feat(dashboard): continue a finished run as a conversation

[lerogo]

feat(dashboard): continue a finished run as a conversation

• Resume a terminal run in place, reusing its round ledger
• Claim operator messages at round start so a stop never delays them
• Order the transcript by time, not by round
• Drop the round copy of a discarded reply
• Add --reasoning-effort per run and per role
• Offer a force stop only after SIGTERM is ignored

[lerogo]

@Upper9527

[Copilot]

Pull request overview

This PR extends the harness + dashboard lifecycle so a terminal run can be resumed “in place” (continuing on its existing round ledger), improves lifecycle idempotency across resume generations, and adds a unified --reasoning-effort surface (global + per-role) that is forwarded to supporting backends and reflected in Web metadata.

Changes:

• Add in-place resume (“continue”) with a resume generation counter (resume_epoch) and generation-scoped lifecycle command IDs.
• Claim queued operator instructions at round start and order the conversation transcript strictly by time.
• Add reasoning_effort across CLI/config/supervisor/adapters and expose richer agent availability + reasoning metadata via the model catalog.

Reviewed changes

Copilot reviewed 38 out of 39 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
tests/webapi/test_resume_routes.py	Adds HTTP tests for /resume, idempotency behavior, snapshot fields, and extra-round validation.
tests/test_resume.py	Adds unit tests for ledger restore, resume epoch merging, supervisor resume modes, and CLI --resume guard.
tests/test_resume_manager_loop.py	Adds async tests validating manager-loop resume semantics (round numbering, prompt history, events, budget).
tests/test_reasoning_effort_chain.py	Verifies reasoning-effort propagation/validation across adapters, CLI resolution, config, and supervisor role configs.
tests/test_opencode_adapter.py	Updates OpenCode availability probing expectations and asserts tri-state availability.
tests/test_model_catalog.py	Updates CLI stubs to return --version output consistent with availability probing.
tests/test_manager_hardening.py	Ensures reopening a run withdraws published final responses the dashboard reads.
tests/test_guard_exclude_paths.py	Adds tests ensuring repeatable CLI flags override (not extend) project-config lists.
tests/test_deepseek_harness_adapter.py	Updates DeepSeek CLI probing expectations and asserts version/availability fields.
tests/test_codex_adapter.py	Updates Codex CLI stub to emit a version string so probing marks it usable.
tests/test_agent_registry.py	Adds tests for the new agent registry, probing tri-state availability, and effort discovery/validation.
src/lh_harness/webapi/snapshot.py	Sanitizes reasoning_effort in role configs surfaced to the browser.
src/lh_harness/webapi/server.py	Adds optional extra_rounds parsing, resume mode selection, and projects resume/lifecycle fields into snapshots.
src/lh_harness/utils/agent_cli.py	Allows probing a caller-provided binary path to keep Web responses self-consistent.
src/lh_harness/supervisor/service.py	Implements resume modes, generation-scoped lifecycle command IDs, unified worker launch transaction, and reasoning-effort forwarding.
src/lh_harness/supervisor/lifecycle.py	Introduces resume_epoch helpers/constants and treats creating as an active lifecycle state.
src/lh_harness/model_catalog.py	Refactors agent metadata to use agent_registry probes and exposes richer reasoning/availability payloads.
src/lh_harness/manager.py	Implements ledger replay on resume, claims operator messages at round start, updates budget semantics, and withdraws discarded replies.
src/lh_harness/dashboard/state.py	Adds approval support for optional extra_rounds with safe normalization and persistence.
src/lh_harness/dashboard/gate.py	Enables budget-related approvals to collect and forward an operator extra-round grant.
src/lh_harness/config.py	Adds project-config support for global + per-role reasoning_effort.
src/lh_harness/cli.py	Adds CLI flags for reasoning effort, adds supervised-only --resume, and fixes repeatable defaults semantics.
src/lh_harness/agent_registry.py	Adds a single declarative agent backend registry with probing, tri-state availability, and effort validation/discovery.
src/lh_harness/adapters/opencode.py	Normalizes/validates reasoning-effort input via shared validator and supports new reasoning_effort param.
src/lh_harness/adapters/deepseek_harness.py	Rejects reasoning effort explicitly for backends without an effort switch.
src/lh_harness/adapters/codex.py	Forwards reasoning effort as a Codex config override (no CLI flag available).
src/lh_harness/adapters/claude_code.py	Forwards reasoning effort via --effort and records it in episode metadata.
README.zh-CN.md	Documents v0.1.7 resume-as-conversation + reasoning-effort and transcript/stop behavior changes.
README.md	Documents v0.1.7 resume-as-conversation + reasoning-effort and transcript/stop behavior changes.
pyproject.toml	Bumps version to 0.1.7.
frontend/web/src/style.css	Adds styling for approval extra-round input and transcript expansion/clamping helpers.
frontend/web/src/api.ts	Extends Web API types and adds payload support for approvals extra rounds and resume mode/options.
frontend/core/test/runFeed.test.ts	Adds core projection tests for resume generation behavior, transcript ordering, approvals extra rounds, and abort command availability.
frontend/core/src/types.ts	Extends snapshot/approval shared types with resume epoch + stop fields + extra rounds.
frontend/core/src/statusView.ts	Adds “awaiting handoff” projection to avoid UI dead-air after approving while worker still reports waiting_approval.
frontend/core/src/runView.ts	Adds sortTranscript to order the conversation strictly by time while keeping untimed entries anchored.
frontend/core/src/runFeed.ts	Adds resume-epoch-aware snapshot preference/merging rules and preserves approval extra-round grants under stale frames.
frontend/core/src/commands.ts	Updates command gating to keep /abort available while stopping and adds /new --effort parsing/validation.

