"""AI REVIEW — let a model decide which stage a card moves to (wave 23, owner ruling R4/R14). ⭐ WHAT THIS IS. A review stage holds a record until somebody decides where it goes next. R4 made that somebody optionally a MODEL: the engine hands over the record's own values, the review's prompt, and the list of stages the review is allowed to send a card to, and gets back ONE of those stage labels plus a one-line reason. Every decision is written to the same `reviews` audit log a human click writes to, tagged `by: "ai"` with the provider and model that made it. ⛔ FAIL-CLOSED IN EVERY DIRECTION, and this is the whole safety story. No key configured, a network failure, a slow answer, a malformed answer, or an answer naming a stage the review does not offer — all return `("", {...})`, and the caller leaves the card exactly where a human would have found it. The feature can be absent, broken or wrong and the worst outcome is a person doing the work. Nothing here can move a card somewhere the review does not already permit. ⭐ CHEAP FIRST (owner R14, verbatim: *"Claude is a bit too expensive"*). The ladder is `groq → cerebras → openrouter → anthropic`; the first CONFIGURED provider wins, and Anthropic is last rather than absent — it is the quality backstop, not the default. `AIOS_AI_REVIEW_PROVIDER` pins one; `AIOS_AI_MODEL` overrides the model. ⭐⭐ THE ANTHROPIC LEG IS ON THE OFFICIAL SDK NOW (D-346, W37-T39, 2026-08-19). ⛔ THIS PARAGRAPH USED TO ARGUE THE OPPOSITE and is rewritten rather than deleted, because a comment left asserting the reverse of its own code is [[two-gates-can-assert-opposite-things]] with no gate to catch it. The retired text said raw HTTP was *"a deliberate deviation … booked as a DEBT line so the integrator can overturn it deliberately rather than by drift"*. This is that overturn. Its two reasons were real and are both answered rather than waved away: 1. *"one SDK leg beside three hand-rolled ones is two implementations of the same call."* It is not, because the leg does not live here: `providers.anthropic_send` owns the wire, prefers the SDK, and normalises BOTH transports to one `(status, body)` pair. This file gained no Anthropic knowledge — it lost some. 2. *"adding a pinned dependency rebuilds the container, a real deploy risk."* Still true, so the import is LAZY and its absence is a FALLBACK, not a crash: with no `anthropic` package the same raw POST runs and `LAST_ANTHROPIC_TRANSPORT` reports `'http'`. The pin is owed in BOTH manifests ([[pin-deps-space-rebuilds]]) and is outside every worker fence this wave. ⚠ The other three legs stay OpenAI-chat-shaped over `requests` — they have no shared SDK, and that half of the original reasoning never expired. The Messages shape below is the current one: `x-api-key` + `anthropic-version: 2023-06-01`, and `stop_reason: "refusal"` is checked BEFORE reading `content` — a refusal answers HTTP 200 with an empty content list, so code that indexes `content[0]` unconditionally breaks on it. """ from __future__ import annotations import json import os import re import requests #: The ladder. Order IS the policy (R14) — cheapest capable first, Anthropic last. PROVIDERS = [ # ⛔⛔ D-345, FIXED W37-T39 (2026-08-19). This read `llama-3.3-70b-versatile` and Groq RETIRED # it: a real POST with the live key answers HTTP 404 `model_not_found`, and the id is absent # from `GET /openai/v1/models`. ⚠ THE `openai/` PREFIX IS PART OF GROQ'S ID and is NOT a typo # for the cerebras rung below, which serves the same family bare as `gpt-oss-120b`. {"name": "groq", "env": "GROQ_API_KEY", "shape": "openai", "url": "https://api.groq.com/openai/v1/chat/completions", "model": "openai/gpt-oss-120b"}, {"name": "cerebras", "env": "CEREBRAS_API_KEY", "shape": "openai", "url": "https://api.cerebras.ai/v1/chat/completions", "model": "gpt-oss-120b"}, {"name": "openrouter", "env": "OPENROUTER_API_KEY", "shape": "openai", "url": "https://openrouter.ai/api/v1/chat/completions", "model": "openai/gpt-4o-mini"}, # ⚠ haiku-class DELIBERATELY, not the default Opus tier: R14 put Anthropic on this ladder as # the backstop for a one-line classification, and this is the cheapest current Claude that # does it well. A bigger model here would be spending the owner's money to pick between two # labels it already has in front of it. {"name": "anthropic", "env": "ANTHROPIC_API_KEY", "shape": "anthropic", "url": "https://api.anthropic.com/v1/messages", # ⚠ AN ENV LEVER, NOT A NEW DEFAULT (W36-T35): haiku stays the choice for the reasons above, # and a deployment whose side rungs are out of credit can raise the tier without a release. "model": os.environ.get("AIOS_AI_REVIEW_ANTHROPIC_MODEL") or "claude-haiku-4-5"}, ] ANTHROPIC_VERSION = "2023-06-01" TIMEOUT_SECONDS = float(os.environ.get("AIOS_AI_REVIEW_TIMEOUT") or 20) MAX_FIELD_CHARS = 200 # per value handed to the model MAX_FIELDS = 30 # columns handed to the model MAX_REASON = 200 MAX_HARNESS_PROMPT_CHARS = 4000 MAX_HARNESS_SUMMARY_CHARS = 240 def ladder(): """The providers that are actually usable here, in order. Empty = the feature is off.""" pin = (os.environ.get("AIOS_AI_REVIEW_PROVIDER") or "").strip().lower() live = [p for p in PROVIDERS if (os.environ.get(p["env"]) or "").strip()] if pin: live = [p for p in live if p["name"] == pin] return live def configured(): return bool(ladder()) def _record_text(row, fields): """The record, as the model sees it. Values are truncated and the column set is bounded — an automation table can carry a 32 KB JSON blob per row (C7) and a review decision does not need it. Machine bookkeeping columns are dropped: a stage cell naming the stage the card is sitting at would be the model reading its own question back.""" keys = [k for k in (fields or list((row or {}).keys())) if not str(k).startswith("stage_")][:MAX_FIELDS] lines = [] for k in keys: v = (row or {}).get(k) if v is None or str(v).strip() == "": continue lines.append(f"{k}: {str(v)[:MAX_FIELD_CHARS]}") return "\n".join(lines) or "(this record has no filled-in values)" def _instruction(prompt, options, label): return ( f"You are deciding what happens to one record waiting at a review step called " f"{label!r} in a workflow.\n\n" f"The person who built this workflow told you: {prompt}\n\n" f"Choose EXACTLY ONE of these next steps, by its exact name:\n" + "\n".join(f"- {o}" for o in options) + "\n\nAnswer with one line of JSON and nothing else:\n" '{"choice": "", "reason": ""}\n' "If the record does not give you enough to decide, answer " '{"choice": "", "reason": "why not"} and a person will decide instead.' ) def _parse(text, options): """The model's line → `(choice, reason)`. A choice that is not one of the offered stages is DISCARDED, not fuzzy-matched: the offered list is a permission boundary, and a near-miss resolved by string distance is how a card ends up somewhere nobody authorised.""" raw = str(text or "").strip() obj = None m = re.search(r"\{.*\}", raw, re.S) if m: try: obj = json.loads(m.group(0)) except (ValueError, TypeError): obj = None if not isinstance(obj, dict): return "", "" choice = str(obj.get("choice") or "").strip() reason = str(obj.get("reason") or "").strip()[:MAX_REASON] for opt in options: if choice.lower() == str(opt).lower(): return str(opt), reason # the OFFERED spelling wins, never the model's return "", reason # ⭐⭐ WAVE 35 · W35-T41 / C7 — THE TWO TRANSPORTS RETURN THE RESPONSE BODY'S `usage`. # # ⛔ THIS FILE'S OWN HEADER USED TO BE THE PRODUCT'S CONFESSION THAT NOTHING COUNTED: `ai_enrich`'s # docstring says *"`ai_review.decide` -- the product's only other LLM entry -- has no token # accounting of ANY kind"*. R9 asks for ONE meter for every AI surface, so the counting has to reach # this transport rather than be bolted onto its caller — the caller never sees the body. # # ⚠ A THIRD RETURN VALUE, NOT A MUTATED ARGUMENT, and both call sites are in `decide` below. The # tuple grew from `(text, err)` to `(text, err, usage)`; `usage` is the raw `usage` OBJECT (or None), # because reading it is `usage_ledger`'s job and a second reader here would be the two-normalizers # shape this repo keeps paying for. def _call_openai(p, model, system, user, timeout): r = requests.post(p["url"], timeout=timeout, headers={"Authorization": f"Bearer {os.environ[p['env']].strip()}", "Content-Type": "application/json"}, json={"model": model, "max_tokens": 300, "temperature": 0, "messages": [{"role": "system", "content": system}, {"role": "user", "content": user}]}) if r.status_code >= 400: # ⚠ NO BODY ON A 4xx/5xx: an error envelope carries no usage, and a provider that refused # before running the model has nothing to bill. The call is still COUNTED by `decide`. return "", f"{p['name']} answered {r.status_code}", None body = r.json() choices = body.get("choices") or [] if not choices: return "", f"{p['name']} returned no choices", body return str(((choices[0] or {}).get("message") or {}).get("content") or ""), "", body #: What transport the last Anthropic call actually used: `'sdk'`, `'http'`, or `''` before any. #: ⚠ PROCESS-LOCAL AND FOR REPORTING ONLY. It exists so an administrator can ask WHICH path ran #: rather than infer it from a requirements file: `GET /api/v1/automations/ai/transport` #: (`routes_automation.ai_transport`, D-384) serves it beside the installed SDK version. LAST_ANTHROPIC_TRANSPORT = "" def _call_anthropic(p, model, system, user, timeout): """One