@linimin/pi-letscook 0.1.65 → 0.1.67

package/scripts/release-check.sh

@@ -5,32 +5,45 @@ ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
 cd "$ROOT"
 export PI_COMPLETION_RUNNING_RELEASE_CHECK=1
 
-echo "[release-check] running control-plane validation, tracked .agent contract coverage, slice-surface parity, explicit-/cook parity, startup/refocus/context regressions, canonical evidence artifact, active-slice contract, observability, legacy cleanup, evaluator calibration, and rubric contract coverage"
+echo "[release-check] running control-plane validation, tracked .agent contract coverage, slice-surface parity, /cook startup-plan parity, startup/refocus/context regressions, canonical evidence artifact, active-slice contract, observability, legacy cleanup, evaluator calibration, and rubric contract coverage"
 bash .agent/verify_completion_control_plane.sh
 git ls-files --error-unmatch .agent/README.md .agent/mission.md .agent/profile.json .agent/verify_completion_stop.sh .agent/verify_completion_control_plane.sh >/dev/null
 
-echo "[release-check] verifying public /cook parity and primary-agent-handoff docs/help"
+echo "[release-check] verifying public /cook parity and startup-plan docs/help"
 python3 - <<'PY'
 from pathlib import Path
 
 checks = {
     "README.md": [
         "You can still implement directly in ordinary chat when you do not need workflow state.",
-        "When you explicitly run `/cook`, it first checks for a fresh explicit primary-agent handoff.",
-        "If one is missing, it calls a same-entry primary-agent handoff synthesis step from the current task context, then asks you to **Start** or **Cancel** before rewriting canonical workflow state.",
-        "Explicit `/cook` capsules are still valid startup intake, but they are no longer the only path because `/cook` can synthesize the primary-agent handoff in the same entry when needed.",
+        "When you explicitly run `/cook`, it first checks for a fresh explicit primary-agent startup-plan preview.",
+        "After **Start**, the approved startup plan is written into `.agent/startup-plan.json` / `.agent/startup-plan.md`, and `completion-regrounder` uses it to derive canonical slices from current repo truth.",
+        "Explicit `/cook` capsules are still valid startup intake, but they are no longer the only path because `/cook` can synthesize the primary-agent startup plan in the same entry when needed.",
     ],
     "CHANGELOG.md": [
-        "changed `/cook` entry again so it now calls a same-entry primary-agent handoff synthesis step when no fresh explicit handoff exists, then proceeds directly to Start / Cancel without requiring a second `/cook` run",
-        "kept `/cook` from guessing startup purely from recent discussion by requiring that same-entry synthesis step to produce explicit primary-agent handoff data before workflow state can start",
+        "rewrote `/cook` startup around an approved startup plan that is captured under `.agent/startup-plan.json` / `.agent/startup-plan.md` after Start instead of leaving startup intent only as advisory intake in `state.json`",
+        "kept `/cook` confirm-first while handing the approved startup plan to `completion-regrounder`, which now derives canonical slices from repo truth instead of treating startup intake like an implementation-ready first-slice contract",
     ],
     "extensions/completion/prompt-surfaces.ts": [
-        '"If the user explicitly runs /cook, the extension should call a primary-agent handoff synthesis step from the current task context, then show Start/Cancel confirmation without making the user rerun /cook."',
-        '"Do not expect /cook to infer or guess startup intent from recent discussion alone; /cook should use explicit primary-agent handoff data, whether it already exists or is synthesized in the same /cook entry."',
+        '"If the user explicitly runs /cook, the extension should call a primary-agent startup-plan synthesis step from the current task context, show Start/Cancel confirmation in the same /cook entry, and only write the approved plan into .agent after Start."',
+        '"When /cook starts, the approved startup plan should be written into .agent and then handed to completion-regrounder so canonical slices can be derived from repo truth."',
+        '"In ordinary chat, do not load or follow completion-protocol, and do not call completion_role."',
     ],
     "extensions/completion/index.ts": [
-        '"/cook failed closed because the primary-agent handoff step could not prepare a concrete startup handoff from the current task context. Clarify the mission, first slice, or verification intent in the main chat, then rerun /cook."',
-        'description: "/cook workflow: start or replace workflow only from an explicit primary-agent handoff, or resume the current workflow from canonical state"',
+        '"/cook failed closed because the startup-plan step could not prepare a concrete workflow startup plan from the current task context. Clarify the mission, scope, acceptance, or verification intent in the main chat, then rerun /cook."',
+        'description: "/cook workflow: capture the approved startup plan into .agent, let completion-regrounder split canonical slices, or resume the current workflow from canonical state"',
+        '"Do not call completion_role from ordinary chat; it is reserved for explicit /cook workflow driver turns."',
+    ],
+    "extensions/completion/policy-guards.ts": [
+        'return "completion_role may only be used from an explicit /cook workflow driver turn.";',
+    ],
+    "skills/cook-handoff-boundary/SKILL.md": [
+        '- load or follow `completion-protocol` while still in ordinary chat',
+        '- call `completion_role` before the user has explicitly entered `/cook`',
+    ],
+    "skills/completion-protocol/SKILL.md": [
+        'Load this skill only after the user explicitly enters `/cook` and you are operating inside the `completion` workflow as the workflow driver or a completion role.',
+        'Do not load or follow this skill from ordinary chat.',
     ],
 }
 

