mirror of
https://github.com/THU-MAIC/OpenMAIC.git
synced 2026-10-02 01:15:18 +08:00
main
12
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Aris-lijinheandwyuc
|
dee421621f |
feat(skills): add 习题课（最近发展区） (#1382)
* feat(skills): add ZPD-based exercise lesson skill * fix(skills): localize exercise lesson title in all workbench locales --------- Co-authored-by: wyuc <wang-yc24@mails.tsinghua.edu.cn> |
||
|
wyuc
|
3b4567076f |
refactor(generation): consume the package everywhere and delete lib/generation (#1090)
Part D of #1057 — the switch. All consumers import @openmaic/generation; lib/generation deleted (net -11,900 lines). PBL wiring preserved (single-call -> classified fallback -> injected loop), streaming outline route deduplicated onto buildOutlinePrompt, serverExternalPackages wired, and the change verified by a real-model four-kind end-to-end run against the production build with a rendered-classroom screenshot check. Closes #1057. |
||
|
wyuc
|
3f5c51bdba |
refactor(pbl-v2): split the planner core from the loop; inject the LLM call path (#1069)
Implements #1062. Pure planner-core extracted from the loop's file; both planners receive their LLM entry as a parameter (call sites pass the real callLLM unchanged); operations/ split into kernel vs runtime with a lint-enforced boundary; planner prompts pinned byte-identical via fixtures generated from the pre-refactor implementation. |
||
|
|
5226f9589f |
fix: tighten PBL v2 planner eval harness (#805)
Co-authored-by: mzb25 <mzb25@mails.tsinghua.edu.cn> Co-authored-by: wu-yx25 <wu-yx25@mails.tsinghua.edu.cn> Co-authored-by: 805813606 <805813606@qq.com> Co-authored-by: zlnn23 <zlnn23@mails.tsinghua.edu.cn> |
||
|
杨慎andzlnn23
|
8f77d61fa4 |
test: add PBL v2 planner eval harness (#803)
Co-authored-by: zlnn23 <zlnn23@mails.tsinghua.edu.cn> |
||
|
wyuc
|
7cb129183d |
chore(packages): publish the @openmaic/* SDK family to npm (#778) (#780)
* chore(packages): publish the @openmaic/* SDK family to npm (#778) Prepares the @openmaic/{dsl,renderer,importer} family for its first npm publish, and moves the SDK packages onto the @openmaic scope. Why the scope move: the @maic org name is unavailable on npm (an unscoped `maic` package already holds the name), so @maic/* is not claimable. @openmaic matches the project name, the scope is free, and the repo already ships an @openmaic/docs package — so the SDK family now lines up with that convention. - rename @maic/{dsl,renderer,importer} -> @openmaic/* across packages, the workspace glob, the package dir, and all import sites; lockfile regenerated - renderer: add publishConfig (public, registry.npmjs.org) — was missing, so a scoped publish would default to the wrong registry / restricted access - importer: add a files allowlist (dist, README, LICENSE) and drop the fragile .npmignore blacklist that shipped src; add an exports map so ESM consumers resolve dist/index.js instead of falling back to the .cjs main - all three: add a prepublishOnly build (+ test/typecheck) guard so a publish can never ship a stale or empty dist - add a tag-triggered publish workflow with npm provenance, pinned by name to the three @openmaic packages so the vendored forks (mathml2omml, pptxgenjs) are never published Refs #778, #720 (Phase 1). * fix(packages): address cross-review on the @openmaic publish prep Cross-review (Claude /code-review + codex) on this PR surfaced: - renderer's advertised CJS entry was broken: it keeps @openmaic/dsl external and imports a runtime enum from it, but dsl is ESM-only (no `require` condition), so `require('@openmaic/renderer')` would throw ERR_PACKAGE_PATH_NOT_EXPORTED. Make renderer ESM-only: drop the `.cjs` rollup output, `main` now points at the ESM build, and the `require` conditions are removed from `exports`. (importer is unaffected — it bundles dsl, so its CJS build still works.) - prepublishOnly re-ran the test suite during `pnpm -r publish`, so a flaky test after dsl had already published gave a non-atomic partial release. Reduce prepublishOnly to a build-only guard (never ship stale/empty dist) and move the real test/typecheck gate into the workflow, before any publish. - document that an @openmaic/* tag publishes the whole family via `pnpm -r` (pnpm skips already-published versions); the tag is a release marker, not a per-package gate. Verified: dsl + renderer + importer build; renderer emits ESM only (0 .cjs), all exports entries resolve; `npm pack` ships dist + README + LICENSE with no src leak; frozen-lockfile passes. Refs #778. * style: reflow @openmaic/dsl type imports past print-width after rename The @maic -> @openmaic rename lengthened two single-line type imports past prettier's 100-col width; prettier --check flagged them. Pure formatting. Refs #778. * docs(importer): mark @openmaic/importer browser-only (cr-loop accepted limitation) codex cross-review flagged that the published @openmaic/importer throws `XMLHttpRequest is not a constructor` when loaded in a pure Node process — its rollup build is browser-targeted (`nodeResolve({browser:true})` + a browser pdf.js build). The app only consumes it client-side ('use client'), so this is by design. Document it as an accepted limitation: prominent browser-only note in the README and a `browser` field in the manifest. Refs #778. |
