v1.21.0 — Session/Streaming Stability + Self-Evolution Closed Loop

[xg-gh-25]

🎉 v1.21.0 — Session/Streaming Stability + Self-Evolution Closed Loop

✅ RELEASED 2026-06-25 — Release v1.21.0 · CI green · deployed.
Kept open as the milestone record for our biggest release to date.

296 commits since v1.20.1 (2026-06-18). A full week of heavy stability work — theme: make the chat experience never lose your work, and make the system fix its own recurring mistakes.

---

💡 What this means for you (user-facing benefits)

Before	After
切 tab / resume 时，完整的回复有时被截断或整段消失	回复永远完整落地——单一渲染源结构性根治了这个 #1 复发 bug
网络/SSE 断开 → 正在生成的长回复丢失	断连不再杀子进程，输出落库可恢复，回来还在
多个 tab 同时跑 → session "乱套"、互相挤占、偶发 OOM	每个 tab 资源独立仲裁、严格隔离，跑满也不互相干扰
Agent 卡住（工具 hang / 等输入）→ 整个会话僵死	多条逃逸路径自动恢复，卡死能自愈
流式输出纯文本到结尾才突然变 markdown，长回复还卡顿	流式过程就是实时 markdown（标题/列表/代码/公式），节流到 ~100ms 既流畅又不卡
headless/后台任务里遇到要确认的问题 → Agent 自问自答蒙混过去	真正阻塞等你回答，不再替你瞎拍板
同一类错误反复犯，只能靠你每次抓	系统自动记录认知偏差、到阈值自动提结构性修复方案

一句话：这一版主要让你不再丢失工作——回复不截断、断连不丢、多 tab 不打架、卡死能自愈、流式实时渲染。

---

🔧 Function summary (technical)

1. Session/Streaming/Recovery 稳定性 (~70 commits) — 最大头

   • R6 session 资源仲裁 + 多 tab 严格隔离（orphan-only eviction、独立 concurrent-streaming cap、TTL 12h→24h）
   • RecoveryCoordinator 统一恢复决策 + crash_to_cold TOCTOU 关闭
   • reconcile-gap 截断渲染根治（单一渲染源，COE07/08/09 类复发 Add Ollama support (local + cloud) with automatic model detection and OpenAI-compatible fallback #1）
   • SSE 断连不杀子进程，输出落库可恢复；reconnect budget ~65s
   • OOM / tool-hang / stuck-WAITING 逃逸路径 + dumb-spawn watchdog
   • 流式 markdown 节流渲染——实时 markdown，O(n²) 重解析降为时间有界（commit dc9d3132）

2. AskUserQuestion block-hook (10) — headless 模式真阻塞等用户答，不再自问自答；断连后自动重发被吞的问题

3. Pipeline 加固 (~20) — completion gate fail-closed、FAILED-vs-ERRORED 区分、adversarial 双闸、universal Understanding Gate (Gate 0)

4. 自进化 v3 + Eval (~40) — OT07 认知缺口闭环（自动记录 + escalation ladder + Governance tab + growth report）、行为轨迹 eval（M3/M4/M5）、噪声闸门、Golden Set self-evolving

5. DDD reclaim (11)、daemon 日志轮转 / zombie-subprocess backoff (~6)、Reversible Context Recall、deps/db/ui 杂项

---

Commit breakdown (v1.20.1..v1.21.0, verified 2026-06-26)

77 feat · 150 fix · 24 docs · 23 test · 7 chore · 5 refactor · 2 governance · 2 reflect · +misc (revert/perf/merge/debug/build) = 296 total

Release checklist ✅

• 修陈旧测试断言（COE01 platform-divergence + tab-bar）— push 前 CI 抓到 5 个陈旧测试失败，已全修
• 全部 commit push 进 origin/main（CI 在 229 个未验证 commit 上跑绿）
• push 后确认 CI 绿
• s_swarm-release → v1.21.0（build + package + smoke + publish）✅
• 部署上线（backend daemon = HEAD healthy；frontend 新 .app 嵌入流式节流修复）
• 收尾：DMG size 判据修复（>30MB → >5MB，6 个 release 实测 ~10.3MB）+ streaming-markdown half-fix 补全

[xg-gh-25]

✅ Pipeline v6 — 3 Gates = 3 Moments of Truth(已实现,随 v1.21.0 发布)

范围:这是本版 "Pipeline 加固"(上面第 3 点)之上的下一步。v1.21.0 在原有 dual-gate(Gate 1 Skeptic + Gate 2 Adversarial)前面加上了 Gate 0 — Understanding Gate。三门齐了。

为什么要第三道门。 原有两道质量门都在解法侧:

• Gate 1(Skeptic + SSA) — BUILD 前,问"方案对不对"
• Gate 2(Adversarial) — DELIVER 前,问"代码对不对"

两道门都默认问题已经框对了,只审之后的东西。所以当 framing 本身错了,没有门能拦——错误一路滑到 Gate 2(整条 pipeline 的代价才发现)或漏给人审。一个 session 里就出现过三次这类失败(编造的 kernel-OOM 根因、把已生效的 [DONE] 当成要修的 no-op、一个永不触发的死判别式),两道门都够不着。其中 "[DONE] no-op" 那次根本不是 bug fix,是对现有代码的理解错了——证明理解错跨越所有工作类型,门必须普适。

