"""Agentic tool-calling traces in the official DeepSeek-V4 chat format. No public dataset exists in this model's DSML tool-call format, so these traces are constructed programmatically. What is real and what is not: REAL - every `read_file` / `grep` / `list_dir` tool result is computed from the actual cloned repository at build time (verbatim file bytes, real regex matches with real line numbers, real directory listings). REAL - `edit_file` old_string values are exact substrings of the real file, so the edits would actually apply. SYNTH - `run_command` outputs (test runners, linters, compilers) are written to match each tool's real output format. SYNTH - the natural-language turns and the reasoning blocks. Files used here are drawn from a reserved partition (see AGENTIC_BUCKET) that is excluded from the plain code/graphics slices, so no file content is duplicated across domains. """ from __future__ import annotations import hashlib import json import os import random import re from dsv4 import encode_chat AGENTIC_BUCKET = 7 # sha1(path) % 10 == 7 -> reserved for agentic traces TOOLS = [ { "type": "function", "function": { "name": "read_file", "description": "Read a file from the repository. Returns the file contents with 1-indexed line numbers.", "parameters": { "type": "object", "properties": { "path": {"type": "string", "description": "Path relative to the repository root"}, "offset": {"type": "integer", "description": "First line to read (1-indexed)"}, "limit": {"type": "integer", "description": "Maximum number of lines to read"}, }, "required": ["path"], }, }, }, { "type": "function", "function": { "name": "edit_file", "description": "Replace an exact string in a file. old_string must match the file exactly and be unique.", "parameters": { "type": "object", "properties": { "path": {"type": "string", "description": "Path relative to the repository root"}, "old_string": {"type": "string", "description": "Exact text to replace"}, "new_string": {"type": "string", "description": "Replacement text"}, }, "required": ["path", "old_string", "new_string"], }, }, }, { "type": "function", "function": { "name": "run_command", "description": "Run a shell command in the repository root and return its combined output.", "parameters": { "type": "object", "properties": { "command": {"type": "string", "description": "The command to run"}, "timeout_s": {"type": "integer", "description": "Timeout in seconds"}, }, "required": ["command"], }, }, }, { "type": "function", "function": { "name": "grep", "description": "Search file contents with a regular expression.", "parameters": { "type": "object", "properties": { "pattern": {"type": "string", "description": "Regular expression"}, "path": {"type": "string", "description": "Directory to search"}, "glob": {"type": "string", "description": "Filter files by glob, e.g. *.ts"}, }, "required": ["pattern"], }, }, }, { "type": "function", "function": { "name": "list_dir", "description": "List the entries of a directory.", "parameters": { "type": "object", "properties": {"path": {"type": "string", "description": "Directory path"}}, "required": ["path"], }, }, }, ] SYSTEM_PROMPTS = [ "You are a coding agent working inside the {repo} repository. Use the provided tools to inspect and modify the codebase. Verify your changes by running the project's checks before reporting completion.", "You are an autonomous software engineering assistant. The working directory is a checkout of {repo}. Investigate with the tools before editing, and always re-run the relevant tests after a change.", "You operate on the {repo} codebase through tools. Prefer reading the surrounding code before making an edit. Never claim a fix works until a command confirms it.", "Coding agent session. Repository: {repo}. Make the smallest change that resolves the request, then validate it.", ] def _numbered(text: str, offset: int = 1, limit: int | None = None) -> str: lines = text.split("\n") if limit is not None: lines = lines[offset - 1 : offset - 1 + limit] else: lines = lines[offset - 1 :] width = len(str(offset + len(lines))) return "\n".join(f"{i + offset:>{width}}\t{ln}" for i, ln in enumerate(lines)) def _real_grep(docs_by_path, pattern, glob_ext=None, limit=25): out = [] try: rx = re.compile(pattern) except re.error: return "No matches found." for path, text in docs_by_path.items(): if glob_ext and not path.endswith(glob_ext): continue for i, ln in enumerate(text.split("\n"), 1): if rx.search(ln): out.append(f"{path}:{i}:{ln.strip()[:200]}") if len(out) >= limit: return "\n".join(out) + f"\n... (truncated at {limit} matches)" return "\n".join(out) if out else "No matches found." def _tool_call(cid, name, args): return { "id": cid, "type": "function", "function": {"name": name, "arguments": json.dumps(args, ensure_ascii=False)}, } # --- synthetic command output, matching each tool's real format ------------- def _vitest_fail(test_name, file_path, expected, actual): return f"""> vitest run RUN v1.6.0 {os.path.dirname(file_path) or '.'