Add project owner handoffs

.gitignore

@@ -16,6 +16,7 @@ loop-engine/state.json
 loop-engine/portfolio/
 loop-engine/pm-reviews/
 loop-engine/approvals/
+loop-engine/handoffs/
 loop-engine/human-feedback/
 loop-engine/evening-scorecards/*
 !loop-engine/evening-scorecards/.gitkeep
@@ -33,3 +34,6 @@ dist/
 build/
 venv/
 .venv/
+
+# Loop runtime handoff artifacts.
+.loop/daily-focus/

AGENTS.md

@@ -6,7 +6,7 @@
 ## What loop is (for you)
 
 `loop` is a daily product operating loop around a private portfolio registry and local Git repos:
-`portfolio intake → portfolio verification → morning review → approval envelope → planner → value/risk gates → worker (isolated worktree) → verification (sandboxed) → reviewer → auto PR + merge → digest → evening scorecard → memory`.
+`portfolio intake → portfolio verification → morning review → project-owner handoff → approval envelope → planner → value/risk gates → worker (isolated worktree) → verification (sandboxed) → reviewer → auto PR + merge → digest → evening scorecard → memory`.
 You operate it through the `loop` CLI (or `python3 loop-engine/bin/loopctl.py <cmd> --project <id>`). You do **not** edit the engine, prompts, or reviewer output to influence outcomes.
 
 ## First, relay these safety facts to your human
@@ -27,9 +27,10 @@ Before the human starts `loop` (especially unattended), tell them, plainly:
   - `waiting_for_human` → `null`, or `{ "reason": <code>, "issue_path": <path> }`
   - `runs[-1].status` → `merged | no_op | needs_human | failed | ...`
 - **Read the recap from a file**, not by re-running digest: `loop-engine/reports/<project>/latest.md`.
-- **Start with portfolio onboarding once:** `loop portfolio init`, then `loop portfolio add ...` using any handle the human has (local path, GitHub repo/URL, Linear project, URL, or plain name). `loop init` also upserts the current repo into the portfolio.
+- **Start with portfolio onboarding once:** `loop portfolio init`, then `loop portfolio add ...` using any handle the human has (local path, GitHub repo/URL, Linear project, URL, or plain name). If the human has a dedicated project conversation/session, capture it with `--owner-thread-name`, `--owner-thread-id`, `--owner-mode`, and `--handoff-policy`. `loop init` also upserts the current repo into the portfolio.
 - **Run CTO catch-up for the portfolio:** `loop portfolio intake [project...]` writes `loop-engine/portfolio/<project>/profile.{md,json}` with end goal, current stage/progress, primary artifacts, verification candidates, readiness blockers, next steps, and task-level risk boundaries. Portfolio stages are fixed to `idea`, `building_mvp`, `mvp_released`, `released_v1`, and `iterating`; do not invent new stage labels.
-- **Start each day with PM review:** `loop morning` refreshes portfolio intake profiles, shows Portfolio Registry Verification and Portfolio Readiness boards, then runs the PM Review Agent and writes `pm-reviews/YYYY-MM-DD.{md,json}`; `loop approve <project>` writes the project's `.loop/daily-focus/latest.md`.
+- **Start each day with PM review:** `loop morning` refreshes portfolio intake profiles, shows Portfolio Registry Verification and Portfolio Readiness boards, then runs the PM Review Agent and writes `pm-reviews/YYYY-MM-DD.{md,json}` plus `handoffs/YYYY-MM-DD/<project>.md`; `loop approve <project>` writes the project's `.loop/daily-focus/latest.md`.
+- **Keep the Secretary/PM thread clean:** after morning review, use `loop handoff [project...]` or read `loop-engine/handoffs/latest/<project>.md`, then move execution commands and debug detail to that project's owner thread. The PM thread should keep only portfolio decisions, approvals, shipped value, PR/issue links, digest paths, and `waiting_for_human` summaries.
 - **Do not ignore non-ready high-value projects:** if a local Git repo lacks `.loop/contract.yaml`, clean baseline, or verification commands, surface readiness work. `init-loop` is a required readiness gate for eligible local Git repos, not a soft suggestion; normalize `blocked_needs_loop_init` projects back to `decision: "init-loop"` even if the PM agent tries to mark them `plan-only` or `hold`. `loop approve <project> --init-loop` is the explicit mutation path for approved loop bootstrap; `loop approve --all-init-loop` applies the latest morning review's required init-loop projects, and `loop portfolio init-loop --all-eligible` bootstraps every eligible local Git repo during onboarding. After any init-loop approval, rerun `loop morning` before execution.
 - **Approve medium risk once per day:** when morning recommends a bounded envelope, `loop approve <project> --approve-medium` approves all same-day medium-risk items that stay inside that envelope. Do not approve medium risk item-by-item inside each cycle.
 - **Start approved projects only:** `loop start-day` reads `approvals/latest.json`; it refuses projects not approved today.
@@ -79,7 +80,8 @@ commands:
   portfolio_stages: { allowed: [idea, building_mvp, mvp_released, released_v1, iterating], default_progress_percent: { idea: 10, building_mvp: 35, mvp_released: 60, released_v1: 80, iterating: 90 } }
   setup:    { in: "operator machine", out: "~/.config/loop/config.json + missing-action prompts" }
   init:     { in: "product repo cwd", out: ".loop/contract.yaml + registry + pilot branch", fail: "LOOP_BLOCKED <reason>" }
-  morning:  { in: "portfolio registry + registered project snapshots", out: "portfolio verification board + pm-reviews/YYYY-MM-DD.{md,json} + latest.{md,json}, PM-agent value-ranked portfolio board" }
+  morning:  { in: "portfolio registry + registered project snapshots", out: "portfolio verification board + pm-reviews/YYYY-MM-DD.{md,json} + latest.{md,json}, PM-agent value-ranked portfolio board + handoffs/YYYY-MM-DD/<project>.md" }
+  handoff:  { in: "latest morning PM review [project...]", out: "handoffs/YYYY-MM-DD/index.{md,json} + handoffs/YYYY-MM-DD/<project>.md + latest mirrors; owner-thread prompt for isolated execution context" }
   approve:  { in: "project [--approve-medium | --medium-envelope ... | --init-loop]", out: ".loop/daily-focus/latest.md + approvals/latest.json, or approved loop bootstrap for readiness work" }
   reject:   { in: "project", out: "approvals/latest.json rejection record" }
   start-day: { in: "today's approved projects", out: "approved loops active; medium first run supervised" }
@@ -127,6 +129,7 @@ runtime_artifacts:          # git-ignored; never commit to a public repo
   runs: loop-engine/runs/<run_id>/
   pm_review: loop-engine/pm-reviews/latest.md
   pm_review_plan: loop-engine/pm-reviews/latest.json
+  handoffs: loop-engine/handoffs/latest/<project>.md
   portfolio: ~/.config/loop/portfolio.json
   portfolio_profiles: loop-engine/portfolio/<project>/profile.{md,json}
   approvals: loop-engine/approvals/latest.json

