[hsliuustc0106]

Fixes #917

These two smoke failures were already red on main (same result at 32f3e7c and on this branch's base). They are not a #903/#908 regression. Harness default and the usage five-field Step 0 work are untouched.

What changed

Admin calls everyone. #894 moved the members line into the group system-prompt section. The mock still looked for Members: … and you (admin). in the turn text, so Chief was never the admin and answered 我觉得可以。 alone. The mock now reads that line from the system prompt (and still from the turn, for older prompts).

Writer condensed in place. The steward did say compact, and the checkpoint was written. DSH's token meter stays anchored on the provider usage from before the span was replaced and only subtracts a 4-characters-per-token estimate of the removed text. The mock reports usage at about 3 characters per token, so after a real checkpoint the anchored total still sat above the check line and the part switched (+3). After a compaction, the part now stays when the heuristic size of the remaining surface plus the tool schemas is under the check line. A surface that is still over it still moves on. Crossing the line in the first place still uses the anchored total.

Verification at the tip

Tip SHA: de7b2e6f0c3a98fd023448a91d6a078eb4db716c

Ran:

• node --import ./test/harness-resolve.mjs --test test/mock-group-roster.test.js test/session-identity.test.js at the unbroken tip: red. Admin answered 我觉得可以。 instead of @all 大家各说一句。. The compact path returned { kind: 'reject' } while the remaining surface was 8,000 tokens.
• Same command after the fix: 10 pass.
• npm test in core/: 1347 pass, 0 fail.
• npm run lint in core/: oxlint warnings are the pre-existing set; no new warning on the edited lines. arch-docs after the commit, with ARCH_DOCS_BASE=origin/main: OK (11 changed files).
• node scripts/ds-bot-patches.mjs generate && node scripts/ds-bot-patches.mjs check: patches apply, core/ matches.
• node .github/scripts/changelog.mjs check: fragments well formed.
• Mock smoke, no API key: DSH_HARNESS pointed at a checkout-shaped install of @deepseek-ai/dsh@0.2.0-rc.2, SLOT=1, sh core/dev/check.sh --only smoke. check passed in 220 s. admin calls everyone and Writer condensed in place, not switched both ok, and the rest of the smoke was green.

Not run (and why):

• npm run verify (plugin suites, interop, live-e2e). This change is core host + the mock.
• Web and desktop shots.
• A live model / DEEPSEEK_API_KEY path.

Re-run smoke on a Linux host with DSH

From the repo root, with Node 22+ and a DeepSeek Harness checkout at dsh-v0.2.0-rc.2 (or DSH_HARNESS set to one), and pnpm on PATH:

cd core
npm ci && npm run harness
cd ..
CONTEXT_WINDOW=100000 CHECK_RATIO=0.45
# dev/check.sh sets those itself for the smoke step
SLOT=1 sh core/dev/check.sh --only smoke

SLOT picks free ports (smoke uses 31n1 / 88n1). The bench data stays under .dsh-test-check-smoke. No API key: DEEPSEEK_API_KEY=mock-key and DEEPSEEK_BASE_URL point at core/dev/mock-llm.mjs.

DSH follow-up

@deepseek-ai/dsh-token-meter (0.2.0-rc.2) reuses the last provider usage as the pressure baseline and applies a heuristic surface delta (CHARS_PER_TOKEN = 4). Compaction does not report usage of its own, so a denser provider report (this mock, or CJK on a real model) can leave measure().totalTokens above the check line after the span is gone. FlyBot compensates only when it decides whether a checkpoint got the part under the line. The meter's own total, which other readers still see, is unchanged. That anchoring belongs in DSH.

  

[Birfy]

Written-by: Dev

Shepherd note (objective; I did not touch this branch, and I did not add Owner: / Reviewer: lines — those are Bot-pipeline fields, and this is a human pull request).