||
|
|
612a1471e7 |
fix(agent): respond to the user's turn before lecturing (+ answer-content eval) (#699)
* test(orchestration): add answer-content eval for agent first-reply answering The existing director question-answering eval only checks that the director ROUTES to the teacher; it never generates the dispatched agent's reply, so it cannot catch the case where routing is correct but the agent's first sentence drifts (greets / opens a lecture / pivots) and only a later turn answers. This adds an answer-content eval that runs the real agent-generate inputs (buildStructuredPrompt + convertMessagesToOpenAI), parses the structured output with the app's runtime parser, and uses an LLM judge for two signals: leads_with_answer (first sentence addresses the literal ask) and answered_anywhere (addressed at all). Their gap quantifies "drift-then-answer". A/B mirrors the answering-runner rule-13 strip: baseline removes the "Answering the User's Question" section from agent-system, with_rule is as-shipped. Scenarios are synthetic/anonymized (opening-lecture override, ignored format/capability/navigation request, ignored correction, frustration re-ask, adjacent pivot, vague-clarify) plus clean controls; no real user data. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * test(orchestration): give answer-content scenarios live-shaped classroom context The minimal store state (empty scenes, null stage, no profile) made the assembled agent prompt far thinner than the live agent-generate path and suppressed the slide-narration "lecture reflex" that drives opener drift. Each scenario now carries a topical slide scene (title + body + latex elements with ids), a stage (name/description/languageDirective), and a student profile, so buildStructuredPrompt emits a "# Current State", "# Student Profile" and "# Language" of the same shape and bulk as production. The runner reads this context from the scenario instead of a hardcoded empty state. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(agent): respond to the user's turn before continuing the lecture Generalize the agent-system "Answering the User's Question" section into "Responding to the User's Turn": the user's latest message always takes priority over continuing the planned lecture. Beyond factual questions it now explicitly covers the interaction requests agents were ignoring — navigation/pacing ("skip to next page"), format/language ("explain in Chinese"), capability requests it cannot fulfill ("make a video" -> say so + offer an alternative), corrections, frustration re-asks, and vague asks (ask one clarifying question). Strengthen the answer-content eval that measures this: - the navigation case now uses a 3-slide deck so "skip to next page" is well-posed; runner supports multi-scene state and strips the renamed section for the A/B; - judge is fairer to legitimate behavior: explaining primarily in the requested language while keeping technical terms standard counts; a verbal transition to the next slide counts as honoring navigation (no page-turn action exists). Result (gemini-3-flash-preview, n=10): mean leads-with-answer rises from 42% (section stripped) to 93% (as-shipped); all 12 scenarios pass. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(agent): don't over-respond to bare acknowledgements; guard empty eval set Code-review follow-ups on the "Responding to the User's Turn" change: - Prompt: the absolute "every user message needs a direct response / treat continuing the lecture as a failure" wording had no carve-out for bare acknowledgements ("ok" / "嗯" / "good" / "thanks"), risking the agent halting the lesson to manufacture a reply, and conflicted with agent-initiated discussions where no user turn exists. Add a bare-acknowledgement carve-out, allow a brief "好的"/"Sure" before the answer, and scope the override to actual requests (and to "if there is a user turn to respond to"). - Eval: answer-content-runner exited 0 with a PASS verdict and NaN% aggregates when the scenario set was empty (e.g. an EVAL_SCENARIO typo). Guard it: error and exit 1 when no scenario matches. Re-run unchanged otherwise (gemini-3-flash-preview, n=10): mean leads-with-answer 42% (stripped) -> 94% (as-shipped), 11/12 scenarios pass; the chem opener is variance-prone around the 70% bar and not over-tuned. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * refactor(eval): cr-loop cleanups on answer-content-runner - ScenarioAgentSpec extends the shared ScenarioAgent (./types) instead of re-declaring id/name/role/priority. - Drop the dead expectedPreFix plumbing from ScenarioResult (it was copied in but never read; it stays as an informational annotation in the scenarios). - extractTexts: replace the hand-rolled JSON fallback with the app's finalizeParser, which recovers plain-prose / unclosed-array output the same way the runtime does. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix: tighten acknowledgement carve-out; don't surface action-JSON as text (cr round 2) Round-2 review follow-ups. Prompt: the bare-acknowledgement carve-out was too loose — it could swallow ack+request hybrids ("嗯，那为什么X?") as acks (re-introducing drift), double-listed "好的" as both an allowed lead-in and a no-response ack, and listed pacing words ("继续"/"go on") that collide with the navigation bullet. Restrict it to a STANDALONE acknowledgement with nothing else attached; route an embedded question/request to that request; route pacing words to the navigation bullet. Eval: extractTexts gated finalizeParser recovery on !state.jsonStarted, so an actions-only / unclosed array yields "no text" (the judge's intended empty case) instead of finalizeParser's raw-buffer fallback surfacing action JSON as fake speech. Re-run (gemini-3-flash-preview, n=5): no regression — 11/12 pass, meta-requests 100/100/80/100; the chem opener remains variance-prone around the 70% bar. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix: address review — honest slide-nav, strict judge booleans, errors fail (cr) Review follow-ups from @cosarah: - Navigation (P2): agents have no slide-navigation action, so "do it" + a verbal transition let "skip to next page" pass the eval while the UI stayed on the same slide. Reframe prompt, judge, and scenario so a correct reply ACKNOWLEDGES the request and is HONEST that it cannot change the slide itself (continue verbally / tell the user how to navigate); pretending it flipped the slide no longer counts. - Judge booleans (P2): answer-content-judge coerced fields with Boolean(...), so a string "false" became true. Parse strictly (real bool or "true"/"false"); malformed output is flagged as an error. - Error accounting (P3): aggregate excluded errored samples from the denominator, letting a scenario pass on one good sample. Errors now count as failures (denominator = all requested samples). Re-run (gemini-3-flash-preview, n=10): navigation 0% -> 100% under the honest framing; aggregate leads-with-answer ~93%, no regression. Opener scenarios (math/chem) remain variance-prone near the 70% bar. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(eval): judge first-sentence lead; use canonical teacher actions (cr round 2) Second review round from @cosarah: - leads_with_answer was judged against texts[0] (the whole first JSON text block), so "Today we'll discuss parabolas. The formula is x=-b/(2a)." counted as leading even though the first sentence drifted. Now extract the first ~2 sentences as the "opening" and have the judge make the call: a brief acknowledgement of the user before the answer counts as leading, but a greeting / self-intro / topic preamble before the answer does not. (Two sentences + semantic judge handles "Sorry, let me clarify. m is inertial mass." vs "Welcome! …" without a brittle keyword list.) - The runner hardcoded TEACHER_ACTIONS, omitting canonical teacher actions (play_video, wb_draw_chart). Use getActionsForRole('teacher') so the available- actions section matches production. Re-run (gemini-3-flash-preview agent, n=10): 11/12 pass, aggregate leads ~90%. The chem opener now correctly fails (~10%): its "Welcome! Today we…" persona greets first, so at sentence granularity it is genuine first-sentence drift (answered-anywhere ~90%) — exactly what the stricter lead measurement should catch. Note: the judge model is deepseek-v4-pro in this run because the anthropic gateway is currently returning Bad Request in our environment; the eval is judge-model-agnostic via EVAL_JUDGE_MODEL. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(eval): judge the opening across all text blocks, not just texts[0] (cr round 3) Review follow-up from @cosarah: leadFromTexts only looked inside the first text block, so a valid multi-block reply like [{"type":"text","content":"Sure."},{"type":"text","content":"The derivative is 2x."}] was judged on "Sure." alone and failed leads_with_answer despite following the allowed acknowledgement-then-answer pattern. Build the opening from the first ~2 speech sentences across ALL ordered text chunks. Re-run (gemini-3-flash-preview, n=10, deepseek judge): 11/12 pass, aggregate leads ~93%. The chem opener stays sub-threshold (~50%) — its "Welcome! Today we…" persona greets before answering about half the time, a genuine first-sentence drift. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com> Co-authored-by: 杨慎 <117187635+cosarah@users.noreply.github.com> |
||
|
wyucandClaude Opus 4.8
|
4dce1f7414 |
refactor(types): import the slide DSL directly from @maic/dsl; drop the shim (#738)
