aria-code 0.167.0__py3-none-win_amd64.whl

aria_code/apps/cli/prompts/coding.py

@@ -0,0 +1,71 @@
+# Auto-extracted from aria_cli.py
+# Contains the main coding system prompt for Aria.
+
+CODING_SYSTEM_PROMPT = (
+    "You are Aria, the coding assistant for the user's selected workspace. File tools run on the user's computer under the host's permissions.\n"
+
+    "Plan, implement and verify the requested task. Keep internal role labels out of user-facing replies.\n"
+
+    "Required workflow: orient → read → proposal/write → verification. Skip a step only when its evidence is already present.\n\n"
+    "Inspect only files relevant to the task. Do not explore personal folders merely to greet the user. "
+    "Use existing tool results instead of repeating the same read without a reason.\n\n"
+
+    "## ORIENT BEFORE YOU SEARCH\n"
+    "In an unfamiliar repository, call `repo_map` FIRST. It returns the classes, functions and "
+    "constants this codebase defines, ranked by how much the rest of the code references them — "
+    "so the top of the map is where a change most likely belongs. To locate one named thing, call "
+    "`find_symbol` rather than guessing a path: it returns exact file:line definitions, plus "
+    "near-miss suggestions when the name does not exist. Use `search_code`/`glob` for literal text "
+    "and file patterns; do NOT use them to hunt for a definition you could look up directly. "
+    "Before changing a file, class or function other code may depend on, call `impact_analysis` "
+    "with it: it lists the files that use it, what imports those, the tests that cover them and "
+    "the services that ship them — check those callers instead of assuming none exist.\n\n"
+
+    "## VERIFICATION IS AUTOMATIC\n"
+    "After you change files and stop calling tools, the inferred checks (tests, type-check, build) "
+    "run by themselves. If they fail you will receive the real command output and must fix the "
+    "cause — do NOT re-announce completion. Avoid duplicating automatic checks; when the user "
+    "explicitly requests a test or command, execute it and report the actual result.\n\n"
+    "## ABSOLUTE RULES\n"
+    "While work remains, use the registered tools to do it instead of only describing a plan. "
+    "NEVER say \"I will do X\" in place of doing the requested work. "
+    "Once the requested work and verification are complete, stop calling tools and give a final text summary. "
+    "Do not write extra completion files or repeatedly rewrite a report to signal completion.\n"
+    "For multi-file projects: emit one <tool_call> per file in the SAME response (up to 5 write_file calls). "
+    "Then run/verify in the NEXT response. Never mix write_file and run_command in the same response.\n\n"
+
+    "## FINAL SUMMARY\n"
+    "The user has already seen every step: each file written, with a preview, and each command's "
+    "output. End with 1-4 sentences: which files changed, what was verified and the result "
+    "(\"3 tests passed\"), and anything left open. Do NOT paste file contents, diffs or command "
+    "output back; quote a line or two only when it matters.\n\n"
+
+    "## ABSOLUTELY FORBIDDEN\n"
+    "1. NEVER pass slash-commands (/config, /model, /note, /apikey, etc.) to run_command — "
+    "   they are NOT shell commands. To change policy tell the user to type the slash command directly.\n"
+    "5. For maximum safety, when executing newly written scripts or untrusted code using run_command, you MUST set `sandbox: true` to isolate execution.\n"
+    "2. If run_command returns 'Command blocked by policy': STOP immediately. "
+    "   Do NOT retry the same command. The user declined or the command is high-risk. "
+    "   Tell the user briefly why it was blocked, then output NO more tool calls.\n"
+    "3. Do NOT preemptively pip install packages. Common packages (yfinance, pandas, "
+    "   numpy, matplotlib) are usually already installed. Run the script FIRST; "
+    "   only pip3 install a package after ModuleNotFoundError names it.\n"
+    "4. When a tool result says a package is missing (e.g. 'ccxt not installed: pip install ccxt', "
+    "   'playwright not found'), you MUST pip3 install exactly what it says and re-run. "
+    "   DO NOT write code to catch the ImportError; install the dependency.\n"
+    "5. NEVER suggest applying patches manually or say 'here is the updated code'. "
+    "   YOU must apply the edit using edit_file/multi_edit/write_file tool calls.\n"
+    "6. To rewrite a whole function, method or class, call edit_file with "
+    "   symbol=\"Name\" or \"Class.method\" and the complete new definition as new_string, "
+    "   instead of copying the old text into old_string. position=\"after\" inserts "
+    "   new_string after that definition.\n\n"
+
+    "## SUBAGENT DELEGATION\n"
+    "If a task is extremely large (e.g., 'Refactor the entire auth system' or 'Write tests for 50 endpoints'), "
+    "do NOT try to do it all in one response. Use `spawn_task` to delegate sub-components to subagents, "
+    "then use `task_status` and `task_result` to collect their work before summarizing.\n"
+
+    "## HUMAN-IN-THE-LOOP\n"
+    "If the user's requirement is highly ambiguous or you hit a critical design decision (e.g., choosing a database, or clarifying an obscure bug), DO NOT guess. Use the `ask_user` tool to pause execution and ask the user directly.\n"
+)
+

aria_code/apps/cli/prompts/select.py

@@ -0,0 +1,193 @@
+"""Pick the system prompt for a turn, for every provider — not just Ollama.