package/scripts/smoke-test.sh

@@ -123,10 +123,10 @@ capsule = {
         "Keep scripts/smoke-test.sh and kickoff-prompt coverage truthful for packaged bootstrap."
     ],
     "risks": [
-        "Smoke-test bootstrap should stay anchored to the fresh explicit handoff."
+        "Smoke-test bootstrap should stay anchored to the fresh explicit startup-plan preview."
     ],
     "notes": [
-        "Keep the smoke fixture aligned with the shipped explicit-handoff-only startup contract."
+        "Keep the smoke fixture aligned with the shipped startup-plan-driven startup contract."
     ],
     "handoff_kind": "implementation_workflow_handoff",
     "first_slice_goal": "Scaffold canonical completion files and verify the packaged startup contract.",
@@ -140,14 +140,14 @@ capsule = {
     "verification_commands": [
         "npm run smoke-test"
     ],
-    "why_this_slice_first": "The packaged explicit-handoff startup path must work before later workflow verification can run.",
+    "why_this_slice_first": "The packaged startup-plan entry path must work before later workflow verification can run.",
     "task_type": "completion-workflow",
     "evaluation_profile": "completion-rubric-v1",
-    "why_cook_now": "The startup handoff is concrete enough to bootstrap canonical workflow files."
+    "why_cook_now": "The startup plan is concrete enough to bootstrap canonical workflow files."
 }
 messages = [
     {"role": "user", "content": "Please prepare the packaged smoke-test bootstrap path and tell me when it is ready for /cook."},
-    {"role": "assistant", "content": "This bootstrap path is ready for /cook. Run /cook to confirm the startup brief.\n\n```cook_handoff\n" + json.dumps(capsule, ensure_ascii=False, indent=2) + "\n```"},
+    {"role": "assistant", "content": "This bootstrap path is ready for /cook. Run /cook to confirm the startup plan.\n\n```cook_handoff\n" + json.dumps(capsule, ensure_ascii=False, indent=2) + "\n```"},
 ]
 print(json.dumps(messages, ensure_ascii=False))
 PY
@@ -175,7 +175,7 @@ chooser = Path(sys.argv[5])
 
