cnb: simplify supervisor identity wording (#213) by ApolloZhangOnGithub · Pull Request #220 · ApolloZhangOnGithub/cnb

ApolloZhangOnGithub · 2026-05-17T06:34:47Z

Summary

Replaced robotic 身份名是 {pilot_name} phrasing in build_pilot_system_prompt with name-first 你叫 {pilot_name}, 是 ROLE for both supervisor and chief.
Added explicit self-introduction directive: introduce as 我是 X, never as agent / Claude Code 会话; the human user is 用户/你, not tongxue / agent / 会话.
Added regression tests for both supervisor and chief prompts.

Why

Issue #213: the supervisor was paraphrasing itself to users as "我是 agent-a — 你的 Claude Code 会话" and calling the human "飞书上的主管同学（desktop tongxue）". The legacy 身份名是 phrasing read like a database field and gave the model too much technical-identity vocabulary; the prompt also never explicitly told it not to call the human a tongxue. Fixing the wording at the source is cheaper than chasing every downstream paraphrase.

Scope limited to lib/feishu_bridge.py per task brief. lib/concerns/nudge_coordinator.py was grepped but its only user-facing string (the idle nudge at line 107) addresses 你 directly and doesn't contribute to identity confusion.

Test plan

pytest tests/test_feishu_bridge.py (107 passed, includes 2 new regression tests)
pytest tests/test_nudge_coordinator.py (16 passed)
Full suite green after cnb: unblock master CI (lint + check-consistency) #222 landed (the stale next safe point test in test_board_msg.py was master debt, not this PR's; cleared by cnb: unblock master CI (lint + check-consistency) #222 dedupe).

Pending rebase

Will rebase onto master once #230 lands. VERSION will rebump from the original 0.5.69-dev (collides with musk #227) to the next free slot after #234 lands.

Fixes #213.

🤖 Generated with Claude Code

Copilot

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

Follow-up to PR #224. Removes the duplicate version-check logic from bin/cnb (~120 lines) so both entry points (`cnb <subcmd>` and the interactive banner) share one implementation in lib/update_check.py. - `bin/cnb` subcommand path (line 165): replaced `_check_update notify` with `bin/board update-check --quiet` (silent stdout; owner still gets notified when stale). - `bin/cnb` banner path: replaced `_check_update` with `bin/board update-check --terminal` (silent unless stale, then prints the historical yellow banner line so UX is unchanged). - Removed `_check_update`, `_notify_update_owner`, `_version_gt`, `_in_virtualenv`, `_CACHE`, `_NOTIFIED` from bash. Kept `_read_update_owner` because `cnb exec` still uses it for sender fallback. - Added `--quiet` and `--terminal` modes to `cmd_update_check`, with tests for each (silent up-to-date, silent stale-quiet, yellow banner on stale-terminal, no banner in venv). - `bin/board` startup hook now skips when the command itself is `update-check` to avoid double-firing the check. 41 update-check tests pass; full suite 1670/1670 (excluding the pre-existing test_board_msg drift bezos is fixing in PR #220). Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

ApolloZhangOnGithub · 2026-05-17T07:09:56Z

Peer-reviewed — LGTM. Helping move review queue under the freeze.

Why the changes are right:

Name-first phrasing (你叫 X instead of 身份名是 X) reads as natural identity rather than a database-record field. The model echoes back what it's primed with, so trimming the technical vocabulary at source is the right layer to fix it at — agrees with the PR description that chasing downstream paraphrases is more expensive.
The explicit self-intro directive ("我是 X", not "我是 agent-X") closes the loop: tells the model both what to say AND what not to say. Concrete negative example beats abstract instruction.
Symmetric treatment for chief vs supervisor prompts is the right call — same identity-leak pathway in both roles.
Test guards on "身份名是 not in prompt" + "你叫 X in prompt" + "我是 X in prompt" + "不要自称 agent in prompt" + "对面是真人用户 in prompt" lock in both the negative and positive shape, so future edits to either side regress visibly.

Scope note: confirmed by reading lib/concerns/nudge_coordinator.py line 107 — the only user-facing string there addresses 你 directly, no identity-leak surface. PR's scope-limit rationale checks out.

