feat: pm 关 resume 降本 — noResume 配置 + 全新会话注入完整最近历史

[lishuceo]

背景

pm bot(土豆儿)在高频私聊场景整天 resume 累积,基线上下文从 60K 一路涨到 600K+,长循环里每轮重读把成本推高,频繁撞 $15 budget 上限、无法使用。根因不是单句话贵,而是长会话逐轮累积。

方案

让 pm 不 resume(每条消息全新 SDK 会话),用"注入最近 N 条聊天历史"替代 SDK session 记忆。配套上调历史预算,让 fresh 会话装得下足够上下文。

改动

1. per-agent noResume 配置字段

• AgentConfig.noResume:开启后该 agent 每条消息全新会话
• schema / loader 透传(agent 级覆盖 defaults);agents.example.json 的 pm 示例开启

2. 统一 resume 判定 canResumeSession(纯函数,可测)

• 无会话 / cwd 变更 / agent noResume → 全新会话
• 接入 executeClaudeTask、executeDirectTask 两条路径
• 修正一处隐患:历史增量去重(afterMsgId)与历史文件附件跳过原本以 activeConversationId 为准,导致不 resume 时只注入增量、SDK 端却无记忆 → 失忆。改为以 canResume 为准,只有真正 resume 才做增量,否则注入完整最近 N 条(cwd 变更场景一并受益)

3. 上调默认历史预算

• CHAT_HISTORY_MAX_COUNT 10 → 20
• CHAT_HISTORY_MAX_CHARS 8000 → 16000
• SELF_BOT_MSG_MAX 150 → 1500(noResume 下历史是唯一自我记忆来源,需保留较完整自身回复;字符硬顶仍防爆)

测试

• 新增 canResumeSession 单测(4 分支)+ noResume 配置透传单测(4 例)
• history-truncation / structured-chat-context 测试适配新预算
• npx vitest run:1757 passed,唯一失败 memory/quality.test.ts > 当前工作重点 在 main 上同样失败(依赖 embedding API 的环境问题,与本 PR 无关)

部署注意

• 本 PR 只提供 noResume 能力;要对 pm 生效,需在部署目录的 config/agents.json 给 pm 加 "noResume": true(该文件不在 git 中)
• 含 src/config.ts 默认值改动,合并后部署目录需 build + 重启生效

🤖 Generated with Claude Code

[github-actions]

Claude finished @lishuceo's task in 2m 58s —— View job

---

PR Review: pm 关 resume 降本

Verdict: ✅ Approved — clean, well-tested, no blocking issues. One trivial doc-drift nit below.

• Read CLAUDE.md & PR intent
• Review canResumeSession extraction + both call sites
• Verify config schema/loader/type passthrough
• Trace history-injection & dedup logic
• Review tests

---

What I checked

canResumeSession (pure fn) + both call sites — Correct. The three not-resumable branches (no activeConversationId / cwd changed / noResume) are right, and both executeClaudeTask (line 2679) and executeDirectTask (line 3344) now route through it. The old inline canResume at the resume decision was removed and resumeSessionId: canResume ? activeConversationId : undefined (line 2905) still gates the actual SDK resume — so history injection and the resume decision are now driven by the same flag. 👍

The amnesia fix is real and correct. Previously afterMsgId and the history-file skip keyed off activeConversationId; on a cwd change (or noResume) the SDK started fresh yet only incremental history was injected → memory loss. Switching both to canResume (lines 2726, 2740, 3367, 3385) fixes this and correctly extends the benefit to the cwd-switch case.

Optional-chaining consistency — executeClaudeTask uses agentCfg?.noResume (agentCfg from .get(), nullable) while executeDirectTask uses agentCfg.noResume (agentCfg from getOrThrow() at line 3261, non-null). Both correct.

Merge precedence — noResume: input.noResume ?? defaults.noResume handles explicit false correctly (agent-level false overrides a defaults: true), matching the test at config-loader.test.ts. Schema adds noResume to both input and defaults schemas ✓.

Budget hard-cap holds — SELF_BOT_MSG_MAX 150→1500 is per-message; historyMaxChars (line 2331–2341) still drops oldest lines to enforce the total cap, so the "字符硬顶仍防爆" claim checks out.

Tests — canResumeSession (all 4 branches + edge cases) and noResume passthrough (4 cases incl. defaults inheritance/override) are thorough and match the implementation.

🟡 Minor nit (confidence 85, non-blocking)

src/feishu/event-handler.ts:2272 — the doc comment still reads 总字符数超 CHAT_HISTORY_MAX_CHARS（默认 8000）, but this PR bumped the default to 16000 (config.ts:168). Stale after this change.

Fix this →

Notes

• Deploy caveat is well-flagged in the PR body: noResume is opt-in per agent via config/agents.json (not in git), and config.ts default changes need build + restart.
  · feat/claude-session-ba6ee8

[lishuceo]

Fixed — 已把 formatHistoryMessages 注释里的 默认 8000 更新为 16000，并顺手修正了同段 自身 bot 150 字符 → 1500（都是本 PR 引入的 drift）。commit 见 HEAD。（回复 review）

[github-actions]

Claude finished @lishuceo's task in 1m 4s —— View job

---

Re-review after doc-comment fix

• Verify the stale-comment fix (默认 8000 → 16000, 150 → 1500)
• Re-scan latest diff for new issues

---

Verdict: ✅ Approved — the fix in a832bb6 is correct and the wiring is unchanged. canResumeSession is still the single source of truth for both the resume decision (resumeSessionId, lines 2905/3350) and the history-injection/dedup gates (lines 2726/2740, 3367/3385). No regressions.

The formatHistoryMessages doc comment (event-handler.ts:2271-2272) now correctly reads 1500 / 16000. 👍

🟡 One more stale comment of the same kind (confidence 90, non-blocking)

src/feishu/event-handler.ts:3565 — buildDirectChatContext's doc comment still says:

 * max 由 CHAT_HISTORY_MAX_COUNT 控制（默认 10）。

but this PR bumped that default to 20 (config.ts:167). Same doc-drift class as the two just fixed — worth catching it in the same pass.

Fix this →

Everything else in the PR holds up — no blocking issues.
· feat/claude-session-ba6ee8