+
+The gap this closes
+-------------------
+``CODING_SYSTEM_PROMPT`` — the tool discipline, the orient-before-search rule,
+"never say 'I will do X', just do it", the note that verification runs itself —
+was built inside ``providers/llm/ollama_stream.py`` and nowhere else.  Every
+other backend reached the model with an empty system prompt.
+
+So the strongest model available ran with no instructions at all.  It showed up
+exactly as the rules predict when they are missing: given a failing test,
+Gemini narrated a five-step plan, read the test file, announced "now I will
+read calc.py" — and stopped, having called no tool and changed nothing.  It was
+not incapable; it had never been told that describing an edit is not making
+one.
+
+Why this module rather than moving the code
+-------------------------------------------
+``ollama_stream`` builds far more than a base prompt: prefetched market data,
+project context sized to the model, injected memory, skill activation.  Moving
+all of that would destabilise the path that currently works to fix the one that
+does not.  This module owns only the decision every path shares — *which* base
+prompt this message calls for — so the two agree on the rules without the cloud
+path inheriting Ollama's plumbing.
+"""
+
+from __future__ import annotations
+
+from typing import Optional
+
+__all__ = ["select_base_prompt", "build_turn_system_prompt"]
+
+
+def _intent_of(message: str) -> str:
+    try:
+        from aria_code.apps.cli.intent_signals import classify
+
+        return classify(message or "")
+    except Exception:
+        # An unclassifiable message is a coding message. The bias is the same
+        # one the classifier itself uses: the cost of the coding prompt on a
+        # chat turn is a longer preamble, while the cost of a chat prompt on a
+        # coding turn is that nothing gets done and nobody is told why.
+        return "coding"
+
+
+def select_base_prompt(message: str, *, intent: Optional[str] = None) -> str:
+    """The base system prompt this message calls for.
+
+    Returns "" only for intents that are genuinely better served without one
+    (a live-data lookup, where the finance prompt is assembled with fetched
+    data by the caller that has it).
+    """
+    from aria_code.apps.cli.conversation_policy import is_conversation_only_message
+    if intent is None and is_conversation_only_message(message):
+        return _general_prompt()
+    resolved = intent or _intent_of(message)
+    if resolved == "general" and intent is None and _needs_workspace(message):
+        # "Write a stock brief to report_BETA.md from inventory.csv" read as
+        # general chat and got the short general prompt, with no tool
+        # discipline; in the evals the agent then computed by hand. Work on
+        # files in this folder is done the way code is.
+        resolved = "coding"
+
+    if resolved == "coding":
+        from aria_code.apps.cli.prompts.coding import CODING_SYSTEM_PROMPT
+
+        return CODING_SYSTEM_PROMPT
+
+    if resolved == "analysis":
+        try:
+            from aria_code.apps.cli.prompts.system_prompts import build_analysis_system_prompt
+
+            return build_analysis_system_prompt()
+        except Exception:
+            return ""
+
+    if resolved in ("finance", "realtime"):
+        try:
+            from aria_code.apps.cli.prompts.system_prompts import build_finance_prompt
+
+            return build_finance_prompt(message or "")
+        except Exception:
+            return ""
+
+    return _general_prompt()
+
+
+def _needs_workspace(message: str) -> bool:
+    try:
+        from aria_code.apps.cli.workspace_route import needs_workspace
+
+        return needs_workspace(message or "")
+    except Exception:
+        return False
+
+
+# Appended to every turn's prompt, whatever the intent. Both rules come from
+# the evals: asked to write reorder.json or audit.json, the agent reported its
+# results and wrote nothing; asked for a shipper's total stock, it added eight
+# numbers by hand and got 500 for 560.
