File size: 14,026 Bytes
98abea7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
"""Run a baseline rollout then N reflection iterations and plot rewards.

Workflow:

  1. Run a baseline rollout (iter 0) using the in-tree ``inference.SYSTEM_PROMPT``.
  2. For each reflection iteration ``i`` in ``1..N``:
     - Read the prior run's artifacts.
     - Build a GEPA-style meta-prompt and call Claude Opus 4.7
       (via OpenRouter) to propose an improved system prompt.
     - Re-run the rollout with the new prompt.
  3. Save every iteration's run dir, the proposed prompts, and the score
     history under ``artifacts/reflection_<ts>/``.
  4. Render a reward-vs-iteration plot (PNG) at the parent dir.

Environment:

  HF_TOKEN              required - Hugging Face Inference Provider key
                        (drives the Qwen3-32B/Groq rollouts)
  OPENROUTER_API_KEY    required - OpenRouter key (drives the Opus 4.7
                        reflection calls)
  ENV_URL               required - URL of the running TradeBench server
                        (e.g. http://127.0.0.1:8773)
  MODEL_NAME            optional - rollout model id (default
                        ``Qwen/Qwen3-32B:groq``)
  REFLECTION_MODEL      optional - reflector model id (default
                        ``anthropic/claude-opus-4.7``)
  MAX_STEPS_PER_TASK    optional - cap on env steps per rollout
"""

from __future__ import annotations

import argparse
import asyncio
import json
import os
import sys
from datetime import UTC, datetime
from pathlib import Path
from typing import Any

PROJECT_ROOT = Path(__file__).resolve().parent.parent
if str(PROJECT_ROOT) not in sys.path:
    sys.path.insert(0, str(PROJECT_ROOT))

import inference  # type: ignore[import-not-found]
import reflection  # type: ignore[import-not-found]


def _ts() -> str:
    return datetime.now(UTC).strftime("%Y%m%dT%H%M%SZ")


async def _run_one_rollout(
    *,
    iteration: int,
    system_prompt: str,
    suffix: str,
) -> dict[str, Any]:
    print(f"\n{'#' * 60}", flush=True)
    print(f"# ROLLOUT iter={iteration} suffix={suffix}", flush=True)
    print(f"{'#' * 60}", flush=True)
    result = await inference.main(system_prompt=system_prompt, run_id_suffix=suffix)
    return result


def _save_proposed_prompt(
    *,
    parent_dir: Path,
    iteration: int,
    new_prompt: str,
    raw_response: str,
    meta_prompt: str,
    reflector_system_prompt: str,
) -> None:
    iter_dir = parent_dir / f"iter_{iteration:02d}__reflection"
    iter_dir.mkdir(parents=True, exist_ok=True)
    (iter_dir / "new_system_prompt.txt").write_text(new_prompt, encoding="utf-8")
    (iter_dir / "reflector_raw_response.txt").write_text(raw_response, encoding="utf-8")
    (iter_dir / "meta_prompt.txt").write_text(meta_prompt, encoding="utf-8")
    (iter_dir / "reflector_system_prompt.txt").write_text(
        reflector_system_prompt,
        encoding="utf-8",
    )


def _plot_rewards(history: list[dict[str, Any]], out_path: Path) -> None:
    try:
        import matplotlib

        matplotlib.use("Agg")
        import matplotlib.pyplot as plt
    except ImportError:
        print(
            "[WARN] matplotlib not available; skipping plot",
            file=sys.stderr,
            flush=True,
        )
        return

    iters = [row["iteration"] for row in history]
    test_scores = [row["test_score"] for row in history]
    cumulative = [row["test_cumulative_reward"] for row in history]
    normalized = [row.get("test_score_normalized", 0.5) for row in history]

    fig, axes = plt.subplots(3, 1, figsize=(8.5, 9.5), sharex=True)
    ax_top, ax_mid, ax_bot = axes

    ax_top.plot(iters, normalized, marker="o", linewidth=2.2, color="#1565c0")
    ax_top.axhline(0.5, color="#888", linewidth=0.8, linestyle="--", label="neutral (0.5)")
    if normalized:
        ax_top.axhline(
            normalized[0],
            color="#d32f2f",
            linewidth=0.8,
            linestyle=":",
            label=f"baseline = {normalized[0]:.3f}",
        )
    ax_top.set_ylim(0.0, 1.0)
    ax_top.set_ylabel("score_normalized in [0, 1]\nsigmoid(cumulative_reward)")
    ax_top.set_title(
        "Reflective prompt optimisation - reward vs iteration (0 = baseline)",
    )
    ax_top.legend(loc="best", fontsize=9)
    ax_top.grid(alpha=0.3)

