diff --git a/skills/ship/SKILL.md b/skills/ship/SKILL.md
index 2e0add17..cedacff4 100644
--- a/skills/ship/SKILL.md
+++ b/skills/ship/SKILL.md
@@ -281,12 +281,39 @@ mid-loop (footguns).
 
 ### 2.3 Implement
 
-Spawn a **fresh** coding agent (`subagent_type: "general-purpose"`, never `fork`,
-`model: "sonnet"` unless the wave plan marks the unit high-risk, in which case omit
-`model` to inherit yours) — do not resume any planner; the canonical plan file, not
-planner memory or ChatGPT chat text, is the handoff. Parallel units get
-`isolation: "worktree"`; a solo unit uses the main tree on `wip/<unit>-<slug>` off the
-current integration HEAD.
+**High risk + Large unit → decompose instead of one implementer.** If the approved
+plan's own risk classification (`per-issue-cycle.md`) states High risk **and** Large,
+do not spawn a single implementation agent for the whole plan — a unit that size
+reliably exhausts one agent's context before finishing (observed live: a ~30-file,
+Docker/live-server/browser-matrix unit hit ~65% context after its first commit). Run:
+
+```
+Workflow {
+  scriptPath: "skills/ship/references/decompose-and-implement-loop.workflow.mjs",
+  args: { planFile: "<abs>", branch: "<wip branch>", issueRef: "#<ISSUE> phase <N>" }
+}
+```
+
+Fable/high reads the plan **and the actual repository state already on the branch**
+(so it excludes whatever a prior partial attempt already committed) and proposes an
+ordered, dependency-respecting sub-task breakdown, sized so one coding agent can
+plausibly finish each in one session. The script then runs one fresh Sonnet agent per
+sub-task, strictly **sequentially** on the same branch (not parallel worktrees —
+sub-task file-scope disjointness is a declared claim, not a verified guarantee, and
+the point is giving each chunk fresh context, not wall-clock speed). Each sub-task
+commits before the next starts. If a running single-agent attempt already exists and
+is approaching its own limit, stop it at a clean commit boundary (`TaskStop`, verify
+`git status` is clean first) before invoking this loop — its committed work becomes
+what the decomposition step reads as already-done. Treat the loop's `error` status
+(a sub-task's agent died) the same as any implementation failure: inspect, fix or
+resume, re-verify — do not silently skip the remaining sub-tasks.
+
+**Otherwise** (the common case), spawn a **fresh** coding agent (`subagent_type:
+"general-purpose"`, never `fork`, `model: "sonnet"` unless the wave plan marks the unit
+high-risk, in which case omit `model` to inherit yours) — do not resume any planner;
+the canonical plan file, not planner memory or ChatGPT chat text, is the handoff.
+Parallel units get `isolation: "worktree"`; a solo unit uses the main tree on
+`wip/<unit>-<slug>` off the current integration HEAD.
 
 The coding-agent prompt must contain, explicitly:
 
