Repository navigation
Releases
v0.16.0 — 部署守卫:线上漂移检测/拉回/回执,杜绝 upload 静默覆盖
[0.16.0] - 2026-11-11
Added — deploy guard(线上直改未同步 → 上传静默覆盖,T-0102 事故)
tower deploy check|init|pull|done|status <local> --remote <path>:对
「本地 ↔ 远端」维护 baseline,三方比较:
remote == local= in_sync;remote == baseline= local_changed(可上传);
否则 remote_drift → 拒绝上传,必须先pull或人工合并。pull先把本地旧版备份到.tower/deploy/backup/,再写入线上内容并推进基线;
done在上传后强制验证「远端 == 本地」,不成立即报 "upload did not land"。- 回执:
.tower/deploy/receipts/<ts>-<key>-<action>.md;--role <role>时追加
roles/<role>/DEPLOY-SYNC.md;全部动作进 ledger(deploy_init/check/pull/done)。 control_tower.deploy_fetch:python -m control_tower.deploy_fetch --cwd <theme-dir> --remote <path>通过项目自带w.py的sync_request read_file取远端内容,对任何
w.py 站点通用;tower.toml 里deploy.fetch配置一次(支持{remote}{local}
{workspace}{project_root}{python}占位符)。- 部署纪律随派单注入:
docs/deploy-rules.md会进入每次 dispatch prompt 的
## Deploy rules段。 - 生产实测:evergocasa
functions.php已纳管(baseline 78c69e1e…),check返回
in_sync,回执与roles/frontend/DEPLOY-SYNC.md已生成。
v0.15.0 — CLI 不再静默回落到全局活跃指针;写操作打印 resolved workspace
[0.15.0] - 2026-11-10
Breaking — the CLI no longer falls back to the global active-project pointer
With three towers alive (evergocasa / prl-ops / offroadleagues), running a command
in a directory that has no tower.toml silently resolved against the active
project: tower say ops deputy "…" answered unknown role 'deputy' (its roles
came from another tower), and the same path would have misrouted dispatch /
accept / relay. Field report 2026-11-10.
- Resolution is now exactly
--workspace>tower.tomlfound walking up from
cwd > a loud error.~/.tower/binding.jsonand thetower projects use
pointer are no longer consulted by the CLI; they exist for the GUI panel and
plugin reconnection only. - Every write command announces its target:
resolved workspace = <path>on
stderr (stdout stays a single JSON document under--json). Covered:
dispatch,accept,rework,relay,say,escalate,batch,restate,
review,attention claim|resolve. - READMEs document the rule and the pointer's scope; regression tests pin that a
registered+active tower elsewhere is never touched from a foreign cwd.
v0.14.1 — 修复 spawn 后误报「无 worker」(等心跳再判死)
[0.14.1] - 2026-11-10
Fixed — false "no worker" alarm right after a spawn
ensure_worker returns the instant the detached process is created, but the
heartbeat lands 1–3 s later (cold Python start). The consumer check added in
0.14.0 trusted the spawn result and could report "worker could not be started"
while the worker was in fact booting (reproduced in a parallel-dispatch smoke
test: 3 tasks queued, worker pid alive seconds later).
- One
_ensure_consumer()path now waits up to 10 s for the heartbeat whenever
a worker is expected, and only then fails (exit 1). Single dispatch, parallel
dispatch andtower reviewall use it. - Tests that fake a spawn now fake
worker_alivetoo (a fake consumer must
answer both questions).
v0.14.0 — 并行派单 + 队列守护(worker 自动拉起/状态/积压告警)
[0.14.0] - 2026-11-10
Added — parallel dispatch (a bot can hold several jobs) + queue supervision
tower dispatch <role> --brief-file A --brief-file B …queues one task per
brief and returns immediately — no more waiting for brief #1's 30-minute adapter
turn before brief #2 is even submitted. Also--parallelfor a single
no-wait submit. Each task runs async; the bot's own session may serialise them.tower worker --statusprints the consumer state for the workspace:
alive/pid/version/heartbeat age, queue depths, oldest pending age, backlog alarm.- enqueue auto-starts a worker for its own workspace — the 2026-11-10
incident: 12 adversarial reviews sat 33 hours because the only worker belonged
to another project and accept-time enqueue never ensured a consumer. - Backlog alarm:
tower status/tower inboxprint a ⚠ line when pending
work has no worker, or the oldest pending job exceedsqueue_alert_minutes
(default 15);tower doctorreports it as a problem. dispatch --async/tower reviewfail loudly (exit 1) when no consumer is
available after the spawn attempt, instead of silently queueing forever.
