runloopable / platform /harness /analyst.py
fsanyoto's picture
Deploy AIOS web (React glide grid + FastAPI slice)
8133df9 verified
Raw History Blame Contribute Delete
67.5 kB
"""harness/analyst.py β€” the AIOS Analyst loop (OM-4, 2026-07-11). A CHEAP small model (via OpenRouter, OpenAI-compatible) drives the compounding tool registry (harness/tools.py) to answer questions, build charts, and spawn dashboards over the semantic layer. The model NEVER writes SQL β€” it picks registry keys, guided by the in-product SKILL recipe index and exact on-demand bodies (model/skills/*.skill.yml). Plan: omni-adoption addendum. Architecture (mirrors the Omni coordinator pattern): ask(question) -> bounded tool loop: chat -> tool_calls -> tools.dispatch() -> results -> chat … until a plain answer or MAX_ITERS. chat_fn is INJECTABLE: live = OpenRouter POST; tests = a scripted double β€” the loop mechanics are provable offline (no key, no tokens). Correctness posture: bounded iterations (never endless); tool errors go BACK to the model as readable envelopes (it self-corrects); the final answer carries the tool trace (auditability + the drill guarantee); exhausted-budget is reported AS exhausted, never as success (loop rules). Cost posture: DEFAULT_MODEL is a cheap tool-calling model; override per-tenant/env; prompt = the skill index + topic summary (small, cacheable), NOT raw data. """ import json import os from pathlib import Path import harness.tools as T # Provider LADDER (owner 2026-07-12: "use Groq and Cerebras"): all OpenAI-compatible chat # endpoints; tried in order, failing over on payment/rate/server errors β€” an exhausted or # rate-limited provider degrades to the next instead of taking the Analyst down (the # OpenRouter-402 incident). Pin one with ANALYST_PROVIDER=groq|cerebras|openrouter; override # its model with ANALYST_MODEL. Every model here must re-pass the eval gate (Β§4b). PROVIDERS = [ {"name": "groq", "env": "GROQ_API_KEY", "url": "https://api.groq.com/openai/v1/chat/completions", "model": "openai/gpt-oss-120b"}, {"name": "cerebras", "env": "CEREBRAS_API_KEY", "url": "https://api.cerebras.ai/v1/chat/completions", "model": "gpt-oss-120b"}, # this account's tool-capable model (from /v1/models) {"name": "openrouter", "env": "OPENROUTER_API_KEY", "url": "https://openrouter.ai/api/v1/chat/completions", # The eval measures this default: analyst_golden.yml `measured` names it, open weights only (R11). "model": "z-ai/glm-4.7-flash"}, ] DEFAULT_MODEL = None # None = the selected provider's default model # A map, YoY or name-resolution ask needs about 8 tool turns with no retry (list, describe, # recipe, field values, query, transform, chart, answer), so 8 exhausted 7 of 26 items (D-8280). # 14 leaves room for 6 error-correction turns. ANALYST_MAX_ITERS overrides it per run. MAX_ITERS = 14 def iter_budget(): """The tool-loop budget: ANALYST_MAX_ITERS when set to a positive integer, else MAX_ITERS.""" try: value = int(os.environ.get("ANALYST_MAX_ITERS") or MAX_ITERS) except ValueError: return MAX_ITERS return value if value > 0 else MAX_ITERS #: Working-context byte ceiling per provider. Groq and Cerebras fail near 8k request tokens. #: OpenRouter tool models take 128k+ tokens. 120000 bytes (about 30k tokens) fits a full #: discovery run and keeps each request cheap. REQUEST_BYTES = {"groq": 32000, "cerebras": 32000, "openrouter": 120000} DEFAULT_REQUEST_BYTES = 32000 def request_budget(provider=None): """The request byte ceiling: ANALYST_MAX_REQUEST_BYTES when a positive integer, else per provider.""" try: value = int(os.environ.get("ANALYST_MAX_REQUEST_BYTES") or 0) except ValueError: value = 0 if value > 0: return value return REQUEST_BYTES.get(provider, DEFAULT_REQUEST_BYTES) #: Set by harness/evals.run to the golden item id, so each ledger line names its item. CALL_ITEM = None #: Running totals of ledger lines written by this process (evals.run reads the delta). CALL_TOTALS = {"calls": 0, "cost": 0.0, "cost_missing": 0} def _log_call(usage): """Append one cost line to ANALYST_CALL_LOG. Never writes message or answer text.""" path = os.environ.get("ANALYST_CALL_LOG") if not path: return import time as _time usage = usage or {} cost = usage.get("cost") cost = float(cost) if isinstance(cost, (int, float)) and not isinstance(cost, bool) else None line = {"ts": _time.strftime("%Y-%m-%dT%H:%M:%SZ", _time.gmtime()), "provider": usage.get("provider"), "model": usage.get("model"), "item": CALL_ITEM, "prompt_tokens": usage.get("prompt_tokens"), "completion_tokens": usage.get("completion_tokens"), "cost": cost} with open(path, "a", encoding="utf-8", newline="") as fh: fh.write(json.dumps(line, separators=(",", ":")) + "\n") CALL_TOTALS["calls"] += 1 if cost is None: CALL_TOTALS["cost_missing"] += 1 else: CALL_TOTALS["cost"] += cost def available_providers(): pin = os.environ.get("ANALYST_PROVIDER") ps = [p for p in PROVIDERS if os.environ.get(p["env"])] if pin: ps = [p for p in ps if p["name"] == pin] or ps return ps def llm_configured(): return bool(available_providers()) SKILLS_DIR = Path(__file__).resolve().parents[1] / "model" / "skills" def _skill_catalog(): """Read only installed recipes; request arguments never become filesystem paths.""" import yaml return {skill["key"]: skill for path in sorted(SKILLS_DIR.glob("*.skill.yml")) for skill in [yaml.safe_load(path.read_text(encoding="utf-8"))]} def _skills_text(): """Complete discovery index; exact bodies are loaded by get_recipe on demand.""" parts = [] for key, skill in _skill_catalog().items(): parts.append(f"{key}: {skill['label']}") parts.extend(f"{i}: {recipe['ask']}" for i, recipe in enumerate(skill.get("recipes", []), 1)) return "\n".join(parts) or "(no skills installed)" def _get_recipe(arguments): """Local read-only metadata tool, with a strict fail-closed argument contract.""" try: args = json.loads(arguments) if isinstance(arguments, str) else arguments if not isinstance(args, dict) or set(args) - {"skill", "recipe", "section"}: raise ValueError("expected an object with skill, optional recipe or section") key = args.get("skill") catalog = _skill_catalog() if not isinstance(key, str) or key not in catalog: raise ValueError("unknown installed skill") skill = catalog[key] data = {k: v for k, v in skill.items() if k != "recipes"} if "recipe" in args and "section" in args: raise ValueError("select either recipe or section") if "recipe" in args: number = args["recipe"] recipes = skill.get("recipes", []) if type(number) is not int or not 1 <= number <= len(recipes): raise ValueError("recipe must be an integer from the skill index") data = {k: v for k, v in data.items() if k not in ("ops", "kinds", "transforms", "metrics")} data["recipe"] = recipes[number - 1] elif "section" in args: section = args["section"] if not isinstance(section, str) or section not in ( "ops", "kinds", "transforms", "metrics") or section not in skill: raise ValueError("unknown metadata section for this skill") data = {"key": key, "context": skill["context"], section: skill[section]} return {"ok": True, "data": data} except (ValueError, TypeError, KeyError) as error: return {"ok": False, "error": str(error)} def _tool_schemas(data_source=None): """Keep complete argument contracts; skill bodies and catalogs load on demand.""" descriptions = { "list_topics": "List topics with metrics, dims and grain. The system prompt's TOPIC " "INDEX already holds this list.", "describe_topic": "Full topic schema, scope, definitions, dims and business context.", "get_field_values": "Resolve real ids/names BEFORE filtering typed names.", "run_semantic_query": "Read governed data using registered measures and dims.", "transform_result": "Chain governed analytics ops from the analytics skill; returns a NEW " "result_id. Never