} ❯ {file_path} (1 test | 1 failed) 12ms × {test_name} 11ms → expected {expected} to be {actual} ⎯⎯⎯⎯⎯⎯⎯ Failed Tests 1 ⎯⎯⎯⎯⎯⎯⎯ FAIL {file_path} > {test_name} AssertionError: expected {expected} to be {actual} ❯ {file_path}:42:24 Test Files 1 failed (1) Tests 1 failed (1) Duration 0.84s """ def _vitest_pass(n=1, file_path="test/index.test.ts"): return f"""> vitest run RUN v1.6.0 ✓ {file_path} ({n} tests) 14ms Test Files 1 passed (1) Tests {n} passed ({n}) Duration 0.79s """ def _pytest_fail(mod, func, err): return f"""============================= test session starts ============================== platform linux -- Python 3.11.8, pytest-8.1.1, pluggy-1.5.0 collected 14 items {mod} .......F...... [100%] =================================== FAILURES =================================== _______________________________ {func} ________________________________ def {func}(): > result = subject() E {err} {mod}:57: {err.split(':')[0]} =========================== short test summary info ============================ FAILED {mod}::{func} - {err} ========================= 1 failed, 13 passed in 0.42s ========================= """ def _cargo_fail(krate, path, line, msg, code="E0308"): return f""" Compiling {krate} v0.1.0 ({path}) error[{code}]: {msg} --> {path}:{line}:17 | {line} | let n: usize = value; | ----- ^^^^^ expected `usize`, found `i64` | | | expected due to this | help: you can convert an `i64` to a `usize` and panic if the converted value doesn't fit | {line} | let n: usize = value.try_into().unwrap(); | ++++++++++++++++++++ error: could not compile `{krate}` (lib) due to 1 previous error """ def _eslint_fail(path, line, rule, msg): return f""" {path} {line}:9 error {msg} {rule} ✖ 1 problem (1 error, 0 warnings) """ TASK_TEMPLATES = [ ("Add a short doc comment to the top of `{path}` explaining what it does, then make sure the lint passes.", "doc"), ("There's a bug somewhere in `{path}` — the {sym} path doesn't behave the way the tests expect. Track it down and fix it.", "bug"), ("Refactor `{sym}` in `{path}` so the early-return case is handled first. Keep behaviour identical and run the tests.", "refactor"), ("I need `{sym}` in `{path}` to also handle the empty-input case without throwing. Add that and a test.", "feature"), ("Find every place `{sym}` is used in this repo and tell me whether `{path}` is the only definition.", "search"), ("Something in `{path}` is slower than it should be. Look at `{sym}` and see if there's an obvious win.", "perf"), ("Can you check `{path}` for the `{sym}` handling? I think there's an off-by-one in there.", "bug"), ("Wire up a small unit test for `{sym}` in `{path}`. It currently has no coverage.", "feature"), ] def _pick_symbol(text, rng): cands = re.findall(r"(?:function|class|const|let|def|fn|struct|export function)\s+([A-Za-z_][A-Za-z0-9_]{3,30})", text) cands = [c for c in cands if not c.startswith("_")] return rng.choice(cands) if cands else "the main entry point" def _cmd_for(repo, lang): if lang in ("typescript", "javascript"): return repo in ("vite",) and "pnpm test" or "npm test" if lang == "python": return "python -m pytest -q" if lang == "rust": return "cargo test" if lang == "cpp": return "cmake --build build && ctest --test-dir build" return "npm test" CODE_LANGS = {"javascript", "typescript", "python", "rust", "cpp", "c", "glsl", "wgsl"} BAD_PATH = re.compile( r"(^|/)(\.storybook|stories|__fixtures__|fixtures|__snapshots__|docs|examples|" r"benchmark|benchmarks|\.github)(/|$)|\.stories\.|\.d\.ts$", re.I ) DEFN_RE = re.compile( r"(^|\n)\s*(export\s+)?(async\s+)?