Fixed — Files section separator ·
acceptverification splits multi-path lines on·/•/tab as well, so
a.json·b.jsonis two paths, not one nonexistent one (T-0062 round-1 case).
v0.13.1 — relay 契约与 CLI 对齐:--to 别名 + 位置参数
[0.13.1] - 2026-11-09
Fixed — relay contract taught a flag the CLI did not accept (B17)
RELAY_CONTRACT (the text delivered to every relayed session) and the module
docstring said tower relay --to <target> -m "...", while the CLI only parsed a
positional target. A peer following the contract got unrecognized arguments: --to and never answered, so --await-receipt waits could hang for nothing.
- The CLI now accepts both:
tower relay <target> -m "…"(canonical) and
tower relay --to <target> -m "…"(alias). A missing target is a clear exit-2. - The contract and docstring now teach the positional form (with
--tonamed as
an accepted alias), so either style a peer copies works. - Regression test covers both forms and the contract wording.
v0.13.0 — B11–B16 字段修复:待总控失真、triage 长等、JSON 管道、多路径/报告缓存/self-review 兼容
[0.13.0] - 2026-11-09
Fixed — prl-ops field batch B11–B16
- B11 「待总控」虚高: queue tasks older than
stale_hours(default 12) are
now 陈旧 — excluded fromneeds_lead()/「待总控」, counted asstaleand listed
separately (/api/snapshot.stale, HQ line 「陈旧 N(不在待总控内)」). Terminal
tasks (accepted,status/state = dropped) appear in no queue at all. Field
check on prl-tower: 待总控 11 → 6, 陈旧 0 (15 hand-dropped tasks closed). - B12
tower triagehung on stale/failed re-dispatches: re-dispatching is now
opt-in (tower triage --retry); the default pass only verifies/accepts. A retry
turn is capped by--budgetseconds (default 300) instead of the role's full
timeout, so one adapter cannot hang the lead for 10+ minutes. - B13
tower status --jsonbroke pipes (Invalid \escape): JSON output is now
pure ASCII (ensure_ascii=True), so PowerShell 5.1's cp936 decoding of native
output can no longer corrupt it. Verified: 1.3 MB output, all bytes < 128. - B14 multi-path Files lines: one line listing
a.json、b.jsonis split on
、 , ; , ; |and every segment is checked (brace patterns stay intact). - B15
tower acceptused the cached report:verify_reportnow prefers the
freshest<task>-REPORT.mdon disk over.tower/tasks/T-….json's cached parse,
so edits to the Markdown take effect on accept. - B16 old reports rejected for missing
## Self-review: compatibility via
self_review_since(reports created before it skip the gate) plus
tower accept --legacyfor a forced one-off override.
v0.12.1 — 交付即 claim:自动收口当场离开「需要你」
[0.12.1] - 2026-11-09
Fixed — delivery to the issues bot IS the claim
The moment a blocked/failed task is auto-delivered to escalation.role, the
attention is claimed on the issues bot's behalf (claimed_by=issues), so the task
leaves 「需要你」 immediately and counts as 处理中 — it no longer waits for the
bot's next turn to claim it. Verified end-to-end with the real CLI: exhausted
task → auto-escalated → attention 0 · attention_claimed 1 → issues ticket
dispatched and accepted.