compute them yourself. Additivity rules apply.", "make_chart": "Render result columns as a chart; follow the charting skill. Map: x=customer " "dim, dot size=y. Heatmap: x/y=dims, value=colour. Histogram bins x, no y.", "make_table": "Render exact figures as a drillable, formatted table with totals; columns " "optionally selects/orders columns.", "make_kpi_card": "Render a scalar KPI; compare_result_id optionally adds YoY delta.", "report_gap": "LAST RESORT after schema/transform/chart checks; follow the GAP PROTOCOL.", "save_view": "Save a named chart view. Confirm first.", "compose_dashboard": "Compose saved views into a named dashboard. Confirm first.", "list_workspace": "List all saved workspace objects and available templates.", "instantiate_template": "Stamp a template as NEW objects, never overwrite. Confirm first.", "update_view": "Patch an existing view/spec/query, re-query before saving; invalid updates " "rejected. Null removes keys. Confirm first.", "update_workbook": "Reorder/add/remove views or rename a dashboard. Confirm first.", "delete_object": "Delete a named workspace object. DESTRUCTIVE: confirm first.", } return [{**tool, "function": {**tool["function"], "description": descriptions.get(tool["function"]["name"], tool["function"]["description"])}} for tool in (data_source if data_source is not None else T).openai_tools()] + [{ "type": "function", "function": { "name": "get_recipe", "description": "Read exact local skill context and recipe/guard by 1-based " "index. Omit recipe for full skill metadata; section selects " "ops, kinds, transforms or metrics. Read before using a recipe.", "parameters": { "type": "object", "additionalProperties": False, "properties": { "skill": {"type": "string", "enum": list(_skill_catalog())}, "recipe": {"type": "integer", "minimum": 1}, "section": {"type": "string", "enum": [ "ops", "kinds", "transforms", "metrics"]}}, "required": ["skill"]}}}] def _topic_index(topics): """One line per queryable topic with its metrics and dims; grid-only topics on one line.""" lines = [f"{t['topic']} ({t.get('label') or t['topic']}): {', '.join(t['metrics'])}; " f"dims {', '.join(t.get('dims') or []) or 'none'}" for t in topics if t.get("metrics")] grids = [t["topic"] for t in topics if not t.get("metrics")] if grids: lines.append("grids with no metrics (no numbers to query): " + ", ".join(grids)) return "\n".join(lines) def _ops_index(): """Every analytics op signature as written in the analytics skill, on one line.""" import re ops = (_skill_catalog().get("analytics") or {}).get("ops") or {} text = " ".join(map(str, ops.values())) if isinstance(ops, dict) else str(ops) sigs = dict.fromkeys(f"{name}{{{params}}}" for name, params in re.findall(r"\b([a-z_0-9]+)\{([^}]*)\}", text)) return " ".join(sigs) or "(read get_recipe(skill=analytics, section=ops))" def system_prompt(data_source=None): """The small model's standing instructions: the platform rules + the skill recipes. Stable content (cache-friendly); per-question content goes in the user turn.""" topics = (data_source if data_source is not None else T).list_topics() topic_summary = _topic_index(topics) team_codes = _team_codes() or "none configured" team_args = " or ".join(f"team_id={tid} ({label})" for tid, label in _team_labels().items()) or "team_id" return f"""AIOS Analyst: wholesale data questions via governed TOOLS. Non-negotiable rules: 1. NEVER write SQL or invent fields/joins/numbers. Use registered schema keys only; recipes and tools version together. The TOPIC INDEX below names every metric and dim: pick the topic from it, describe_topic it once for its scope rules, then query. Wholesale sales, margin, units and distinct customers come from sales_lines; order counts from sales_orders. Invoice-line basis, commission agent, allocated or unallocated sales: invoice_lines. An explicit revenue ask uses revenue, including a short ask such as "Revenue for June". Resolve its stated window and query it even when there is no earlier conversation. 2. Resolve every typed name or place (customer, product, agent, salesperson, city, state) with get_field_values BEFORE filtering, even when the spelling looks exact. Every filter value of any dim comes from get_field_values or an earlier result, never typed from the ask. A metric named for a subset (agent_allocated_sales, unallocated_line_sales) already applies that subset: add no filter for it. Once a name is resolved, EVERY query of the ask filters on its id: a customer list for an agent is filters {{"agent": [<id>]}}, not the whole book. Business units: {team_codes}; never mix them in one series unless explicitly comparing. 3. Every presented number comes from a tool result with drillable result_id. Read tool errors and correct calls; never fabricate numbers, including rounded approximate aggregates. 4. Artifact actions (save_view, compose_dashboard, schedule_report, create_alert) need the user's explicit confirmation first. Follow every recipe guard, including updates/deletes. 5. Lead plainly with the finding; state scope and the result's 'window', never guess it. The final answer names the finding in words with its numbers (for a ranking: the top names and values), even when a chart or table is also shown. Never only point to a chart. No dates means ALL recorded history, no hidden default. Time-bound asks require date_from/date_to. 6. Single-number asks: give the window/scope TOTAL first, then breakdowns. Consolidated asks: query WITHOUT group_by or sum the exact returned rows; never give only monthly/per-BU components. 7. Totals require a scalar query or adding EXACT fetched rows; otherwise show rows without a total. NEVER join/diff result sets by hand. transform_result yoy adds <m>_ly/_delta/_yoy_pct; if no transform can compute the comparison, say you cannot. 8. GAP PROTOCOL: check list_topics/describe_topic AND transforms/chart kinds first. Only if no registered dim/metric/transform/chart can answer, report_gap(kind,missing,question,workaround) ONCE; state what is missing in one honest sentence and offer the nearest answerable ask. YoY/decline/growth, rankings (top_n/sort/head), shares, running totals and distributions are answerable via transform_result. Refusing without report_gap OR reporting supported asks fails. Share or concentration of the top N customers is ANSWERABLE (concentration transform). Balances at a point in time (stock on hand, units in stock now, inventory level) have NO metric: a movement topic such as stock_moves sums flows inside a window and is never a balance. For such an ask call report_gap once, then offer the movement figure as the nearest ask. Never present a sum of movements as what is on hand. A window is never a gap: every dated topic takes date_from/date_to, with no date dim. 9. Charts ONLY via make_chart/make_table; say "the chart below". NEVER emit markdown images, data: URIs, chart links, HTML or vega. Derived analytics (top-N/YoY/shares/running totals/ moving averages/ranks/bins) ONLY via transform_result, never mental arithmetic. 10. Write plain text. Never use an em dash or an en dash; write "to" for a range. QUERY ARGUMENTS (run_semantic_query; a wrong argument costs a turn): - Window: date_from and date_to as YYYY-MM-DD, never inside filters. - Business unit: {team_args} only when the ask names one; omit it otherwise. Never filters.team. - Monthly or weekly series: grain="month" or "week" with NO group_by returns one row per period. group_by takes only the topic's dims; there is no date dim to group by. - filters: {{"<dim>": [values]}} with the exact ids get_field_values returned (agent, partner, product). For a city or state pass EVERY matching spelling it returned in one list (case and trailing-space variants are separate values). - sort: "-<measure>" is largest first (top, biggest, most); "<measure>" is smallest first. - A total: no group_by. Rows: a result shows at most 100 preview rows, and row_count is the true count. Never count preview rows. Transforms and charts read every row. - To count rows that meet a condition (customers in a tier, above a value), apply transform_result {{"op":"filter_rows","col":<column>,"cmp":"==","value":<v>}} (cmp: ==, !