classification turn on the Messages API, through `providers.anthropic_send`. ⭐ D-346 (W37-T39): this used to hand-roll the POST. It now goes through the shared wire, which prefers the OFFICIAL `anthropic` SDK and falls back to the same raw POST when the package is absent. The shape of what comes back is unchanged, deliberately: `anthropic_send` normalises both transports to `(status, body)`, so every line below this call is untouched. """ global LAST_ANTHROPIC_TRANSPORT import providers as _prov # noqa: PLC0415 # ⚠ NO TOOLS ON THIS PATH. The builder omits `tool_choice` for exactly that case, so there is # nothing to strip here — the rule lives in `anthropic_request`, beside the thing it constrains. req = _prov.anthropic_request( model=model, key=os.environ[p["env"]].strip(), system=system, messages=[{"role": "user", "content": user}], tools=[], max_tokens=300) status, body, transport = _prov.anthropic_send(req, timeout=timeout) LAST_ANTHROPIC_TRANSPORT = transport if status >= 400 or status == 0: # ⭐ THE SENTENCE COMES FROM THE LADDER, NOT FROM HERE. `refusal_sentence` already names the # vendor and the action for every status R4 enumerated, so a bare "anthropic answered 402" # (which is what this line used to say, and what the owner quoted back at us) cannot recur. return "", _prov.refusal_sentence("anthropic", status, json.dumps(body)[:400]), None body = body or {} # ⛔ stop_reason FIRST. A safety refusal is a successful 200 with an EMPTY content list, so # reading content[0] before this check turns a refusal into an IndexError inside a run. # ⚠ A refusal IS billed and its body carries `usage`, so the body rides back on this branch too. if body.get("stop_reason") == "refusal": return "", "anthropic declined to answer this record", body parts = [b.get("text") or "" for b in (body.get("content") or []) if isinstance(b, dict) and b.get("type") == "text"] if not parts: return "", "anthropic returned no text", body return "".join(parts), "", body def decide(*, prompt, options, row, fields=(), label="Review", timeout=None, st=None, user=""): """Pick this record's next stage. Returns `(choice, meta)`. `choice` is "" whenever a person should decide — which is every failure mode there is. `meta` carries `provider`, `model`, `reason` on success, and `problem` on refusal to answer. ⭐ W35-T41 / C7 — `st` and `user` are the USAGE LEDGER's target. Both default to absent because this function's caller is `automation_engine.ai_decide(rt, ...)`, in another lane's fence: it HAS the runtime and does not pass it yet, so until it does, a review's tokens are counted as unattributed and REPORTED by `GET /usage` rather than dropped. See `usage_ledger`'s header for why they cannot be resolved implicitly (measured: a contextvar does not survive a FastAPI dependency). """ import usage_ledger # noqa: PLC0415 opts = [str(o) for o in (options or []) if str(o).strip()] if not opts: return "", {"problem": "the review offers no next stages"} if not str(prompt or "").strip(): return "", {"problem": "the review has no prompt for the model to follow"} live = ladder() if not live: return "", {"problem": "no AI provider is configured on this deployment"} system = _instruction(prompt, opts, label) user = "Here is the record:\n\n" + _record_text(row, fields) tmo = float(timeout or TIMEOUT_SECONDS) override = (os.environ.get("AIOS_AI_MODEL") or "").strip() problems = [] for p in live: model = override or p["model"] body = None try: text, err, body = (_call_anthropic if p["shape"] == "anthropic" else _call_openai)( p, model, system, user, tmo) except Exception as e: # noqa: BLE001 text, err = "", f"{p['name']} failed: {type(e).__name__}" # ⭐⭐ C7 — THE LEDGER LINE, BEFORE ANY BRANCH BELOW READS THE ANSWER. # ⛔ IT IS WRITTEN ON EVERY OUTCOME THAT REACHED A PROVIDER, including a refusal and an # unusable answer. A meter that books only successes reports a cheap week for a run that # spent its budget being declined — and a declined call is billed. The only path that does # NOT record is a transport that never reached the vendor (`body is None`), which spent # nothing. if body is not None: ins, outs = usage_ledger.tokens_from(body) usage_ledger.record("ai_review", p["name"], model, ins, outs, total=usage_ledger.total_from(body), st=st, user=user) if err: problems.append(err) continue # ladder: a dead provider degrades to the next one choice, reason = _parse(text, opts) if not choice: # The provider ANSWERED and declined (or answered unusably). That is a decision about # this record, not a fault in the provider, so it does NOT fall through to a more # expensive one — the card goes to a human, which is what the model just asked for. return "", {"provider": p["name"], "model": model, "problem": reason