    ax_mid.plot(iters, cumulative, marker="s", linewidth=2, color="#2e7d32")
    ax_mid.axhline(0, color="#888", linewidth=0.8, linestyle="--")
    if cumulative:
        ax_mid.axhline(
            cumulative[0],
            color="#d32f2f",
            linewidth=0.8,
            linestyle=":",
            label=f"baseline = {cumulative[0]:+.4f}",
        )
        ax_mid.legend(loc="best", fontsize=9)
    ax_mid.set_ylabel("test cumulative composite reward (raw)")
    ax_mid.grid(alpha=0.3)

    ax_bot.plot(iters, test_scores, marker="^", linewidth=2, color="#6a1b9a")
    ax_bot.axhline(0, color="#888", linewidth=0.8, linestyle="--")
    if test_scores:
        ax_bot.axhline(
            test_scores[0],
            color="#d32f2f",
            linewidth=0.8,
            linestyle=":",
            label=f"baseline = {test_scores[0]:+.4f}",
        )
        ax_bot.legend(loc="best", fontsize=9)
    ax_bot.set_xlabel("iteration")
    ax_bot.set_ylabel("test ROI (portfolio_value / initial - 1)")
    ax_bot.grid(alpha=0.3)
    ax_bot.set_xticks(iters)

    fig.tight_layout()
    fig.savefig(out_path, dpi=130)
    plt.close(fig)


async def run_reflection_loop(
    *,
    n_iterations: int,
    parent_dir: Path,
    task_ids: list[str],
    rollout_model: str | None,
    from_run: Path | None = None,
    initial_prompt_file: Path | None = None,
) -> list[dict[str, Any]]:
    parent_dir.mkdir(parents=True, exist_ok=True)
    inference.TASK_IDS = task_ids
    if rollout_model:
        os.environ["MODEL_NAME"] = rollout_model

    history: list[dict[str, Any]] = []
    if initial_prompt_file is not None:
        current_prompt = initial_prompt_file.read_text(encoding="utf-8")
        print(
            f"\n[CONTINUE] Using prompt from {initial_prompt_file} as iter 0 baseline "
            f"({len(current_prompt)} chars)",
            flush=True,
        )
    else:
        current_prompt = inference.SYSTEM_PROMPT
    (parent_dir / "iter_00__baseline").mkdir(parents=True, exist_ok=True)
    (parent_dir / "iter_00__baseline" / "system_prompt.txt").write_text(
        current_prompt,
        encoding="utf-8",
    )

    if from_run is not None:
        from_run = from_run.resolve()
        print(
            f"\n[REUSE] Skipping baseline rollout; using existing run as iter 0:\n"
            f"  {from_run}",
            flush=True,
        )
        baseline = _baseline_from_existing_run(from_run)
    else:
        baseline = await _run_one_rollout(
            iteration=0,
            system_prompt=current_prompt,
            suffix="iter00_baseline",
        )
    history.append(_history_row(0, baseline, prior_run=None))
    _write_history(parent_dir, history)

    for it in range(1, n_iterations + 1):
        prior = history[-1]
        prior_artifacts = prior.get("artifacts_root")
        if not prior_artifacts:
            print(
                f"[WARN] prior iteration {it - 1} has no artifacts_root; aborting reflection loop",
                file=sys.stderr,
                flush=True,
            )
            break
        ctx = reflection.load_run_context(
            Path(prior_artifacts),
            iteration=it,
            history_snapshot=list(history),
        )
        client = reflection.make_openrouter_client()
        new_prompt, raw_response, reflector_system_prompt, meta_prompt = (
            reflection.propose_improved_prompt(ctx, client=client)
        )
        print(
            f"\n[REFLECTION] iter={it} meta_prompt={len(meta_prompt)} chars  "
            f"new_prompt={len(new_prompt)} chars  budget={ctx.target_prompt_chars}",
            flush=True,
        )
        if not new_prompt or len(new_prompt) < 200:
            print(
                f"[WARN] reflector returned a too-short prompt ({len(new_prompt)} chars); "
                "keeping prior prompt for this iteration",
                file=sys.stderr,
                flush=True,
            )
            new_prompt = current_prompt
        _save_proposed_prompt(
            parent_dir=parent_dir,
            iteration=it,
            new_prompt=new_prompt,
            raw_response=raw_response,
            meta_prompt=meta_prompt,
            reflector_system_prompt=reflector_system_prompt,
        )
        current_prompt = new_prompt
        result = await _run_one_rollout(
            iteration=it,
            system_prompt=new_prompt,
            suffix=f"iter{it:02d}_reflect",
        )
        history.append(_history_row(it, result, prior_run=prior_artifacts))
        _write_history(parent_dir, history)

    plot_path = parent_dir / "scores.png"
    _plot_rewards(history, plot_path)
    print(f"\n[DONE] history -> {parent_dir / 'history.json'}", flush=True)
    print(f"[DONE] plot    -> {plot_path}", flush=True)
    return history


def _baseline_from_existing_run(run_dir: Path) -> dict[str, Any]:
    """Reconstruct an inference.main()-style result dict from a run on disk.