Files not reviewed (1)

• frontend/web/src/style.css: Generated file

Suppressed comments (1)

src/lh_harness/manager.py:1182

• The budget extension can exceed MAX_ROUNDS: when round_index == MAX_ROUNDS, max(round_index + 1, min(round_index + extra, MAX_ROUNDS)) evaluates to MAX_ROUNDS + 1. That breaks the global round ceiling and can also discard the just-written terminal reply even though the run cannot be extended safely. Consider ensuring ctx.round_budget never exceeds MAX_ROUNDS and handling the edge case (already at the ceiling) by rejecting/overriding a continue decision instead of synthesizing an out-of-range budget.

        extra = _extra_rounds(decision.get("extra_rounds")) or max(1, ctx.config.max_total_episodes or 1)
        # Always grant at least one more round: clamping to MAX_ROUNDS must not
        # produce a budget below the current round, which would end the run
        # immediately after the operator asked to continue.
        ctx.round_budget = max(round_index + 1, min(round_index + extra, MAX_ROUNDS))

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[TON14]

@lerogo Following your note in #54 — here is a test report for the v0.1.7 reasoning-effort feature (a1dd930), run end-to-end against real providers on the three backends I have installed. Every run below is a full Manager→Executor→Auditor loop to completion, with the run artifacts inspected afterwards.

What works

• claude_code (--reasoning-effort low): task completed and audited; all 10 episodes carry --effort low in their commands and "claude_reasoning_effort": "low" in metadata.
• opencode (1.18.19, real DeepSeek API, --reasoning-effort deep where deep is a custom variant defined in the workspace opencode.jsonc): task completed and audited; all 10 episodes carry --variant deep and "opencode_variant": "deep". I also pointed OpenCode at a request-capturing endpoint to confirm the variant's options genuinely reach the provider: the request body carries reasoning_effort with the variant's value. Free-form variant names passing the charset check is exactly right for OpenCode, since variants can be user-defined.
• deepseek_harness (@deepseek-ai/dsh 0.1.0-rc.7, real DeepSeek API, no effort set): base adapter works — task completed, audit passed.
• Unit suite on a clean checkout: 404 passed, 2 skipped.

Two observations

1. An explicit role effort on a backend without a reasoning switch disappears silently. lh-harness run --agent deepseek_harness --executor-reasoning-effort high ... completes normally, but the value leaves no trace at all — not in the spawned commands, not in episode metadata, and no warning on the console. _resolve_role_reasoning_effort returns None for an unsupported backend before the adapter can object, so the adapter's own "Rejected rather than ignored" guard never fires. Dropping an inherited global value across a backend switch makes sense; dropping an explicitly typed --executor-reasoning-effort with no message might surprise the operator — a console warning (or an error for the explicit flag) would make the behaviour visible.