or "the model did not choose one of the stages"} return choice, {"provider": p["name"], "model": model, "reason": reason} return "", {"problem": "; ".join(problems)[:300] or "no provider answered"} # ══════════════════════════════════════════════════════════════════ the FLOW WRITER # ⭐⭐ WAVE 33 · W33-T54/T55 (owner item 7, ruling R3) — A PROMPT BECOMES A DRAFT AUTOMATION. # # Owner, verbatim: *"lay out the foundation with custom tools we provide"* — and the tools are the # action catalog, which has been machine-readable since wave 23. So this writes no new vocabulary: # it hands the model the SAME `ACTION_CATALOG` the menu paints, the SAME `ACTION_REQUIRED` the # runner blocks on, and the SAME trigger keys `clean_trigger` accepts, and asks for one JSON object # in that vocabulary. # # ⛔ WHY IT LIVES IN `ai_review.py` RATHER THAN A NEW FILE. Two reasons, both about ownership. This # is the one module in the API package that already holds an LLM ladder and its credentials, so a # second one would be a second place a rotated key has to be noticed; and a new `routes_*.py` needs # a `main.py` line from another lane (contract C2) to be reachable at all, which is how wave 23 # shipped three finished routers 404-dead behind green gates. The DOOR is a route on # `routes_automation.py`, which is already mounted. # # ⛔⛔ CONTRACT C7 IS THE DESIGN, NOT A CHECK AT THE END: *"the flow writer may emit only kinds # present in `ACTION_CATALOG`, and its output must pass `clean_actions` unchanged."* Both halves are # enforced mechanically rather than asked for politely — the kind list is an `enum` in the schema # the model fills, and the caller runs `clean_actions` and DIFFS the result. That diff matters # because `clean_actions` has no disclosure channel (D-75): it drops a key it does not recognise # and answers 200, so a draft accepted without the diff would show a person a flow that is not the # flow the model described, with nothing anywhere saying so. # # ⚠ `chat` IS INJECTABLE, exactly as `routes_query._call_model`'s is, and it is the reason this door # can be proven end to end with no API key and no spend. A gate that can only run where a # credential exists is a gate that never runs. #: Cerebras first, NOT the cheap-first order `ladder()` uses. Same reason `routes_query` inverts it: #: this contract includes a REFUSAL, and refusing honestly ("I cannot build that from the steps you #: have") is a model property. A cheaper rung answers an impossible request with a plausible flow, #: which is worse than no flow. Override with `AIOS_FLOW_PROVIDER`. FLOW_PROVIDER_ORDER = ("cerebras", "groq", "openrouter", "anthropic") MAX_DRAFT_ACTIONS = 12 # a draft a person reads in one screen; the engine's own cap is 25 MAX_PROMPT_CHARS = 2000 def flow_providers(pin=None): """The rungs usable here, in THIS module's refusal-first order. Empty = the feature is off. ⭐ `pin` IS ASK D-18 (2026-08-18): the Agent chat's model toggle must configure something, and the draft door used to read `prompt` off the body and nothing else — so the key the client sent was accepted and dropped, and the picker was a control over nothing. ⚠ AN UNKNOWN OR UNCONFIGURED PIN FALLS BACK TO THE LADDER rather than refusing. A model the ladder stopped offering must not turn every later draft into an error; the caller is told which rung actually answered, which is the honest half. """ by_name = {p["name"]: p for p in PROVIDERS} live = [n for n in FLOW_PROVIDER_ORDER if n in by_name and (os.environ.get(by_name[n]["env"]) or "").strip()] wanted = str(pin or os.environ.get("AIOS_FLOW_PROVIDER") or "").strip().lower() if wanted and wanted in live: return [by_name[wanted]] return [by_name[n] for n in live] def flow_schema(kinds, trigger_keys, table_keys): """The tool schema the model fills — the catalog's OWN key lists as enums. ⛔ `kind` AND `trigger` ARE ENUMS, not free strings, and that is C7's first half enforced by the transport rather than by a check afterwards. A model asked for "any action name" invents `send_slack_message` on a deployment that has no such kind, and the failure then surfaces as a 400 from `clean_actions` carrying a sentence about a word the person never typed. ⚠ `required` IS `["kind"]` ALONE, deliberately: a REFUSAL carries no name and no actions, and a schema demanding them turns an honest refusal into a provider-side 400 that reads exactly like a transport failure (`routes_query._spec_schema` carries the same note, for the same measured reason). """ return { "type": "object", "properties": { "kind": {"type": "string", "enum": ["flow", "refused"], "description": "refused = this cannot be built from the steps