v0.12.0 — 问题同步:issues bot 自动收口 + attention 生命周期 + 秘书链路
[0.12.0] - 2026-11-09
Added — 问题同步:issues bot + 自动收口 + 秘书链路(John 高优先级)
看板上「需要你 5」的任务其实早已被后续工作取代/属外部额度,但没人收口,一直挂到
John 亲自来问。这一版把收口变成塔台自动行为:
-
自动投递(关键):
dispatch/worker 结束若state=blocked或status=failed
(重试用尽),塔台把该任务作为问题工单投给escalation.role(默认issues),
而不是留在「需要你」等人。同一任务幂等(task.escalation记录 delivered/issues_task/
attempts),工单未关闭前不重复投递。escalation: enabled: true role: issues max_attempts: 1
-
attention 生命周期:
open → claimed(by issues bot) → resolved(reason)。
tower attention claim <task> --by issues --note …后任务进「处理中」,不再计入
「需要你」;tower attention resolve <task> --as dropped|rework|accepted --note …
落终态并清 attention。已 resolve 的任务,triage 不会再抛回 blocked。 -
秘书链路可审计:
tower escalate <task> --to deputy|planner|… --note …生成承接
任务并记录escalation.chain(issues → deputy → planner),原任务 ledger/events 留痕。 -
面板同步:
/api/snapshot新增attention_claimed/attention_resolved;
HQ 顶部状态行显示「处理中 N · 已处理 M」,与needs_person()实时一致。 -
工单模板(kind: tower-bug|exec|infra|policy,severity,回写要求)随投递 brief 一起
下发;链路与判断标准见各塔docs/issues-flow.md。
Wired
- evergocasa 塔:新增
issues角色(Evergo 问题处理,session 已接)、
roles/issues/AGENT.md(来自_charters/issues.md)、escalation配置;
tower doctor= problems: none。
v0.11.0 — B1/B2/B3/B9/B10 字段修复:worker 配置哈希、测试不泄漏、unstick、inbox 容错、--brief-file
[0.11.0] - 2026-11-09
Fixed — the prl-ops field batch (B1–B3, B9, B10)
- B1 — a long-running worker dispatched with a stale tower.toml (swapped role
session ids, tasks still went to the old sessions). Two layers: the worker
re-parses config per job (0.10.4) and the heartbeat now carries a config
hash —ensure_workerretires a worker whose tower.toml changed and spawns a
fresh one. - B2 — pytest leaked detached workers (44 immortal
pytest-of-Administrator
workers at full CPU). The autousenever_spawn_workerfixture blocks real
spawns in every test; workers also exit when their workspace disappears. - B3 — tasks stuck in
claimed/in_progresswith no live worker had no
escape hatch. Newtower unstick <task> [--as pending|dropped](pending
re-queues, dropped closes as failed; ledgered), plus a⧗stalemarker on
tower statuswhen a claim sits longer thanstaleness_minutes(default 30). - B9 —
tower inboxcrashed witherror: 'role'on entries that predate
the role field: the printer now defaults missing ts/task/role and falls back
to a top-levelsummary. - B10 — long briefs fought PowerShell quoting.
tower dispatch --brief-file <path>(UTF-8) —'<'/'>'no longer get eaten as redirects.
Noted — DSH-side gaps (B7/B8, documented, not tower bugs)
session/create accepting only cwd leaves sessions ungrouped and
workspace/move-invalid has no workspace/accountSession escape; empty-arg
session/create returns ok and creates a stray session; no session/archive/
session/delete/workspace/list. Workarounds recorded in
docs/bug-report-protocol.md.
v0.10.9 — 修复 tower_* 中文乱码:任何代码页下输出强制 UTF-8
[0.10.9] - 2026-11-09
Fixed — tower_* 中文乱码(plugin channel forced cp936 → UTF-8 decode)
John's field report: every tower_status / tower_inbox / tower_roles call
from the plugin returned Chinese briefs as U+FFFD mush (G:\新网站项目 →
G:\??????), because the spawned tower CLI encoded stdout with the Windows
code page (cp936) while the plugin decodes UTF-8.
- CLI entry guard (
_force_utf8_stdio): piped stdout/stderr are
reconfigured to UTF-8 on every run —tower.exe,python -m, the plugin
channel, scripts; interactive consoles (isatty) keep their native wrapper, so
GUI display is untouched. EvenPYTHONIOENCODING=cp936cannot change piped
bytes any more. - Plugin spawn env: every CLI/panel child process now runs with
PYTHONUTF8=1+PYTHONIOENCODING=utf-8(belt over the braces). - Regression tests: piped CLI output under a forced cp936 environment decodes as
UTF-8 with the Chinese brief intact and no U+FFFD; tty streams stay native.