming its anchor
@@ -1003,7 +1003,7 @@ Feature: the refactor slice is reviewed on its own before the feature is impleme
         """
       When gtd lands
       And the process is driven
-      Then the process rests at "build.review.composing"
+      Then the process rests at "build.review.reviewing"
       And "src/calc.ts" contains "subtract"
 
     Scenario: a praise-only refactor review still signs off and the feature slice follows
@@ -1038,7 +1038,7 @@ Feature: the refactor slice is reviewed on its own before the feature is impleme
         """
       When gtd lands
       And the process is driven
-      Then the process rests at "build.review.composing"
+      Then the process rests at "build.review.reviewing"
       And "src/calc.ts" contains "subtract"
 
     Scenario: a nit fix in the refactor slice that changes a non-stray is refused, though a glob stray widens the write access
@@ -1228,11 +1228,11 @@ Feature: the refactor slice is reviewed on its own before the feature is impleme
       And the resting step's content contains the hash of the commit "gtd(check): refactor.build.review.closing → implement.building"
       Given step "build.quality.correctness.reviewing" changes nothing
       When the process is driven
-      Then the process rests at "build.review.composing"
+      Then the process rests at "build.review.reviewing"
       And the commit subjects end with:
         """
         gtd(check): implement.closing → build.quality.correctness.reviewing
-        gtd(agent): build.quality.correctness.reviewing → build.review.composing
+        gtd(agent): build.quality.correctness.reviewing → build.review.reviewing
         """
 
   Rule: a refactor package never carries a test file, and demoted headings stop at ######

package/src/workflows/feature/review-answers.feature

@@ -1,7 +1,7 @@
 @inmem
 Feature: every note in the review is answered with one tick
 
-  Each note gtd writes to `.gtd/REVIEW.md` ends in three boxes — fix as
+  Each quality note in `.gtd/REVIEW.md` ends in three boxes — fix as
   suggested, won't fix, or the human's own answer written over the
   placeholder. The gate refuses to land until every note holds exactly one
   tick. A fix goes to the nit batch, a won't-fix is dropped and recorded in
@@ -74,7 +74,41 @@ Feature: every note in the review is answered with one tick
       Should subtract round? The criterion does not say.
       """
     And step "build.quality.sorting" runs its script
-    And step "build.review.composing" runs its script
+    And step "build.review.reviewing" writes ".gtd/REVIEW.md" with:
+      """
+      # Review: abc1234
+
+      <!-- base: 0000000 -->
+
+      ## Subtract
+
+      - [ ] ./src/calc.ts#2-2 — subtract(a, b)
+      """
+    And step "build.review.reviewing", when prompted with "- `## AC-1 ignores floating point`", writes ".gtd/REVIEW.md" with:
+      """
+      # Review: abc1234
+
+      <!-- base: 0000000 -->
+
+      ## Subtract
+
+      - [ ] ./src/calc.ts#2-2 — subtract(a, b)
+
+      ## AC-1 ignores floating point
+
+      - [ ] ./src/calc.ts#2-2 — returns the raw difference
+      """
+    And step "build.review.reviewing", when prompted with "worked on the human's nits", writes ".gtd/REVIEW.md" with:
+      """
+      # Review: abc1234
+
+      <!-- base: 0000000 -->
+
+      ## Subtract
+
+      - [ ] ./src/calc.ts#2-2 — now rounds the difference
+      """
+    And step "build.review.carrying" runs its script
     And judge step "build.quality.triage" answers:
       """
       [{"id":"Q1","answer":"noise","p":0.9},{"id":"Q2","answer":"noise","p":0.9},{"id":"Q3","answer":"design","p":0.9}]
@@ -90,6 +124,10 @@ Feature: every note in the review is answered with one tick
 
       Should subtract round? The criterion does not say.
 
+      In the code:
+
+      - [ ] ./src/calc.ts#2-2 — returns the raw difference
+
       Tick exactly one:
 
       - [ ] fix as suggested
@@ -141,7 +179,7 @@ Feature: every note in the review is answered with one tick
     And the process is driven
     Then the process rests at "build.review.await-review"
     And no commit body contains "build.review.triage"
-    And ".gtd/REVIEW.md" contains "## Fixed from your notes"
+    And ".gtd/REVIEW.md" contains "now rounds the difference"
     And ".gtd/REVIEW.md" does not contain "AC-1 ignores floating point"
     And ".gtd/TODO.md" does not exist
     When gtd lands
@@ -176,8 +214,9 @@ Feature: every note in the review is answered with one tick
       gtd(judge): build.review.triage → build.review.fix.nits.fixing
       gtd(agent): build.review.fix.nits.fixing → build.health.check
       gtd(check): build.health.check → build.review.closing
-      gtd(check): build.review.closing → build.review.composing
-      gtd(check): build.review.composing → build.review.await-review
+      gtd(check): build.review.closing → build.review.reviewing
+      gtd(agent): build.review.reviewing → build.review.carrying
+      gtd(check): build.review.carrying → build.review.await-review
       """
     And ".gtd/REVIEW.md" does not contain "AC-1 ignores floating point"
 

package/src/workflows/feature/review-gate.feature

@@ -57,7 +57,7 @@ Feature: a feature process ends at the human review
       """
       [{"id":"AC-1","answer":"yes","p":0.9}]
       """
-    And step "build.review.composing" writes ".gtd/REVIEW.md" with:
+    And step "build.review.reviewing" writes ".gtd/REVIEW.md" with:
       """
       # Review: abc1234
 
@@ -78,8 +78,8 @@ Feature: a feature process ends at the human review
     And the commit subjects end with:
       """
       gtd(judge): implement.coverage → implement.closing
-      gtd(check): implement.closing → build.review.composing
-      gtd(check): build.review.composing → build.review.await-review
+      gtd(check): implement.closing → build.review.reviewing
+      gtd(agent): build.review.reviewing → build.review.await-review
       """
 
   Scenario: a sign-off ends the process at idle
@@ -102,7 +102,7 @@ Feature: a feature process ends at the human review
     And ".gtd/REVIEW.md" does not exist
     And the commit subjects end with:
       """
-      gtd(check): build.review.composing → build.review.await-review
+      gtd(agent): build.review.reviewing → build.review.await-review
       gtd(human): build.review.await-review → build.review.closing
       gtd(check): build.review.closing → idle
       """
@@ -214,8 +214,8 @@ Feature: a feature process ends at the human review
         gtd(agent): implement.building → implement.health.check
         gtd(check): implement.health.check → implement.coverage
         gtd(judge): implement.coverage → implement.closing
-        gtd(check): implement.closing → build.review.composing
-        gtd(check): build.review.composing → build.review.await-review
+        gtd(check): implement.closing → build.review.reviewing
+        gtd(agent): build.review.reviewing → build.review.await-review
         """
 
     Scenario: a design numbering new behaviour at or below an earlier ID is refused