README.md

@@ -108,6 +108,8 @@ loop doctor                       # checks gh / sandbox-exec / the agent CLI you
 # 4) First onboarding: create your portfolio registry
 loop portfolio init
 loop portfolio add /path/to/your-product-repo
+loop portfolio add /path/to/owner-product \
+  --owner-thread-name "Owner thread name" --owner-mode codex_thread
 loop portfolio add https://github.com/owner/repo --mode plan-only
 loop portfolio add --linear-project "Newsletter" --mode hold
 loop portfolio status             # private source of truth for daily PM review
@@ -124,6 +126,8 @@ loop status
 # 6) Morning PM review across the portfolio
 loop morning                         # starts with a Portfolio Registry Verification board,
                                      # then writes pm-reviews/YYYY-MM-DD.{md,json}
+                                     # and handoffs/YYYY-MM-DD/<project>.md
+loop handoff <project>               # prints the owner-thread execution handoff
 loop approve <project>               # approve low-risk automatic work only
 # To approve every medium-risk item today inside the PM-recommended envelope:
 loop approve <project> --approve-medium
@@ -166,7 +170,14 @@ By default **only the top-ranked auto-runnable task executes per cycle** (`max_t
 
 ## Daily rhythm: morning & evening / 每日节奏
 
-`loop` now owns the full daily routine: **portfolio verification → morning review → approvals → day loop → evening recap**. The loop still keeps the human in charge of direction: morning review proposes and ranks work; only `loop approve` turns it into execution input.
+`loop` now owns the full daily routine: **portfolio verification → morning review → project-owner handoff → approvals → day loop → evening recap**. The loop still keeps the human in charge of direction: morning review proposes and ranks work; only `loop approve` turns it into execution input.
+
+The recommended operating model is a clean split between:
+
+- **Secretary/PM thread**: portfolio review, value ranking, approvals, evening recap, and only short product-level status.
+- **Project owner thread**: project execution, loop commands, PR/debug details, and `waiting_for_human` handling for one product.
+
+Morning review writes durable handoff prompts so the Secretary/PM thread does not get polluted by worker logs, diffs, install noise, or project-specific debugging.
 
