diff --git a/agentd/cli.py b/agentd/cli.py index 42d120c..90784ec 100644 --- a/agentd/cli.py +++ b/agentd/cli.py @@ -121,15 +121,17 @@ def cmd_smoke(cfg, a): body["labels"] = [labels[x] for x in a.labels.split(",")] issue = pl.create_issue(p["id"], body) key = "%s-%s" % (p.get("key_prefix", a.project), issue["sequence_id"]) - print("created %s; ticking every %ss (timeout %s min)" % (key, a.interval, a.timeout)) - d = Agentd(cfg) + print("created %s; %s every %ss (timeout %s min)" % ( + key, "watching (the scheduled dispatcher runs it)" if a.watch else "ticking", a.interval, a.timeout), flush=True) + d = None if a.watch else Agentd(cfg) states = pl.states(p["id"]) done_states = {states[p["states"][k]] for k in ("review", "waiting")} deadline = time.time() + a.timeout * 60 state = None while time.time() < deadline: try: - d.tick(project=a.project, only_issue=issue["id"]) + if d: + d.tick(project=a.project, only_issue=issue["id"]) except Busy as exc: print("skip: %s" % exc) state = pl.get_issue(p["id"], issue["id"]).get("state") @@ -199,6 +201,8 @@ def parser(): s.add_argument("--interval", type=int, default=30) s.add_argument("--timeout", type=int, default=30, help="minutes") s.add_argument("--labels", default="", help="comma-separated Plane labels, e.g. model:sol") + s.add_argument("--watch", action="store_true", + help="do not tick here; only wait for the scheduled dispatcher (use when it runs on another host)") s.add_argument("--as", dest="as_") s.set_defaults(fn=cmd_smoke) sub.add_parser("check-config", help="validate the config and list token env vars").set_defaults(fn=cmd_check_config) diff --git a/agentd/config.py b/agentd/config.py index e5acff0..052c3af 100644 --- a/agentd/config.py +++ b/agentd/config.py @@ -155,8 +155,18 @@ class Config: return p if p.is_absolute() else ROOT / p def helpers_dir(self, terminal): - d = self.terminals[terminal].get("helpers_dir") or self.agentd.get("helpers_dir") or str(ROOT / "tools") - return Path(os.path.expanduser(d)) + d = self.terminals[terminal].get("helpers_dir") or self.agentd.get("helpers_dir") or "tools/win10" + p = Path(os.path.expanduser(d)) + return p if p.is_absolute() else ROOT / p + + @property + def log_dir(self): + """agentd.log lives here (default: state_dir).""" + d = self.agentd.get("log_dir") + if not d: + return self.state_dir + p = Path(os.path.expanduser(d)) + return p if p.is_absolute() else ROOT / p def token_env(self, terminal): """Environment variable that carries the terminal's Plane token (== the Vault key name).""" diff --git a/agentd/store.py b/agentd/store.py index 7037937..289996f 100644 --- a/agentd/store.py +++ b/agentd/store.py @@ -17,8 +17,9 @@ class Busy(Exception): class Store: - def __init__(self, state_dir, echo=True): + def __init__(self, state_dir, echo=True, log_dir=None): self.dir = Path(state_dir) + self.log_dir = Path(log_dir) if log_dir else self.dir self.echo = echo self.path = self.dir / "agentd-state.json" @@ -59,9 +60,9 @@ class Store: self._atomic_write(self.dir / "heartbeat.json", info) def log(self, msg): - self.dir.mkdir(parents=True, exist_ok=True) + self.log_dir.mkdir(parents=True, exist_ok=True) line = time.strftime("%Y-%m-%dT%H:%M:%S") + " " + msg if self.echo: print(line, flush=True) - with open(self.dir / "agentd.log", "a", encoding="utf-8") as fh: + with open(self.log_dir / "agentd.log", "a", encoding="utf-8") as fh: fh.write(line + "\n") diff --git a/config/agentd.example.yaml b/config/agentd.example.yaml index 0c94d51..4853536 100644 --- a/config/agentd.example.yaml +++ b/config/agentd.example.yaml @@ -7,7 +7,7 @@ vault: {client: ~/.local/share/server-access-client/agent-access.py, config: htt agentd: state_dir: state # relative to the repo root; JSON state, logs, heartbeat, local run dirs - helpers_dir: /home/ubuntu/Documents/wjj/project/gwent-design/tools # win10-ps.sh / win10-put.sh (Windows terminals only) + helpers_dir: tools/win10 # win10-ps.sh / win10-put.sh (Windows terminals only); relative = repo interval: 120 # seconds between ticks in --loop mode max_rounds_per_thread: 8 # rebuild the executor thread (with a handoff summary) after this many rounds comment_max_chars: 6000 diff --git a/config/agentd.oregon.yaml b/config/agentd.oregon.yaml new file mode 100644 index 0000000..5be331c --- /dev/null +++ b/config/agentd.oregon.yaml @@ -0,0 +1,89 @@ +# plane-agentd dispatcher configuration for oregon -- the ROLLBACK target since the dispatcher moved to the LXC +# agentd-01 (PAD-11; live config there: config/agentd.yaml). Used by cron/agentd-tick.sh (not in oregon's crontab +# while the LXC dispatches). No secrets: token_key values are Vault key names. +plane: {base_url: https://plane.kappiee.com, workspace: kappiee} +vault: {client: ~/.local/share/server-access-client/agent-access.py, config: https://nas.tail9531b.ts.net:8443/identity/agent.json} + +agentd: + state_dir: ~/.gwent-board/agentd # JSON state, agentd.log, heartbeat.json, local run dirs (runs/) + helpers_dir: tools/win10 # win10-ps.sh / win10-put.sh (Windows terminals only); relative = repo + interval: 120 + max_rounds_per_thread: 8 + comment_max_chars: 6000 + cli_terminal: oregon + +terminals: + win10: + transport: warpgate-ssh + target: win10 + os: windows + launch: schtasks + repo: 'C:\Users\Administrator\Documents\GwentAssistantProject\gwent' + run_root: 'C:\Users\Administrator\Documents\GwentAssistantProject\gwent\runtime\agentd' + bins: {codex: 'C:\Users\Administrator\AppData\Local\Programs\OpenAI\Codex\bin\codex.exe'} + plane_user: window10 + token_key: PLANE_WINDOW10_API_TOKEN + concurrency: 1 + busy_processes: [codex] # do not start while a codex.exe agentd did not start is running + capabilities: [android, game] + executors: [codex] + oregon: + # remote even though agentd runs on oregon today: the dispatcher is moving to an LXC without agent CLIs + # (design §9), so this goes through Warpgate like any other host. `transport: local` + `launch: subprocess` + # still works for a dispatcher co-located with its executors. + transport: warpgate-ssh + target: oregon # agent-access.py target name + os: linux + launch: nohup + repo: /home/ubuntu/Documents/wjj/project/gwent + run_root: /home/ubuntu/.gwent-board/agentd-runs # absolute: the remote shell does not expand ~ inside quotes + plane_user: oregon + token_key: PLANE_OREGON_API_TOKEN + concurrency: 