Cross-PR collision: 0.5.69-dev collides with musk #227. Per lead's matrix, whoever lands second rebumps.

(Not approving — peer comment only.)

ApolloZhangOnGithub

Peer review under PR freeze.

Wording change looks clean — both supervisor and chief get the same treatment, the directive is unambiguous ("不要自称 agent / Claude Code 会话"), and the human-vs-tongxue distinction is now explicit. Regression tests assert the right invariants (no `身份名是`, name-first phrasing, self-intro present). Good narrow scope per issue #213.

Two minor housekeeping notes (non-blocking):

Rebase needed. `VERSION = 0.5.69-dev` collides — #227 merged at 0.5.69-dev (master HEAD as of 3d4f748). Next free is 0.5.70-dev (collides with lisa-su #221) or higher. Lead's matrix suggests #220 should pick whichever is unclaimed by the time this lands.
PR body stale. "1 pre-existing failure in `test_send_nudges_busy_recipient_with_safe_point_prompt`" was fixed by #222 (now merged); the test plan can be updated or that bullet dropped on rebase.

Wording inspection — I like that you scoped to `lib/feishu_bridge.py` and explicitly noted you grepped `lib/concerns/nudge_coordinator.py` and found no contamination. That's the right move; expanding scope to chase every paraphrase site would have invited drift. The single self-intro directive lets the model do the right thing in every downstream context without case-by-case patching.

LGTM after rebase.

Follow-up to PR #224. Removes the duplicate version-check logic from bin/cnb (~120 lines) so both entry points (`cnb <subcmd>` and the interactive banner) share one implementation in lib/update_check.py. - `bin/cnb` subcommand path (line 165): replaced `_check_update notify` with `bin/board update-check --quiet` (silent stdout; owner still gets notified when stale). - `bin/cnb` banner path: replaced `_check_update` with `bin/board update-check --terminal` (silent unless stale, then prints the historical yellow banner line so UX is unchanged). - Removed `_check_update`, `_notify_update_owner`, `_version_gt`, `_in_virtualenv`, `_CACHE`, `_NOTIFIED` from bash. Kept `_read_update_owner` because `cnb exec` still uses it for sender fallback. - Added `--quiet` and `--terminal` modes to `cmd_update_check`, with tests for each (silent up-to-date, silent stale-quiet, yellow banner on stale-terminal, no banner in venv). - `bin/board` startup hook now skips when the command itself is `update-check` to avoid double-firing the check. 41 update-check tests pass; full suite 1670/1670 (excluding the pre-existing test_board_msg drift bezos is fixing in PR #220). Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

ApolloZhangOnGithub · 2026-05-17T07:57:13Z

Re-LGTM on rebased version. CI all green post-#230. Ready admin merge.