available"}, "refusal": {"type": "string", "description": "when kind=refused: ONE plain sentence naming what is " "missing, in the words a non-technical person would use"}, "name": {"type": "string", "description": "a short title for the automation"}, "trigger": {"type": "string", "enum": sorted(trigger_keys), "description": "what starts this automation"}, "table": {"type": "string", "enum": sorted(table_keys), "description": "the database whose records this flow walks, if any"}, "actions": { "type": "array", "maxItems": MAX_DRAFT_ACTIONS, "items": { "type": "object", "properties": { "kind": {"type": "string", "enum": sorted(kinds)}, "why": {"type": "string", "description": "one short sentence: why this step is here"}, "config": {"type": "object", "additionalProperties": True, "description": "the step's settings, using ONLY the keys named " "for that kind in the system message"}, }, "required": ["kind"], }, }, }, "required": ["kind"], } def flow_system_prompt(catalog, required, triggers, tables): """What the model is told it may build with — DERIVED, never written down twice. Every list here is the server's own: the catalog rows the menu paints, the required-key table the runner blocks on, the trigger keys `clean_trigger` accepts, and this tenant's real databases with their real columns. Nothing about the vocabulary is restated by hand, so a kind added to the catalog is offered here on the same deploy and a kind removed stops being offered. """ lines = ["You build small automations for a business tool. You are given the EXACT set of " "steps this tool can perform. You may use nothing else.", "", "THE STEPS YOU MAY USE:"] # ⚠ BOTH HALVES, `key (phrase)`. The KEY is what the model must write into `config` and the # PHRASE is the only human wording of that requirement anywhere — the one the runner's own # refusal sentence is built from. Keys alone leave the model guessing what `field` means on a # `web_read` (it is a column to write into, not a form field); phrases alone leave it guessing # what to call the setting. Handing it one and hoping is how a draft comes back configured # against a key the validator drops. req = {str(k): [f"{key} ({phrase})" for phrase, key in v] for k, v in (required or {}).items()} for row in catalog or []: if not row.get("ready"): continue kind = str(row.get("kind") or "") need = req.get(kind) or [] lines.append(f"- {kind}: {row.get('label')} - {row.get('detail') or ''}" + (f" REQUIRED settings: {', '.join(need)}" if need else "")) lines += ["", "WHAT CAN START AN AUTOMATION:"] for t in triggers or []: if t.get("planned") or not t.get("ready", True): continue lines.append(f"- {t.get('key')}: {t.get('label')}") lines += ["", "THE DATABASES THIS PERSON HAS:"] for t in (tables or [])[:40]: cols = ", ".join(str(f.get("key")) for f in (t.get("fields") or [])[:25]) lines.append(f"- {t.get('key')} ({t.get('label')}): {cols or 'no columns yet'}") if not tables: lines.append("- (none - do not name a database)") lines += [ "", "RULES:", "1. Use ONLY the step kinds listed above. If what is asked needs a step that is not there, " "answer kind=refused and say plainly which capability is missing.", "2. Fill in every REQUIRED setting you can from what the person told you. Leave one blank " "rather than inventing a web address, a CSS selector or a column name.", "3. A web address must start with http:// or https://.", "4. To use a value from the record the flow is walking, write it as {{Column name}}.", "5. Name a database only from the list above, by its key.", "6. Keep it short. Fewer steps that work beat more steps that guess.", ] return "\n".join(lines) def _flow_from_tool_call(obj): """The model's tool arguments -> `(draft, refusal)`. A shape error is a refusal, never a crash.""" if not isinstance(obj, dict): return None, "the assistant's answer could not be read" if str(obj.get("kind") or "") == "refused": return None, (str(obj.get("refusal") or "").strip() or "the assistant could not build this from the steps available") acts = obj.get("actions") if not isinstance(acts, list) or not acts: return None, ("the assistant did not produce any steps - try describing what should " "happen, one action at a time") out = [] for a in acts[:MAX_DRAFT_ACTIONS]: if not isinstance(a, dict) or not str(a.get("kind") or "").strip(): continue cfg = a.get("config") out.append({"kind": str(a["kind"]).strip(), "config": cfg if isinstance(cfg, dict) else {}, "why": str(a.get("why") or "").strip()[:200]}) if not out: return None, "the assistant's steps could not be read" return {"name": str(obj.get("name") or "").strip()[:80] or "New automation", "trigger": str(obj.get("trigger") or "").strip(), "table": str(obj.get("table") or "").strip(), # ⛔⛔ THE TRUNCATION IS REPORTED, and it is a STANDING RULE that it must be (owner, # 2026-08-12, W30/R6 second sentence): *"if there is lag or it can't be done, you need # to explicitly tell me why and recommend a fix"* — **a silent truncation IS the # violation, not the limit.