Minbyul/AgentMercury-SWE-construction

AgentMercury — SWE construction strata (private)

⚠️ Private repository. Scrubbed to a hard gate (see Scrub below) but not cleared for public release — three documented classes survive by design. The blocker list is at the end.

The environment-construction code differences behind AgentMercury: Your Agent Can Synthesize Verifiable Environments for Business Scenarios at Scale.

AgentMercury's central claim is that a world is built, and that the build is itself a learnable act. That makes environment construction a software-engineering task: a diff, against a tree, judged by tests. These three strata are that task, cut three ways.

	What one row is	Rows	Schema
S-C S-C_env_build_traces/	Complete the build of a synthetic company world from stage k onward	30	SWE-bench_Verified, verbatim
S-A S-A_commit_mining/tier_a_tasks.jsonl	Reproduce a real commit to the environment-synthesis codebase	31	SWE-bench-shaped
S-A S-A_commit_mining/trajectories.jsonl	The multi-turn agent session that produced such a change	731	chat + gold_diff
S-B S-B_intent_to_diff/intent_to_diff.jsonl	Design intent + target path → author the file	1,434	chat (messages)

Together these are the supervision behind the paper's §4.4 result: fine-tuning on construction traces lifts the oracle-pass rate on held-out business briefs from 3.3% → 83.3%, level with the strongest off-the-shelf API models, while the base model with a hand-written recipe reaches only 20.0%.

---

S-C — environment build traces as SWE-bench instances

The one that is literally "environment construction code differences". A synthesized world is built in five staged artifacts; an instance freezes the build partway and asks for the rest.

base tree (stage k)  = the env with stage-k-and-later artifacts removed
target tree          = the full env
patch                = the unified diff between them   ← the gold construction diff
FAIL_TO_PASS(k)      = validators that fail on base and pass on target

stage	artifacts the patch must add
S1_profile	step1_company_profile.json, step1c_grounding.json
S2_service_graph	step2_service_graph.json
S3_state_schema	step3_state_schema.json, step3c_data_profiles.json
S3_5_seed_db	step3_5_seed_stats.json, db/** (schema + one seed file per collection)
S4_constraints_compile	step4_constraints.json, env_config.json

6 worlds × 5 stages = 30 instances, one world per country (Argentina, Canada, Qatar, Saudi Arabia, Spain, Switzerland). Because the target is always the complete world, patches are nested and shrink monotonically:

stage	mean |FAIL_TO_PASS|	mean |PASS_TO_PASS|	mean patch bytes
S1_profile	19.0	0.0	793,561
S2_service_graph	17.0	2.0	790,656
S3_state_schema	16.0	3.0	786,919
S3_5_seed_db	12.0	7.0	703,693
S4_constraints_compile	7.0	12.0	99,084

What is verified. 19 validators, all active and none vacuous. The base-failure matrix is exactly stage-monotone — every validator flips at its own stage boundary, in every world. Target pass rate 1.0, zero consistency violations (no validator passes on a base and fails on its target). And the property that matters most for a diff dataset: 30/30 gold patches git apply onto the materialized base tree and reproduce the target tree byte-for-byte — re-verified after anonymization by scripts/verify_patches.py, not inherited from the original run.

What it is not. repo/base_commit/environment_setup_commit are schema placeholders, not a real git backing. test_patch is empty: the "tests" are pure-Python validators named by the FAIL_TO_PASS ids and defined inside the exporter, not in-tree pytest files. The validators check artifact consistency, not runtime behaviour — SQL verification queries are checked for referencing known collections, not executed. Treat this as a proof-of-concept stratum: the exporter (scripts/planet_swe_export.py, stdlib-only, deterministic, $0) scales to the full library, which holds 4,783 complete build traces ≈ 24K instances.

S-A — commit mining

Real changes to the environment-synthesis codebase itself, mined and reproduced against their base commits. tier_a_tasks.jsonl carries problem_statement, code_patch, test_patch and FAIL_TO_PASS/PASS_TO_PASS as pytest node ids — 31 tasks, mean 2.3 changed code files, 15 of them with a single failing test. trajectories.jsonl is the same work as agent sessions: 731 trajectories, mean 3.86 turns, with real tool observations (651 tool turns over 731 rows) and the gold_diff as the target.

S-B — intent → diff

1,434 pairs of (repository context + design intent + target file path) → the file that implements it. 740 are edit-diff shaped, 694 author a file from scratch. Five repositories: solar-coder 787, office-universe 291, solar-mcp 156, solar-search 115, tool-dataset-generation 85.

Each row carries public_safe (see below). 1,191 of 1,434 are true.

---

Scrub

Three token classes were treated differently, because prose, patches and credentials do not have the same truth conditions.

