Add embodied agent evaluation pipeline with real simulator support:
- ReAct agent with multi-action buffering and PromptBuilder protocol
- EB-Alfred task support (6 splits via multi-split YAML discovery)
- AI2-THOR v2.1.0 bridge with skill-based actions and state tracking
- EvaluationRunner with structured output: <output_dir>/<task>/<run_id>/
- Per-episode artifacts: result.json, trajectory.jsonl, rgb_*.png
- Centralized logging (print -> logger), --verbosity CLI option
- Subprocess observability: bridge output streaming, Ctrl+C cleanup
- LLM API client (OpenAI-compatible) and dummy LLM server
- 106 tests passing

Split the monolithic bridge into a generic AI2ThorBridge base class
(simulator layer) and an EBAlfredBridge subclass (task layer), so that
future benchmarks using ai2thor==2.1.0 can reuse the simulator bridge
without rewriting controller management, IPC, or navigation helpers.

- Extract EB-Alfred goal evaluation and task loading into
  easi/tasks/ebalfred/thor_utils.py
- Trim easi/simulators/ai2thor/v2_1_0/thor_utils.py to generic-only
  constants and object query utilities
- Refactor bridge.py from EBAlfredBridge (1062 lines) to generic
  AI2ThorBridge (~314 lines) with configurable simulator_kwargs
- Create easi/tasks/ebalfred/bridge.py with EBAlfredBridge subclass
  containing all skill execution, state tracking, and goal evaluation
- Add get_bridge_script_path() and simulator_kwargs to BaseTask and
  TaskProtocol; override in EBAlfredTask
- Update EvaluationRunner to prefer task-specific bridge paths and
  forward simulator_kwargs
- Add simulator_kwargs to all EB-Alfred YAML configs
- Add 29 tests covering imports, inheritance, method separation,
  bridge path resolution, and simulator_kwargs

Unified client with generate() and generate_structured() methods,
lazy imports, and cumulative usage tracking (tokens + cost).

Manages start/stop, port checking, health polling with timeout,
and context manager support. Extensible for future backends.

New arguments for `easi run` to select LLM backend and configure
inference server. Backward compatible with existing --llm-url.

Runner now resolves backend, auto-starts vLLM when needed,
creates LLMClient for non-legacy backends, wraps structured
output, and tracks LLM usage per-episode and per-run.

- Track usage in generate_structured() via instructor's _raw_response
- Fix log file handle leak in ServerManager (store and close in stop())
- Remove duplicate agent_config computation in runner._create_agent()

…port

Replace the tightly-coupled agent/prompt design with a memory-based
architecture where AgentMemory holds shared state, PromptBuilder reads
from memory to construct prompts and parse responses, and the agent is
a thin orchestrator.

Key changes:
- AgentMemory + StepRecord dataclasses as shared agent state
- New PromptBuilderProtocol: build_messages(memory) + parse_response(response, memory)
- Simplified BaseAgent (removed _chat_history, abstract stubs, default act())
- ReActAgent rewritten as thin orchestrator delegating to builder
- EBAlfredPromptBuilder gains chat_history=True mode with VLMPlanner parity
- json_repair moved to easi/utils/ (old location re-exports)
- Removed stateless flag from agent config (builder controls mode)

Builder-owned schema enforcement: prompt builders can now optionally
implement get_response_format() to provide a JSON schema dict that gets
passed through to litellm. ReActAgent handles fallback automatically
when the backend doesn't support response_format.

- LLMClient.generate() accepts optional response_format param
- ReActAgent._generate_with_fallback() tries schema, caches on failure
- EBAlfredPromptBuilder.get_response_format() returns vlm_generation_guide
- Remove dead code: instructor dep, Pydantic schemas, monkey-patching

Add BaseTask.on_episode_reset() hook for task-specific post-reset setup.
EBAlfredTask overrides it to update agent action space from bridge metadata,
removing EB-Alfred-specific logic from the general EvaluationRunner.