#707 left lib/types/slides.ts as a thin `export * from '@maic/dsl'` shim so the ~100 existing `@/lib/types/slides` import sites kept working unchanged. Point them at `@maic/dsl` directly and delete the shim, so the app consumes the package contract with no indirection and the legacy module path is gone. Pure module-specifier migration — every symbol imported from `@/lib/types/slides` was already a re-export of the same `@maic/dsl` symbol, so behavior is identical. Also redirects the two `lib/types` relative importers (`./slides`), the e2e fixture's relative import, and the dynamic `import('../types/slides').Slide` type alias in stage-storage; merges the now same-module import pair in slide-edit-elements. Verified: tsc --noEmit clean, eslint clean, vitest 804/804, next build OK, e2e (Playwright) 19/19. Part of #720 (Phase 1). Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com> |
||
|
wyuc
|
f064590a5e |
test+fix(orchestration): rip maxTurns; restore answer-the-user's-question behavior (#598 follow-up) (#599)
* test(orchestration): add premature-END regression eval (#598 follow-up) Adds eval/orchestration/ following the outline-language pattern: a runner that A/B-tests the director against the same scenario with two prompt variants — current main (post-#554) and a synthesised pre-#554 baseline that strips the role-aware summary labels AND the new system.md rules 10/11/12 together. Pass criterion is now framed as a regression guard rather than a fixture discrimination test: every scenario's post-fix END rate must stay below EVAL_END_THRESHOLD (default 20%). The pre-vs-post Δ is reported as informational data, since #554's reviewer feedback was that earlier fixtures didn't discriminate. Empirical finding across 7 shipped model configs (gpt-4.1-mini, gpt-4o-mini, gpt-5.4-nano, qwen-plus, qwen3.5-flash, deepseek-chat, deepseek-v4-flash, gemini-2.5-flash, claude-haiku-4-5) with 5 scenarios modelled on #511 (incl. the exact Tiananmen-3D objection trace, soft pushback after a long resolved-looking discussion, topic pivot, brief acknowledgement, and explicit teacher-signals-end-then-user-objects): no scenario produced a non-zero END rate in either variant. Useful data for the #554/#598 discussion — the prompt-layer rules don't measurably change behavior on these shipped models with these prompts, but the eval is now wired so future regressions show up. * test(orchestration): add director question-answering eval (#598 follow-up) Adds an A/B eval probing whether the director routes correctly when the conversation contains an unanswered user question — both in the first-turn drift case (agents started answering but drifted onto adjacent topics, no frustration yet) and in the escalated frustration case (user has already complained multiple times). The prod symptom this targets ("答非所问"): agents reply on adjacent topics rather than answering the literal question, and the director keeps picking peer agents for "variety" instead of routing back to the teacher to actually answer. Scenarios: 5 synthesized cases across math, biology, English grammar, physics, and calculus. 3 are first-turn drift (no frustration yet), 2 include frustration signals. Each captures: user asks specific question, agents reply on adjacent topics, director is asked to pick the next move. Decision rule-judged into USER | TEACHER | OTHER_AGENT | END: - USER → ✓ valid (cue user to clarify) - TEACHER → ✓ valid (re-route to teacher to answer) - OTHER_AGENT → ✗ wrong (peer-agent "variety" routing — the bug) - END → ✗ wrong A/B: - baseline : current director template with rule 13 stripped - with_rule : current director template as-shipped Pass = with_rule.correctRate ≥ EVAL_PASS_THRESHOLD (default 0.7). Pre-vs-post Δ is reported as informational only. Note this commit only adds the eval; the rule it strips/restores in the A/B is injected in the next commit. * fix(orchestration): inject director rule for answering user questions Adds rule 13 to lib/prompts/templates/director/system.md so the director recognizes when the conversation contains an unanswered user question and routes appropriately — to the teacher (literal role match) by default, or to USER cue when the question is ambiguous or the user has already complained multiple times. This addresses the "答非所问" symptom: when agents drift onto adjacent topics, the director keeps picking peer agents for "variety" rather than routing back to the teacher to actually answer. The rule is preventive (applies whenever the question is unanswered, including the first turn after drift) rather than reactive-only (waiting for the user to express frustration). Rule wording uses constrained-output framing: literal `role: teacher` match (not "best fit" / "highest priority" / "most knowledgeable"), and an explicit list of forbidden choices (peer agents, END). Earlier softer wordings let the model pick the assistant when it seemed more topically relevant; the literal-match wording prevents that. Validated by the question-answering eval added in the previous commit. On google:gemini-3-flash-preview, 5 samples per variant across 5 synthesized scenarios: | Scenario | baseline | with_rule | |----------------------------------------------------|----------|-----------| | math_quadratic_axis_drift_no_frustration | 100% | 100% | | bio_dark_reaction_drift_no_frustration | 80% | 100% | | english_team_isare_drift_no_frustration | 20% | 100% | | physics_inertial_mass_escalated_frustration | 0% | 100% | | calculus_product_rule_drift_no_frustration | 60% | 100% | | **mean** | **52%** | **100%** | 3 of 5 scenarios reproduce the bug at baseline (correct rate ≤ 60%); the rule lifts every scenario to 100%. The remaining 2 baselines are already high — the rule doesn't regress them. * fix(orchestration): teach agents to answer the user's literal question Adds an "Answering the User's Question" section to the agent system prompt. Before this, the agent SP had no rule for direct-answering — the persona, role guideline, current-state context, and length guidelines all pushed the agent to advance the lesson and inspire thought, with no explicit directive to first answer the user's literal question. The result observed in prod ("答非所问"): when the user asks a specific question (a formula, a yes/no, a term), agents reply with adjacent concepts, examples, or follow-up questions — but never the literal answer. The new section directs the agent, on every response where the user message contains a question, to: - lead with the concrete answer to the literal question, - not pivot to an adjacent topic even when it seems pedagogically richer, - treat "inspire thought" and peer-differentiation guidance as applying only AFTER the literal question is answered, - say "I don't know" directly when uncertain rather than answering a different question, - on frustration signals from prior turns, look back at the original question (before the frustration) and answer THAT specifically. This is the agent-side counterpart to the director rule landed in the previous commit. Director routes to the teacher when the question is unanswered; the teacher now has explicit guidance for what to do when dispatched. * refactor+fix(orchestration): rip maxTurns end-to-end; distinguish errors from completion Two coupled cleanups in one commit since the outcome-handler rewrite naturally falls out of removing the client-side max-turns loop. The director graph used `maxTurns: turnCount + 1` to force "one director→agent cycle per request" — a hack expressing the single-round contract as a turn limit. This commit removes the hack and the surrounding user-facing setting, replacing implicit intent with explicit graph topology. Server: - `director-graph.ts`: drop `maxTurns` from state, drop the hard-cap check at the top of `directorNode`, change `agent_generate → director` edge to `agent_generate → END`. Single-round contract is now expressed by topology. - `lib/types/chat.ts`: drop `maxTurns` and `currentTurn` from `SessionConfig`; drop `maxTurns?` from `CreateSessionRequest.trigger`. Client: - `lib/store/settings.ts`: drop `maxTurns` state, `setMaxTurns` action, the `'10'` default, and the localStorage migration path. - `components/agent/agent-bar.tsx`: drop the entire compact-stepper UI block plus the unused MessageSquare/Minus/Plus lucide imports. - `components/settings/agent-settings.tsx`: drop the AgentSettings prop and the multi-agent-only Input block. - `lib/chat/agent-loop.ts`: drop `maxTurns` parameter, change `while (turnCount < maxTurns)` to `while (true)` controlled by inner exits, drop `'max_turns'` reason from `AgentLoopOutcome`. - `components/chat/use-chat-sessions.ts`: drop the maxTurns lookup, drop the trailing arg to runAgentLoop, drop vestigial `maxTurns: 0` / `currentTurn: 0` from session creation. - `eval/whiteboard-layout/runner.ts`: drop unused MAX_AGENT_TURNS plus the trailing arg. i18n: drop `settings.maxTurns` and `settings.maxTurnsDesc` from six locales (en-US, zh-CN, zh-TW, ja-JP, ar-SA, ru-RU). Motivation for removing the user-facing setting: the LLM director controls round length via cue_user / END decisions, so a numeric "max discussion turns" knob is redundant with rule-driven routing. When the director prompt was buggy, the cap masked the bug by force- ending; once director routing is fixed, the cap just removes user agency. Before, every non-`cue_user` outcome of the agent loop mapped to `status: 'completed'` in `use-chat-sessions.ts`. That conflated four distinct paths: - `end`: director chose END (legitimate completion) - `empty_turns`: two consecutive agents returned no text or actions (parser failure, model timeout, etc.) - `no_done`: SSE stream ended without a `done` event (infrastructure failure) - `aborted`: user pressed stop (handled elsewhere) The new switch routes `end` to completed, `empty_turns` and `no_done` through `clearLiveSessionAfterError` so the user sees an inline System message describing what went wrong instead of a quiet "completed". - `lib/types/chat.ts`: add `'error'` to `SessionStatus`. - `components/chat/use-chat-sessions.ts`: rewrite the outcome handler as a switch; have `clearLiveSessionAfterError` set `status: 'error'` so all entry points (the runAgentLoop outcome AND existing throw paths from fetchChat / SSE error events) converge on the same visible state. - `components/chat/session-list.tsx`: add `AlertCircle` red icon for the new error status. - i18n: add `chat.error.emptyAgentResponses` and `chat.error.streamInterrupted` strings to all six locales. This is the in-product fix that the prod investigation surfaced as an independent root cause of the #511 "discussion ended unexpectedly" symptom — separate from the director and agent prompt fixes in the earlier commits of this PR. * fix(chat): drop dead settingsState binding; wire eval:orchestration:answering script Two minor CR follow-ups: - components/chat/use-chat-sessions.ts: the `settingsState` lookup at the top of runAgentLoopFn was the entry point for the maxTurns lookup ripped in the previous commit; with maxTurns gone, the binding is unused and ESLint flags it. Drop the line. (Other uses of useSettingsStore.getState() elsewhere in this file are unaffected.) - package.json: the answering eval runner had no convenience script, while the premature-end runner did. Add eval:orchestration:answering for symmetry and discoverability. * fix(orchestration): route ambiguous/unaddressed user turns to teacher, never USER cue Addresses reviewer issues 2 & 3 (ambiguous "帮我看一下这个" / "Can you take a look at this?"): the director was cueing USER for vague input, which makes no agent speak — the user faces dead air with nothing to react to ("nobody responds"); and when an agent did speak it guessed a topic instead of asking what the user meant. Root cause was the previous rule 13, which offered `{"next_agent":"USER"}` as a valid response to an ambiguous user turn. A USER cue is the wrong tool here: it returns control to the user without any agent speaking. The right behavior is to dispatch the teacher, who either answers or — if the message is too vague — asks a clarifying question out loud. Director (`director/system.md` rule 13): - When the most recent user turn is unaddressed (clear question, vague request, OR frustration signal), the only acceptable output is the literal `role: teacher` agent id. - USER cue is now explicitly forbidden as a response to a user turn (reserved for rule 7, where an agent has just asked the user something). - Tightened the END prohibition: a high turn count or a thoroughly-discussed topic does NOT mean the specific question was answered. Added rule 5 (don't-drag-on) to the override list, since that pressure was driving the director to END on well-discussed-but-unanswered turns. Agent (`agent-system/system.md`): - Added a clause to "Answering the User's Question": if the user's message is too vague to answer (no clear referent), ask ONE short clarifying question with a concrete option or two — do not guess a topic and lecture, do not stay silent. Eval (`answering-runner.ts` + `scenarios/answering.json`): - Reclassified USER as wrong (dead air); only TEACHER counts as correct now (teacher answers, or asks a clarifying question for vague input). - Added two ambiguity scenarios: a bare "帮我看一下这个" opener, and a vague "再帮我看看这个吧" follow-up after a real Q&A. - Made the rule-13 strip regex match by rule number rather than heading text, so it survives heading rewrites. On google:gemini-3-flash-preview, 10 samples per variant, all 7 scenarios PASS (with_rule mean ~97%, baseline ~57%). The two ambiguity scenarios: ambiguous_look_at_this 100%/100%, ambiguous_vague_followup 20%→100%. Manually verified the agent side: for "帮我看一下这个" the teacher now replies "…你想讨论的是具体的代码逻辑，还是 Agent 的决策流程图呢?" (clarifying question + concrete options) instead of guessing. Note: reviewer issue 1 (follow-up not answered) could not be reproduced on this branch across 7 configurations; it is consistent with the deployed environment running older code and/or provider API flakiness, which the C7 error-vs-completion distinction in this PR already surfaces instead of a silent 已结束. * docs(eval): fix stale physics scenario description after USER→wrong reclassification The physics_inertial_mass scenario description still said "USER-cue signal — but TEACHER re-answer is also valid", which contradicts the new classifier (USER is dead air = wrong; only TEACHER is correct). Cosmetic only — the description field is printed into the report, never used in classification. |
||
|
|
fc6b186929 |
refactor(whiteboard): file-based prompts + geometry conflict detection (#485)
* feat(prompts): declare whiteboard role prompt IDs + whiteboard-reference snippet * feat(prompts): add whiteboard role template skeletons + reference snippet stub * refactor(orchestration): move buildWhiteboardGuidelines to role-specific markdown templates * docs(prompts): whiteboard guidelines now in markdown templates * feat(prompts): whiteboard-reference canvas + JSON output sections * feat(prompts): whiteboard-reference action schemas for text/shape/line * feat(prompts): whiteboard-reference action schemas for latex/chart/table * feat(prompts): whiteboard-reference action schemas for code/delete/clear/close * feat(prompts): whiteboard-reference LaTeX JSON escape section * feat(prompts): whiteboard-reference bounds + overlap section * feat(prompts): whiteboard-reference font + latex height tables * feat(prompts): whiteboard-reference pre-output checklist * feat(prompts): expand teacher whiteboard template * feat(prompts): expand assistant whiteboard template * feat(prompts): expand student whiteboard template * test(prompts): assert whiteboard-reference sections reach every role prompt * tune(prompts): cap teacher element budget at 4/turn, encourage wb_clear on crowded board * tune(prompts): simplify teacher template — single conservative rule + state-awareness emphasis * tune(prompts): add LaTeX-height × text-fontSize pairing table for visual weight consistency * feat(orchestration): preserve image parts from UIMessage to AI SDK ModelMessage * feat(eval): attach prior-turn whiteboard screenshot as user message image part Gated behind EVAL_ATTACH_PRIOR_SCREENSHOT=1 env. When on, captures a screenshot after every turn (not just checkpoints) and attaches it as a file part with mediaType:'image/png' on the next user message. Teacher prompt gets a new 'Prior-state image' section telling the agent to treat it as direct visual feedback. * fix(orchestration): strip data: prefix before passing image to AI SDK Vercel AI SDK's streamText/generateText treats ImagePart.image as a URL to fetch when it's a string. data: URLs fail the http/https scheme check and throw AI_DownloadError. Strip the data URL prefix and pass raw base64 with mediaType separately — AI SDK treats base64 strings as data content. * refactor(whiteboard): add conflict summarizer, drop prior-screenshot feature Eval across flash (image-on vs off × repeat-3) and pro (same) showed the prior-screenshot image-input feature gives ~0 net benefit: +0.4 overall on flash and −0.6 on pro. The image feedback loop's theoretical upside (agent self-correcting from visual) didn't materialize — weak models couldn't act on the image and strong models "found things to fix" and over-corrected. Meanwhile programmatic geometry detection on the raw JSON cleanly lifted overlap from 6.3 → 8.1 on flash. So: remove the image-input plumbing, keep a programmatic equivalent. ## Removed (reverts |
||
|
|
7d47b0ad3a |
refactor(eval): unify outline-language and whiteboard-layout harness (#453)
* feat(eval): add resolveEvalModel helper with fail-fast Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(eval): add createRunDir helper with path sanitization Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * feat(eval): add markdown report helpers * refactor(eval): move language test cases under eval/outline-language * feat(eval): add outline-language types * feat(eval): add outline-language LLM judge * feat(eval): add outline-language reporter * feat(eval): add outline-language runner entry * chore(eval): add eval:outline-language pnpm script * refactor(eval): adopt shared createRunDir and drop gpt-4o fallbacks in whiteboard runner * refactor(eval): drop redundant non-null assertions after narrowing Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * refactor(eval): remove SCORER_MODEL_DEFAULT hardcoded gpt-4o fallback * chore(eval): delete tests/generation and clean up vitest/gitignore config * docs(eval): explain why outline-language runner pre-validates env vars Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(eval): escape pipe chars in markdown summary table cells LLM judge output may contain | which breaks GFM table rendering. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> Co-authored-by: 杨慎 <117187635+cosarah@users.noreply.github.com> |
||
|
|
ea0e8126a5 |
feat: whiteboard layout quality eval harness (#425)
* feat(eval): add state manager bridging ActionEngine for eval Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat(eval): add shared types for whiteboard layout eval harness * feat(eval): add SSE chat client for whiteboard eval * feat(eval): add Playwright capture module for whiteboard screenshots * feat(eval): add VLM scorer for whiteboard layout evaluation * feat(eval): add report generator for whiteboard eval results * feat(eval): add 8 constructed scenarios for whiteboard layout eval * feat(eval): add minimal whiteboard render page for Playwright screenshots Creates app/eval/whiteboard/page.tsx — a headless client page that seeds the stageStore with a synthetic slide scene, exposes window.__setElements() for Playwright to inject PPTElement[], and renders them via ScreenElement inside a 1000×562.5px white canvas. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat(eval): add main runner for whiteboard layout eval Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat(eval): add eval:whiteboard script, install tsx, gitignore results * fix(eval): fix TS errors, lint, and prettier formatting * fix(eval): address code review — add cue_user/empty turn guards, validate VLM output, fix empty array crash * refactor(eval): replace synthetic scenarios with realistic ones - Replace 8 generic scenarios with 6 that match real usage patterns - Multi-agent discussion with short user replies (嗯, 明白了, 继续) - Include real slide scene data as initialStoreState - Generated agent configs with Chinese names and proper roles - Cover: physics, math, finance, primary school, economics, medical Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * refactor: extract shared agent loop from use-chat-sessions Extract the core agent loop logic into lib/chat/agent-loop.ts as a pure async function with callback injection. Both the frontend React hook and the eval harness now share the same loop — SSE parsing, exit conditions (END/cue_user/empty turns/max turns), and director state accumulation. The frontend wires StreamBuffer callbacks for UI pacing; the eval wires ActionEngine + message accumulation for headless execution. If loop logic changes in the shared module, both consumers automatically stay in sync. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix(eval): use project LLM infrastructure, fix eval page and model config - Rewrite scorer to use resolveModel() + generateText() from AI SDK instead of raw fetch — supports all providers (OpenAI, Google, Anthropic) - Model config via env vars (EVAL_CHAT_MODEL, EVAL_SCORER_MODEL), matching the pattern from outline-language eval - Fix eval page: bootstrap store before SceneProvider mounts - Fix __dirname for tsx CJS mode - Remove --api-key/--scorer-model CLI args (use env vars instead) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat(eval): organize results by model/timestamp * fix: remove double turnCount increment in shared agent loop The extracted agent-loop.ts had both `turnCount = directorState?.turnCount ?? turnCount + 1` (line 190) and a redundant `turnCount++` (line 215), causing multi-agent scenarios to hit maxTurns at half the expected number of iterations. Also removes unused processSSEStream import from use-chat-sessions.ts. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat(eval): organize screenshots by scenario subdirectory Results structure: results/<model>/<timestamp>/<scenario>/run0_turn1.png Report files stay at the timestamp level. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat(eval): revise scorer rubric and add rescore mode - Replace space_utilization with rendering_correctness and content_completeness - Rubric now evaluates from a teacher's perspective (empty space is normal) - readability emphasizes font size consistency - Add --rescore flag to re-score existing screenshots without re-running chat - Increase maxOutputTokens to 2000, add JSON parse error recovery - Score errors no longer abort the entire scenario Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat(eval): sharpen scorer rubric with teacher-perspective examples The rubric now catches specific classroom whiteboard failure modes: - overlap now explicitly penalizes writing over existing content when empty space is available (spatial planning failure) - rendering_correctness calls out diagram accuracy (e.g., parabola drawn as V-shape), raw subscripts (G_x), Chinese inside LaTeX math mode - content_completeness emphasizes canvas edge clipping and bare unlabeled diagrams - readability penalizes text styled as UI components (gray card backgrounds) - overall instructed to weight overlap and rendering_correctness more heavily - explicit note to ignore the "N" page UI element Also increases maxOutputTokens to 3000 since longer rubric produces longer justifications. Reporter now guards against null scores (scorer failures no longer crash report generation). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: address code review findings Critical fixes: - use-chat-sessions: restore agent_end handling, currentMessageId fallback for text_delta/action events with missing messageId, and re-throw on SSE error events (previously silently pushed to buffer only). - eval runner: serialize ActionEngine executions via promise chain. void-fire-and-forget raced with ensureWhiteboardOpen's 2s delay and could insert elements out of order or before the first element was committed to the store. Important fixes: - CHAT_MODEL default: 'openai/gpt-4o-mini' -> 'openai:gpt-4o-mini' (parseModelString splits on ':', not '/'). - CheckpointResult.score is now VlmScore | null; removed the 'as unknown as' cast that hid the null contract from consumers. - Delete dead code: eval/whiteboard-layout/chat-client.ts and components/chat/process-sse-stream.ts (both unused after the shared agent loop refactor). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Co-authored-by: 杨慎 <117187635+cosarah@users.noreply.github.com> |