 assert not Path('.agent').exists(), 'startup /cook inline-args rejection should leave canonical state untouched'
 assert not routing.exists(), 'startup /cook inline-args rejection should not open active-workflow routing before a workflow exists'
-assert not proposal.exists(), 'startup /cook inline-args rejection should not emit a startup-brief proposal snapshot'
+assert not proposal.exists(), 'startup /cook inline-args rejection should not emit a startup-plan proposal snapshot'
 assert not chooser.exists(), 'startup /cook inline-args rejection should not open the existing-workflow chooser before a workflow exists'
 assert '/cook no longer accepts inline arguments.' in output, 'startup /cook inline-args rejection should explain the bare-only entry contract'
 PY
@@ -189,7 +189,7 @@ PI_COMPLETION_TEST_DRIVER_PROMPT_PATH="$KICKOFF_PROMPT" \
 pi --session "$BOOTSTRAP_SESSION" -e "$PKG_ROOT" -p "/cook" \
   >"$TMPDIR/pi-completion-smoke-bootstrap.out" 2>"$TMPDIR/pi-completion-smoke-bootstrap.err"
 
-for file in .agent/profile.json .agent/state.json .agent/plan.json .agent/active-slice.json .agent/verification-evidence.json; do
+for file in .agent/profile.json .agent/state.json .agent/startup-plan.json .agent/startup-plan.md .agent/plan.json .agent/active-slice.json .agent/verification-evidence.json; do
   [[ -f "$file" ]] || { echo "missing canonical bootstrap file: $file" >&2; exit 1; }
 done
 
@@ -209,6 +209,8 @@ profile = json.loads(Path('.agent/profile.json').read_text())
 state = json.loads(Path('.agent/state.json').read_text())
 plan = json.loads(Path('.agent/plan.json').read_text())
 active = json.loads(Path('.agent/active-slice.json').read_text())
+startup_plan = json.loads(Path('.agent/startup-plan.json').read_text())
+startup_plan_md = Path('.agent/startup-plan.md').read_text()
 evidence = json.loads(Path('.agent/verification-evidence.json').read_text())
 kickoff = Path(sys.argv[1]).read_text()
 