 **First onboarding — build the private portfolio registry once:**
 1. `loop portfolio init` creates `~/.config/loop/portfolio.json`.
@@ -175,6 +186,7 @@ By default **only the top-ranked auto-runnable task executes per cycle** (`max_t
    - GitHub repo or URL: `loop portfolio add owner/repo` or `loop portfolio add https://github.com/owner/repo`
    - Linear project: `loop portfolio add --linear-project "Project Name"`
    - plain product name or URL for early ideas
+   - optional owner routing: `--owner-thread-name "TokenPulse owner" --owner-thread-id <thread-id> --owner-mode codex_thread`
 3. `loop init` also upserts the current repo into the portfolio as an executable `loop` project.
 4. `loop portfolio intake [project...]` performs the CTO catch-up pass for each portfolio entry: where the project lives, what it appears to be building, current stage/progress, primary artifact candidates, verification candidates, loop readiness, blockers, next steps, and task-level risk boundaries. It writes:
    - `loop-engine/portfolio/<project>/profile.md`
@@ -203,23 +215,29 @@ For local Git repos, `init-loop` is a required readiness gate, not a soft sugges
    - `loop-engine/pm-reviews/latest.md`
    - `loop-engine/pm-reviews/YYYY-MM-DD.json`
    - `loop-engine/pm-reviews/latest.json`
-2. `loop approve <project>` writes that project's approved focus to:
+   - `loop-engine/handoffs/YYYY-MM-DD/index.md`
+   - `loop-engine/handoffs/YYYY-MM-DD/<project>.md`
+   - `loop-engine/handoffs/latest/<project>.md`
+2. `loop handoff [project...]` rebuilds and prints the latest handoff index. Send each `<project>.md` to that project's owner thread. The handoff contains the day's objective, expected user value, risk envelope, value-ranked work, exact commands, and a reporting contract.
+3. `loop approve <project>` writes that project's approved focus to:
    - `<product-repo>/.loop/daily-focus/YYYY-MM-DD.md`
    - `<product-repo>/.loop/daily-focus/latest.md`
    - `loop-engine/approvals/YYYY-MM-DD.{json,md}`
-3. Medium-risk work is approved once in the morning, for the whole day, inside a bounded envelope:
+4. Medium-risk work is approved once in the morning, for the whole day, inside a bounded envelope:
    - recommended: `loop approve <project> --approve-medium`
    - `loop approve <project> --medium-envelope <name> --allowed-file ... --verification-command ...`
    - PM-recommended verification commands are clipped to the project's trusted `verification_commands`; untrusted suggestions are shown in the morning review but are not written into the approved envelope.
-4. Readiness work can be the highest-value work. If an important local Git repo is not loop-ready, morning review treats loop init as a required setup gate: define its artifact contract, verification commands, and baseline, then explicitly approve loop init with `loop approve <project> --init-loop`. That command mutates the repo by running loop bootstrap, then asks you to run morning review again before execution.
+5. Readiness work can be the highest-value work. If an important local Git repo is not loop-ready, morning review treats loop init as a required setup gate: define its artifact contract, verification commands, and baseline, then explicitly approve loop init with `loop approve <project> --init-loop`. That command mutates the repo by running loop bootstrap, then asks you to run morning review again before execution.
    - To approve every `init-loop` project recommended by the latest morning review: `loop approve --all-init-loop`
    - During first onboarding, to initialize every eligible local Git repo in the portfolio: `loop portfolio init-loop --all-eligible`
 
 **Day — execute only approved work:**
-1. `loop start-day [project...]` starts approved projects only.
-2. Low-risk work may run unattended after value/verification gates.
-3. Medium-risk work may run unattended only when it matches the morning-approved envelope. The first medium-risk execution runs with `supervised=true`, then continues hourly only if that first cycle does not leave `waiting_for_human`.
-4. Budgets and stop rules come from daily focus: `recommended_cycles`, `stop_condition`, `value_threshold`, `max_noop_cycles`.
+1. Run execution commands from the project owner thread, using `loop-engine/handoffs/latest/<project>.md` as the prompt.
+2. `loop start-day [project...]` starts approved projects only.
+3. Low-risk work may run unattended after value/verification gates.
+4. Medium-risk work may run unattended only when it matches the morning-approved envelope. The first medium-risk execution runs with `supervised=true`, then continues hourly only if that first cycle does not leave `waiting_for_human`.
+5. Budgets and stop rules come from daily focus: `recommended_cycles`, `stop_condition`, `value_threshold`, `max_noop_cycles`.
+6. The owner thread reports back only shipped value, PR/issue links, digest/report path, `waiting_for_human` items, and a continue/stop/move recommendation.
 