…onfig

- trajectory.jsonl: add llm_response field to each step entry
- result.json: add instruction field for each episode
- config.json: include all CLI options and full task YAML config

Retry: LLMClient passes num_retries to litellm.completion() for
automatic exponential backoff on transient errors (timeouts, rate
limits). Configurable via --max-retries (default 3).

Resume: --resume <run_dir> loads config.json from a previous run,
skips completed episodes, clears and re-runs the last episode (which
may have been interrupted), then continues the remaining episodes.
All CLI options are restored from config.json so only --resume is needed.

…ps bug

Fix max_steps mismatch where YAML configured 50 but vendor EBAlfEnv
hardcoded 30. Now max_steps flows from YAML through simulator_kwargs
to the bridge and vendor env.

Add per-episode retry in EvaluationRunner: on crash (e.g. AI2-THOR
Unity segfault), the episode dir is cleared, the simulator is
re-launched, and the episode is retried up to max_retries times.
If all retries are exhausted the episode is recorded as failed and
the runner continues to the next episode.

Integrate EmbodiedBench EB-Navigation into EASI with vendored env,
task bridge, prompt builder, and 5 split configs (ai2thor v5.0.0).

Remove action_space field from all YAML configs and TaskEntry. Tasks
now define their action space via _build_action_space() override with
caching, eliminating the confusing pattern of empty YAML fields.

…atform

Replace stub bridge with working AI2ThorV5Bridge class that starts a
real controller, handles scene reset and discrete navigation actions.
Switch platform from CloudRendering to Linux64. Increase sim test
timeout default from 30s to 200s for THOR startup.

Fix OUTPUT_TEMPLATE trailing spaces on 3 lines and regenerate
navigation_examples.json from source to fix line continuation artifact
and curly quote mismatch. Verified character-level parity.

- Add habitat_sim:v0_3_0 simulator registration (conda env + manifest)
- Vendor EBHabEnv from EmbodiedBench with fixed imports
- Add EBHabitatTask with dynamic action space via on_episode_reset hook
- Add EBHabitatPromptBuilder matching VLMPlanner prompt construction
- Add 6 per-split YAML configs (base, common_sense, complex_instruction,
  spatial_relationship, visual_appearance, long_horizon)
- Add 26 offline tests for actions, task, prompts, and registry
- Move EB-Habitat-specific deps (gym, hydra-core, omegaconf, imageio,
  habitat-lab) from simulator requirements.txt to task YAML additional_deps

Add --redownload to 'easi task download' and 'easi run' to force
re-download of cached HuggingFace datasets. Useful when a previous
download was interrupted or incomplete.

…v level)

…prompt fix

- react_agent: force stop after N consecutive fallbacks (configurable via
  max_consecutive_fallbacks in YAML, default 0 = disabled)
- react_agent: reprompt correction message now comes from prompt builder
  (fixes SFT builder getting JSON-format correction prompt)
- metrics: add early_stop_rate to generic_aggregate
- runner: record forced_early_stop in per-episode results
- llm/client: support skip_special_tokens via extra_body for vLLM backend
- llm/models/internvl3: bypass model.chat() when skip_special_tokens=False
  to preserve SFT action tokens that get stripped by hardcoded decode
- llm/models/qwen3_vl: read skip_special_tokens from kwargs
- sft prompt builder: add get_reprompt_message() for action token format
- _base.yaml: add max_consecutive_fallbacks: 3

- prompt_builder: validate PNG with PIL before base64 encoding, retry
  with exponential backoff (0.1s, 0.2s, 0.4s) if truncated, send
  anyway after retries to trigger episode restart
- sft prompt builder: switch to interleaved content blocks (image_url
  at correct positions instead of all-first with <image> in text),
  matching InternVL chat template convention
- test: use real PIL image instead of fake PNG header bytes

… stage advance

… not ./datasets)

…not ~/.cache)

…s.py (./datasets)

ReActAgent captures the flattened prompt preview (`--- {role} ---`
blocks with `[img_N]` markers) alongside llm_response each time it
queries the LLM. Runner threads it into trajectory.jsonl under a new
"prompt" field next to "llm_response". Buffered / fallback-only
steps keep prompt = null.