@@ -228,11 +230,20 @@ assert brief['source'] == 'primary_agent_handoff', 'smoke bootstrap should recor
 assert brief['mission'] == state['mission_anchor'], 'advisory startup brief mission should match the canonical mission anchor after bootstrap'
 assert brief['scope'] == ['Materialize the canonical completion control-plane files.', 'Keep the smoke test on supported /cook startup behavior.'], 'advisory startup brief should preserve scope items'
 assert brief['constraints'] == ['Keep startup proposal confirmation approval-only.'], 'advisory startup brief should preserve constraints'
+assert startup_plan['artifact_type'] == 'completion-startup-plan', 'startup-plan.json should persist the approved startup plan canonically'
+assert startup_plan['mission_anchor'] == state['mission_anchor'], 'startup-plan.json mission anchor should match canonical workflow state'
+assert startup_plan['source'] == 'primary_agent_handoff', 'startup-plan.json should preserve the primary-agent startup source'
+assert startup_plan['scope'] == brief['scope'], 'startup-plan.json should preserve approved scope items'
+assert startup_plan['constraints'] == brief['constraints'], 'startup-plan.json should preserve approved constraints'
+assert startup_plan['planned_surfaces'] == ['.agent/README.md', 'scripts/smoke-test.sh'], 'startup-plan.json should preserve planned surfaces derived from startup hints'
+assert startup_plan['verification_intent'] == ['npm run smoke-test'], 'startup-plan.json should preserve verification intent derived from startup hints'
+assert '## Planned surfaces' in startup_plan_md, 'startup-plan.md should render planned surfaces as a readable section'
+assert '.agent/README.md' in startup_plan_md and 'scripts/smoke-test.sh' in startup_plan_md, 'startup-plan.md should mirror planned surfaces'
 assert brief['acceptance'] == [
     'Scaffold .agent/profile.json, .agent/state.json, .agent/plan.json, .agent/active-slice.json, and .agent/verification-evidence.json for the smoke fixture.',
     'Keep scripts/smoke-test.sh and kickoff-prompt coverage truthful for packaged bootstrap.'
 ], 'advisory startup brief should preserve acceptance'
-assert brief['risks'] == ['Smoke-test bootstrap should stay anchored to the fresh explicit handoff.'], 'advisory startup brief should preserve handoff risks'
+assert brief['risks'] == ['Smoke-test bootstrap should stay anchored to the fresh explicit startup-plan preview.'], 'advisory startup brief should preserve startup-plan risks'
 assert 'First slice goal: Scaffold canonical completion files and verify the packaged startup contract.' in brief['notes'], 'advisory startup brief should preserve the first_slice_goal in notes'
 assert 'Verification commands: npm run smoke-test' in brief['notes'], 'advisory startup brief should preserve verification_commands in notes'
 assert evidence['artifact_type'] == 'completion-verification-evidence', 'verification-evidence.json artifact_type mismatch after bootstrap'
@@ -269,9 +280,11 @@ assert 'ordinary main chat unless the user explicitly runs /cook' in handoff_tex
 assert 'directly implement requested repo changes, including multi-file work' in handoff_text, 'ordinary handoff reminder should allow direct ordinary-chat implementation'
 assert 'Do not proactively tell the user to run /cook' in handoff_text, 'ordinary handoff reminder should keep ordinary chat neutral until explicit /cook entry'
 assert '/cook is optional workflow mode' in handoff_text, 'ordinary handoff reminder should position /cook as optional workflow mode'
+assert 'In ordinary chat, do not load or follow completion-protocol, and do not call completion_role.' in handoff_text, 'ordinary handoff reminder should forbid workflow-role routing before explicit /cook'
 assert 'If the user wants direct implementation now, stay in ordinary chat and help directly instead of blocking on /cook.' in handoff_text, 'ordinary handoff reminder should avoid blocking implementation on /cook'
-assert 'the extension should call a primary-agent handoff synthesis step from the current task context' in handoff_text, 'ordinary handoff reminder should describe same-entry primary-agent handoff synthesis for /cook'
+assert 'the extension should call a primary-agent startup-plan synthesis step from the current task context' in handoff_text, 'ordinary handoff reminder should describe same-entry primary-agent startup-plan synthesis for /cook'
 assert 'Do not expect /cook to infer or guess startup intent from recent discussion alone' in handoff_text, 'ordinary handoff reminder should forbid /cook-side guessing'
+assert 'the approved startup plan should be written into .agent and then handed to completion-regrounder' in handoff_text, 'ordinary handoff reminder should mention startup-plan persistence and regrounder slice derivation'
 assert 'do not silently rewrite discussion into canonical workflow state' in handoff_text, 'ordinary handoff reminder should preserve non-canonical ordinary-chat behavior'
 assert not auto_resume.exists(), 'ordinary non-/cook turn should not queue auto-resume before /cook activation'
 assert 'Skipped completion workflow auto-resume prompt (test mode)' not in output, 'ordinary non-/cook turn should not attempt auto-resume'
@@ -375,6 +388,54 @@ active['evaluation_profile'] = profile['evaluation_profile']
 active_path.write_text(json.dumps(active, indent=2) + '\n')
 PY
 
+python3 - <<'PY'
+import json
+from pathlib import Path
+path = Path('.agent/startup-plan.json')
+startup = json.loads(path.read_text())
+startup.pop('task_type', None)
+path.write_text(json.dumps(startup, indent=2) + '\n')
+PY
+
+if bash .agent/verify_completion_control_plane.sh >/dev/null 2>&1; then
+  echo "expected control-plane verification to fail when startup-plan.json omits task_type" >&2
+  exit 1
+fi
+
+python3 - <<'PY'
+import json
+from pathlib import Path
+profile = json.loads(Path('.agent/profile.json').read_text())
+startup_path = Path('.agent/startup-plan.json')
+startup = json.loads(startup_path.read_text())
+startup['task_type'] = profile['task_type']
+startup_path.write_text(json.dumps(startup, indent=2) + '\n')
+PY
+
+python3 - <<'PY'
+import json
+from pathlib import Path
+path = Path('.agent/startup-plan.json')
+startup = json.loads(path.read_text())
+startup['mission_anchor'] = 'drifted-startup-plan-mission'
+path.write_text(json.dumps(startup, indent=2) + '\n')
+PY
+
+if bash .agent/verify_completion_control_plane.sh >/dev/null 2>&1; then
+  echo "expected control-plane verification to fail when startup-plan.json mission_anchor drifts from canonical workflow state" >&2
+  exit 1
+fi
+
+python3 - <<'PY'
+import json
+from pathlib import Path
+state = json.loads(Path('.agent/state.json').read_text())
+startup_path = Path('.agent/startup-plan.json')
+startup = json.loads(startup_path.read_text())
+startup['mission_anchor'] = state['mission_anchor']
+startup_path.write_text(json.dumps(startup, indent=2) + '\n')
+PY
+
 python3 - <<'PY'
 import json
 from pathlib import Path