 **Evening — stop, recap, score:**
 1. `loop evening [project...]` pauses the named projects. With no project args, it pauses **all active registered loops**, even loops that were started manually outside the approval flow.
@@ -270,6 +288,7 @@ Risk is task-level, not project-level. A trading project is not automatically hi
 Quick orientation for an operating agent:
 
 - **Read state, don't poll commands:** `loop status --json`, or read `loop-engine/state.json` directly. Key fields: `loop_job.state` (`active|paused|stopped`), `current_phase`, `waiting_for_human` (`{reason, issue_path}` or null), `runs[-1].status`.
+- **Keep Secretary/PM context clean:** after `loop morning`, read `loop-engine/handoffs/latest/<project>.md` and move execution to that project owner thread. Report back only product-level outcome and links.
 - **Recap is a file:** read `loop-engine/reports/<project>/latest.md`; don't call `loop digest` in a loop.
 - **Escalate, don't bypass:** anything in `waiting_for_human` needs a human decision (see the reason-code table in [AGENTS.md](AGENTS.md) and [Troubleshooting](#troubleshooting--排错)). Never edit prompts/reviewer output to force a pass.
 
@@ -281,6 +300,7 @@ Quick orientation for an operating agent:
 - [examples/portfolio.example.json](examples/portfolio.example.json) — private portfolio registry shape (auto-written under `~/.config/loop/portfolio.json`)
 - [examples/portfolio-profile.example.json](examples/portfolio-profile.example.json) — CTO intake profile shape (auto-written by `loop portfolio intake`)
 - `loop-engine/portfolio/<project>/profile.{md,json}` — private CTO intake profiles (auto-written by `loop portfolio intake`)
+- `loop-engine/handoffs/<date>/<project>.md` — owner-thread execution prompt generated by `loop morning` / `loop handoff`
 - [examples/contract.example.yaml](examples/contract.example.yaml) — the `.loop/contract.yaml` schema
 - [examples/daily-focus.example.md](examples/daily-focus.example.md) — daily focus + preapproved envelope
 
@@ -364,7 +384,7 @@ Morning/day-start `LOOP_BLOCKED` reasons: `portfolio_missing` means run `loop po
 
 ## Runtime artifacts / 运行产物
 
-Git-ignored by design — they hold local paths, private strategy, and agent transcripts. **Keep them out of public repos:** `loop-engine/runs/`, `reports/`, `registry.json`, `state.json`, `pm-reviews/`, `approvals/`, `evening-scorecards/`, `human-feedback/`, `knowledge/`, `logs/`, `locks/`, `worktrees/`.
+Git-ignored by design — they hold local paths, private strategy, and agent transcripts. **Keep them out of public repos:** `loop-engine/runs/`, `reports/`, `registry.json`, `state.json`, `pm-reviews/`, `approvals/`, `handoffs/`, `evening-scorecards/`, `human-feedback/`, `knowledge/`, `logs/`, `locks/`, `worktrees/`.
 
 ## Repository layout / 仓库结构
 

examples/portfolio.example.json

@@ -11,6 +11,12 @@
         "github_repo": "operator/tokenpulse",
         "linear_project": "TokenPulse",
         "loop_project_id": "tokenpulse"
+      },
+      "owner_thread": {
+        "id": "thread-tokenpulse-owner",
+        "name": "TokenPulse owner thread",
+        "mode": "codex_thread",
+        "handoff_policy": "manual_copy"
       }
     },
     "newsletter": {

loop-engine/bin/loop_facade.py

@@ -18,6 +18,7 @@ def usage(code: int = 2) -> int:
         "  /loop doctor [project]\n"
         "  /loop portfolio init|add|status|intake|init-loop [project|--all-eligible]\n"
         "  /loop morning [project...]\n"
+        "  /loop handoff [project...] [--json]\n"
         "  /loop approve [project] [--approve-medium | --medium-envelope NAME | --init-loop | --all-init-loop]\n"
         "  /loop reject [project]\n"
         "  /loop start-day [project...]\n"
@@ -113,10 +114,10 @@ def main() -> int:
         loopctl_args = [command] + passthrough
         if project:
             loopctl_args += ["--project", project]
-    elif command in {"morning", "start-day", "evening"}:
+    elif command in {"morning", "start-day", "evening", "handoff"}:
         loopctl_args = [command]
         for arg in rest:
-            if arg == "--start":
+            if arg in {"--start", "--json"}:
                 loopctl_args.append(arg)
             else:
                 loopctl_args += ["--project", arg]