=, >, >=, <, <=, contains) and report the row_count of THAT new result. abc_classify{{of}} adds abc_class A/B/C. add_total appends a total row: never count rows after it. - A map needs no coordinates: group_by the customer dim, then make_chart kind="map" with x=the customer dim and y=the measure. The platform places each customer. - Same window last year, YoY, decline or growth per customer: ONE query of the current window with group_by, then transform_result [{{"op":"yoy"}}, {{"op":"sort","by":"revenue_delta", "direction":"asc"}}, {{"op":"head","n":10}}]. Never query the prior year yourself: two results cannot be joined by hand. - Share held by the top 1, 5 or 10 customers: query grouped by customer, then transform_result [{{"op":"concentration","of":"revenue"}}] ALONE on that full result (no sort, head or add_total before it). Report its top1/top5/top10_share_pct as given. - TRANSFORM OPS, exact names and params (no others exist): {_ops_index()} - Never repeat a call with the same arguments: read its error, change the call, or answer. CHART PICKER: trend=line (few periods=bar); seasonal/two-dim intensity=heatmap; YoY by period=yoy transform then yoy_bars; ranking=top_n then sorted/labeled ranked_bar; parts=pie/donut <=6 slices (top_n first), over time=stacked_bar, shifting shares=stacked_pct, hierarchy=treemap; change/contribution=waterfall (signed steps, platform adds Total); concentration=pareto; correlation=scatter, third measure=bubble size; distribution=histogram or bin transform then bar; stages=funnel; actual/target=bullet (y=actual,y2=target); level/rate=combo (y=bars left,y2=line right); small multiples=facet on line/bar/area/scatter; exact figures/many columns/lists=make_table with totals; customer geography=map. Ties: simpler. TOPIC INDEX (topic (label): metrics; dims. describe_topic gives definitions and scope; an ask that no metric here measures, such as stock on hand, is a gap: call report_gap once, never a prose-only refusal): {topic_summary} SKILL RECIPE INDEX (get_recipe(skill,recipe) loads the exact plan, context and guard; read the relevant recipe BEFORE acting, then adjust its [bracketed] parameters. get_recipe(skill) loads full metadata; analytics ops and charting kinds are available via section="ops"/"kinds". Never invent an op or arguments; read definitions on demand. All skill context/guards remain mandatory, subordinate to the standing rules above): {_skills_text()} TOOL CONTEXT: exact_cache receipts replace stale tool bodies in this ask only. To read one, call its original tool with the EXACT arguments shown: this restores the complete cached response, without re-querying or repeating an action. Never infer missing fields or numbers from a receipt. Loaded context, scope and guards remain mandatory in RETAINED RULES. artifact_receipt confirms rendering: full exact rows remain in the returned artifact, outside this prompt. Use its spec's result_id for further tools; use query/transform rows for numbers, never infer totals from a render receipt.""" _DEAD_STATUS = {401, 402, 403} # payment/auth: skip the provider immediately, no retry _RETRY_STATUS = {429, 500, 502, 503} # rate/server: retry with backoff before failing over _MAX_ANSWER_TOKENS = 1500 # answers are short; also keeps low-credit providers usable #: Reasoning models spend completion tokens on thought: 1500 cut glm-4.7-flash off mid-answer. _MAX_TOKENS_BY_PROVIDER = {"openrouter": 6000} _COOLDOWN = {} # provider name -> unix ts until which we skip it (in-proc) #: Live turns left at which the model is told to stop reading and answer. _BUDGET_WARN_TURNS = 3 #: Re-asks per ask for an empty reply or a tool call written as text. _MAX_REPLY_RETRIES = 2 #: An identical successful read may run twice: once, then once to restore its receipt. _MAX_SAME_READS = 2 _TOOL_MARKUP = ("<tool_call>", "<arg_key>", "<|tool", "<function=") def _unusable_reply(content): """True for an empty reply or one that writes a tool call as text instead of calling it.""" text = (content or "").strip() return not text or any(mark in text for mark in _TOOL_MARKUP) def _team_labels(): """{team_id: label} for the active tenant's teams, {} when this process cannot name them.""" try: import core.odoo as _odoo return _odoo.active_team_names() except Exception: return {} def _team_codes(): """The active tenant's teams as `<label>=<id>` pairs, comma separated.""" return ", ".join(f"{label}={tid}" for tid, label in _team_labels().items()) def _bu_filter_error(): """The tool error for a business unit passed in filters.team instead of team_id.""" return (f"A business unit is the team_id argument ({_team_codes()}), not filters.team. " "Move it to team_id, drop team from filters, and run the query again.") def _bu_in_filters(arguments): """True when a query filters team by a business-unit id instead of passing team_id.""" try: args = json.loads(arguments) if isinstance(arguments, str) else arguments except (TypeError, ValueError): return False filters = args.get("filters") if isinstance(args, dict) else None if not isinstance(filters, dict) or "team" not in filters or args.get("team_id") is not None: return False values = filters["team"] if isinstance(filters["team"], list) else [filters["team"]] bu_values = _bu_values() return bool(values) and all(str(v).strip().lower() in bu_values for v in values) def _bu_values(): """Every id and lower-cased label that names one of the active tenant's teams.""" teams = _team_labels() return {str(tid) for tid in teams} | {str(label).strip().lower() for label in teams.values()} #: Ops whose answer is a share of the WHOLE result, and ops that cut rows away first. _WHOLE_OPS = {"concentration", "abc_classify", "share_of_total", "cum_share"} _CUT_OPS = {"head", "top_n", "bottom_n", "filter_rows", "add_total"} _CUT_BEFORE_WHOLE_ERROR = ("concentration, abc_classify, share_of_total and cum_share measure the " "WHOLE result, so a head, top_n, bottom_n, filter_rows or add_total " "before them gives a wrong share. Run them first, on the full grouped " "result, and cut rows only after them.") _DATE_FILTER_ERROR = ("A window is the date_from and date_to arguments (YYYY-MM-DD), never a " "filter. Move the dates there, drop them from filters, and run it again.") def _date_in_filters(arguments): """True when a query puts a date or a window inside filters.""" import re try: args = json.loads(arguments) if isinstance(arguments, str) else arguments filters = args.get("filters") except (TypeError, ValueError, AttributeError): return False if not isinstance(filters, dict): return False for key, value in filters.items(): values = value if isinstance(value, list) else [value] if ("date" in str(key).lower() or str(key).lower() in ("month", "period", "year") or any(re.search(r"\b(19|20)\d\d-(0[1-9]|1[0-2])-(0[1-9]|[12]\d|3[01])\b", str(v)) for v in values)): return True return False _RECEIPT_ROWS = 3 _RECEIPT_ROW_BYTES = 800 _WHOLE_ON_CUT_ERROR = ("That result_id holds only part of the rows (a limit, head, top_n, " "bottom_n or filter_rows cut it), so a share or tier on it is wrong. Run " "the grouped query again with no limit, then transform that result_id.") _YOY_BY_HAND_ERROR = ("This is the same query one year earlier. Never compare two results by " "hand: run transform_result on result_id {result_id} with " "[{{\"op\":\"yoy\"}}, {{\"op\":\"sort\",\"by\":\"<m>_delta\"," "\"direction\":\"asc\"}}, {{\"op\":\"head\",\"n\":10}}]. yoy adds <m>_ly, " "<m>_delta and <m>_yoy_pct per row.") def _note_field_values(name, arguments, res, field_values): """Remember the text values get_field_values returned for each (topic, dim).""" if name != "get_field_values" or not res.get("ok") or not isinstance(res.get("data"), list): return args = _parsed(arguments) ids = [row.get("id") for row in res["data"] if isinstance(row, dict)] key = (args.get("topic"), args.get("dim")) # Ids are kept as text too: an id the model echoes as "9959" is still a resolved value. field_values.setdefault(key, set()).update(str(v) for v in ids if v is not None) def _expand_variants(arguments, field_values): """Add every returned spelling of a text filter value: 'BROOKLYN' also takes 'Brooklyn '.""" args = _parsed(arguments) filters = args.get("filters") if not isinstance(filters, dict): return arguments fixed = dict(filters) for dim, value in filters.items(): known = field_values.get((args.get("topic"), dim)) values = value if isinstance(value, list) else [value] if not known or not all(isinstance(v, str) for v in values): continue want = {v.strip().casefold() for v in values} extra = sorted(k for k in known if k.strip().casefold() in want and k not in values) if extra: fixed[dim] = list(values) + extra if fixed == filters: return arguments return json.dumps({**args, "filters": fixed}, ensure_ascii=False) _YOY_REASK_NOTE = ("\nCHECK: you queried the same grouping for two windows one year apart and " "have not run yoy. Two results are never compared by hand. Call " "transform_result on result_id {result_id} (the LATER window) with " "[{{\"op\":\"yoy\"}}, {{\"op\":\"sort\",\"by\":\"<m>_delta\",\"direction\":\"asc\"}}, " "{{\"op\":\"head\",\"n\":10}}], then answer from that result.") _GAP_REASK_NOTE = ("\nCHECK: your draft says the data is not available, but report_gap was never " "called. If no registered metric answers the ask, call report_gap now, once, " "then answer. If a listed metric or transform does answer it, call that tool.") _UNAVAILABLE = (r"\b(not available|unavailable|not tracked|isn't tracked|is not tracked|" r"no (metric|topic|data|dim|dimension|field) (for|that|to|on|which)|" r"(cannot|can't|can not|unable to) (answer|provide|determine|report|compute|" r"calculate|give|tell)|does not (track|hold|record|provide|contain|have a)|" r"doesn't (track|hold|record|provide|contain|have a)|not (in|part of) the " r"(governed|available|registered))") _DATE_GROUP_ERROR = ("There is no date dim to group by. A window is date_from and date_to; a " "monthly or weekly series is grain=\"month\" or \"week\" with no group_by.") _WINDOW_GAP_ERROR = ("A window is never a gap: every dated topic takes date_from and date_to, and " "grain gives months or weeks. Run the query with them instead.") _TYPED_FILTER_ERROR = ("The filter value for '{dim}' was typed, not resolved. Call get_field_values " "for that dim (with search) first, then filter on the exact values it returns.") _WORDS_REASK_NOTE = ("\nCHECK: your draft does not state the finding. Write the answer in words " "with the exact names and numbers from the tool results (for a ranking, the " "top names and values). A chart or table alone is not an answer, and " "placeholders such as [Customer 1] are never allowed.") _DATE_WORDS = ("date", "month", "period", "year", "week", "day", "time") def _date_group(arguments, topic_dims): """True when a query groups by a date-like name that is not a dim of its topic.""" args = _parsed(arguments) group = args.get("group_by") or [] group = group if isinstance(group, list) else [group] dims = topic_dims.get(args.get("topic")) or set() return any(str(g) not in dims and any(w in str(g).lower() for w in _DATE_WORDS) for g in group) def _window_gap(arguments): """True when a gap report claims a missing date dimension or window.""" missing = str(_parsed(arguments).get("missing") or "").strip().lower() return missing == "date" or any(w in missing for w in ("date dim", "date dimension", "no date", "date field", "time dimension", "date filter", "filtered by 20", "filter by year", "by date")) def _typed_filter(arguments, field_values, entries): """The dim whose text filter value no earlier tool result in this ask returned, else None.""" args = _parsed(arguments) filters = args.get("filters") if not isinstance(filters, dict): return None seen = None for dim, value in filters.items(): values = value if isinstance(value, list) else [value] texts = [v for v in values if isinstance(v, str)] if not texts: continue known = field_values.get((args.get("topic"), dim)) or set() if all(v in known for v in texts): continue if seen is None: seen = " ".join(e["content"] for e in entries if e["response"].get("ok")) if not all(json.dumps(v, ensure_ascii=False) in seen for v in texts): return dim return None def _answer_lacks_finding(content): """True when a reply carries no figure, or fills a template with placeholders.""" import re text = content or "" return (not re.search(r"\d", text) or bool(re.search(r"\[(customer|amount|name|value|revenue)[^\]]*\]", text, re.I))) def _says_unavailable(content): """True when a reply says the data cannot be had: the case the gap protocol covers.""" import re return bool(re.search(_UNAVAILABLE, content or "", re.IGNORECASE)) def _has_op(arguments, op): """True when a transform_result chain contains `op`.""" ops = _parsed(arguments).get("transforms") or [] return any(isinstance(t, dict) and t.get("op") == op for t in ops) def _parsed(arguments): """Tool arguments as a dict, or {} when they do not parse.""" try: args = json.loads(arguments) if isinstance(arguments, str) else arguments except (TypeError, ValueError): return {} return args if isinstance(args, dict) else {} def _note_result(name, arguments, res, cut_results, queries): """Record which results are cut, and each query's window, for the guards in ask().""" data = res.get("data") if res.get("ok") else None rid = data.get("result_id") if isinstance(data, dict) else None if not rid: return args = _parsed(arguments) if name == "run_semantic_query": limit = args.get("limit") if isinstance(limit, int) and isinstance(data.get("row_count"), int) \ and data["row_count"] >= limit: cut_results.add(rid) rest = {k: v for k, v in args.items() if k not in ("date_from", "date_to", "limit", "sort")} queries.append((json.dumps(rest, sort_keys=True, default=str), args.get("date_from"), args.get("date_to"), rid)) elif name == "transform_result": ops = [str(t.get("op")) for t in args.get("transforms") or [] if isinstance(t, dict)] if args.get("result_id") in cut_results or any(op in _CUT_OPS for op in ops): cut_results.add(rid) def _whole_on_cut(arguments, cut_results): """True when a whole-result share op runs on a result that a limit or cut made partial.""" args = _parsed(arguments) ops = [str(t.get("op")) for t in args.get("transforms") or [] if isinstance(t, dict)] return args.get("result_id") in cut_results and bool(ops) and ops[0] in _WHOLE_OPS def _prior_year_of(arguments, queries, years=-1): """The result_id of an earlier query that this one repeats exactly `years` years off, else None. years=-1 finds an earlier query one year later than this one; years=1 one year earlier. """ args = _parsed(arguments) lo, hi = args.get("date_from"), args.get("date_to") if not (isinstance(lo, str) and isinstance(hi, str) and args.get("group_by")): return None rest = json.dumps({k: v for k, v in args.items() if k not in ("date_from", "date_to", "limit", "sort")}, sort_keys=True, default=str) def shift(day, years): return f"{int(day[:4]) + years:04d}{day[4:]}" if len(day) >= 4 and day[:4].isdigit() else None for sig, qlo, qhi, rid in queries: if sig == rest and isinstance(qlo, str) and isinstance(qhi, str) and \ ((shift(qlo, years), shift(qhi, years)) == (lo, hi)): return rid return None def _cut_before_whole(arguments): """True when a transform chain cuts rows away before a whole-result share op.""" try: args = json.loads(arguments) if isinstance(arguments, str) else arguments ops = [str(t.get("op")) for t in args.get("transforms") or [] if isinstance(t, dict)] except (TypeError, ValueError, AttributeError): return False cut = next((i for i, op in enumerate(ops) if op in _CUT_OPS), None) return cut is not None and any(op in _WHOLE_OPS for op in ops[cut + 1:]) def _coerce_args(name, arguments, tools): """Repair array parameters the model sent as a string, such as "[opex]" or "opex".""" try: args = json.loads(arguments) if isinstance(arguments, str) else arguments except (TypeError, ValueError): return arguments schema = next((t["function"].get("parameters") or {} for t in tools if t.get("function", {}).get("name") == name), {}) props = schema.get("properties") or {} if not isinstance(args, dict): return arguments fixed = dict(args) for key, value in args.items(): if (props.get(key) or {}).get("type") != "array" or not isinstance(value, str): continue text = value.strip() try: parsed = json.loads(text) except ValueError: parsed = None if isinstance(parsed, list): fixed[key] = parsed else: items = [x.strip().strip("'\"") for x in text.strip("[]").split(",")] fixed[key] = [x for x in items if x] if fixed == args: return arguments return json.dumps(fixed, ensure_ascii=False) _EMPTY_REPLY_NOTE = ("\nYour last reply was empty or wrote a tool call as text. Write the final " "answer in plain words from the tool results above, or make a real tool call.") _REPEAT_ERROR = ("repeat: this exact call already ran in this ask and its result is above. " "Change the arguments, or write the final answer from the results you have.") def _budget_note(left): """The working-context note that tells the model how many turns it has left.""" if left <= 1: return ("\nBUDGET: this is your LAST turn and tools are off. Write the final answer now " "from the tool results above, with the exact numbers and names they returned.") return (f"\nBUDGET: {left} turns left including the final answer. Call only a tool the ask " "still needs (its query, transform or requested chart), then answer. Skip re-reads, " "schema lookups and tables nobody asked for.") def _live_chat(messages, tools, model, tool_choice="auto"): """POST down the provider ladder with LATENCY-FIRST failover (a chat user is waiting): a rate-limited provider is SKIPPED IMMEDIATELY in favor of the next one β€” we only sleep when the ENTIRE ladder is limited (then short ladder-level backoffs). A LONG Retry-After (daily quota, e.g. Groq TPD) puts the provider in an in-process COOLDOWN; 401/402/403 skip the provider outright.""" import time as _time import requests provs = available_providers() if not provs: raise RuntimeError("no LLM provider configured; set GROQ_API_KEY / CEREBRAS_API_KEY / " "OPENROUTER_API_KEY in platform/.env") last_err = None # Two patience profiles: CHAT (default) fails over fast β€” a user is waiting; BATCH # (ANALYST_PATIENT=1, set by the eval gate / insight runner) waits out per-minute rate # windows across more ladder rounds instead of exhausting the ladder in seconds. patient = os.environ.get("ANALYST_PATIENT") == "1" round_sleeps = (5, 10, 20, 30, 45, 60) if patient else (2, 6, 12) for rnd in range(len(round_sleeps)): # ladder rounds saw_retryable = False for p in provs: if _COOLDOWN.get(p["name"], 0) > _time.time(): last_err = RuntimeError(f"{p['name']}: cooling down (quota)") continue sent_model = model or os.environ.get("ANALYST_MODEL") or p["model"] body = {"model": sent_model, "messages": messages, "tools": tools, "tool_choice": tool_choice, "temperature": 0.1, "max_tokens": _MAX_TOKENS_BY_PROVIDER.get(p["name"], _MAX_ANSWER_TOKENS)} if tools is None: # an answer-only turn sends no tool list body.pop("tools") body.pop("tool_choice") if p["name"] == "openrouter": body["usage"] = {"include": True} # OpenRouter returns usage.cost in USD try: r = requests.post(p["url"], timeout=120, headers={"Authorization": f"Bearer {os.environ[p['env']]}", "HTTP-Referer": "https://aios.local", "X-Title": "AIOS Analyst"}, json=body) if r.status_code in _DEAD_STATUS: last_err = RuntimeError(f"{p['name']}: HTTP {r.status_code}") _COOLDOWN[p["name"]] = _time.time() + 300 # don't re-poke dead auth/payment continue if r.status_code == 400: # β›” WAVE 32 (W32-T50/T54) β€” A 400 FROM THIS LADDER IS NOT A TRANSPORT FACT, # AND THIS FUNCTION USED TO THROW ITS EXPLANATION AWAY. Groq validates the # model's tool ARGUMENTS against the tool schema server-side and answers # `400 tool_use_failed` with the generation it rejected in `failed_generation`. # Measured 2026-08-13: a correct refusal that omitted a `required` property # arrived here, `raise_for_status()` turned it into "400 Client Error", and the # caller was told `all LLM providers failed` β€” for a request the model had # answered. The control flow is unchanged (still fail over); what changes is # that the message NAMES the cause, because a schema this ladder sent is the # one thing the operator can actually fix. detail = "" try: e400 = (r.json() or {}).get("error") or {} detail = str(e400.get("message") or "")[:300] if e400.get("failed_generation"): detail += f" | rejected generation: " \ f"{str(e400['failed_generation'])[:300]}" except Exception: detail = (r.text or "")[:300] last_err = RuntimeError(f"{p['name']}: HTTP 400 {detail}") continue if r.status_code in _RETRY_STATUS: last_err = RuntimeError(f"{p['name']}: HTTP {r.status_code}") try: wait = float(r.headers.get("retry-after") or 0) except ValueError: wait = 0 if wait > 60: # daily/long quota β€” cool down _COOLDOWN[p["name"]] = _time.time() + wait else: saw_retryable = True # NO sleep β€” next provider NOW continue r.raise_for_status() payload = r.json() usage = dict(payload.get("usage") or {}) usage["provider"] = p["name"] usage["model"] = payload.get("model") or sent_model return payload["choices"][0]["message"], usage except requests.RequestException as e: last_err = e continue # network blip β€” next provider if not saw_retryable: break # nothing retryable left anywhere _time.sleep(round_sleeps[rnd]) # whole ladder limited β€” pause a round raise RuntimeError(f"all LLM providers failed (last: {last_err})") #: Wave 15 R2 β€” what a restricted account is told instead of an answer. ANALYST_DENIED = ("The Analyst is not available on your account. It answers from the full " "governed model, which is wider than what your account is permitted to see.") def may_ask(user): """β›” R2 β€” THE AI WALL, AND IT IS DELIBERATELY BLUNT. Owner, verbatim: *"If the AI is triggered by a user, the AI ITSELF should not be able to see the data."