    Used when ``--from-run`` is supplied so the reflection loop can treat
    an already-completed rollout as iter 0 instead of re-running it.
    """

    test_summary_path = run_dir / "test" / "summary.json"
    manifest_path = run_dir / "manifest.json"
    test_summary = json.loads(test_summary_path.read_text(encoding="utf-8"))
    manifest = json.loads(manifest_path.read_text(encoding="utf-8"))
    cum = float(test_summary.get("cumulative_reward", 0.0))
    bars = int(test_summary.get("bars_completed", 0))
    score_norm = float(test_summary.get("score_normalized", 0.5))
    final_pv = float(test_summary.get("final_portfolio_value", 0.0))
    initial_pv = float(test_summary.get("initial_portfolio_value", 1.0)) or 1.0
    return {
        "run_id": manifest.get("run_id", run_dir.name),
        "artifacts_root": str(run_dir.resolve()),
        "scores": dict(manifest.get("scores", {})),
        "test_score": (final_pv / initial_pv) - 1.0 if initial_pv > 0 else 0.0,
        "test_cumulative_reward": cum,
        "test_bars_completed": bars,
        "test_score_normalized": score_norm,
        "test_action_counts": dict(manifest.get("action_counts", {}).get("test", {})),
    }


def _history_row(
    iteration: int,
    result: dict[str, Any],
    *,
    prior_run: str | None,
) -> dict[str, Any]:
    return {
        "iteration": iteration,
        "run_id": result.get("run_id"),
        "artifacts_root": result.get("artifacts_root"),
        "test_score": float(result.get("test_score", 0.0)),
        "test_cumulative_reward": float(result.get("test_cumulative_reward", 0.0)),
        "test_bars_completed": int(result.get("test_bars_completed", 0)),
        "test_score_normalized": float(result.get("test_score_normalized", 0.5)),
        "test_action_counts": dict(result.get("test_action_counts", {})),
        "scores": dict(result.get("scores", {})),
        "prior_run": prior_run,
    }


def _write_history(parent_dir: Path, history: list[dict[str, Any]]) -> None:
    (parent_dir / "history.json").write_text(
        json.dumps(history, indent=2, ensure_ascii=False),
        encoding="utf-8",
    )


def main() -> int:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--iters", type=int, default=5, help="Number of reflection iterations after the baseline")
    parser.add_argument(
        "--task-ids",
        nargs="+",
        default=["train", "test"],
        help="Tier sequence to run each iteration (default: train test)",
    )
    parser.add_argument(
        "--out",
        type=Path,
        default=None,
        help="Parent dir under artifacts/ (default artifacts/reflection_<ts>/)",
    )
    parser.add_argument(
        "--rollout-model",
        default=None,
        help="Override MODEL_NAME for the rollout (e.g. Qwen/Qwen3-32B:groq)",
    )
    parser.add_argument(
        "--from-run",
        type=Path,
        default=None,
        help=(
            "Use an existing run directory as iter 0 instead of running a "
            "fresh baseline. Path should point at artifacts/runs/<run_id>/."
        ),
    )
    parser.add_argument(
        "--initial-prompt-file",
        type=Path,
        default=None,
        help=(
            "Override iter-0 baseline prompt with the contents of this file. "
            "Pair with --from-run to continue from a prior reflection's "
            "final iter (use that iter's new_system_prompt.txt and run dir)."
        ),
    )
    args = parser.parse_args()

    artifacts_root = Path(
        os.environ.get(
            "TRADEBENCH_ARTIFACTS_ROOT",
            str(PROJECT_ROOT / "artifacts"),
        ),
    )
    parent_dir = args.out or (artifacts_root / f"reflection_{_ts()}")

    if "OPENROUTER_API_KEY" not in os.environ:
        print(
            "[ERROR] OPENROUTER_API_KEY env var is required (Opus 4.7 reflection calls).",
            file=sys.stderr,
            flush=True,
        )
        return 2

    history = asyncio.run(
        run_reflection_loop(
            n_iterations=args.iters,
            parent_dir=parent_dir,
            task_ids=list(args.task_ids),
            rollout_model=args.rollout_model,
            from_run=args.from_run,
            initial_prompt_file=args.initial_prompt_file,
        ),
    )

    print(f"\n{'=' * 60}", flush=True)
    print("REFLECTION HISTORY", flush=True)
    print(f"{'=' * 60}", flush=True)
    for row in history:
        tag = "baseline" if row["iteration"] == 0 else f"iter {row['iteration']:02d}"
        print(
            f"  {tag:8s}  test_score={row['test_score']:+.4f}  "
            f"cum_reward={row['test_cumulative_reward']:+.4f}  "
            f"bars={row['test_bars_completed']}  run_id={row['run_id']}",
            flush=True,
        )
    return 0


if __name__ == "__main__":
    sys.exit(main())