(function|class|def|fn|struct|impl|const\s+\w+\s*=\s*\()", re.M ) # A line we are willing to anchor an edit on: looks like a statement, not markup. GOOD_LINE = re.compile(r"^[\t ]*[A-Za-z_@#/][^\n]*[;{)\],:]\s*$") MARKUP_LINE = re.compile(r"|^\s*[<*]|^\s*//\s*$") def _anchor_lines(text, limit=None): src = text.split("\n")[:limit] if limit else text.split("\n") out = [ln for ln in src if 20 < len(ln.strip()) < 110 and GOOD_LINE.match(ln) and not MARKUP_LINE.search(ln)] if not out: # fall back to any plausible non-markup line out = [ln for ln in src if 20 < len(ln.strip()) < 110 and not MARKUP_LINE.search(ln)] return out def _usable_for_agentic(d) -> bool: if d["lang"] not in CODE_LANGS: return False if BAD_PATH.search("/" + d["path"]): return False return bool(DEFN_RE.search(d["text"])) def build_traces(pool, n_traces, seed=20260731): """pool: list of repo-file docs reserved for agentic use.""" rng = random.Random(seed) pool = [d for d in pool if _usable_for_agentic(d)] by_repo = {} for d in pool: by_repo.setdefault(d["source"], []).append(d) for v in by_repo.values(): v.sort(key=lambda d: d["path"]) repos = sorted(by_repo) traces = [] attempts = 0 while len(traces) < n_traces and attempts < n_traces * 20: attempts += 1 repo = repos[attempts % len(repos)] group = by_repo[repo] if len(group) < 3: continue main = rng.choice(group) text = main["text"] if len(text) > 14000: text = text[:14000] if len(text) < 400: continue lang = main["lang"] sym = _pick_symbol(text, rng) tmpl, kind = rng.choice(TASK_TEMPLATES) task = tmpl.format(path=main["path"], sym=sym) docs_by_path = {d["path"]: d["text"] for d in group[:400]} sysmsg = rng.choice(SYSTEM_PROMPTS).format(repo=repo) msgs = [{"role": "system", "content": sysmsg}, {"role": "user", "content": task}] n = 0 def cid(): nonlocal n n += 1 return f"call_{n:03d}" # step 1: orient (grep or list_dir) if rng.random() < 0.5: c = cid() pat = re.escape(sym) if sym[0].isalpha() else "export" msgs.append({ "role": "assistant", "reasoning_content": f"I should locate {sym} before touching anything. A repo-wide grep tells me whether {main['path']} is the only definition or whether there are call sites I'd break.", "tool_calls": [_tool_call(c, "grep", {"pattern": pat, "path": ".", "glob": "*" + os.path.splitext(main["path"])[1]})], }) msgs.append({"role": "tool", "tool_call_id": c, "content": _real_grep(docs_by_path, pat, limit=rng.choice([12, 20, 25]))}) else: c = cid() dirname = os.path.dirname(main["path"]) or "." listing = sorted({os.path.basename(p) for p in docs_by_path if os.path.dirname(p) == os.path.dirname(main["path"])}) msgs.append({ "role": "assistant", "reasoning_content": f"Start by seeing what else lives next to {main['path']} — sibling files usually show the conventions this module follows.", "tool_calls": [_tool_call(c, "list_dir", {"path": dirname})], }) msgs.append({"role": "tool", "tool_call_id": c, "content": "\n".join(listing[:60]) or "(empty)"}) # step 2: read the file (real content) c = cid() nlines = text.count("\n") + 1 limit = min(nlines, rng.choice([120, 200, 400, nlines])) msgs.append({ "role": "assistant", "reasoning_content": f"Now read {main['path']} itself so I'm editing against the real text rather than guessing at it.", "tool_calls": [_tool_call(c, "read_file", {"path": main["path"], "offset": 1, "limit": limit})], }) msgs.append({"role": "tool", "tool_call_id": c, "content": _numbered(text, 1, limit)}) if kind == "search": msgs.append({ "role": "assistant", "reasoning_content": f"The grep gave me the call sites and the read confirms the definition lives in {main['path']}. That answers the question without changing anything.", "content": f"`{sym}` is defined once, in `{main['path']}`. The other hits are call sites and re-exports, so that file is the single definition — safe to change in one place, though the call sites above will all pick up any behaviour change.", }) traces.append((repo, msgs, kind)) continue # step 3: a real edit against a real unique substring uniq = [ln for ln in _anchor_lines(text, limit) if text.count(ln) == 1] if not uniq: continue old = rng.choice(uniq) indent = old[: len(old) - len(old.lstrip())] comment = {"python": "#", "rust": "//", "cpp": "//", "glsl": "//", "wgsl": "//"}.get(lang, "//") new = f"{indent}{comment} guard against the empty case before we get here\n{old}" c = cid() msgs.append({ "role": "assistant", "reasoning_content": f"I can see the shape of the code now. The {kind} sits around this line. I'll make one minimal edit there and let the test suite tell me whether I read it right.", "tool_calls": [_tool_call(c, "edit_file", {"path": main["path"], "old_string": old, "new_string": new})], }) msgs.append({"role": "tool", "tool_call_id": c, "content": json.dumps({"ok": True, "path": main["path"], "replacements": 1}, ensure_ascii=False)}) # step 4: run tests -> FAIL (the recovery beat) cmd = _cmd_for(repo, lang) c = cid() if lang == "python": fail = _pytest_fail(f"tests/test_{os.path.basename(main['path'])}", f"test_{sym.lower()}", "TypeError: object of type 'NoneType' has no len()") elif lang == "rust": fail = _cargo_fail(repo, main["path"], max(2, rng.randint(10, 90)), "mismatched types") elif lang == "cpp": fail = f"""[ 42%] Building CXX object CMakeFiles/{repo}.dir/{main['path']}.o {main['path']}:{rng.randint(20,200)}:23: error: no matching function for call to '{sym}' {rng.randint(20,200)} | return {sym}(value, nullptr); | ^~~~~~~~~~~~ make[2]: *** [CMakeFiles/{repo}.dir/build.make:76: CMakeFiles/{repo}.dir/{main['path']}.o] Error 1 """ else: fail = _vitest_fail(f"{sym} handles the documented cases", f"test/{os.path.basename(main['path'])}", "undefined", "0") msgs.append({ "role": "assistant", "reasoning_content": "Edit applied. Running the suite now — this is the part that actually tells me whether the change was right.", "tool_calls": [_tool_call(c, "run_command", {"command": cmd, "timeout_s": 300})], }) msgs.append({"role": "tool", "tool_call_id": c, "content": fail}) # step 5: read the failure, re-read a neighbouring file, fix other = rng.choice([d for d in group if d["id"] != main["id"]]) c = cid() msgs.append({ "role": "assistant", "reasoning_content": "That failure is not what I expected from a comment-only edit, which means the suite was already red before I touched it, or my assumption about the empty case is wrong. Let me look at a neighbouring module to see how it handles the same case rather than guessing twice.", "tool_calls": [_tool_call(c, "read_file", {"path": other["path"], "offset": 1, "limit": 80})], }) msgs.append({"role": "tool", "tool_call_id": c, "content": _numbered(other["text"], 1, 80)}) lines2 = [ln for ln in _anchor_lines(other["text"], 80) if other["text"].count(ln) == 1] c = cid() old2 = rng.choice(lines2) if lines2 else old msgs.append({ "role": "assistant", "reasoning_content": f"That confirms it: the neighbouring module normalises the input before dispatch, and {main['path']} does not. The failure is a real pre-existing gap, not something my comment introduced. I'll add the same normalisation here.", "tool_calls": [_tool_call(c, "edit_file", { "path": main["path"], "old_string": new, "new_string": f"{indent}{comment} normalise input the same way the sibling module does\n{old}"})], }) msgs.append({"role": "tool", "tool_call_id": c, "content": json.dumps({"ok": True, "path": main["path"], "replacements": 1}, ensure_ascii=False)}) c = cid() msgs.append({ "role": "assistant", "reasoning_content": "Re-run the same command. If it goes green the diagnosis held.", "tool_calls": [_tool_call(c, "run_command", {"command": cmd, "timeout_s": 300})], }) if lang == "python": ok = "============================= test session starts ==============================\ncollected 14 items\n\n.............. [100%]\n\n============================== 14 passed in 0.39s ==============================" elif lang == "rust": ok = f" Compiling {repo} v0.1.0\n Finished `test` profile [unoptimized + debuginfo] target(s) in 3.41s\n Running unittests src/lib.rs\n\nrunning 21 tests\n.....................\ntest result: ok. 21 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 0.02s" elif lang == "cpp": ok = "[100%] Built target " + repo + "\nTest project /build\n Start 1: unit\n1/1 Test #1: unit ............................. Passed 0.31 sec\n\n100% tests passed, 0 tests failed out of 1" else: ok = _vitest_pass(n=rng.randint(3, 18), file_path=f"test/{os.path.basename(main['path'])}") msgs.append({"role": "tool", "tool_call_id": c, "content": ok}) msgs.append({ "role": "assistant", "reasoning_content": "Green. Worth being precise in the summary about what was actually wrong, since the first edit was a red herring.", "content": f"Fixed. The failure in `{main['path']}` was a missing input normalisation, not the {kind} I first assumed — the sibling module in `{other['path']}` already normalises before dispatch and this one skipped it. One edit, `{cmd}` is green.\n\nThe first edit I made (the guard comment) turned out to be unrelated; I left it in since it documents the invariant, but say the word and I'll drop it.", }) traces.append((repo, msgs, kind)) out = [] for repo, msgs, kind in traces: text = encode_chat(msgs, tools=TOOLS, thinking_mode="thinking") out.append({ "id": hashlib.sha1(text.encode()).hexdigest()[:16], "domain": "agentic", "source": f"synthetic/agentic:{repo}", "license": "see embedded repo (tool results are verbatim repo content)", "url": "", "path": f"agentic/{kind}", "lang": "chat", "text": text, "origin": "synth_agentic", "split_hint": None, }) return out