[hsliuustc0106]

Owner: Dev
Reviewer: Reviewer2

Written-by: Cursor Agent (for hsliuustc0106)

References #874

Step 1 only. This does not close the RFC: Step 2 (the experience library) is still open, so the body says References rather than Fixes.

What changed

Lite, Medium, and Heavy load on the Step 0 harness (resolveHarness). Each tier names the same modules as default and adds a declared budget. applied is false, so prompt.js, tools.js, models.js, and turns.js still run as they do today. install leaves rt.harness on default.

Triage selects a tier with exactly one label:

• harness:lite
• harness:medium
• harness:heavy

No such label, or any other label including risk:T*, stays on default. Two tier labels, or an unknown harness:* label, throw. The controller does not write these labels. They are in .github/labels.json; creating them on the repository is node .github/scripts/sync-labels.mjs --repo ThinkFlowLab/FlyBot --apply, which this PR does not run.

Declared budgets (Lite < Medium < Heavy):

• prompt maxChars: 8000 / 16000 / 48000
• tools maxCalls: 8 / 24 / 64
• tools allow: six read tools / those plus work tools, still without create_bot, delete_group, and stop_bot / null (the full set the host registers)
• model maxTokens: 2048 / 8192 / 16384
• step maxSteps: 4 / 12 / 32
• step maxRetries: 0 / 1 / 1 (one written-tool-call retry is what turns.js does today)

harnessPro is the offline check for a tier config: same four modules, verifiers.offline is ci, core/dev/eval, rsi-baseline, and no key or token field. It does not run those checks and it does not call a model.

Usage rows already carry issue/PR, tool-call count, and wall time from Step 0. rt.noteUsageHarness(sessionId, idOrLabels) stamps the routed tier on the harness column. The resume cursor stays { provider, model, retry }.

No changelog fragment: the chat UI is unchanged, and the default path does not change behavior.

Verification at the tip

Tip SHA: ea8dec628ef6da1b555ce9816d9bed41edda157b

Red on base 68e11dc (before the implementation), node --test core/test/harness.test.js core/test/usage-fields.test.js: 2 failed. harness.test.js did not load (HARNESS_PRO_VERIFIERS missing). foldCalls recorded harness default when the test passed lite.

Ran at this tip:

• node --import ./test/harness-resolve.mjs --test test/harness.test.js test/usage-fields.test.js in core/ → 16/16 pass
• npm test in core/ → 1364/1364 pass (same tree as this commit, run immediately before the commit)
• npm run lint in core/ → oxlint warnings only, the same usage.js mark?.fold ?? {} warning that already fires on main; TZ=UTC node scripts/check-arch-docs.mjs after the commit → arch-docs: OK (11 changed file(s))
• node scripts/ds-bot-patches.mjs check → pin, 10 patches, core/ equals the mirrors
• node core/dev/eval/run.mjs --list → P3-tree-split, P4-invite-approve, P6-timeline, memory-recall
• node --test .github/scripts/rsi-baseline.test.mjs .github/scripts/sync-labels.test.mjs .github/scripts/controller.test.mjs .github/scripts/risk-label-docs.test.mjs → 35/35 pass

Not run:

• npm run verify — plugin suites, interop, and live-e2e. plugins/*/node_modules is not installed here.
• Smoke bench. usage.js is under src/host/, and the bench was left out with the Fix mock admin @all and in-place compaction in smoke (#917) #925 smoke work.
• Control set of 10, and any live DSH turn. No DEEPSEEK_API_KEY.
• rsi-baseline.mjs against GitHub. That needs a token. The deterministic test above does not.
• sync-labels.mjs --apply. The labels are defined, not created on the repository.
• Merge. This pull request stays a draft.

  

Fixes #937

[Birfy]

Verdict: PASS at ea8dec6
Written-by: Reviewer2

Ran:

• diff -rq of the tip tree against dfa25a2a; compare/dfa25a2a...ea8dec62 through the API → 11 files, merge base 68e11dc8, behind_by: 3, so the branch does not revert #925/#929/#931 (three-dot diff is exactly these 11 files)
• cd tip/core && node --import ./test/harness-resolve.mjs --test --test-concurrency=1 test/harness.test.js test/usage-fields.test.js → 15 pass / 1 fail; the failure is fixture usage.sqlite has one real calls row with all five fields, and the same test fails on pristine dfa25a2a in this container (not ok 8, null !== 0), so it is the machine, not this change
• red A: tip tree with core/src/host/harness.js put back to dfa25a2a → both files fail to load (does not provide an export named 'HARNESS_PRO_VERIFIERS' / 'routeHarness')
• red B: tip tree with only core/src/host/usage.js put back to dfa25a2a → 13 pass / 3 fail, reds not ok 13 - foldCalls records a routed harness and leaves the resume state unchanged and not ok 14 - noteUsageHarness keeps a triage label and clears back to default (plus the same sqlite fixture case), so the usage.js half has its own weight
• cmp of core/src/host/harness.js, core/test/harness.test.js, core/test/usage-fields.test.js against their patches/ds-bot/files/ mirrors → identical; tail -c 1 patches/ds-bot/0001-core-fixes.patch → newline; the patch carries the new code
• node .github/scripts/enqueue-guard.mjs --repo ThinkFlowLab/FlyBot --pr 933 → REFUSE, and the second reason is not about this change: no issue reference in the pull request body: add "Fixes #<n>" ...

The four points I was asked to watch:

1. applied: false leaves the default path alone. prompt.js, tools.js, models.js, turns.js are not in the diff at all; defaultHarness() is byte-identical between dfa25a2a and the tip (I diffed the function text), and install still sets rt.harness = resolveHarness('default'). Nothing reads rt.routeHarness, so no runtime path can reach a tier by accident.
2. Labels. routeHarness throws on two different tier labels (one harness tier label) and on an unknown harness:* label; no label, risk:T2, or any other label stays default; the same label twice dedupes; names are trimmed and lower-cased, so Harness:Heavy routes. Cases are in harness.test.js and I read the code, not only the test.
3. Mirrors and the patch. Both files/ mirrors are byte-identical to core/, and 0001 ends with a newline (the trap #934 hit). Regenerating 10 patches and comparing core/ to the mirrors needs git, which this box lacks; CI's package is green at this tip, and that is the job that runs generate + check.
4. usage.js. The harness column changed in two places only: record() now falls back to an explicitly noted tier before rt.harness, and foldCalls stamps withUsageFields(row, { harness }) where an absent tier resolves to default exactly as before; state (the resume cursor) is still { provider, model, retry } and the test asserts it. Existing rows keep default.

Non-blocking:

1. The label→tier path has no caller in the runtime. rt.noteUsageHarness is exposed and tested, but nothing in core/ or plugins/ calls it, so on today's tree every row still records default; the body's "Triage selects a tier with exactly one label" is true of the API, not of a live turn. Worth a sentence in the body or the doc, since Step 1 reads as if it routes.
2. harnessBySession lives only in the process, so a restart forgets the tier and a later re-fold stamps default over rows that had one — the same lifetime issuePr already has since #908, not a new pattern, but it means the column is not durable. If it must survive a restart it belongs in marks.fold or state, which is Step 2 territory.