package/skills/completion-protocol/SKILL.md

@@ -1,11 +1,13 @@
 ---
 name: completion-protocol
-description: Shared completion workflow protocol for long-running coding tasks with canonical .agent state, one-slice execution, mandatory role dispatch, recovery, and final stop criteria. Use when driving persistent end-to-end completion work across multiple sessions.
+description: Shared completion workflow protocol for long-running coding tasks with canonical .agent state, one-slice execution, mandatory role dispatch, recovery, and final stop criteria. Use only after the user explicitly enters `/cook` and the main session becomes the completion workflow driver.
 ---
 
 # Completion Protocol
 
-Load this skill whenever you are operating inside the `completion` workflow.
+Load this skill only after the user explicitly enters `/cook` and you are operating inside the `completion` workflow as the workflow driver or a completion role.
+
+Do not load or follow this skill from ordinary chat.
 
 This skill defines shared protocol facts only. Role-specific behavior belongs in the dedicated completion agents:
 
@@ -19,7 +21,9 @@ This skill defines shared protocol facts only. Role-specific behavior belongs in
 ## Shared Rules
 
 - Current repo truth beats stale notes, stale summaries, and conversation memory.
-- `plan.json` is the persistent machine-readable slice backlog. Rebuild it during every re-grounding wave and keep it truthful after every committed slice.
+- `startup-plan.json` is the approved workflow startup plan captured at `/cook`. Treat it as planning input only: `completion-regrounder` must reconcile it against current repo truth and derive canonical slices instead of copying it blindly into `plan.json`.
+- `startup-plan.md` is the human-readable rendering of that same approved startup plan and must remain in parity with `startup-plan.json`.
+- `plan.json` is the persistent machine-readable slice backlog. Rebuild it during every re-grounding wave from repo truth plus the approved startup plan, and keep it truthful after every committed slice.
 - `state.json` is the persistent machine-readable workflow controller. Keep `current_phase`, `continuation_policy`, `continuation_reason`, `next_mandatory_role`, and `next_mandatory_action` truthful after every transition.
 - Every slice in `plan.json` must have non-empty `acceptance_criteria` — concrete, verifiable conditions that define done. A slice without acceptance criteria is invalid and must not be selected.
 - Acceptance criteria are immutable after lock except for removing a criterion already satisfied with evidence or adding a missing criterion discovered during implementation.
@@ -74,7 +78,7 @@ If the workflow driver detects that the next mandatory action belongs to a compl
 
 1. If tracked protocol contract files are missing or first-time onboarding is required, invoke `completion-bootstrapper`.
 2. If canonical `.agent` execution state is missing, invalid, contradictory, stale, or ambiguous after compaction or recovery, invoke `completion-regrounder`.
-3. If no slice is selected, invoke `completion-regrounder` to reconcile `.agent/plan.json` and return the next exact handoff payload.
+3. If no slice is selected, invoke `completion-regrounder` to read `.agent/startup-plan.json`, reconcile `.agent/plan.json` against current repo truth, and return the next exact handoff payload.
 4. If a slice is `selected` or `in_progress` and no new slice commit exists yet, invoke `completion-implementer`.
 5. If the latest committed slice lacks a review result, invoke `completion-reviewer`.
 6. If the latest committed slice lacks an audit result, invoke `completion-auditor`.
@@ -97,6 +101,8 @@ Tracked repo-contract files:
 Ignored canonical execution-state files:
 