** `acts[:MAX_DRAFT_ACTIONS]` above dropped everything past # the twelfth and said nothing, so a model that answered with a twenty-step journey had # eight steps deleted between the answer and the screen, with no key anywhere in the # response naming them. `MAX_ACTIONS` is 20, so those steps were STORABLE — this # ceiling is the draft door's own, which makes reporting it the whole obligation. # ⚠ Found by the verifier that checked this ticket, not by a gate. "asked": len([a for a in acts if isinstance(a, dict)]), "actions": out}, "" def harness_proposal_schema(max_body_chars): """One typed answer for an agent-authored harness proposal. ``kind`` alone is required so a refusal remains representable. The caller validates the proposed body in bytes against the store's real cap; JSON Schema's ``maxLength`` is a character count and cannot safely stand in for that UTF-8 boundary. """ return { "type": "object", "properties": { "kind": {"type": "string", "enum": ["proposal", "refused"]}, "refusal": {"type": "string", "description": "when refused: one plain sentence saying what is missing"}, "summary": {"type": "string", "description": "one short sentence describing the proposed change"}, "body": {"type": "string", "maxLength": int(max_body_chars), "description": "the complete replacement body, never a patch or excerpt"}, }, "required": ["kind"], } def harness_system_prompt(index, loaded): """Build the bounded metadata index + ONE exact body the proposal model receives.""" rows = [] for row in index or []: rows.append( f"- {row.get('path')} | {row.get('name')} | v{int(row.get('version') or 0)} | " f"{row.get('description')}" ) source = loaded.get("source") if isinstance(loaded.get("source"), dict) else {} return "\n".join([ "You maintain one stored AIOS Agent's harness. The harness index below contains metadata " "only. Exactly one file body has been loaded because the person opened that file and asked " "for a proposal. Do not claim to have read any other file.", "", "Return a complete replacement body for the loaded file, not a diff. Preserve instructions " "the request does not ask to change. Never widen data access, invent credentials, or turn " "a fact into an instruction. This is a proposal only; a person must approve it before it " "becomes a harness version.", "", "BOUNDED HARNESS INDEX:", *(rows or ["- (no other readable harness files)"]), "", "LOADED EXACT FILE:", f"path: {loaded.get('path')}", f"version: {int(loaded.get('version') or 0)}", f"author: {source.get('author') or ''}", f"author kind: {source.get('authorKind') or ''}", f"updated: {source.get('updated') or ''}", "--- BEGIN EXACT BODY ---", str(loaded.get("body") or ""), "--- END EXACT BODY ---", ]) def _harness_from_tool_call(obj, max_body_bytes): if not isinstance(obj, dict): return None, "the agent's harness proposal could not be read" if str(obj.get("kind") or "") == "refused": return None, (str(obj.get("refusal") or "").strip() or "the agent could not propose a safe harness change") if str(obj.get("kind") or "") != "proposal" or not isinstance(obj.get("body"), str): return None, "the agent did not return a complete harness file" body = obj["body"] if len(body.encode("utf-8")) > int(max_body_bytes): return None, "the proposed harness file is over the size limit" return {"body": body, "summary": " ".join(str(obj.get("summary") or "").split())[ :MAX_HARNESS_SUMMARY_CHARS]}, "" def propose_harness_version(*, prompt, index, loaded, max_body_bytes, chat=None, timeout=None, st=None, user="", model=None): """Return ``(proposal, refusal, provider)`` without writing the proposal or the harness. This is the production model turn behind the Harness panel. The host supplies a bounded metadata index and one exact, provenance-bearing body; this function has no store access and therefore cannot silently widen that scope or approve its own output. """ global LAST_ANTHROPIC_TRANSPORT text = str(prompt or "").strip()[:MAX_HARNESS_PROMPT_CHARS] if not text: return None, "say what the agent should improve in this file", None loaded = loaded if isinstance(loaded, dict) else {} if not isinstance(loaded.get("body"), str) or not loaded.get("path"): return None, "the exact harness file could not be loaded", None schema = harness_proposal_schema(max(1, int(max_body_bytes))) tools = [{"type": "function", "function": { "name": "propose_harness_version", "description": "Propose one complete replacement body, or