三门 = 三个真相时刻,各守一个,互相结构上补不了:

Gate	守的真相	时机	防的失败	状态
Gate 0 — Understanding	现状/根因懂了吗	EVALUATE → THINK	framing 错	✅ v1.21.0 已发布
Gate 1 — Skeptic + SSA	方案 sound 吗	PLAN 后,BUILD 前	修症状、修错层	✅ live
Gate 2 — Adversarial	代码对吗	DELIVER 内	self-authored bug	✅ live

经济学就是论据:framing 错在 EVALUATE 拦 = 1 个 sub-agent;漏到 Gate 2 拦 = 整条 pipeline。最便宜犯的错,却因为没门,最贵才发现。现在有门了。

Gate 0 查什么。 在谈"怎么修/怎么建"(THINK)之前,必须先产出一个关于当前世界状态的可证伪断言,用观察(非推断)支撑,且扛过一次反驳。普适于所有工作类型,证据形态随类型变(bugfix→ps/log/复现;existing-feature→code-trace;refactor→characterization;greenfield/research→premortem;docs→代码引用)。

三道机械子门(已在 validator 落地,pipeline_validator.py:263-385):

1. M1 隔离墙 — EVALUATE 只描述现状,结构上禁止提解法(解法属于墙另一侧的 THINK)。
2. M2 hedge-word 扫描 — 断言/证据里有未消解的 hedge("似乎/可能/probably/should be")且无观察支撑 → BLOCK。把过去被动的"先观察再断言"规则变成机械过门动作(那条规则曾在一个 session 里静默失效四次)。
3. M3 fresh-context skeptic — 零作者推理的 sub-agent 尝试反驳理解。判决:SUPPORTED / UNSUPPORTED / ALREADY-SATISFIED(抓 no-op)/ WRONG-FRAME。只有 SUPPORTED 才开 THINK。profile 分级,trivial/docs 不付 sub-agent 成本。

实现证据(code + behavior,非文档):

• 2479fed2 feat(pipeline): universal Understanding Gate at EVALUATE→THINK boundary
• 58bb0287 feat(pipeline): REPRO gate + diagnostic-challenge — diagnosis before bug-class build
• 58923b50 fix(pipeline): Understanding Gate adversarial convergence — fail-closed profile + M1 coverage
• 自指验收达成:设计时立的标准是"实现 Gate 0 的那个 run 必须先过 Gate 0"——run_862fa4e0 做到了(work_type=existing-feature,证据=现有 gate 的 code-trace)。
• 实战验证:本版发布后两个 pipeline run(DMG 判据修复、streaming-markdown throttle)都先过了 Gate 0——每个 run 在 THINK 之前产出了 observation-backed 的 understanding block(throttle 那个:"O(n²) 根因 = 每 token 全量重解析",code-trace 支撑,扛过反驳)。门确实在拦"框对没框对",不只走过场。

理论依据(两个 AWS repo 殊途同归 → "诊断↔解法墙"是对的结构,不是一拍脑袋):

• awslabs/aidlc-workflows v2 — P3(同一 model 既生成又自评 NL 判据会收敛到"满足字面",所以纯 LLM 验证的阶段不能自停,需外部验证器=fresh-context skeptic)+ P6(分阶段分解优于单次压缩,理解与求解是两个不同的搜索)
• aws-samples/sample-ai-plc — envision↔solution 分文件 + 硬门;overconfidence-prevention 反转默认("除非异常清楚否则必问")

3 gates, 3 moments of truth:understanding → plan → code. Pipeline 终于三门全守。完整设计详述见 Discussion #68(https://github.com/xg-gh-25/SwarmAI/discussions/68#discussioncomment-17437617)。

[xg-gh-25]

🚀 Shipped — v1.21.0 is live

Release v1.21.0 · published 2026-06-25 · CI green · deployed to daemon + desktop .app.

Our biggest release to date — 296 commits in one week. The headline isn't a single feature; it's that the chat experience stops losing your work and the system starts fixing its own recurring mistakes.

How it actually shipped (the process was the story)

• The tag was 8 days stale → 229 commits had never run CI. Pushing to origin before tagging triggered CI on all of them and caught 5 stale-test failures that would otherwise have shipped inside the released DMG (a COE01 macOS-vs-Linux platform divergence + tab-bar assertions). Lesson banked: push → CI-green → then tag.
• Adversarial review on a "trivial" CI-unblock found 2 more latent test files that maxfail had hidden downstream — the failure class was broader than the 3 visible failures.

Session close-out (post-release polish, same day)

• Streaming markdown, properly fixed. The earlier perf fix rendered streaming text as plaintext to kill O(n²) reparse jank — but that was a half-fix (lost live formatting + end-of-stream jump). Now markdown renders live during streaming, with the parse throttled to ~100ms so it's both rich and smooth (dc9d3132). Meta-lesson: a perf fix that abandons a UX capability is a half-fix — bound the frequency, don't drop the capability.
• DMG verify criterion fixed — s_swarm-release checked >30MB, but every release is ~10.3MB (the PyInstaller backend ships separately; the DMG is the Rust shell). Corrected to a >5MB broken-build floor.

Keeping this issue open as the milestone record. 🐝