When MIRROR_DEBUG_DIR is set, the mirror prompt builder writes every
flipped PNG it emits (current 3 slots + full history buffer) under that
directory with step+slot naming. Lets us A/B against sim-rendered
originals to verify the mirror transform and slot swap. Paired with
debug_mirror_one_episode.sh which runs one episode + stages the dumps
into the episode dir next to the originals.

SFT checkpoints emit action tokens <|forward|>/<|left|>/etc. vLLM's
OpenAI chat endpoint strips them at decode by default, which causes
_parse_sft_response to return empty and every step falls back to
move_forward. Set skip_special_tokens=false in _base.yaml
generation_kwargs so the tokens survive.

Adds LHPRVLNEnhancedSFTPromptBuilder that applies fantasy-vln's
display_env preprocessing (contrast 1.5 + resize 366) at prompt-build
time. Subclass of LHPRVLNSFTPromptBuilder; kwargs enhance_contrast and
resize_to are yaml-tunable. Ships new tasks for val_filtered and
test_filtered splits plus 9 unit tests covering contrast math, resize,
RGBA passthrough, and structural parity with the parent builder.

Adds save_rgba + enhance_images toggles to the LHPR-VLN bridge so the
enhancement (contrast 1.5 + resize 366) happens BEFORE the PNG hits
disk, matching fantasy-vln's display_env pipeline exactly. Paired with
the baseline SFT prompt builder (no in-prompt transform), the on-disk
image and the VLM input are both the 366x366 RGBA-enhanced frame.
Defaults off — existing yamls unaffected.

LHPRVLNBridge.reset() called self._scene_sim.actor('move_forward') to
fetch an initial observation. That call applied a second move_forward
(on top of the one already inside SceneSimulator.__init__), shifting
the agent's starting pose by 0.25 m along its facing direction.
Because actor() on step=-1 returns early without refreshing self.info,
the reset's logged pose stayed stale while the agent had actually
moved, so step 0 -> step 1 in trajectory.jsonl showed a phantom 0.25
m translation on a pure turn_left.

SceneSimulator.__init__ already captures the initial observation; use
it directly. Matches fantasy-vln's agent loop which starts with
task_sim.observations without calling actor() first.

Impact: EASI episodes now start at the same pose as fantasy for a
given scene/seed, enabling apples-to-apples SR comparison. Prior EASI
eval numbers were computed with the +0.25 m bias and are not directly
comparable.

Duplicates the bridge-enhanced val task under a short name so autoeval
treats it as a fresh task key and reruns checkpoints against the
post-bridge-fix code without colliding with prior bridge_enhanced
summaries. Temporary — safe to delete when done.

SceneSimulator.actor() had an early return on self.step == -1 that
absorbed the first call without incrementing bookkeeping or running
the max_step check. Combined with b6eb3c6 (bridge no longer calls
actor on reset), the runner's for step in range(max_steps) loop left
self.step one short of max_steps, so the timeout branch never fired
and nav_errors / nav_steps were never appended for episodes that ran
to the wall clock. aggregate_results then died with IndexError in
NavigationMetrics.TAR() because gt_path outlasted error_length.

This also silently masked a pre-existing bug: a first-action 'stop'
was dropped (skipped sim.step and took the early return), never
advancing the stage or recording a nav_error.

Fix: initialise self.step = 0 at end of __init__ and remove the
early-return block from actor(). Every call now goes through the full
step-increment + info-refresh + oracle/stop/max pipeline. Semantic
shift: nav_steps[0] is now +1 vs pre-fix runs because it includes the
init move_forward; fantasy-vln's nav_steps already carry the same +1
offset (their while-not-episode-over loop makes one extra actor call
past max_steps), so post-fix EASI is more comparable to fantasy.

Adds tests/test_lhpr_vln_scene_step.py covering timeout length
invariants, all-subtasks-success accounting, first-action-stop being
honoured, and mixed partial success + timeout — all via a
habitat_sim stub so they run offline in the main venv.