[two-tiger]

Summary

• add a shared a11y/CDP semantic element table for slow planner and fast-model candidates
• recover labels from AT-SPI metadata, nearby text and DOM; keep field values separate from names
• merge DOM/a11y targets and classify browser_chrome/web_content/dialog/menu/native scopes
• restrict alternatives to the proposal scope, reject mismatched focused typing, and omit unavailable visual fallback
• compact JEV history/options while preserving semantic source and target IDs
• add protocol 0.2 browser_snapshot references and episode browser/step_XX.json artifacts
• retry transient slow-model ConnectionError and require a submit milestone before filtered search results
• no screen OCR or document-evidence extraction; no third-party runtime dependency

Follow-up: schema and unbound recovery

• scroll_up/scroll_down prompts explicitly require target_name and target_role
• invalid-stage retry feedback identifies the exact milestone/proposal and missing field
• stage binding tries exact, normalized, and compatible-role matches only when unique
• fully unbound stages now reach JEV with bounded same-scope recovery candidates and RETURN_TO_PLANNER
• stage_replan records unbound_proposals, available_targets, and recovery_scopes for the next repair

Validation

• ruff check src tests: passed
• pytest: 329 passed
• tandem check: 17 golden samples passed
• mock rjob smoke: semantic-recovery-mock-smoke, score 1.0
• Task 059 five-step smoke: first unbound proposal reached JEV recovery and executed a click, followed by Escape and Alt+F10; ended at max_steps rather than no_action_limit
• CDP Task 008 probes: port 1337, visible-page selection, fragmented responses and episode persistence verified

Task 008 acceptance

Run: osw21-semantic-008-ebcba05 (stage/c10, max 371 steps, 14400 s)

• completed a scored run with 16 actions; stop_reason=no_action_limit, score=0.0
• remaining blocker: planner repeatedly requested Desktop/list_item, which was absent from the a11y targets in Files; current main has no visual grounder fallback. This is not reported as a task pass.
• search now proceeds through input -> Enter -> filtered results instead of stopping at target_unbound_or_ambiguous
• 16 a11y+dom candidates across 4 of 13 GUI decisions
• zero web_content -> browser_chrome candidate leaks and zero visual_fallback offers
• zero generic Focused entry labels and zero verification blobs in JEV history
• median JEV state: 2,493 bytes, versus 10,854 bytes in the reference trajectory
• 16 CDP snapshots preserved in the episode

Run artifacts: /mnt/shared-storage-gpfs2/evobox-share-gpfs2/chenxinquan/tandem/rjob_logs/osw21-semantic-008-ebcba05

Latest corner-case fixes

• bound-but-unexecutable GUI proposals now enter bounded recovery with PageDown/PageUp and arrow-key alternatives
• CLI milestones require a top-level command; operation=cli receives dedicated repair feedback
• mock smoke: semantic-unexecutable-recovery-mock, score 1.0

JEV service recovery

• recover rank calls after URL timeouts, JSON failures, and transient HTTP 408/429/500/502/503/504
• honor Retry-After, use 30-second health checks, and keep the 1200-second recovery bound
• refresh endpoint discovery during recovery and log endpoint changes for redeployments

Named experiment runner

• scripts/run_experiment.sh records manifest, batch, driver, worker, model and result logs under rjob_logs//
• experiment names are unique lowercase/digit/hyphen identifiers; duplicate experiment directories are rejected
• dry-run and artifact-layout tests included; rjob submission is unchanged