git:20260923.d406817 to git:20260923.1734b64

9 added, 9 removed. Audit A to A.

# CLAUDE.md

Entry pointer for Claude Code in kcn's investment workspace. Same workflow as `AGENTS.md` (openclaw entry) — content lives in the canonical files below, this is just the road map.

## Identity & user

- Investment chat: you're `Rick` (see `IDENTITY.md`). Coding sessions in this

repository keep the coding-agent identity; they read Rick's files only when the

task is about that behaviour. (OpenClaw's own claude-cli backend starts Claude

with `--setting-sources user`, so WeChat chat does not load this file.)

- User is `kcn` / Shengyu Li (see `USER.md`)

## Context on demand

- Use already-injected files instead of rereading them. OpenClaw's session profile owns

- runtime bootstrap; Claude Code separately owns native project auto-memory and compaction.

- Do not add a second startup scan or memory pipeline over either mechanism.

+ Session start rules (which bootstrap files are injected, when to read `INVESTMENT_SOP.md`,

+ group contexts never load main-session memory, isolated jobs follow their profile,

+ delegation) live once in `AGENTS.md` § Every Session. Claude-specific:

- - For investment chat, use `SOUL.md`, `USER.md`, `MEMORY.md` and the `TOOLS.md` route already supplied;

- read missing relevant files, then `INVESTMENT_SOP.md` and the selected skill.

- - Load `portfolio.json` only when the question needs positions or prices.

- - Coding work and ordinary delegation do not require investment persona, SOP or holdings reads.

- Follow coding-agent guidance and the task's scope; a delegate performs its own investigation.

- - Group/shared contexts must not load main-session memory. Isolated jobs follow their own payload/profile.

+ - OpenClaw's session profile owns runtime bootstrap; Claude Code separately owns native

+ project auto-memory and compaction. Do not add a second startup scan or memory pipeline

+ over either mechanism.

+ - Coding work and ordinary delegation need no persona, SOP or holdings reads; skill and

+ tool routing is in `TOOLS.md`.

+ - For an investment question, load `portfolio.json` only when it needs positions or prices.

## What lives where

| Topic | File |

|---|---|

| Disposition / persona | `SOUL.md`, `IDENTITY.md` |

| User profile / preferences | `USER.md` |

| Iron rules / traps | `MEMORY.md` |

| Scripts / fallback / skill routing / cron | `TOOLS.md` |

| Skill bodies | `skills/{name}/SKILL.md` |

| Portable workflows / market / portfolio tools | installed `clawock` CLI; source in `src/clawock/` |

| Host / publish / CI / growth wiring | `ops/{host,publish,ci,growth}/` |

| Repository automation | installed `clawock-*` commands and named `ops/*` entry points |

| Daily deep brief output | `memory/{date}-pre-open.md` + `memory/{date}-plan.json` |

| Daily portfolio snapshots | `memory/snapshots/{date}.json` |

| Preflight context (gitignored) | `memory/.tmp/` |

| Startup sequence | `INVESTMENT_SOP.md` |

| Heartbeat workflow | `HEARTBEAT.md` (heartbeat poll only) |

| Auto-commit rules | `AGENTS.md` |

| Interactive code PR/worktree rules | `AGENTS.md` § Interactive Codex/Claude PR workflow |

| Getting a merged fix onto this host (no release needed) | `ops/host/refresh_live.sh`; rule in `docs/operations/release.md` § Running the latest code on this host |

| Pages dashboard input | data-plane generation refreshed by KCNyu postflight and `ops/publish/publish_dashboard.sh` |

| Risk metrics snapshot | `assets/data/risk.json` (built by `clawock portfolio-risk`, refreshed daily via brief preflight) |

| Decision execution marking | `memory/decisions.jsonl` `execution.status`; manual override via `clawock mark-followed DECISION_ID [--no]` |

| Pages source | `site/index.html` (dashboard) + `site/briefs.md` (daily briefs index) |

## Cron run loop (what openclaw fires)

Each cron job's prompt tells you which harness 4-step to execute:

```

Step 1 preflight clawock {brief|report|intraday} preflight [args]

→ writes memory/.tmp/{brief|report|intraday}-context-*.json

Step 2 read context inside that .json: raw_wechat_block, anomalies, title, signals, etc.

Step 3 LLM synthesis you write the analysis prose + (brief only) plan.json

following the SKILL.md Mode template. Mode 6 reports: prose ONLY —

report_postflight prepends title + raw_wechat_block itself

Step 4 postflight clawock {brief|report|intraday} postflight [args]

validates report, computes wechat_prefix, then handles its scoped publish

(all three refresh the complete dashboard generation; intraday publishes only semantic diffs)

```

Don't ask permission for internal reads/edits. Action over confirmation.