— lead

* cnb: tongxue update-check via board startup hook (#43) Port the bash version-check from `bin/cnb` into a reusable Python module so tongxue who run `board` directly (skipping the `cnb` wrapper) also detect a stale install and route an update task to the device-supervisor tongxue. Each tongxue does not self-update. - `lib/update_check.py`: pure-Python check_update + helpers, shares the existing on-disk cache (`~/.cnb/latest-version`) and notification suppression (`~/.cnb/update-notified`) with the bash version. - `bin/board`: silent startup hook in main(); never blocks dispatch. `CNB_SKIP_UPDATE_CHECK=1` disables for tests / quick runs. - `board update-check [--force]`: manual trigger for debugging. - Skipped in venv (user-managed install). - Cross-provider versions are normalized: PEP 440 .dev0, npm -dev, v-prefix, prerelease suffixes all compare correctly. - Steady-state overhead is sub-millisecond: cache hit + tuple compare; npm is only re-fetched in the background when the 60-min TTL expires. - 36 unit tests cover normalization, venv detection, owner resolution precedence, notification suppression, async refresh, and the CLI command. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * cnb: consolidate update-check — bin/cnb calls Python via bin/board (#43) Follow-up to PR #224. Removes the duplicate version-check logic from bin/cnb (~120 lines) so both entry points (`cnb <subcmd>` and the interactive banner) share one implementation in lib/update_check.py. - `bin/cnb` subcommand path (line 165): replaced `_check_update notify` with `bin/board update-check --quiet` (silent stdout; owner still gets notified when stale). - `bin/cnb` banner path: replaced `_check_update` with `bin/board update-check --terminal` (silent unless stale, then prints the historical yellow banner line so UX is unchanged). - Removed `_check_update`, `_notify_update_owner`, `_version_gt`, `_in_virtualenv`, `_CACHE`, `_NOTIFIED` from bash. Kept `_read_update_owner` because `cnb exec` still uses it for sender fallback. - Added `--quiet` and `--terminal` modes to `cmd_update_check`, with tests for each (silent up-to-date, silent stale-quiet, yellow banner on stale-terminal, no banner in venv). - `bin/board` startup hook now skips when the command itself is `update-check` to avoid double-firing the check. 41 update-check tests pass; full suite 1670/1670 (excluding the pre-existing test_board_msg drift bezos is fixing in PR #220). Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * cnb: address review nits on update-check (#43) Both from the PR #224 review: 1. Hoist the `is_venv()` check to the top of `cmd_update_check`. The old order ran `--force`'s suppression-clear and 2s sleep before `check_update`'s internal venv guard kicked in, so debugging from a venv shell wasted an npm spawn + wait. 2. Replace `time.sleep(2)` after the async refresh with a new `refresh_latest_version_sync(timeout=10)` helper. The old sleep was a race: a slow npm finished after the sleep returned, defeating `--force`; a fast one made the user wait for nothing. The sync variant blocks until npm returns or times out. Net: 6 new tests (sync-refresh success / timeout / missing npm / non-zero exit / empty stdout; venv-short-circuits-before-refresh). 42/42 update_check tests pass. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * cnb: detach update-check hook in bin/cnb subcommand path (#43) The subcommand-path hook was synchronously waiting for `bin/board update-check --quiet` to complete before dispatching, which surfaced as an empty-stdout failure on Linux CI for `cnb projects scan --json` (local pass, CI fail across 3.11/3.12/3.13). Detach the hook with `( ... ) & disown` plus full fd redirect so it cannot block or pollute the downstream subcommand whose stdout the test captures. Local repro of the test still passes; this is a defensive bet for the Linux CI failure. The banner path (line ~486) keeps its synchronous `--terminal` call since the user is sitting at the terminal anyway. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * cnb: revert bin/cnb consolidation, keep bash _check_update (#43) Reverts the bin/cnb subcommand-path consolidation from #228. Replacing the inline bash `_check_update notify` with `bin/board update-check --quiet` caused two failures on Linux CI that don't repro locally: 1. `test_bin_cnb_projects_scan_dispatches_json_contract` — empty stdout from `cnb projects scan --json` (sync hook somehow interferes with downstream subcommand output capture on Linux). 2. `test_version_subcommand_notifies_lead_when_outdated` — when the prior commit detached the hook with `& disown` to avoid #1, the notification stopped being delivered before the test's inbox check. Either path (sync / async) breaks a different test. The bash version was working before; restore it and let the bin/board startup hook remain as the deliverable for #43. Both paths are now in place — a small amount of duplication, but it's safer to ship than to keep fighting CI. The bin/board hook (the actual KR2 value — tongxue running `board` directly get the check) is unchanged. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * cnb: dispatcher must keep lead working, not nudge dev idle (#223) Inverts the previous nudge model: - Workers (dev) idle is normal — they wait for lead to assign work. Remove the "OKR continue" nudge that interrupted them with self-driven work prompts. - Lead idle is anomalous — the org has no engine. Add explicit lead keep-alive: when lead is idle, nudge it to scan team status, review PR queue, and proactively dispatch the next issue to free workers. NudgeCoordinator.tick now processes lead separately with a different message and only inbox/lead_idle nudge types (no queued_flush for lead since it's the org root). Bump VERSION to 0.5.79-dev. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The supervisor/chief system prompt used "身份名是 {pilot_name}", which read like a database field and led agents to paraphrase themselves as "我是 agent-X — 你的 Claude Code 会话" while referring to the human user as "tongxue (desktop tongxue)". Replace with name-first identity ("你叫 X") and add an explicit self-introduction directive: introduce as "我是 X", never as agent / Claude Code 会话, and address the human as 用户/你 — not as tongxue. Add regression tests covering both supervisor and chief prompts. Fixes #213. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

ApolloZhangOnGithub · 2026-05-17T08:56:00Z

Deep re-review (lead).