2. A claude_code typo runs the whole task at default depth. claude -p ... --effort ultrathink prints Warning: Unknown --effort value 'ultrathink' — ignoring it and using the default effort to stderr and exits 0, so in a harness run the warning is buried in episode stderr and every round quietly runs at the default. The ReasoningSpec already documents this as validation="silently_ignored" — surfacing that warning once on the run console at startup would close the gap cheaply.

Offer: dsh reasoning depth

DeepSeek Harness does document a reasoning dial: the llm-deepseek plugin takes reasoningEffort (low/high/max). I have a practically verified mechanism from my earlier branch: append an - id: llm-deepseek layer to the --patch overlay the runner already writes. Verified against the real API on dsh 0.1.0-rc.7 — dsh --dump-config shows the value merged into the plugin config, and a full MEA run carried it in every episode. If you'd like dsh covered, I'm happy to send a small PR shaped around the ReasoningSpec registry (transport/source fields and all); just say the word.

PS: Neither English nor Chinese is my first language, so I've been posting in both to make reading easier for the team — both halves are translations anyway. Tell me which you prefer for future messages, English-only or bilingual, and I'll stick to that.

---

@lerogo 根据你在 #54 的留言，这是 v0.1.7 思考强度功能（a1dd930）的测试报告：在我本地安装的三个后端上，对接真实服务端到端测试。以下每次运行都是完整的 Manager→Executor→Auditor 循环跑到结束，并在事后检查了运行工件。

正常工作的部分

• claude_code（--reasoning-effort low）：任务完成并通过审计；全部 10 个 episode 的命令带 --effort low，元数据为 "claude_reasoning_effort": "low"。
• opencode（1.18.19，真实 DeepSeek API，--reasoning-effort deep，其中 deep 是工作区 opencode.jsonc 中自定义的 variant）：任务完成并通过审计；全部 10 个 episode 带 --variant deep 和 "opencode_variant": "deep"。我还让 OpenCode 指向一个记录请求的端点，确认 variant 的选项确实到达服务方：请求体中带有对应值的 reasoning_effort。字符校验放行自由命名的 variant 对 OpenCode 来说完全正确，因为 variant 可以由用户自定义。
• deepseek_harness（@deepseek-ai/dsh 0.1.0-rc.7，真实 DeepSeek API，未设置强度）：基础适配器正常——任务完成，审计通过。
• 干净检出上的单元测试：404 通过，2 跳过。

两个观察

1. 在没有思考强度开关的后端上，显式指定的角色强度会无声消失。 lh-harness run --agent deepseek_harness --executor-reasoning-effort high ... 正常跑完，但该值没有留下任何痕迹——生成的命令里没有，episode 元数据里没有，控制台也没有警告。_resolve_role_reasoning_effort 对不支持的后端直接返回 None，适配器里"Rejected rather than ignored"的防线根本没有机会触发。跨后端切换时丢弃继承来的全局值是合理的；但对操作者显式敲入的 --executor-reasoning-effort 无提示丢弃可能让人意外——加一条控制台警告（或对显式 flag 直接报错）就能让这个行为可见。

2. claude_code 的强度拼写错误会让整个任务以默认深度运行。 claude -p ... --effort ultrathink 只向 stderr 打印 Warning: Unknown --effort value ... 然后以 0 退出，因此在 harness 运行中这条警告被埋在 episode 的 stderr 里，每一轮都在默认深度下安静运行。ReasoningSpec 已把它记录为 validation="silently_ignored"——在运行开始时把这条警告在控制台显示一次，就能低成本补上这个缺口。

提议：dsh 的思考深度

DeepSeek Harness 其实有文档化的强度开关：llm-deepseek 插件接受 reasoningEffort（low/high/max）。我在之前的分支里有一个实测可行的机制：在 runner 已经写入的 --patch 覆盖层中追加一个 - id: llm-deepseek 层。已在 dsh 0.1.0-rc.7 上对真实 API 验证——dsh --dump-config 显示该值合并进插件配置，完整 MEA 运行的每个 episode 都携带它。如果你希望覆盖 dsh，我很乐意按 ReasoningSpec 注册表的架构（transport/source 字段等）提交一个小 PR，说一声即可。

另注：英语和中文都不是我的母语，此前双语发帖只是为了方便团队阅读——反正两个版本都是翻译。今后的消息您希望只用英文还是继续双语？我会照此执行。