* Today the Analyst's tools run UNSCOPED β€” `harness/tools.run_semantic_query` takes no subject and no filter, so every question it answers is answered over the whole tenant. A restricted account asking "what were our top customers?" would therefore receive, in prose, precisely the rows the grid's permanent filter exists to withhold. Prose is not a lesser leak than a table; it is a harder one to notice. So the wall is REFUSAL, not narrowing: an account with ANY declared restriction β€” or any non-admin on a migrated record β€” does not reach the model at all. An AI that never runs cannot leak what it never read. Fail-closed on an unreadable record for the same reason. β­’ NAMED DEBT, owner-mandated (R2, doc Β§Debt, log 2026-08-02). The RIGHT answer is to thread the subject's permanent filter into every semantic query the Analyst runs β€” `filter_sql` already compiles that tree to SQL, so the shape exists; what does not exist is a per-subject query path through `harness/semantic.py`, which takes `team_id` and nothing else. Until that lands this refusal is the whole enforcement, and it must NOT be relaxed into "narrow the prompt" or "tell the model what it may not say" β€” a prompt is not a permission boundary. """ import core.perms as perms import core.perm_scope as perm_scope if not isinstance(user, dict): return False # no identity resolved -> no Analyst if perms.is_admin(user): return True if perm_scope.is_migrated(user): # Migrated and not an admin: the account is governed by a wall the Analyst cannot honour. return False # Legacy record: restricted the old way is still restricted. return (perms.allowed_modules(user) is None and perms.scope_team_id(user) is None and not perms.scope_agent(user)) _PREVIEW_NOTE = ("rows holds only the first {shown} of {total} rows. Never count, rank or total " "from it. Count rows with transform_result filter_rows and read its row_count; " "find the largest or smallest with sort{{by,direction}} then head{{n}}; get " "shares and totals from a transform or a query with no group_by.") def _tool_content(name, response, tool_call_id, metadata, artifact_index=None): """Lossless JSON and receipts for retained renders. Working-context retrieval is owned by _AskContext, never an earlier message pointer. Rendering keeps the full artifact in ask's return; replaying its rows into the model would duplicate query data (or expand a 100-row query preview into thousands of rows). Errors, warnings, exact numbers and query scope stay intact. """ def serialize(value): return json.dumps(value, default=str, ensure_ascii=False, separators=(",", ":")) data = response.get("data") if isinstance(response, dict) else None if (name in ("run_semantic_query", "transform_result") and isinstance(data, dict) and isinstance(data.get("rows"), list) and isinstance(data.get("row_count"), int) and data["row_count"] > len(data["rows"])): # A preview read as the whole result gave wrong counts and totals: label it in place. response = {**response, "data": {**data, "preview": _PREVIEW_NOTE.format( shown=len(data["rows"]), total=data["row_count"])}} content = serialize(response) if artifact_index is not None and name in ("make_chart", "make_table"): data = response["data"] kind = "chart" if name == "make_chart" else "table" spec = data.get(kind) if isinstance(spec, dict) and isinstance(spec.get("rows"), list): # Keep every other spec/envelope field, including notes and warnings. The exact # original data object has already been retained in artifacts before this call. receipt = {"artifact_index": artifact_index, "result_id": spec.get("result_id"), "row_count": len(spec["rows"]), "note": "Full exact rows retained in the returned rendered artifact; " "this receipt is not a numeric total."} content = serialize({**response, "data": { **data, kind: {k: v for k, v in spec.items() if k != "rows"}, "artifact_receipt": receipt}}) return content class _AskContext: """Per-ask exact read cache; only whole bodies leave the working conversation. Discovery is not an append-only prompt. Keep the current topic schema, newly requested metadata body and newest data response, plus every loaded rule. Older bodies remain retrievable by their original tool/arguments. No process-global or cross-user cache. The byte ceiling bounds requests even without a tokenizer; it is set per provider by request_budget(), below each provider's measured request limit. Oversize protected material stops explicitly rather than being truncated. Artifacts/results are never row-capped. """ _READS = frozenset(("list_topics", "describe_topic", "get_recipe", "get_field_values", "run_semantic_query", "transform_result")) _METADATA = frozenset(("list_topics", "describe_topic", "get_recipe")) # Corpus: initial prompt/tools ~3.2k tokens; schema ~1.8k. Groq's observed 8k # request limit counts input, not max_tokens. Leave room for exact ops/schema together. # The default only. ask() passes the provider's ceiling from request_budget(). MAX_REQUEST_BYTES = DEFAULT_REQUEST_BYTES def __init__(self, messages, tools, max_request_bytes=None, data_source=None): self.messages, self.tools = messages, tools self.data_source = data_source if data_source is not None else T if max_request_bytes: self.MAX_REQUEST_BYTES = max_request_bytes self.cache, self.entries, self.rules, self.results = {}, [], {}, {} self.current_topic = None @staticmethod def _json(value): return json.dumps(value, default=str, ensure_ascii=False, separators=(",", ":")) @classmethod def _key(cls, name, args): try: parsed = json.loads(args) if isinstance(args, str) else args return name, json.dumps(parsed, sort_keys=True, ensure_ascii=False, separators=(",", ":")) except (TypeError, ValueError): return name, cls._json(args) @staticmethod def _rules(value): """Exact rule fields at every depth, including metric/field and recipe cautions.""" if isinstance(value, dict): out = {} for key, item in value.items(): if key in ("context", "ai_context", "scope", "guard", "warning", "note"): out[key] = item else: nested = _AskContext._rules(item) if nested: out[key] = nested return out if isinstance(value, list): return {str(i): rules for i, item in enumerate(value) if (rules := _AskContext._rules(item))} return {} def dispatch(self, name, args): from copy import deepcopy key = self._key(name, args) if name in self._READS and key in self.cache: return deepcopy(self.cache[key]) res = (_get_recipe(args) if name == "get_recipe" else self.data_source.dispatch(name, args)) if name in self._READS and res.get("ok"): self.cache[key] = deepcopy(res) return res def append(self, name, args, response, tool_call_id, artifact_index): from copy import deepcopy key = self._key(name, args) content = _tool_content(name, response, tool_call_id, {}, artifact_index) message = {"role": "tool", "tool_call_id": tool_call_id, "content": content} self.messages.append(message) entry = {"tool": name, "args": args, "response": deepcopy(response), "message": message, "content": content, "key": key, "shown": False} self.entries.append(entry) if response.get("ok"): if name in self._METADATA: rules = self._rules(response) if rules: self.rules[key] = rules data = response.get("data") if isinstance(data, dict) and data.get("result_id") in self.data_source._RESULTS: rid = data["result_id"] self.results[rid] = deepcopy(self.data_source._RESULTS[rid]) if name in ("describe_topic", "run_semantic_query"): parsed = json.loads(args) if isinstance(args, str) else args self.current_topic = parsed.get("topic") def _receipt(self, entry): receipt = {"ok": True, "exact_cache": { "tool": entry["tool"], "arguments": json.loads(entry["key"][1])}} data = entry["response"].get("data") if isinstance(data, dict): # Scope/warnings and result identity survive; receipts never assert totals. receipt.update({k: v for k, v in data.items() if k in ("result_id", "row_count", "window", "warning", "note")}) rows = data.get("rows") # A small result keeps its exact rows, so a later turn never loses an earlier figure. if (isinstance(rows, list) and len(rows) <= _RECEIPT_ROWS and data.get("row_count", len(rows)) == len(rows) and len(self._json(rows)) <= _RECEIPT_ROW_BYTES): receipt["rows"] = rows return self._json(receipt) def working_messages(self): schema = next((e for e in reversed(self.entries) if e["tool"] == "describe_topic" and e["response"].get("ok") and e["response"]["data"].get("topic") == self.current_topic), None) metadata = next((e for e in reversed(self.entries) if e["tool"] in self._METADATA - {"describe_topic"} and e["response"].get("ok")), None) if metadata is not None and metadata["shown"]: metadata = None data = next((e for e in reversed(self.entries) if e["tool"] in self._READS - self._METADATA and e["response"].get("ok")), None) keep = {id(e) for