 - `.agent/state.json`
+- `.agent/startup-plan.json`
+- `.agent/startup-plan.md`
 - `.agent/plan.json`
 - `.agent/active-slice.json`
 - `.agent/slice-history.jsonl`
@@ -112,6 +118,7 @@ Read these when making completion decisions:
 - `.agent/README.md`
 - `.agent/profile.json`
 - `.agent/state.json`
+- `.agent/startup-plan.json`
 - `.agent/plan.json`
 - `.agent/active-slice.json`
 - `.agent/slice-history.jsonl`
@@ -138,6 +145,7 @@ Canonical truth remains in `.agent/**`.
 After context compaction, suspected memory loss, stalled-role recovery, or any ambiguous completion state, the workflow driver must re-read:
 
 - `.agent/state.json`
+- `.agent/startup-plan.json`
 - `.agent/plan.json`
 - `.agent/active-slice.json`
 - `.agent/verification-evidence.json`
@@ -155,7 +163,7 @@ The exact implementer handoff now includes implementation-scope surfaces and exp
 
 The workflow driver must not continue implementation, review, audit, or stop evaluation from compacted conversation memory alone.
 
-After compaction or recovery, `completion-implementer` must also re-read canonical `.agent/state.json`, `.agent/plan.json`, `.agent/active-slice.json`, and `.agent/verification-evidence.json` before resuming work. If `.agent/active-slice.json` still contains a truthful exact handoff snapshot, continue from canonical state rather than asking the user to resend the original caller payload.
+After compaction or recovery, `completion-implementer` must also re-read canonical `.agent/state.json`, `.agent/startup-plan.json`, `.agent/plan.json`, `.agent/active-slice.json`, and `.agent/verification-evidence.json` before resuming work. If `.agent/active-slice.json` still contains a truthful exact handoff snapshot, continue from canonical state rather than asking the user to resend the original caller payload.
 
 ## Shared Report Header
 

package/skills/completion-protocol/references/completion.md

@@ -13,6 +13,8 @@
 ## Ignored Canonical Execution State
 
 - `.agent/state.json`
+- `.agent/startup-plan.json`
+- `.agent/startup-plan.md`
 - `.agent/plan.json`
 - `.agent/active-slice.json`
 - `.agent/slice-history.jsonl`
@@ -28,6 +30,7 @@ Read these when making completion decisions:
 - `.agent/README.md`
 - `.agent/profile.json`
 - `.agent/state.json`
+- `.agent/startup-plan.json`
 - `.agent/plan.json`
 - `.agent/active-slice.json`
 - `.agent/slice-history.jsonl`
@@ -130,6 +133,37 @@ Rules:
 4. `continuation_policy == paused` means the user explicitly paused the workflow.
 5. `continuation_policy == done` means canonical final stop reconciliation is complete and the workflow may stop.
 
+## Fixed Startup Plan Model
+
+`startup-plan.json` carries the approved workflow startup plan captured at `/cook`.
+
+Required fields:
+
+- `schema_version`
+- `artifact_type` where the value is `completion-startup-plan`
+- `status` where the value is `approved`
+- `source`
+- `captured_at`
+- `mission_anchor`
+- `goal_text`
+- `task_type`
+- `evaluation_profile`
+- `scope`
+- `constraints`
+- `acceptance`
+- `risks`
+- `notes`
+- `planned_surfaces`
+- `verification_intent`
+- `sequencing_hints`
+
+Rules:
+
+1. `startup-plan.json` is canonical startup intake, not canonical slice selection.
+2. `completion-regrounder` must treat `startup-plan.json` as planning input, reconcile it against current repo truth, and derive canonical slices in `plan.json`.
+3. `startup-plan.json` must stay aligned with the current canonical mission anchor in `.agent/state.json`, `.agent/plan.json`, and `.agent/active-slice.json` until a later `/cook` refocus or next-round startup rewrites it.
+4. `startup-plan.md` is the human-readable rendering of the same approved startup plan and must remain in parity with `startup-plan.json`.
+
 `plan.json` carries the ordered persistent slice backlog.
 