1 # one shared worktree: no parallel edits + bins: {claude: /home/ubuntu/.local/bin/claude} # the SSH non-login shell has the system PATH only; no codex on oregon + executors: [claude-code] + +executors: + codex: + adapter: codex + run: "codex exec --json -C '{repo}' -s workspace-write -m {model} -c 'model_reasoning_effort=\"{effort}\"' {tier} -o '{out}' -" + # `exec resume` takes no -C/-s: the runner already cd's into the repo, the sandbox goes in as a -c override + resume: "codex exec resume {thread} --json -m {model} -c 'model_reasoning_effort=\"{effort}\"' -c 'sandbox_mode=\"workspace-write\"' {tier} -o '{out}' -" + tier_arg: "-c 'service_tier=\"{tier}\"'" + tier_defaults: {gpt-6.1-sol: fast} # tier when no label sets one (`tier:`, `tier:none` = no tier) + thread_from: events.jsonl:thread.started.thread_id + last_message: '{out}' + models: [gpt-6.1-sol, gpt-6-astra] + default_effort: medium + claude-code: + adapter: claude-code + run: "claude -p --model {model} --output-format json --permission-mode {perm}" + resume: "claude -p --resume {thread} --model {model} --output-format json --permission-mode {perm}" + permission_mode: acceptEdits + thread_from: json:session_id + last_message: json:result + models: [sonnet, opus, haiku] + +projects: + GWENT: + id: 79cb5a60-aa4f-4138-9864-1dd70426e665 + key_prefix: GW + states: {backlog: Backlog, todo: Todo, running: In Progress, waiting: Waiting, review: Review, done: Done} + reviewer: oregon + answerer: oregon + protocol_file: config/protocols/gwent-board-protocol.md # relative = repo; refresh from gwent/design/board-protocol.md + routing: + by_assignee: {window10: win10, oregon: oregon} + by_label: + 'model:astra': {model: gpt-6-astra, effort: high} + 'model:sol': {model: gpt-6.1-sol, tier: fast} + 'effort:high': {effort: high} + 'needs-game': {serialize_group: game} + default_executor: {win10: codex, oregon: claude-code} + retry: + transient_max: 3 + backoff_minutes: [2, 5, 10] + timeout_hours: 4 + max_rounds_per_thread: 8 + max_progress_rounds: 12 + # notify (ntfy) is phase 3: left out on purpose diff --git a/config/agentd.yaml b/config/agentd.yaml index 69f6442..4a417ec 100644 --- a/config/agentd.yaml +++ b/config/agentd.yaml @@ -1,11 +1,13 @@ -# plane-agentd live configuration on oregon (phase 1). No secrets: token_key values are Vault key names; the tokens -# are injected by `agent-access.py exec --secret-env KEY=KEY` (see cron/agentd-tick.sh and docs/runbook.md). +# plane-agentd live configuration on the LXC agentd-01 (/opt/plane-agentd, runs as root; docs/deploy-lxc.md). +# No secrets: token_key values are Vault key names; the tokens are injected by +# `agent-access.py exec --secret-env KEY=KEY` (cron/agentd-tick-lxc.sh). Rollback to oregon: config/agentd.oregon.yaml. plane: {base_url: https://plane.kappiee.com, workspace: kappiee} vault: {client: ~/.local/share/server-access-client/agent-access.py, config: https://nas.tail9531b.ts.net:8443/identity/agent.json} agentd: - state_dir: ~/.gwent-board/agentd # JSON state, agentd.log, heartbeat.json, local run dirs (runs/) - helpers_dir: /home/ubuntu/Documents/wjj/project/gwent-design/tools # win10-ps.sh / win10-put.sh (Windows terminals only) + state_dir: /var/lib/plane-agentd # JSON state, heartbeat.json, lock + log_dir: /var/log/plane-agentd # agentd.log (cron output: cron.log, same dir) + helpers_dir: tools/win10 # win10-ps.sh / win10-put.sh (Windows terminals only); relative = repo interval: 120 max_rounds_per_thread: 8 comment_max_chars: 6000 @@ -27,9 +29,7 @@ terminals: capabilities: [android, game] executors: [codex] oregon: - # remote even though agentd runs on oregon today: the dispatcher is moving to an LXC without agent CLIs - # (design §9), so this goes through Warpgate like any other host. `transport: local` + `launch: subprocess` - # still works for a dispatcher co-located with its executors. + # the dispatcher (LXC, no agent CLIs) reaches oregon through Warpgate like any other host (design §9) transport: warpgate-ssh target: oregon # agent-access.py target name os: linux @@ -70,7 +70,7 @@ projects: states: {backlog: Backlog, todo: Todo, running: In Progress, waiting: Waiting, review: Review, done: Done} reviewer: oregon answerer: oregon - protocol_file: /home/ubuntu/Documents/wjj/project/gwent/design/board-protocol.md + protocol_file: config/protocols/gwent-board-protocol.md # relative = repo; refresh from gwent/design/board-protocol.md routing: by_assignee: {window10: win10, oregon: oregon} by_label: diff --git a/config/protocols/gwent-board-protocol.md b/config/protocols/gwent-board-protocol.md new file mode 100644 index 0000000..b9635d2 --- /dev/null +++ b/config/protocols/gwent-board-protocol.md @@ -0,0 +1,47 @@ +# 看板协作协议(Plane 作为任务板与信箱) + +看板 = https://plane.kappiee.com/kappiee/projects/79cb5a60-aa4f-4138-9864-1dd70426e665/issues/ ,工作项编号 `GW-k`。Gitea 只管代码(仓库、提交、PR),不再放任务。 + +## 账号 = 终端 +- `kappiee` 用户(所有者,做决定);`oregon` Claude(设计、派活、验收);`window10` win10 Codex(执行);`mac` 之后接入的 Mac 终端。 +- 凭据键(Agent Vault 注入内存,不落盘、不进 win10、不进评论):`PLANE_OREGON_API_TOKEN`、`PLANE_WINDOW10_API_TOKEN`(worker 以 window10 发帖)。API 头 `X-API-Key`。 + +## 状态机 +`Backlog → Todo → In Progress → Waiting → Review → Done / Cancelled` +- Backlog:想法/未排期。Todo + assignee `window10`:交给 Codex,worker 下一个 tick 取走。 +- In Progress:worker 已启动 codex 运行(worker 自动设置)。 +- Waiting + assignee `oregon`:Codex 有待回答的问题或被阻塞(`[question]`/`[blocked]`),等 Claude 回答或升级给 kappiee。 +- Review + assignee `oregon`:产出就绪(`[review]`),Claude 验收。通过 → Done;不通过 → 写意见,改回 Todo + `window10`。 +- `[progress]`:worker 自动改回 Todo + `window10`,下一轮续跑。 +- 需要用户决定时 Claude 把 assignee 改为 `kappiee`(state 保持 Waiting)。 + +## 工作项 +- 标题 `T-xxx 简述`;描述 = 任务书(目标、允许的工具、不做的事、产出、验收线)。 +- 评论 = 消息,作者由账号区分;正文首行保留角色标记 `[claude]` / `[codex]` / `[worker]`,worker 构造 prompt 时跳过 `[worker]` 评论。 +- 结果标签(Codex 最后一条消息首行必须是其一): + - `[review]` 产出完成,附:改动落点、产出路径、SHA、自检结果、如何重跑。 + - `[question]` 需要回答才能继续,附"若无人回答,我默认会怎么做"。 + - `[blocked]` 无法继续,附原因与已尝试的办法。 + - `[progress]` 做了一部分,需要下一轮继续。 + - 未标注首行按 `[review]` 处理。 + +## 标签、模块、关系、视图 +- 标签:`model:astra`(gpt-6-astra,high)/ `model:sol`(gpt-6.1-sol fast,medium,默认)/ `effort:high` / `needs-game`(需要游戏在线,别与其他占用游戏的任务并行)。 +- 模块(Modules)= 任务模块(T-007-0 …);一个模块完成后 Claude 亲自验收再关闭。 +- 依赖用关系 "blocked by",被阻塞的项留在 Backlog,前置完成后再转 Todo。 +- 视图:`等我决定`(assignee=kappiee 或 Waiting)、`Codex 执行中`(In Progress / Todo + window10)、`待验收`(Review)。 + +## Codex 侧(worker 自动化) +worker 跑在 Claude 所在的 Linux 机(`gwent-design/tools/board-worker-plane.py`,cron 每 2 分钟一次,token 只由 Vault 注入内存)。 +1. 取 state=Todo 且 assignee=`window10` 的工作项,按 `GW-k` 从小到大取一个;win10 上已有 codex 进程则本轮跳过(避免抢游戏/ADB)。 +2. 名称 + 描述 + 评论转成文本拼成 prompt 跑 `codex exec`;同一工作项第二次起 `codex exec resume ` 只追加新评论。 +3. 结束后把最后一条消息贴回评论,按结果标签改状态与 assignee(见状态机)。 + +## Claude 侧 +- 工具:`gwent-design/tools/plane-claude.sh`(`show` / `comment` / `state` / `assign` / `new` / `done`)。 +- 轮询 Waiting 与 Review(assignee=`oregon`):回答问题、验收、改任务书或升级给 kappiee。新任务一律建工作项,不再用临时 brief 文件(紧急修复除外)。 + +## 通用规则 +- 消息要短、要有结论;路径用仓库相对路径;大图只贴缩略拼图(≤1600px)。 +- 不在评论里贴 token、密码、完整日志。 +- 有疑问优先按任务书里的默认值继续并在评论里说明,而不是停下等人。 diff --git a/config/schema.json b/config/schema.json index f08e047..8b912d6 100644 --- a/config/schema.json +++ b/config/schema.json @@ -28,6 +28,7 @@ "additionalProperties": false, "properties": { "state_dir": {"type": "string"}, + "log_dir": {"type": "string"}, "helpers_dir": {"type": "string"}, "interval": {"type": "integer", "minimum": 10}, "max_rounds_per_thread": {"type": "integer", "minimum": 1}, diff --git a/cron/agentd-tick-lxc.sh b/cron/agentd-tick-lxc.sh new file mode 100755 index 0000000..20bac7d --- /dev/null +++ b/cron/agentd-tick-lxc.sh @@ -0,0 +1,11 @@ +#!/usr/bin/env bash +# LXC agentd-01 cron entry (*/2, root): one agentd tick under server-access exec (Plane tokens only in the +# process env), with a lock. The outer flock skips a tick while the previous one still runs; agentd also holds +# its own state lock. Logs: /var/log/plane-agentd/{agentd.log,cron.log}; state: /var/lib/plane-agentd. +exec flock -n /var/lib/plane-agentd/cron.lock \ + timeout 900 python3 -I /root/.local/share/server-access-client/agent-access.py \ + --config https://nas.tail9531b.ts.net:8443/identity/agent.json exec \ + --secret-env PLANE_OREGON_API_TOKEN=PLANE_OREGON_API_TOKEN \ + --secret-env PLANE_WINDOW10_API_TOKEN=PLANE_WINDOW10_API_TOKEN --timeout 850 -- \ + python3 /opt/plane-agentd/cli/agentd --config /opt/plane-agentd/config/agentd.yaml \ + tick --once >> /var/log/plane-agentd/cron.log 2>&1 diff --git a/cron/agentd-tick.sh b/cron/agentd-tick.sh index e79e6be..0c384a8 100755 --- a/cron/agentd-tick.sh +++ b/cron/agentd-tick.sh @@ -1,9 +1,12 @@ #!/usr/bin/env bash -# cron entry (*/2): one agentd tick under server-access exec (Plane tokens in the process env only), with a lock. -# The outer flock skips a tick while the previous one still runs; agentd also holds its own state lock. +# oregon cron entry (*/2) -- ROLLBACK ONLY since PAD-11: the dispatcher runs on the LXC agentd-01 +# (cron/agentd-tick-lxc.sh). Never run both: two dispatchers double-dispatch. See docs/runbook.md "回滚到 oregon". +# One agentd tick under server-access exec (Plane tokens in the process env only), with a lock. exec flock -n /home/ubuntu/.gwent-board/agentd.lock \ timeout 900 python3 -I /home/ubuntu/.local/share/server-access-client/agent-access.py \ --config https://nas.tail9531b.ts.net:8443/identity/agent.json exec \ --secret-env PLANE_OREGON_API_TOKEN=PLANE_OREGON_API_TOKEN \ --secret-env PLANE_WINDOW10_API_TOKEN=PLANE_WINDOW10_API_TOKEN --timeout 850 -- \ - python3 /home/ubuntu/Documents/wjj/project/plane-agentd/cli/agentd tick --once >> /home/ubuntu/.gwent-board/agentd-cron.log 2>&1 + python3 /home/ubuntu/Documents/wjj/project/plane-agentd/cli/agentd \ + --config /home/ubuntu/Documents/wjj/project/plane-agentd/config/agentd.oregon.yaml \ + tick --once >> /home/ubuntu/.gwent-board/agentd-cron.log 2>&1 diff --git a/tools/win10/win10-ps.sh b/tools/win10/win10-ps.sh new file mode 100755 index 0000000..707a6c2 --- /dev/null +++ b/tools/win10/win10-ps.sh @@ -0,0 +1,18 @@ +#!/usr/bin/env bash +# Run a PowerShell script (passed as $1 text, or a file with -f) on win10 through server-access, +# encoded as -EncodedCommand so quoting never crosses the relay. Output is UTF-8; progress noise off. +# AGENT_ACCESS_HOME = directory of the vault client (default ~/.local/share/server-access-client); +# AGENT_ACCESS_CONFIG = identity URL/file. +set -euo pipefail +CLIENT="${AGENT_ACCESS_HOME:-$HOME/.local/share/server-access-client}/agent-access.py" +CFG="${AGENT_ACCESS_CONFIG:-https://nas.tail9531b.ts.net:8443/identity/agent.json}" +TIMEOUT="${WIN10_TIMEOUT:-120}" +if [ "${1:-}" = "-f" ]; then SCRIPT="$(cat "$2")"; else SCRIPT="$1"; fi +SCRIPT="[Console]::OutputEncoding=[Text.Encoding]::UTF8 +\$ErrorActionPreference='Continue' +\$ProgressPreference='SilentlyContinue' +$SCRIPT" +B64="$(printf '%s' "$SCRIPT" | iconv -f UTF-8 -t UTF-16LE | base64 -w0)" +if [ "${#B64}" -gt 7600 ]; then echo "encoded command too long (${#B64} > 7600)" >&2; exit 2; fi +exec python3 -I "$CLIENT" --config "$CFG" run --target win10 --timeout "$TIMEOUT" -- \ + powershell.exe -NoProfile -NonInteractive -EncodedCommand "$B64" 2> >(grep -v -E '^#< CLIXML|^&2) diff --git a/tools/win10/win10-put.sh b/tools/win10/win10-put.sh new file mode 100755 index 0000000..6ada438 --- /dev/null +++ b/tools/win10/win10-put.sh @@ -0,0 +1,26 @@ +#!/usr/bin/env bash +# Upload a local text/binary file to win10: gzip+base64, appended in chunks through the relay, +# then decoded remotely and verified by SHA-256. +# usage: win10-put.sh +set -euo pipefail +HERE="$(cd "$(dirname "$0")" && pwd)" +LOCAL="$1"; REMOTE="$2" +TMP="$REMOTE.b64" +SHA="$(sha256sum "$LOCAL" | cut -d' ' -f1)" +B64="$(gzip -9c "$LOCAL" | base64 -w0)" +CHUNK=2400 +"$HERE/win10-ps.sh" "New-Item -ItemType Directory -Force -Path (Split-Path -Parent '$REMOTE') | Out-Null; Set-Content -Path '$TMP' -Value '' -NoNewline -Encoding Ascii; 'reset'" >/dev/null +TOTAL=$(( (${#B64} + CHUNK - 1) / CHUNK )) +for ((i=0; i<${#B64}; i+=CHUNK)); do + PART="${B64:i:CHUNK}" + n=$(( i / CHUNK + 1 )) + "$HERE/win10-ps.sh" "Add-Content -Path '$TMP' -Value '$PART' -NoNewline -Encoding Ascii; 'chunk $n/$TOTAL'" | tail -1 +done +"$HERE/win10-ps.sh" "\$b=[Convert]::FromBase64String((Get-Content -Raw -Encoding Ascii '$TMP')) +\$ms=New-Object IO.MemoryStream(,\$b) +\$gz=New-Object IO.Compression.GZipStream(\$ms,[IO.Compression.CompressionMode]::Decompress) +\$out=New-Object IO.MemoryStream; \$gz.CopyTo(\$out) +[IO.File]::WriteAllBytes('$REMOTE',\$out.ToArray()) +Remove-Item '$TMP' +'remote sha256 ' + (Get-FileHash '$REMOTE' -Algorithm SHA256).Hash.ToLower() +'expected $SHA'"