e in (schema, metadata, data) if e is not None} for entry in self.entries: if entry["tool"] in self._READS and entry["response"].get("ok"): entry["message"]["content"] = (entry["content"] if id(entry) in keep else self._receipt(entry)) working = [dict(m) for m in self.messages] retained = [{"tool": k[0], "arguments": json.loads(k[1]), "rules": v} for k, v in self.rules.items() if not any(id(e) in keep and e["key"] == k for e in self.entries)] if retained: # Add to the existing system message: no tool-role adjacency is disturbed. working[0]["content"] += "\nRETAINED RULES (exact, still mandatory):\n" + self._json(retained) def fits(): return len(self._json({"messages": working, "tools": self.tools}).encode("utf-8")) \ <= self.MAX_REQUEST_BYTES if not fits(): raise ValueError("complete current schema, result, rules or history exceed the " "working-context budget; no content was truncated") for entry in self.entries: if any(m.get("tool_call_id") == entry["message"]["tool_call_id"] and m.get("content") == entry["content"] for m in working): entry["shown"] = True return working def retained(self): return {"tool_results": [{k: e[k] for k in ("tool", "args", "response")} for e in self.entries], "results": self.results} def ask(question, history=None, chat_fn=None, model=DEFAULT_MODEL, max_iters=None, telemetry_kind="analyst", user=None, data_source=None): """Run one Analyst turn. Returns {answer, artifacts, tool_trace, iterations, exhausted, usage}. chat_fn(messages, tools, model) -> (assistant_message_dict, usage_dict). `user` is the SUBJECT asking. Passing it engages the R2 wall (`may_ask`); omitting it keeps the unscoped behaviour that `harness/evals.py` and `harness/routines.py` rely on β€” those run as the SYSTEM, not on behalf of a person, and gating them on a user record they do not have would break the eval ladder to no security benefit. Every call that IS on behalf of somebody passes it, and `app.py`'s chat entry point does. """ if user is not None and not may_ask(user): return {"answer": ANALYST_DENIED, "artifacts": [], "tool_trace": [], "iterations": 0, "exhausted": False, "denied": True, "usage": {"prompt_tokens": 0, "completion_tokens": 0}} if max_iters is None: max_iters = iter_budget() chat = chat_fn or _live_chat backend = data_source if data_source is not None else T tools = _tool_schemas(data_source) messages = [{"role": "system", "content": system_prompt(data_source)}] messages += list(history or []) # The model has no clock: anchor relative windows ("last week", "this month") to today. # Lives in the USER turn so the system prompt stays byte-stable (prompt-cache friendly). import datetime as _dt today = backend.today if data_source is not None else _dt.date.today().isoformat() messages.append({"role": "user", "content": f"(today is {today})\n{question}"}) trace, artifacts, usage_total = [], [], {"prompt_tokens": 0, "completion_tokens": 0} # The live ladder's first provider sets the ceiling. An injected chat_fn has no provider. provider = (available_providers() or [{}])[0].get("name") if chat is _live_chat else None context = _AskContext(messages, tools, request_budget(provider), data_source) calls, failed_calls, retries = {}, set(), 0 cut_results, queries, field_values = set(), [], {} yoy_pending, yoy_done, reasked = set(), False, set() try: topic_dims = {t["topic"]: set(t.get("dims") or []) for t in backend.list_topics()} except Exception: # noqa: BLE001 - guard only topic_dims = {} for i in range(max_iters): try: working = context.working_messages() except ValueError: return {"answer": "(stopped: complete required context exceeds the working-context " "budget; narrow the question or history; no data was truncated)", "artifacts": artifacts, "tool_trace": trace, "iterations": i, "exhausted": True, "context_exhausted": True, "usage": usage_total, **context.retained()} left = max_iters - i final_turn = chat is _live_chat and left == 1 and bool(trace) if chat is _live_chat and left <= _BUDGET_WARN_TURNS and trace: # Working copy only: the stored system message stays byte-stable. working[0] = dict(working[0], content=working[0]["content"] + _budget_note(left)) if final_turn: # The last turn answers from the results already read, so a long ask ends in an answer. msg, usage = _live_chat(working, tools, model, tool_choice="none") else: msg, usage = chat(working, tools, model) _log_call({"provider": "chat_fn", "model": model, **(usage or {})}) for k in usage_total: usage_total[k] += usage.get(k, 0) or 0 tool_calls = msg.get("tool_calls") or [] # Free-tier models INTERMITTENTLY answer a data question without calling any tool # (~1 in 4 observed on gpt-oss-120b β€” the "no chart" defect). Deterministic fix: if # the VERY FIRST turn used no tools, redo it once with tool_choice='required'. Meta # questions just get a harmless schema lookup before answering. if i == 0 and not tool_calls and chat is _live_chat: msg, usage = _live_chat(working, tools, model, tool_choice="required") _log_call(usage) for k in usage_total: usage_total[k] += usage.get(k, 0) or 0 tool_calls = msg.get("tool_calls") or [] while (chat is _live_chat and retries < _MAX_REPLY_RETRIES and not tool_calls and _unusable_reply(msg.get("content"))): # An empty reply, or a tool call written as text, is a broken turn: ask again. retries += 1 nudge = dict(working[0], content=working[0]["content"] + _EMPTY_REPLY_NOTE) msg, usage = _live_chat([nudge] + working[1:], None if final_turn else tools, model) _log_call(usage) for k in usage_total: usage_total[k] += usage.get(k, 0) or 0 tool_calls = msg.get("tool_calls") or [] if chat is _live_chat and not tool_calls and not final_turn: # A final answer that skipped a required step gets one corrected re-ask. note = None if yoy_pending and not yoy_done and "yoy" not in reasked: reasked.add("yoy") note = _YOY_REASK_NOTE.format(result_id=sorted(yoy_pending)[-1]) elif ("gap" not in reasked and _says_unavailable(msg.get("content")) and not any(t["tool"] == "report_gap" for t in trace)): reasked.add("gap") note = _GAP_REASK_NOTE elif artifacts and "words" not in reasked and _answer_lacks_finding(msg.get("content")): reasked.add("words") note = _WORDS_REASK_NOTE if note: nudge = dict(working[0], content=working[0]["content"] + note) msg, usage = _live_chat([nudge] + working[1:], tools, model) _log_call(usage) for k in usage_total: usage_total[k] += usage.get(k, 0) or 0 tool_calls = msg.get("tool_calls") or [] messages.append({"role": "assistant", "content": msg.get("content") or "", **({"tool_calls": tool_calls} if tool_calls else {})}) if not tool_calls: answer = (msg.get("content") or "").strip() if not answer: # A model occasionally returns NO content and NO tool calls (observed live # 2026-07-16: the reply bubble was simply empty and the next turn hallucinated). # Never surface silence β€” report the failure honestly instead. answer = ("(the model returned an empty answer; nothing was computed; " "please rephrase or ask again)") import harness.telemetry as _tel # Wave 19 (R4): ATTRIBUTED. The token count was previously logged against nobody, so # "which customer is spending model budget, and who inside it?" had no answer even in # principle. Read defensively off `user` because this function is deliberately # callable WITHOUT one β€” `harness/evals.py` and `harness/routines.py` run as the # system, and telemetry omits both keys for them rather than inventing a subject. log = data_source.log if data_source is not None else _tel.log log(telemetry_kind, question=question[:200], user=(user or {}).get("username") if isinstance(user, dict) else None, tenant=(user or {}).get("tenant") if isinstance(user, dict) else None, model=model or (available_providers() or [{}])[0].get("name", "auto"), iterations=i + 1, tools=[t["tool"] for t in