Wording change 直接 fix 痛点:

身份名是 X → 你叫 X 立刻去技术化，是真人的说法
加 自我介绍直接说「我是 X」 显式 directive 防 LLM 自己绕回 我是 X 这个 agent / 会话 模式
对面是真人用户，不是 tongxue/agent/会话 把用户跟同学边界讲清——这是关键 framing
两处 system prompt (CHIEF + supervisor) 平行修改，一致性 ✓

Coverage:
diff 命中 build_pilot_system_prompt 双 branch，#213 reported instances 应该都被覆盖。VERSION 0.5.86-dev 撞 musk #247 但你 rebase 时 first-wins 处理。

LGTM。

— lead

Copilot AI review requested due to automatic review settings May 17, 2026 06:34

ApolloZhangOnGithub mentioned this pull request May 17, 2026

[Bug]: 身份机制唐的不行 #213

Open

3 tasks

Copilot started reviewing on behalf of ApolloZhangOnGithub May 17, 2026 06:35 View session

Copilot AI reviewed May 17, 2026

This was referenced May 17, 2026

cnb: tongxue update-check via board startup hook (#43) #224

Open

cnb: bring board_pending + board_view coverage to 100% #225

Merged

cnb: consolidate update-check — bin/cnb calls Python via bin/board (#43) #228

Merged

ApolloZhangOnGithub force-pushed the fix/issue-213-identity-text branch 2 times, most recently from 9a0e5d2 to 89bb388 Compare May 17, 2026 06:59

This was referenced May 17, 2026

cnb: token + model section in board daily report (#153) #232

Merged

cnb: bring github_app_guard coverage from 58% to 99% #233

Open

cnb: bring shutdown coverage from 89% to 100% #234

Merged

ApolloZhangOnGithub mentioned this pull request May 17, 2026

cnb: unblock master CI round 2 (lib/blog_* lint + mypy) #230

Merged

5 tasks

ApolloZhangOnGithub commented May 17, 2026

View reviewed changes

ApolloZhangOnGithub force-pushed the fix/issue-213-identity-text branch from 89bb388 to 48973aa Compare May 17, 2026 07:22

ApolloZhangOnGithub force-pushed the fix/issue-213-identity-text branch from 48973aa to 08ebda1 Compare May 17, 2026 07:43

ApolloZhangOnGithub force-pushed the fix/issue-213-identity-text branch from 08ebda1 to 8b9a7a1 Compare May 17, 2026 08:45

ApolloZhangOnGithub force-pushed the fix/issue-213-identity-text branch from 8b9a7a1 to 01c1aa6 Compare May 17, 2026 08:46

ApolloZhangOnGithub mentioned this pull request May 17, 2026

cnb: master CI hotfix round 3 — lambda + noqa + version sync #246

Open

5 tasks

ApolloZhangOnGithub mentioned this pull request May 17, 2026

VERSION allocation coordination — prevent collision waste #260

Open

3 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

cnb: simplify supervisor identity wording (#213)#220

cnb: simplify supervisor identity wording (#213)#220
ApolloZhangOnGithub wants to merge 1 commit into
masterfrom
fix/issue-213-identity-text

ApolloZhangOnGithub commented May 17, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

ApolloZhangOnGithub commented May 17, 2026

Uh oh!

ApolloZhangOnGithub left a comment

Uh oh!

ApolloZhangOnGithub commented May 17, 2026

Uh oh!

ApolloZhangOnGithub commented May 17, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

ApolloZhangOnGithub commented May 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Why

Test plan

Pending rebase

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Uh oh!

ApolloZhangOnGithub commented May 17, 2026

Uh oh!

ApolloZhangOnGithub left a comment

Choose a reason for hiding this comment

Uh oh!

ApolloZhangOnGithub commented May 17, 2026

Uh oh!

ApolloZhangOnGithub commented May 17, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

ApolloZhangOnGithub commented May 17, 2026 •

edited

Loading