 Required fields:
@@ -338,7 +372,7 @@ It must not, while a slice is selected or in progress:
 
 1. If tracked protocol contract files are missing or first-time onboarding is required, invoke `completion-bootstrapper`.
 2. If canonical `.agent` execution state is missing, stale, invalid, contradictory, or ambiguous after compaction or recovery, invoke `completion-regrounder` first.
-3. If no slice is selected, invoke `completion-regrounder` to reconcile `.agent/plan.json` and return the next exact handoff payload.
+3. If no slice is selected, invoke `completion-regrounder` to read `.agent/startup-plan.json`, reconcile `.agent/plan.json` against current repo truth, and return the next exact handoff payload.
 4. If a slice is `selected` or `in_progress` and no new commit exists for it yet, invoke `completion-implementer`.
 5. If the latest committed slice lacks review, invoke `completion-reviewer`.
 6. If the latest committed slice lacks audit, invoke `completion-auditor`.
@@ -351,6 +385,7 @@ It must not, while a slice is selected or in progress:
 After context compaction, suspected memory loss, stalled-role recovery, or any ambiguous completion state, the workflow root must re-read:
 
 - `.agent/state.json`
+- `.agent/startup-plan.json`
 - `.agent/plan.json`
 - `.agent/active-slice.json`
 - `.agent/verification-evidence.json`
@@ -366,7 +401,7 @@ The workflow root must invoke `completion-regrounder` before continuing whenever
 
 The workflow root must not continue implementation, review, audit, or stop evaluation from compacted conversation memory alone.
 
-After compaction or recovery, `completion-implementer` must also re-read canonical `.agent/state.json`, `.agent/plan.json`, `.agent/active-slice.json`, and `.agent/verification-evidence.json` before resuming work. If `.agent/active-slice.json` still contains a truthful exact handoff snapshot, continue from canonical state rather than asking the user to resend the original caller payload.
+After compaction or recovery, `completion-implementer` must also re-read canonical `.agent/state.json`, `.agent/startup-plan.json`, `.agent/plan.json`, `.agent/active-slice.json`, and `.agent/verification-evidence.json` before resuming work. If `.agent/active-slice.json` still contains a truthful exact handoff snapshot, continue from canonical state rather than asking the user to resend the original caller payload.
 
 ## Default Priority Policy
 

package/skills/cook-handoff-boundary/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: cook-handoff-boundary
-description: Ordinary-chat contract for treating `/cook` as an optional workflow mode while requiring `/cook` to use primary-agent-authored handoff data instead of guessing from recent discussion.
+description: Ordinary-chat contract for treating `/cook` as an optional workflow mode while requiring `/cook` to capture a primary-agent-authored startup plan instead of guessing from recent discussion.
 ---
 
 # /cook Handoff Boundary
@@ -34,6 +34,8 @@ The primary agent should not:
 
 - proactively tell the user to run `/cook` just because the task looks workflow-worthy
 - proactively emit a `cook_handoff` capsule by default
+- load or follow `completion-protocol` while still in ordinary chat
+- call `completion_role` before the user has explicitly entered `/cook`
 - act as though workflow has already started when it has not
 - silently rewrite ordinary-chat discussion into canonical workflow state
 
@@ -56,20 +58,20 @@ But even in those cases:
 
 If the user explicitly runs or clearly chooses `/cook` workflow mode, the system behavior should be:
 