trace], tokens=usage_total["prompt_tokens"] + usage_total["completion_tokens"]) return {"answer": answer, "artifacts": artifacts, "tool_trace": trace, "iterations": i + 1, "exhausted": False, "usage": usage_total, **context.retained()} for tc in tool_calls: name = tc["function"]["name"] args = tc["function"].get("arguments") or "{}" args = _coerce_args(name, args, tools) if name == "run_semantic_query": args = _expand_variants(args, field_values) key = _AskContext._key(name, args) seen = calls.get(key, 0) if key in failed_calls or (seen and name not in _AskContext._READS) \ or seen >= _MAX_SAME_READS: # A repeat cannot change the outcome: say so instead of running it again. res = {"ok": False, "error": _REPEAT_ERROR} elif name == "run_semantic_query" and _bu_in_filters(args): res = {"ok": False, "error": _bu_filter_error()} elif name == "run_semantic_query" and _date_in_filters(args): res = {"ok": False, "error": _DATE_FILTER_ERROR} elif name == "run_semantic_query" and _date_group(args, topic_dims): res = {"ok": False, "error": _DATE_GROUP_ERROR} elif name == "report_gap" and _window_gap(args): res = {"ok": False, "error": _WINDOW_GAP_ERROR} elif name == "run_semantic_query" and (dim := _typed_filter(args, field_values, context.entries)): res = {"ok": False, "error": _TYPED_FILTER_ERROR.format(dim=dim)} elif name == "transform_result" and _cut_before_whole(args): res = {"ok": False, "error": _CUT_BEFORE_WHOLE_ERROR} elif name == "transform_result" and _whole_on_cut(args, cut_results): res = {"ok": False, "error": _WHOLE_ON_CUT_ERROR} elif name == "run_semantic_query" and (prior := _prior_year_of(args, queries)): res = {"ok": False, "error": _YOY_BY_HAND_ERROR.format(result_id=prior)} yoy_pending.add(prior) else: earlier = (_prior_year_of(args, queries, years=1) if name == "run_semantic_query" else None) res = context.dispatch(name, args) if earlier and res.get("ok") and isinstance(res.get("data"), dict) \ and res["data"].get("result_id"): # The prior year ran first: point the model at yoy on THIS result. rid = res["data"]["result_id"] res = {**res, "data": {**res["data"], "yoy_note": _YOY_BY_HAND_ERROR.format(result_id=rid)}} yoy_pending.add(rid) if name == "transform_result" and res.get("ok") and _has_op(args, "yoy"): yoy_pending.clear() yoy_done = True _note_result(name, args, res, cut_results, queries) _note_field_values(name, args, res, field_values) calls[key] = seen + 1 if not res.get("ok"): failed_calls.add(key) trace.append({"tool": name, "args": args if isinstance(args, str) else json.dumps(args), "ok": res["ok"], "error": res.get("error")}) artifact_index = None if res["ok"] and isinstance(res["data"], dict) and ( "chart" in res["data"] or "kpi" in res["data"] or "table" in res["data"] or "dashboard" in res["data"]): artifact_index = len(artifacts) artifacts.append(res["data"]) context.append(name, args, res, tc.get("id", name), artifact_index) # budget exhausted β€” report honestly (never as success) return {"answer": "(stopped: tool-call budget exhausted before a final answer; " "narrow the question or raise max_iters)", "artifacts": artifacts, "tool_trace": trace, "iterations": max_iters, "exhausted": True, "usage": usage_total, **context.retained()}
"""harness/analyst.py β€” the AIOS Analyst loop (OM-4, 2026-07-11).
A CHEAP small model (via OpenRouter, OpenAI-compatible) drives the compounding tool registry
(harness/tools.py) to answer questions, build charts, and spawn dashboards over the semantic
layer. The model NEVER writes SQL β€” it picks registry keys, guided by the in-product SKILL
recipe index and exact on-demand bodies (model/skills/*.skill.yml). Plan: omni-adoption addendum.
Architecture (mirrors the Omni coordinator pattern):
ask(question) -> bounded tool loop: chat -> tool_calls -> tools.dispatch() -> results -> chat …
until a plain answer or MAX_ITERS. chat_fn is INJECTABLE: live = OpenRouter POST; tests = a
scripted double β€” the loop mechanics are provable offline (no key, no tokens).
Correctness posture: bounded iterations (never endless); tool errors go BACK to the model as
readable envelopes (it self-corrects); the final answer carries the tool trace (auditability +
the drill guarantee); exhausted-budget is reported AS exhausted, never as success (loop rules).
Cost posture: DEFAULT_MODEL is a cheap tool-calling model; override per-tenant/env; prompt = the
skill index + topic summary (small, cacheable), NOT raw data.
"""
import json
import os
from pathlib import Path
import harness.tools as T
# Provider LADDER (owner 2026-07-12: "use Groq and Cerebras"): all OpenAI-compatible chat
# endpoints; tried in order, failing over on payment/rate/server errors β€” an exhausted or
# rate-limited provider degrades to the next instead of taking the Analyst down (the
# OpenRouter-402 incident). Pin one with ANALYST_PROVIDER=groq|cerebras|openrouter; override
# its model with ANALYST_MODEL. Every model here must re-pass the eval gate (Β§4b).
PROVIDERS = [
{"name": "groq", "env": "GROQ_API_KEY",
"url": "https://api.groq.com/openai/v1/chat/completions",
"model": "openai/gpt-oss-120b"},
{"name": "cerebras", "env": "CEREBRAS_API_KEY",
"url": "https://api.cerebras.ai/v1/chat/completions",
"model": "gpt-oss-120b"}, # this account's tool-capable model (from /v1/models)
{"name": "openrouter", "env": "OPENROUTER_API_KEY",
"url": "https://openrouter.ai/api/v1/chat/completions",
# The eval measures this default: analyst_golden.yml `measured` names it, open weights only (R11).
"model": "z-ai/glm-4.7-flash"},
]
DEFAULT_MODEL = None # None = the selected provider's default model
# A map, YoY or name-resolution ask needs about 8 tool turns with no retry (list, describe,
# recipe, field values, query, transform, chart, answer), so 8 exhausted 7 of 26 items (D-8280).
# 14 leaves room for 6 error-correction turns. ANALYST_MAX_ITERS overrides it per run.
MAX_ITERS = 14
def iter_budget():
"""The tool-loop budget: ANALYST_MAX_ITERS when set to a positive integer, else MAX_ITERS."""
try:
value = int(os.environ.get("ANALYST_MAX_ITERS") or MAX_ITERS)
except ValueError:
return MAX_ITERS
return value if value > 0 else MAX_ITERS
#: Working-context byte ceiling per provider. Groq and Cerebras fail near 8k request tokens.
#: OpenRouter tool models take 128k+ tokens. 120000 bytes (about 30k tokens) fits a full
#: discovery run and keeps each request cheap.
REQUEST_BYTES = {"groq": 32000, "cerebras": 32000, "openrouter": 120000}
DEFAULT_REQUEST_BYTES = 32000
def request_budget(provider=None):
"""The request byte ceiling: ANALYST_MAX_REQUEST_BYTES when a positive integer, else per provider."""
try:
value = int(os.environ.get("ANALYST_MAX_REQUEST_BYTES") or 0)
except ValueError:
value = 0
if value > 0:
return value
return REQUEST_BYTES.get(provider, DEFAULT_REQUEST_BYTES)
#: Set by harness/evals.run to the golden item id, so each ledger line names its item.
CALL_ITEM = None
#: Running totals of ledger lines written by this process (evals.run reads the delta).
CALL_TOTALS = {"calls": 0, "cost": 0.0, "cost_missing": 0}
def _log_call(usage):
"""Append one cost line to ANALYST_CALL_LOG. Never writes message or answer text."""
path = os.environ.get("ANALYST_CALL_LOG")
if not path:
return
import time as _time
usage = usage or {}
cost = usage.get("cost")
cost = float(cost) if isinstance(cost, (int, float)) and not isinstance(cost, bool) else None
line = {"ts": _time.strftime("%Y-%m-%dT%H:%M:%SZ", _time.gmtime()),
"provider": usage.get("provider"), "model": usage.get("model"),
"item": CALL_ITEM,
"prompt_tokens": usage.get("prompt_tokens"),
"completion_tokens": usage.get("completion_tokens"),
"cost": cost}
with open(path, "a", encoding="utf-8", newline="") as fh:
fh.write(json.dumps(line, separators=(",", ":")) + "\n")
CALL_TOTALS["calls"] += 1