feat: 14 vendor-derived agent optimizations

[Mingye-Lu]

Summary

Implements 14 optimizations derived from analysis of 10 competitor products (browser-use, Stagehand, crawl4ai, Skyvern, Spider, nanobrowser, ZeroClaw, Gottem). All features are disabled by default and opt-in via settings.optimization.*.

Optimizations

Token Savings

• HTML Diff Mode (html_diff_mode) — On repeated visits to the same URL, returns only changed content sections with [unchanged: N sections] markers. 50–70% token reduction on multi-turn sessions.
• Content-Aware Cleaning Profiles (content_aware_profiles) — Auto-selects ReadingMode / Minimal / Aggressive cleaning profile based on task keyword and content size.

Loop Prevention

• Action Loop Detection (loop_detection) — Rolling-window action hash with escalating nudges (soft at 5, medium at 8, strong at 12 repeats) + page stagnation detection.
• Page Fingerprinting (page_fingerprinting) — Lightweight FNV-1a fingerprint used by loop detection and action caching.
• Planning Interval (planning_interval) — Every N steps injects a planning-checkpoint or execution-mode guidance into the dynamic prompt.

Reliability

• Failure Classification (ailure_classification) — 16-category keyword-based error taxonomy (SelectorNotFound, CaptchaDetected, RateLimited, …). Zero LLM cost.
• Self-Healing Selectors (self_healing) — On SelectorNotFound/SelectorAmbiguous, text-matches a replacement element ref from a fresh page_map. Logs [healed: @Eold → @Enew]. Zero LLM calls.

Performance

• Action Caching (�ction_caching) — In-memory cache for read-only tools keyed by tool + input + page fingerprint. TTL-based expiry, invalidated on page change.
• Confidence Tracking (confidence_tracking) — Parses [confidence: HIGH/MEDIUM/LOW] from assistant responses; 2 consecutive LOWs trigger a stagnation alert.
• Compound Component Enrichment (compound_enrichment) — Adds enrichment metadata to complex form controls (date/range/select/file/textarea). Max 200 bytes/element.

Budget

• Budget Enforcement (�udget_max_session_cost_usd + �udget_enforcement) — Warn or block when session cost limit is reached.
• Per-Agent Cost Attribution (per_agent_cost_tracking) — /cost shows a per-child-agent cost breakdown.

Foundation

• Dynamic System Prompt Infrastructure — DynamicPromptContext struct with four optional nudge fields injected as section 9 of the system prompt via a shared Arc<Mutex<>> slot.
• Optimization Settings Schema — 18 fields in nested settings.optimization block, all Option, all default OFF.

Verification

• \cargo test --workspace\ — 1,097 tests pass, 0 failures
• \cargo clippy --workspace --all-targets -- -D warnings\ — clean
• \cargo fmt --check\ — clean
• \cargo build --release\ — clean

Commits

Commit	Description
\8370d6e\	feat(agent): dynamic system prompt infrastructure + settings schema
\d5936f\	feat(agent): HTML diff, loop detection, page fingerprinting, planning interval
\2689427\	feat(agent): failure classification, self-healing, action caching, confidence tracking
\8f94f78\	feat(browser,runtime): component enrichment, cleaning profiles, budget enforcement, cost attribution
\699f7f\	chore: CHANGELOG update
\7e4b9a5\	fix: clippy --all-targets warnings
\231b2f9\	docs: optimization settings + performance features documentation