What it changed — 11 files, +317/−10:

• core/dev/mock-llm.mjs +12/−4: the members/admin line is read from the group's system-prompt section as well as from the turn (Group turns re-send and store the group notice and fixed room rules on every turn #894 moved it there); decide is exported; the server binds only on a direct run;
• core/src/host/parts.js +26/−3: pressureLeft — after a compaction, the stay/move decision reads the heuristic size of what remains (surface + tool schemas at 4 characters per token) instead of the anchored meter total;
• new core/test/mock-group-roster.test.js, core/test/session-identity.test.js +44/−2, one line each in docs/subsystems/core-host.md and docs/subsystems/ds-bot-core.md, changelog.d/917.fixed.md;
• generated: patches/ds-bot/0001-core-fixes.patch +65, 0014-flybot-branding.patch +56, 0016-room-collaboration-smoke.patch +35, files/test/mock-group-roster.test.js.

Scope — the two smoke failures in #917 and nothing else. #903/#908's harness and the usage fields are untouched.

Verified on this machine (no git here, so no npm run verify):

• at the tip de7b2e6f: node --import ./test/harness-resolve.mjs --test test/mock-group-roster.test.js test/session-identity.test.js → 10 pass / 0 fail;
• the same two test files on the base b38783df → 2 fail: the compaction case (a checkpoint stays in place when the remaining surface fits and only the usage anchor reads high) is a real behavioural red, and the new mock test cannot load on base at all — importing mock-llm.mjs there starts the server and keeps the process alive, which the tip fixes with the export plus the "only on a direct run" guard. So the admin half's red proof is at module level, not at an assertion; the issue's claim that both were red on main holds, but only the compaction half is red as a value.

Risk to weigh — pressureLeft only lowers the number the post-compaction decision reads, and it prices the remaining surface at 4 characters per token. Real (CJK, or a dense model) text is denser than that, so a part can now stay where today it would move on. The change is gated to baseline.kind === 'usage' && surfaceDeltaTokens < 0, and "a surface still over the line still moves the part on" is covered — but that case is the mock's density, not a dense model's. core/src/host/parts.js is the autonomous-loop part switching, a known area, and this is the only behavioural change in the diff.

Mergeable alone — not as it stands: the base is b38783df, one merge behind (main is 0b22aa53 after #908), and #908 edited patches/ds-bot/0001-core-fixes.patch and docs/subsystems/core-host.md, both of which this pull request also edits. It needs an update-branch or a rebase before anyone judges it; GitHub still answers mergeable: unknown.

[cursor]

Bugbot needs on-demand usage enabled

Bugbot uses usage-based billing for this team and requires on-demand usage to be enabled.

A team admin can enable on-demand usage in the Cursor dashboard.

[hsliuustc0106]

Verdict: PASS at de7b2e6
Written-by: Reviewer
Tip SHA: de7b2e6

Ran:

• diff review: mock roster from system section (#894/#917); pressureLeft after compaction so dense usage anchors do not force a part switch when remaining surface fits
• unit coverage: mock-group-roster.test.js, session-identity checkpoint stay/switch cases
• hk2 (via 开发工程师): SLOT=1 sh core/dev/check.sh --only smoke → 42/42, including admin calls everyone and Writer condensed in place, not switched
• CI at tip: tests-ok / Node 24 · DSH 0.2.0-rc.2, judge, merge-gate green; labels agent-pr, risk:T3
• PR body: patches check + core mirrors claimed green

Not run:

• full npm run verify / live model path — smoke+CI cover the #917 scope; live deferred
• re-run hk2 personally — accepted 开发工程师 report at tip de7b2e6

Findings: no actionable blockers for #917.

• Soft: changelog fragment covers compaction stay; admin/@ALL mock fix is only in PR body (optional follow-up fragment).
• Soft: until #926 lands on main, author self-PASS may still be ignored by enqueue-guard; Birfy or owner path may still be needed to enqueue this tip.

This report is a Verifier result, not merge authority on its own; merge follows AGENTS.md step 6.