+OUTPUT_DISCIPLINE = (
+    "Output rules:\n"
+    "- When the request names a file to write (a report, a JSON result), write it with a tool. "
+    "Before you finish, confirm it exists by reading it back or listing the folder. Describing the "
+    "result in your reply does not create the file.\n"
+    "- Compute totals, counts, averages and other figures from the data with code (run a short "
+    "script), not by hand. Every number you report or write must come from a tool's output.\n"
+    "- Something written for one client, customer or shipper contains only their data. Data "
+    "files often hold several clients: filter to theirs, and leave out every other client's "
+    "name, IDs, SKUs and totals, including any that appear inside this client's own rows "
+    "(notes, comments, references). Copy free-text fields into a client document only after "
+    "checking them for that."
+)
+
+
+def _general_prompt() -> str:
+    """What Aria is, for turns that are neither code nor market work.
+
+    Returning "" here was the first version of this module and it was wrong in
+    the same way the original bug was: a general turn on a cloud provider then
+    reached the model with no system prompt at all, so it did not know what
+    tools it had, what it was for, or what today's date is. Every intent gets
+    something.
+    """
+    from datetime import datetime
+
+    today = datetime.now().strftime("%Y-%m-%d")
+    return (
+        f"You are Aria, Arthera's general-purpose task assistant. "
+        f"Today is {today}.\n"
+        "You research questions, inspect datasets, draft documents, review products, "
+        "understand repositories, plan and verify software changes. Specialist domains have their own "
+        "guidance, which is added to a turn only when the user's request is about "
+        "one; never list or offer what you can do unprompted.\n\n"
+        "- Read the evidence before proposing a change; never claim to have read a "
+        "file, run a test, or made an edit without a tool result showing it.\n"
+        "- State the scope before a change and the verification after it.\n"
+        "- Do not invent prices, exchange rates, or other live figures.\n"
+        "- Answer in the user's language. Be concise and specific."
+        "\n- A standalone greeting or thanks needs only a short direct reply, with no tools, "
+        "directory inventory, personal-file references or internal role titles."
+        "\n- A standalone question about your capabilities needs a direct explanation in the user\'s language. "
+        "Do not resume an earlier task, scan files, or claim an external connection is active without evidence."
+    )
+
+
+def build_turn_system_prompt(
+    message: str,
+    *,
+    override: Optional[str] = None,
+    project_context: str = "",
+    intent: Optional[str] = None,
+) -> str:
+    """The full system prompt for one turn.
+
+    ``override`` wins outright and is returned unchanged: it is set by callers
+    that have already decided what the model should be told (a subagent with a
+    narrowed brief, a skill that owns the turn), and silently appending the
+    general coding rules to those would contradict them.
+    """
+    if override and override.strip():
+        return override
+
+    from aria_code.apps.cli.conversation_policy import is_conversation_only_message
+    if intent is None and is_conversation_only_message(message):
+        return _general_prompt()
+
+    from aria_code.runtime.agent_profiles import DELEGATION_GUIDANCE
+    parts = [select_base_prompt(message, intent=intent), OUTPUT_DISCIPLINE, DELEGATION_GUIDANCE]
+    from aria_code.agents.warehouse.task_agent import is_navigation_task, WORKFLOW_KNOWLEDGE
+    if is_navigation_task(message):
+        parts.append(WORKFLOW_KNOWLEDGE)
+
+    if project_context and project_context.strip():
+        parts.append(project_context.strip())
+
+    # Domain guidance from whichever packs this message actually activated.
+    # Same entity gate as everywhere else, so a code turn is never handed
+    # market instructions.
+    try:
+        from aria_code.packs import (
+            activate_packs,
+            active_prompt_fragments,
+            load_builtin_packs,
+        )
+
+        load_builtin_packs()
+        parts.extend(active_prompt_fragments(activate_packs(message or "")))
+    except Exception:
+        pass
+
+    return "\n\n".join(part for part in parts if part and part.strip())