-1. check for a fresh explicit primary-agent `cook_handoff`
-2. if none exists, call a primary-agent handoff synthesis step immediately from the current task context
-3. use that handoff to show Start / Cancel confirmation in the same `/cook` entry
-4. write canonical workflow state only after Start
+1. check for a fresh explicit primary-agent `cook_handoff` preview when one already exists
+2. if none exists, call a primary-agent startup-plan synthesis step immediately from the current task context
+3. use that startup plan to show Start / Cancel confirmation in the same `/cook` entry
+4. after Start, write the approved startup plan into `.agent/startup-plan.json` / `.agent/startup-plan.md`, then let `completion-regrounder` derive canonical slices from repo truth
 
 That means:
 
-- `/cook` must not infer or guess the startup slice from recent discussion alone
-- `/cook` should use primary-agent-authored handoff data
-- `/cook` should not require a manual rerun just to consume a handoff it can synthesize immediately from the primary-agent view
+- `/cook` must not infer or guess canonical slices from recent discussion alone
+- `/cook` should use primary-agent-authored startup-plan data
+- `/cook` should not require a manual rerun just to consume a startup plan it can synthesize immediately from the primary-agent view
 
 ## Optional Preview Behavior
 
-Only if the user explicitly asks for a preview startup brief or handoff capsule in ordinary chat may the primary agent provide one.
+Only if the user explicitly asks for a preview startup plan or handoff capsule in ordinary chat may the primary agent provide one.
 
 Optional preview capsule format:
 
@@ -80,7 +82,7 @@ Optional preview capsule format:
   "source": "primary_agent",
   "captured_at": "<ISO-8601 timestamp>",
   "source_turn_id": "<current assistant turn id>",
-  "mission": "<startable implementation mission>",
+  "mission": "<approved workflow mission>",
   "scope": ["..."],
   "constraints": ["..."],
   "non_goals": ["..."],
@@ -103,31 +105,33 @@ Optional preview capsule format:
 Notes:
 
 - `constraints` may be replaced or supplemented by `non_goals` when clearer.
-- `first_slice_goal`, `first_slice_non_goals`, `implementation_surfaces`, `verification_commands`, and `why_this_slice_first` are required for an implementation-ready handoff.
+- `first_slice_goal`, `first_slice_non_goals`, `implementation_surfaces`, `verification_commands`, and `why_this_slice_first` are optional sequencing hints. They help `completion-regrounder` split slices later when they are already obvious, but the approved startup plan may still be startable without them.
 - Any capsule is startup intake for `/cook` only. It is not canonical `.agent/**` state, not active-slice state, and not a second repo contract source.
 
 Suggested wording:
 
-> We can continue directly in ordinary chat if you want. If you prefer workflow mode, run `/cook` and it should use a primary-agent handoff for Start / Cancel confirmation rather than guessing from recent discussion.
+> We can continue directly in ordinary chat if you want. If you prefer workflow mode, run `/cook` and it should capture an approved startup plan for Start / Cancel confirmation, then hand that plan to `completion-regrounder` for slice derivation.
 
 ## Forbidden Behaviors
 
 Before the user explicitly runs `/cook`, the primary agent must not:
 
 - pretend `/cook` has already been invoked
+- load or follow `completion-protocol`
+- call `completion_role` or any completion subagent
 - silently rewrite ordinary-chat discussion into active workflow state
 - claim canonical `.agent/**` startup state exists when it does not
 - refuse ordinary-chat implementation solely because `/cook` would also be possible
 
 When the user does explicitly choose `/cook`, the system must not:
 
-- let `/cook` invent the startup mission from recent discussion alone
-- let `/cook` replace missing handoff data with generic transcript guessing
-- require a second `/cook` invocation when same-entry primary-agent handoff synthesis is possible
+- let `/cook` invent canonical slices from recent discussion alone
+- let `/cook` replace missing startup-plan data with generic transcript guessing
+- require a second `/cook` invocation when same-entry primary-agent startup-plan synthesis is possible
 
 ## Relationship To `completion-protocol`
 
-This skill is only about pre-`/cook` ordinary-chat behavior and `/cook` handoff expectations.
+This skill is only about pre-`/cook` ordinary-chat behavior and `/cook` startup-plan expectations.
 
 After the user explicitly enters `/cook`, the separate `completion-protocol` skill governs: