feat(agentic): add Hidden Intent proactivity tracking framework (pi-Bench)

[harryfan1985]

Overview

This PR adds the platform-neutral groundwork for Hidden Intent / proactivity tracking in BitFun agent sessions, inspired by the pi-Bench paper (arXiv 2605.14678).

The current implementation does not claim to fully reproduce pi-Bench's hidden-intent evaluator. Instead, it introduces the session/config/data contracts, prompt guidance, runtime evidence capture, and report fields needed to evaluate whether an agent proactively handles latent requirements once real hidden-intent assignments are available.

Paper Source

pi-Bench: Evaluating Proactive Personal Assistant Agents in Long-Horizon Workflows
Zhang et al., arXiv 2605.14678, May 2026
https://arxiv.org/abs/2605.14678

The paper evaluates proactive personal assistants through hidden intents with three terminal states:

• Completed: the agent directly satisfies the hidden intent without the user explicitly providing it.
• Inferred: the agent asks a targeted clarification and the user reveals the hidden intent.
• Provided: the user must proactively supply the hidden intent.

Proactivity Score = (Completed + Inferred) / Total Hidden Intents.

Task completeness is a separate final-output/task-requirement judgment in the paper, not something that should be inferred from the hidden-intent terminal states.

What This PR Changes

1. Prompt Guidance

• agentic_mode.md: adds guidance for coding agents to infer likely requirements from workspace context and ask targeted questions when information is missing.
• claw_mode.md: strengthens personal-assistant proactivity guidance, including preference/context recovery across longer workflows.
• facet_extraction.md: extends the session-insights extraction prompt with proactivity and completeness fields. This is prompt/schema guidance only; it is not yet wired as the authoritative hidden-intent assignment grader.

2. Platform-Neutral Data Contracts

• hidden_intent_types.rs: adds DTOs/enums for hidden intents, persistent intents, terminal statuses, session-level tracking, raw turn evidence, and score/report value types.
• session/types.rs: adds persisted intent_assignments and intent_evidence on dialog turns, plus session metadata fields for intent tracking and optional score snapshots.
• core/session.rs: adds enable_intent_tracking to SessionConfig, defaulting to false.
• services-core owns the shared contracts so the logic stays platform-agnostic and can be exposed through desktop/web/server adapters.

3. Runtime Evidence Collection

• intent_evidence.rs: collects lightweight per-turn trajectory signals such as targeted user-question usage, question topics, proactive tool calls, output production, and round count.
• round_executor.rs: detects AskUserQuestion tool usage and extracts question-topic hints from tool-call arguments.
• execution_engine.rs: accumulates evidence during the turn and persists a snapshot after the dialog loop completes.
• coordinator.rs: creates an IntentEvidenceCollector only when enable_intent_tracking=true.
• session_manager.rs: persists raw evidence to both session metadata and the dialog-turn file without converting it into hidden-intent terminal assignments.

4. Session Usage Report Surface

• session_usage/types.rs: adds optional proactivity and completeness report fields.
• session_usage/service.rs: aggregates real hidden-intent assignments into a proactivity report when such assignments exist.
• Completeness remains unset until a dedicated checklist/rubric/rule-based grader is implemented.
• Legacy proxy-style turn-* assignments generated from raw evidence are ignored so old heuristic data is not reported as real hidden-intent evaluation.

5. Frontend / Adapter Plumbing

• Desktop API request DTOs accept enable_intent_tracking and pass it into SessionConfig.
• Flow chat config types and session creation paths propagate enableIntentTracking.
• Session API/report typings include optional proactivity/completeness report fields.
• Session history typings include optional intentEvidence, so future UI/report features can inspect raw evidence separately from hidden-intent assignments.

Incremental Refactor / pi-Bench Alignment

A follow-up refactor tightened the implementation against the paper's functional model:

• Separated evidence from assignment: runtime signals are stored as IntentTurnEvidence, not synthetic IntentAssignment rows.
• Corrected proactivity semantics: score helpers use the full hidden-intent count as denominator and return unavailable while any hidden intent lacks a terminal status.
• Removed false completeness derivation: completeness is no longer computed from Completed/Inferred/Provided; it is reserved for an independent final-task grader.
• Preserved compatibility: serde defaults/aliases keep older session files readable, and usage reports filter old proxy assignments.
• Kept boundaries clean: shared DTOs live in services-core, execution evidence collection lives in bitfun-core, and UI code consumes typed API/session-history data.

Current Limitations / Follow-ups

• This PR does not yet implement the full hidden-intent discovery/assignment evaluator described by pi-Bench.
• Persistent-intent DTOs are present, but cross-session memory recovery and application are not fully wired.
• Completeness needs a separate grader over the final trajectory/artifacts and task requirements.
• AskUserQuestion topic extraction depends on the current tool-call argument shape and should be revisited if the tool schema changes.
• Proactivity reporting is meaningful only when real hidden-intent assignments are produced by a future evaluator or imported evaluation data.

Follow-up TODO: Validation and Optimization

Validation TODO

• Add fixture-based tests with hand-authored hidden intents and expected terminal statuses to verify Completed, Inferred, and Provided classification semantics independently from runtime evidence collection.
• Add replay tests for saved sessions with enable_intent_tracking=true, covering metadata persistence, turn-file intentEvidence, and usage report aggregation after reload.
• Add backward-compatibility tests for older session files that contain no intent fields, plus sessions that contain legacy proxy-style turn-* assignments.
• Add remote-workspace coverage to confirm evidence persistence and usage-report generation behave the same for local and remote sessions.
• Add a small evaluator-golden set once the assignment evaluator exists, with cases for direct satisfaction, targeted clarification, generic clarification that should not count as Inferred, and user-provided hidden intent.
• Add completeness-grader tests separately from proactivity, using task checklists/rubrics over final artifacts so completeness does not regress into hidden-intent status counting.

Optimization TODO

• Introduce a post-hoc hidden-intent assignment evaluator that compares concrete hidden intents against the full trajectory in two stages: direct satisfaction first, then targeted elicitation.
• Wire persistent intents into the workspace/session memory layer so preferences can be recovered across sessions without coupling core logic to desktop-specific storage.
• Replace fragile question-topic extraction with a typed AskUserQuestion tool contract or versioned parser to avoid silent drift when the tool schema changes.
• Consider batching or coalescing post-turn evidence persistence if high-volume tracking sessions show measurable metadata I/O overhead.
• Add UI affordances for raw evidence vs evaluated assignments, making it clear when a report is evidence-only, fully evaluated, or unavailable.
• Add report coverage metadata for proactivity/completeness so consumers can distinguish not evaluated, partially evaluated, and fully evaluated states.
• Revisit proactivity thresholds after collecting real coding-assistant sessions; pi-Bench-style labels may need calibration for BitFun's coding and desktop workflows.

Risk Assessment

Low Risk

• enable_intent_tracking defaults to false, so evidence collection is opt-in.
• New persisted fields use Option, Vec, serde defaults, and aliases for backward-compatible deserialization.
• Raw evidence and terminal assignments are stored separately, reducing the risk of misleading reports.

Medium Risk

• Enabling tracking adds post-turn metadata/turn-file persistence work.
• The current report surface includes reserved completeness fields before the completeness grader exists.
• Future evaluator work must be careful to distinguish targeted clarification (Inferred) from generic questions or passive waiting.

Verification

• cargo test -p bitfun-services-core hidden_intent -- --nocapture
• cargo test -p bitfun-core intent_evidence -- --nocapture
• cargo test -p bitfun-core report_ -- --nocapture
• cargo check --tests -p bitfun-services-core
• cargo check --tests -p bitfun-core
• cargo check --tests -p bitfun-desktop
• pnpm run type-check:web
• pnpm run lint:web
• pnpm --dir src/web-ui run test:run (139 files / 744 tests passed)

---

Generated with BitFun

[GCWing]

This PR involves significant changes and affects the Agentic agent; it will be considered for merging after verification.

[harryfan1985]

This PR involves significant changes and affects the Agentic agent; it will be considered for merging after verification.

sure!

[harryfan1985]

Follow-up: 代码审查问题修复 (commit 4ffbb24)

针对此前审查中发现的问题，已在最新提交中修复并补充回归测试。

修复内容

问题	文件	修复
AskUserQuestion 漏报	round_executor.rs	detect_ask_user_question 此前在 questions[].header 缺失时返回 (false, [])，导致工具调用被静默丢弃。改为用独立的 called 标记记录调用本身，topic 提取保持 best-effort。
未使用参数	session_manager.rs / execution_engine.rs	删除 record_intent_evidence 中从未使用的 _turn_id 参数（实际通过 evidence.turn_index 定位），同步更新唯一调用方。
Mutex 中毒静默丢失	execution_engine.rs	证据收集器锁中毒时由静默跳过改为输出 warn! 日志，便于排查。
阈值三处重复	hidden_intent_types.rs / intent_evidence.rs / service.rs	将 0.8/0.5/0.2 等级阈值统一收敛到 ProactivityLevel::from_score() 与 as_str()，另外两处改为代理调用。
代理赋值检测脆弱	hidden_intent_types.rs / service.rs	IntentAssignment 新增显式 is_proxy: bool 字段（serde 默认 false，向后兼容）。is_legacy_proxy_intent_assignment 优先读取该字段，原 intent_id.starts_with("turn-") 字符串启发式保留为旧数据兜底，避免误判真实意图。
单条 Provided 过滤无说明	service.rs	补充注释，解释为何单条 Provided(total=1) 不构成有意义的报告而需抑制。

新增测试

• round_executor.rs：6 个 detect_ask_user_question 用例，覆盖有 header / 无 header / 空数组 / 缺 key / 不存在 / 混合工具调用。
• session_usage/service.rs：2 个代理过滤用例 —— is_proxy=true 必须排除（无论 intent_id），以及 turn- 前缀的真实意图在 is_proxy=false 时不被误过滤。

验证

cargo test -p bitfun-services-core hidden_intent   # 10 passed
cargo test -p bitfun-core intent_evidence          # 12 passed
cargo test -p bitfun-core report_                   # 22 passed (含 2 新增)
cargo test -p bitfun-core detect_ask_user_question  # 6 passed (新增)
cargo check --tests -p bitfun-services-core / bitfun-core / bitfun-desktop  # 全部通过

说明：问题 #2（从 trigger_description 自由文本解析 proactive_tools=）属于设计层面，建议随后续的结构化评估器一并替换为专用字段，本次未改动以控制范围。