refuse.", "parameters": schema, }}] messages = [{"role": "system", "content": harness_system_prompt(index, loaded)}, {"role": "user", "content": text}] if chat is not None: proposal, refusal = _harness_from_tool_call(chat(messages, tools), max_body_bytes) return proposal, refusal, "injected" provs = flow_providers(model) if not provs: return None, ("the assistant is not configured on this deployment, so it cannot propose " "a harness version yet"), None tmo = float(timeout or TIMEOUT_SECONDS) problems = [] import providers as _prov # noqa: PLC0415 import usage_ledger # noqa: PLC0415 for p in provs: try: if p["shape"] == "anthropic": req = _prov.anthropic_request( model=p["model"], key=os.environ[p["env"]].strip(), system=None, messages=messages, tools=tools, max_tokens=6000, tool_choice="required") status, response, transport = _prov.anthropic_send(req, timeout=tmo) LAST_ANTHROPIC_TRANSPORT = transport else: wire = requests.post( p["url"], timeout=tmo, headers={"Authorization": f"Bearer {os.environ[p['env']].strip()}", "Content-Type": "application/json"}, json={"model": p["model"], "messages": messages, "tools": tools, "tool_choice": "required", "temperature": 0.1, "max_tokens": 6000}) status = wire.status_code response = (wire.json() if wire.content and wire.status_code == 200 else {"_text": wire.text}) except Exception as exc: # noqa: BLE001 problems.append(f"{p['name']}: {type(exc).__name__}") continue if status != 200: detail = json.dumps(response)[:600] if isinstance(response, dict) else str(response)[:600] if _prov.is_credit_failure(status, detail): _prov.mark_no_credit(p["name"]) problems.append(_prov.refusal_sentence(p["name"], status, detail)) continue try: response = response or {} if p["shape"] == "anthropic": _text, args, declined = _prov.anthropic_read(response) if declined: problems.append(f"{p['name']}: {declined}") continue else: calls = (((response.get("choices") or [{}])[0].get("message") or {}) .get("tool_calls") or []) args = json.loads(calls[0]["function"]["arguments"]) if calls else None except Exception as exc: # noqa: BLE001 problems.append(f"{p['name']}: unreadable answer ({type(exc).__name__})") continue ins, outs = usage_ledger.tokens_from(response) usage_ledger.record("agent_harness_proposal", p["name"], p["model"], ins, outs, total=usage_ledger.total_from(response), st=st, user=user) proposal, refusal = _harness_from_tool_call(args, max_body_bytes) if proposal is not None: return proposal, "", p["name"] if isinstance(args, dict) and str(args.get("kind")) == "refused": return None, refusal, p["name"] problems.append(f"{p['name']}: {refusal}") return None, ("; ".join(problems)[:300] or "no provider answered"), None def draft_flow(*, prompt, catalog, required, triggers, tables, chat=None, timeout=None, st=None, user="", model=None): """A sentence -> `(draft, refusal_sentence, provider)`. ⛔ NOTHING IS SAVED HERE. Exactly one of `draft` and `refusal_sentence` is truthy — the same contract `web_agent.run_step` keeps, so a caller has no third case to get wrong. ⭐ W35-T41 / C7 — `st`/`user` are the usage ledger's target, exactly as on `decide` above and for the same reason: both of this function's callers (`automation_engine._ai_agent_plan` and `routes_automation`'s draft door) are in lane D's fence. C7 says E adds the ledger line here and D asserts it; the two keywords are what D has to pass for the line to be attributable. """ # ⚠ DECLARED AT THE TOP OF THE FUNCTION, not beside the assignment inside the provider loop. # It parses either way; a reader scanning for the declaration does not look inside a `for`. global LAST_ANTHROPIC_TRANSPORT import usage_ledger # noqa: PLC0415 text = str(prompt or "").strip()[:MAX_PROMPT_CHARS] if not text: return None, "type what you want the automation to do", None kinds = sorted({str(r.get("kind")) for r in (catalog or []) if r.get("ready")}) if not kinds: return None, "this deployment offers no automation steps to build with", None trigger_keys = sorted({str(t.get("key")) for t in (triggers or []) if t.get("key") and not t.get("planned")}) or ["manual"] table_keys = sorted({str(t.get("key")) for t in (tables or []) if t.get("key")}) or [""] tools = [{"type": "function", "function": { "name": "build_automation", "description": "Emit the automation, or refuse.", "parameters": flow_schema(kinds, trigger_keys, table_keys)}}] messages = [{"role": "system", "content": flow_system_prompt(catalog, required, triggers, tables)}, {"role": "user", "content": text}] if chat is not None: # ⚠ THE INJECTED PATH IS THE PROVEN PATH. It runs the SAME parse and the SAME refusal # branches as a live call; only the transport is replaced. draft, refusal = _flow_from_tool_call(chat(messages, tools)) return