Identity (S-C). Every world is seeded from a real company profile, and the identity leaked through the wrapper — env id, slug, config company_profile, init.sql header — even though the world's own content was already fictional. scripts/anonymize_env_identity.py rewrites each env to its own step1c_grounding.fictional_name and the traces were re-exported from the anonymized trees, so no patch is a rewritten diff. Two things the obvious implementation gets wrong, both found by residual scanning rather than by inspection:

• the library truncates slugs to 30 characters and keeps &, so the on-disk directory name (qatar_cinema_&_film_distributi) is not derivable from the company name — it has to be passed in as its own surface, or it survives;
• the exchange symbol is stored under stock_ticker, not ticker; looking for the wrong key left all six envs' real exchange symbols in the first export (six occurrences each), and a ticker maps one-to-one onto the firm — caught by scanning the output for the symbols rather than by re-reading the rule list, which is why the census scans shipped files and not just inputs;
• the system word list carries capitalized proper nouns, so testing brand tokens against it case-insensitively classifies a firm's short name as an ordinary word and leaves it in place — one of the six envs leaked exactly this way. Only lowercase dictionary entries count as ordinary words.

Residual identity hits after the final pass: 0.

Patches (S-A). A banned token inside code_patch/test_patch/gold_diff drops the row. Rewriting a unified diff would destroy the single property that makes it worth shipping — that it is the real change — so 2 of 33 tasks and 51 of 782 trajectories were dropped rather than papered over.

Prose and credentials (all strata). Redacted in place: commit trailers deleted, credentials → <REDACTED_KEY>, gateway hosts → <GATEWAY_HOST>, absolute cluster paths → <ROOT>/. In S-B, secret- or gateway-bearing assistant turns drop the row (55 of 1,489); infrastructure paths are substituted, since a path string is incidental to the code.

RESIDUAL_CENSUS.json records the final counts; the gate verdict is CLEAN. Per-stratum detail is in each SCRUB_REPORT.json, including every dropped row id.

What the census still shows

Three classes survive on purpose, and each is a reason this repo is private:

1. Teacher/judge model names inside source code — 2,055 occurrences in S-B, 27 in S-A. These are the mined repositories' own code and comments (anthropic/sonnet-4.6 in an LLM wrapper, .claude/ in a .gitignore). Redacting them would corrupt the SFT target and teach the model to emit placeholders. S-B rows are flagged public_safe=false when their assistant turn contains one — filter on it and 1,191 rows remain.
2. UpstageAI/… in S-A repo fields (31 + 14) — genuine provenance for private repositories.
3. Structural facts of the seed firm in S-C problem_statement — industry, employee count, revenue, HQ city, founding year are kept as the scenario parameters that define the world; only the identity was replaced. For a distinctive firm these remain a re-identification vector.

The false positives a naive scan reports, checked individually and left alone: the English adverb timely (18 in S-C — "ensures timely assessment"), the bare identifier api_key (538 in S-B, a variable name, not a value), a synthetic Swiss employee named Jean-Claude Wegmann, a synthetic supplier "Qatar Cinema Supply Co.", and the brand denylist inside the anonymizer's own source — which is anti-leak machinery, not a leak.

Files

S-C_env_build_traces/  instances.jsonl · verification_report.json · EXPORT_REPORT.json
S-A_commit_mining/     tier_a_tasks.jsonl · trajectories.jsonl · SCRUB_REPORT.json
S-B_intent_to_diff/    intent_to_diff.jsonl · SCRUB_REPORT.json
scripts/               planet_swe_export.py    S-C exporter (stdlib, deterministic)
                       anonymize_env_identity.py  identity anonymizer
                       verify_patches.py       byte-exact patch check
                       scrub_strata.py / scrub_tierd.py   the release scrub
RESIDUAL_CENSUS.json   final gate

from datasets import load_dataset
import json

swe = load_dataset("Minbyul/AgentMercury-SWE-construction", "S-C_env_build_traces")["train"]
r = swe[0]
print(r["instance_id"], r["meta"]["stage"])
print(json.loads(r["FAIL_TO_PASS"]))       # validator ids, JSON-encoded string
print(r["patch"][:400])                    # the construction diff

d2d = load_dataset("Minbyul/AgentMercury-SWE-construction", "S-B_intent_to_diff")["train"]
safe = d2d.filter(lambda r: r["public_safe"])          # 1,191 of 1,434

Before this can go public

1. Decide the teacher-name policy for S-B/S-A code (ship public_safe only, or rewrite).
2. Replace UpstageAI/… in S-A repo fields with a neutral placeholder.
3. Decide whether S-C keeps the seed firm's revenue / HQ / founding year, or jitters them.
4. Re-run RESIDUAL_CENSUS with those classes promoted to blocking.

Citation

@article{jeong2026agentmercury,
  title  = {AgentMercury: Your Agent Can Synthesize Verifiable Environments for Business Scenarios at Scale},
  author = {Jeong, Minbyul and Yoon, Chanwoong},
  year   = {2026}
}

License: Apache-2.0. Related: Minbyul/AgentMercury-corpus (RL corpus, private) · Minbyul/AgentMercury-corpus-sample (public sample) · Minbyul/AgentMercury-Qwen3.5-4B