diff --git a/skills/ship/references/decompose-and-implement-loop.workflow.mjs b/skills/ship/references/decompose-and-implement-loop.workflow.mjs
new file mode 100644
index 00000000..4a04868c
--- /dev/null
+++ b/skills/ship/references/decompose-and-implement-loop.workflow.mjs
@@ -0,0 +1,78 @@
+export const meta = {
+  name: 'ship-decompose-and-implement-loop',
+  description: 'For a High-risk, Large approved /ship unit plan: decompose the remaining work into sub-tasks, then implement each via a fresh sequential coding agent',
+  whenToUse: 'Invoked by the /ship coordinator at step 2.3 only when the approved plan classifies itself High risk AND Large',
+  phases: [
+    { title: 'Decompose', detail: 'Fable/high proposes an ordered, dependency-respecting sub-task breakdown of the remaining plan work, grounded against what is already committed' },
+    { title: 'Implement', detail: 'one fresh Sonnet coding agent per sub-task, in order, each committing locally before the next starts' },
+  ],
+}
+
+// args: { planFile, branch, issueRef } — planFile is the approved plan's absolute path
+// (unchanged from the plan-review loop); branch is the already-checked-out wip branch
+// (main tree for a solo unit, or the unit's worktree); issueRef is a human label like
+// "#447 phase 2" used in prompts and commit messages (its leading "#<n>" is reused
+// verbatim in every sub-task's commit message, so pass the exact issue number here).
+const runArgs = typeof args === 'string' ? JSON.parse(args) : args
+if (!runArgs || !runArgs.planFile || !runArgs.branch || !runArgs.issueRef) {
+  throw new Error('args {planFile, branch, issueRef} required')
+}
+const { planFile, branch, issueRef } = runArgs
+const issueTag = (issueRef.match(/#\d+/) || ['#?'])[0]
+
+const DECOMPOSE_SCHEMA = {
+  type: 'object', additionalProperties: false,
+  required: ['subtasks', 'alreadyDoneSummary'],
+  properties: {
+    alreadyDoneSummary: { type: 'string' },
+    subtasks: {
+      type: 'array',
+      items: {
+        type: 'object', additionalProperties: false,
+        required: ['id', 'title', 'planSections', 'description', 'fileScope', 'dependsOn', 'doneWhen'],
+        properties: {
+          id: { type: 'string' },
+          title: { type: 'string' },
+          planSections: { type: 'array', items: { type: 'string' } },
+          description: { type: 'string' },
+          fileScope: { type: 'array', items: { type: 'string' } },
+          dependsOn: { type: 'array', items: { type: 'string' } },
+          doneWhen: { type: 'string' },
+        },
+      },
+    },
+  },
+}
+const READ_ONLY = 'Strictly read-only: no Edit or Write, no git or gh mutations, no task or memory writes, no chatgpt-review invocation.'
+
+phase('Decompose')
+const decomposition = await agent(
+  `Branch ${branch} is mid-implementation of the approved plan at ${planFile} (issue ${issueRef}). Inspect the ACTUAL current repository state on this checked-out branch: run \`git log --oneline origin/main..HEAD\` and \`git show --stat\` on each commit found, and read the actual files already created/modified to see exactly what's genuinely done versus merely planned. Read the full plan file. Then propose an ORDERED list of sub-tasks covering ONLY the plan's requirements NOT yet satisfied by the existing commit(s) — do not re-propose anything already done. Each sub-task must: name the specific plan section numbers it covers; declare a fileScope (paths/globs it will create or touch) that does not overlap any earlier sub-task's fileScope (list dependsOn by id if it genuinely needs another sub-task's output first, and order the list so every sub-task appears after everything in its dependsOn); and state a concrete doneWhen (a test, measurement, or gate command whose result proves it's actually complete, not just attempted). Size each sub-task so ONE fresh coding agent can plausibly complete it in one focused session (prefer more, smaller sub-tasks over few large ones) — this repository's local gate and any environment constraints (Docker/$TMPDIR, browser availability, etc.) are documented in skills/ship/references/repo-footguns.md and CLAUDE.md; read them and account for them. Also return a 2-4 sentence alreadyDoneSummary of what the existing commit(s) actually accomplished, so downstream agents don't redo it. ${READ_ONLY}`,
+  { label: 'decompose remaining work', phase: 'Decompose', schema: DECOMPOSE_SCHEMA, model: 'fable', effort: 'high' },
+)
+if (!decomposition) return { status: 'error', reason: 'decomposition agent died' }
+log(`Decomposed remaining work into ${decomposition.subtasks.length} sub-task(s): ${decomposition.subtasks.map(t => t.id).join(', ')}`)
+
+phase('Implement')
+const results = []
+for (const [index, task] of decomposition.subtasks.entries()) {
+  log(`Sub-task ${index + 1}/${decomposition.subtasks.length} — ${task.id}: ${task.title}`)
+  const priorSummaries = results.map(r => `- ${r.task.id} (${r.task.title}): ${r.summary ?? '(agent died / no summary)'}`).join('\n')
+  const result = await agent(
+    `You are implementing ONE sub-task of the approved plan at ${planFile} for issue ${issueRef}, on branch ${branch} (already checked out — stay on it, do not switch branches; ${decomposition.alreadyDoneSummary}).\n\n` +
+    `Your sub-task: ${task.id} — ${task.title}\n` +
+    `Covers plan sections: ${task.planSections.join(', ')}\n` +
+    `Description: ${task.description}\n` +
+    `Your file scope (create/modify only within this — if the plan genuinely requires touching something outside it, explain why in your final report rather than silently expanding scope): ${task.fileScope.join(', ')}\n` +
+    `Definition of done: ${task.doneWhen}\n\n` +
+    (priorSummaries ? `Earlier sub-tasks already completed on this branch, in order:\n${priorSummaries}\n\n` : '') +
+    `Read the full plan file for complete context (architecture, invariants, sabotage checks, and the exact commands/conventions it specifies) before implementing — your sub-task description is a pointer into it, not a replacement for it. Follow skills/ship/references/per-issue-cycle.md steps 2-3 and skills/ship/references/repo-footguns.md. Run the full local gate (` + '`npm run check:types && npm run check:arch && npm run check:schemas && npm run check:examples && npm test && npm run build`' + `, captured to a file, tail only on completion) before considering your sub-task done, plus whatever sub-task-specific test/measurement command your doneWhen names.\n\n` +
+    `Mutation boundary: Edit/Write + local \`git commit\` on ${branch} ONLY (this working tree — no push, no PR, no other gh mutations, no issue edits, no ship-log writes, no memory writes, no CHANGELOG.md beyond your own entry if this sub-task warrants one, no TaskCreate/TaskUpdate, never invoke chatgpt-review). Commit message format: \`<type>(${issueTag}): <summary>\`, ending with the repo's standard commit footer.\n\n` +
+    `Return: what you implemented, the gate's pass/fail tail, your doneWhen command's actual result, files touched, and your commit(s) (\`git log --oneline\` for just your new commits). If you get stuck or the sub-task turns out to need a decision not in the plan, stop and report exactly what's missing rather than guessing.`,
+    { label: task.id, phase: 'Implement', model: 'sonnet' },
+  )
+  results.push({ task, summary: result })
+  if (!result) return { status: 'error', reason: `sub-task ${task.id} agent died`, completedSubtasks: results.map(r => r.task.id) }
+}
+
+return { status: 'done', alreadyDoneSummary: decomposition.alreadyDoneSummary, subtasks: decomposition.subtasks, results }
diff --git a/skills/ship/references/per-issue-cycle.md b/skills/ship/references/per-issue-cycle.md
index 4abf3461..0d2ba28c 100644
--- a/skills/ship/references/per-issue-cycle.md
+++ b/skills/ship/references/per-issue-cycle.md
@@ -44,6 +44,15 @@ without migration, shared API/type changes, or non-trivial fixture/e2e impact.
 dependency/framework swap, complex ordering/concurrency/cancellation/rollback, large
 cross-cutting refactor, or under-determined architecture.
 