draft, refusal, "injected" provs = flow_providers(model) if not provs: # ⛔ SAY SO. An AI feature that silently does nothing is indistinguishable from one that was # never built [[flag-shipped-without-its-writer]]. return None, ("the assistant is not configured on this deployment, so an automation cannot " "be drafted from a description yet"), None tmo = float(timeout or TIMEOUT_SECONDS) problems = [] import providers as _prov for p in provs: # ⭐⭐ W36-T35 / R4 — THE OWNER QUOTED THIS LINE BACK AT US. It used to read # `problems.append(f"{p['name']}: tool calls are not wired for this shape")` and `continue`, # so the sentence *"anthropic: tool calls are not wired for this shape"* appeared under the # Agent module verbatim. R4: *"Anthropic becomes the tool-calling path that always works."* # It is wired now, through the ONE wire in `providers`, and it matters more than it looks: # every side rung on this account is refusing today (cerebras 402, groq 404, openrouter # 402), so without this branch the drafter cannot answer at all. # ⚠ NO `effort` ON THIS DOOR — `output_config.effort` errors on Haiku 4.5, the tier this # ladder runs. The parameter is the caller's to send, which is why the wire takes it. # ⭐ D-346 (W37-T39): BOTH shapes now yield the SAME `(status, body)` pair, so every branch # below reads one thing. The anthropic side gets there through `anthropic_send` (official # SDK when installed, the identical raw POST when not); the openai side still posts here # because there is no shared SDK across three different vendors on that wire. global LAST_ANTHROPIC_TRANSPORT try: if p["shape"] == "anthropic": req = _prov.anthropic_request( model=p["model"], key=os.environ[p["env"]].strip(), system=None, messages=messages, tools=tools, max_tokens=1500, tool_choice="required") status, body, transport = _prov.anthropic_send(req, timeout=tmo) LAST_ANTHROPIC_TRANSPORT = transport else: r = requests.post(p["url"], timeout=tmo, headers={"Authorization": f"Bearer {os.environ[p['env']].strip()}", "Content-Type": "application/json"}, json={"model": p["model"], "messages": messages, "tools": tools, "tool_choice": "required", "temperature": 0.1, "max_tokens": 1500}) status = r.status_code body = r.json() if (r.content and r.status_code == 200) else {"_text": r.text} except Exception as e: # noqa: BLE001 problems.append(f"{p['name']}: {type(e).__name__}") continue if status != 200: # ⛔ A SENTENCE, NOT `HTTP 402` (R4's third clause). The owner read the status codes off # this very door. `refusal_sentence` also decides whether this was a CREDIT failure, and # the memo makes the next turn SKIP the empty account instead of paying for it again. # ⚠ The body is re-serialised for the substring test because the SDK path never had a # `.text` — that is the one thing this normalisation costs, and it costs nothing else. detail = json.dumps(body)[:600] if isinstance(body, dict) else str(body)[:600] if _prov.is_credit_failure(status, detail): _prov.mark_no_credit(p["name"]) problems.append(_prov.refusal_sentence(p["name"], status, detail)) continue try: body = body or {} if p["shape"] == "anthropic": _text, args, _refused = _prov.anthropic_read(body) if _refused: problems.append(f"{p['name']}: {_refused}") continue else: calls = (((body.get("choices") or [{}])[0].get("message") or {}) .get("tool_calls") or []) args = json.loads(calls[0]["function"]["arguments"]) if calls else None except Exception as e: # noqa: BLE001 problems.append(f"{p['name']}: unreadable answer ({type(e).__name__})") continue # ⭐⭐ C7 — THE LEDGER LINE. Placed after the parse rather than before it so an UNREADABLE # answer is not double-counted by the `continue` above... ⚠ which means an unreadable answer # is NOT counted at all, and that is a deliberate, disclosed loss: a body this code cannot # parse is a body whose `usage` it also cannot trust, and `body` is out of scope in that # branch by construction. The 200-with-junk case is rare and named here rather than # silently rounded to zero. ins, outs = usage_ledger.tokens_from(body) usage_ledger.record("automation_draft", p["name"], p["model"], ins, outs, total=usage_ledger.total_from(body), st=st, user=user) draft, refusal = _flow_from_tool_call(args) if draft is None and isinstance(args, dict) and str(args.get("kind")) == "refused": # ⛔ A REFUSAL IS AN ANSWER, NOT A FAULT, so it does NOT fall through to a more # expensive rung — the model has just said this cannot be built. `decide()` takes the # same posture for the same reason. return None, refusal, p["name"] if draft is not None: return draft, "", p["name"] problems.append(f"{p['name']}: {refusal}") return None, ("; ".join(problems)[:300] or "no provider answered"), None