+**If High, also state whether the unit is Large** — likely to exceed what one coding
+agent can complete well in a single session (a wide file footprint, several
+independent subsystems, external infra such as Docker/live servers/real browsers) —
+with a one-line justification. This flag, not risk alone, decides whether the
+coordinator decomposes implementation (`SKILL.md` step 2.3,
+`references/decompose-and-implement-loop.workflow.mjs`) instead of spawning one
+implementer. Most units are not Large; do not default to it, and do not infer it from
+risk alone — a High-risk unit can still be small (e.g. a narrow auth fix).
+
 ### Invariant map
 
 For medium- and high-risk work, include:
diff --git a/skills/ship/tests/workflow-contract.test.mjs b/skills/ship/tests/workflow-contract.test.mjs
index 17d24e18..532be894 100644
--- a/skills/ship/tests/workflow-contract.test.mjs
+++ b/skills/ship/tests/workflow-contract.test.mjs
@@ -31,3 +31,34 @@ test('the canonical plan path remains the authoring-session identity', async ()
   assert.doesNotMatch(workflow, /planFile\s*=/);
   assert.match(workflow, /--output-file \$\{shellQuote\(runArgs\.planFile\)\}/);
 });
+
+test('the decompose-and-implement loop grounds decomposition in the real branch, runs sub-tasks sequentially, and propagates failure', async () => {
+  const source = await fs.readFile(path.join(root, 'references/decompose-and-implement-loop.workflow.mjs'), 'utf8');
+  assert.doesNotThrow(() => new Function(`return async function workflowSyntaxCheck() {\n${source.replace('export const meta', 'const meta')}\n}`));
+  // Decomposition must read what's ALREADY committed, not just the plan in the abstract.
+  assert.match(source, /git log --oneline origin\/main\.\.HEAD/);
+  assert.match(source, /model: 'fable', effort: 'high'/);
+  // Sequential, not parallel — a for-loop with one agent() awaited per iteration, no
+  // parallel()/pipeline() call fanning sub-tasks out concurrently.
+  assert.match(source, /for \(const \[index, task\] of decomposition\.subtasks\.entries\(\)\)/);
+  assert.doesNotMatch(source, /parallel\(/);
+  assert.doesNotMatch(source, /pipeline\(/);
+  assert.match(source, /model: 'sonnet'/);
+  // A dead sub-task agent must stop the loop, not silently skip the rest.
+  assert.match(source, /if \(!result\) return \{ status: 'error'/);
+});
+
+test('required args are validated and the issue tag is derived for commit messages', async () => {
+  const source = await fs.readFile(path.join(root, 'references/decompose-and-implement-loop.workflow.mjs'), 'utf8');
+  assert.match(source, /throw new Error\('args \{planFile, branch, issueRef\} required'\)/);
+  assert.match(source, /issueRef\.match\(\/#\\d\+\/\)/);
+});
+
+test('SKILL.md wires the High+Large decomposition branch into step 2.3, and per-issue-cycle.md defines the flag', async () => {
+  const skill = await fs.readFile(path.join(root, 'SKILL.md'), 'utf8');
+  assert.match(skill, /High risk \+ Large unit/);
+  assert.match(skill, /decompose-and-implement-loop\.workflow\.mjs/);
+  const cycle = await fs.readFile(path.join(root, 'references/per-issue-cycle.md'), 'utf8');
+  assert.match(cycle, /also state whether the unit is Large/);
+  assert.match(cycle, /Most units are not Large/);
+});