General-purpose coding policy for Baruch's AI agents
74
93%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
Medium
Suggest reviewing before use
"""Operator-opted minimum-adequate pair selection, not a new qualification programme.
Config owns launch/account facts; capabilities owns sourced qualification;
measure owns capacity. None supplies another owner's evidence. Rows are
assessed against this task's requirements. Declared resource weights are
ordering proxies, never a currency or separate-quota claim.
"""
import copy
from datetime import timedelta
from . import capabilities, successors
from .billing import billing_window, effective_multiplier
from .chronology import timestamp
from .errors import ConfigError, UsageError
from .tiers import JUDGMENT_ROUNDS, ROLE_ROUNDS, canonical_role
MODES = frozenset({"minimum_adequate", "pinned"})
FACT_MAX_AGE = capabilities.INTERVAL
CAPACITY_MAX_AGE = timedelta(minutes=15)
EFFORT_ORDER = {None: 0, "low": 1, "medium": 2, "high": 3, "xhigh": 4, "max": 5}
def parse(value, tiers):
"""Validate schema-8 operator input; missing means legacy configured-row routing."""
if value is None:
return None
if (not isinstance(value, dict) or set(value) != {"mode", "evidence"}
or not isinstance(value["mode"], str) or value["mode"] not in MODES
or not isinstance(value["evidence"], dict)):
raise ConfigError("tier_routing needs mode (minimum_adequate or pinned) and an evidence object keyed by configured row.", {})
for name, proof in value["evidence"].items():
if name not in tiers or not isinstance(proof, dict) or set(proof) != {"model", "effort", "launch", "access"}:
raise ConfigError("tier_routing evidence names a configured row and carries model, effort, launch and access.", {})
if not isinstance(proof["model"], str) or not isinstance(proof["effort"], (str, type(None))):
raise ConfigError("tier_routing evidence needs the observed model and effort, not an inferred pair.", {})
for kind, statuses in (("launch", {"supported", "unsupported", "unknown"}),
("access", {"accessible", "unavailable", "unknown"})):
fact = proof[kind]
fields = {"status", "ref", "checked_at"} | ({"window_group"} if kind == "access" else set())
if (not isinstance(fact, dict) or set(fact) != fields
or not isinstance(fact["status"], str) or fact["status"] not in statuses
or not isinstance(fact["ref"], str) or not fact["ref"].strip()
or (kind == "access" and not isinstance(fact["window_group"], str))):
raise ConfigError("tier_routing {} fact needs status, a non-secret evidence ref, checked_at{}.".format(
kind, " and window_group" if kind == "access" else ""),
{"row": name, "fact": kind})
try:
timestamp(fact["checked_at"], "Tier-routing fact")
except UsageError as exc:
raise ConfigError(exc.message, {"row": name, "fact": kind}) from None
return copy.deepcopy(value)
def _fresh(value, at, age):
if value is None or at is None:
return False
try:
delta = timestamp(at, "Tier-routing checkpoint") - timestamp(value, "Tier-routing fact")
except UsageError:
return False
return timedelta(0) <= delta < age
def _fact(proof, kind, row, agent, at):
fact = proof.get(kind) if isinstance(proof, dict) else None
status = fact.get("status", "unknown") if isinstance(fact, dict) else "unknown"
if proof and (proof.get("model"), proof.get("effort")) != (row["model"], row.get("effort")):
status = "pair_mismatch"
elif kind == "access" and fact and fact["window_group"] != agent.window_group:
status = "account_mismatch"
elif fact and not _fresh(fact["checked_at"], at, FACT_MAX_AGE):
status = "stale"
return {"status": status, "evidence": copy.deepcopy(fact)}
def decide(agent, role, tier, needs, table, headroom, measured_at, at, capacity_group=None, worker_kind=None):
"""Return the selected tier plus audit facts; judgment and explicit pins stay fixed."""
policy = agent.tier_routing
if policy is None:
return tier, None
mode = policy["mode"]
override = "operator_pin" if mode == "pinned" else None
if tier["round"] in JUDGMENT_ROUNDS or tier["tier_row"] in JUDGMENT_ROUNDS:
override = "judgment_floor"
elif tier["de_escalated"] or (tier["model"], tier.get("effort")) != (
agent.tiers[tier["tier_row"]]["model"], agent.tiers[tier["tier_row"]].get("effort")):
override = "risk_escalation"
capacity_status = "unknown"
if headroom is not None:
capacity_status = "available" if headroom > 0 else "exhausted"
if not _fresh(measured_at, at, CAPACITY_MAX_AGE):
capacity_status = "stale"
elif capacity_group != agent.window_group:
capacity_status = "account_unknown"
capacity = {"status": capacity_status, "remaining_pct": headroom,
"measured_at": measured_at, "window_group": capacity_group}
allowed = ROLE_ROUNDS.get(canonical_role(role), frozenset())
candidates = []
for name, row in sorted(agent.tiers.items()):
proof = policy["evidence"].get(name)
launch = _fact(proof, "launch", row, agent, at)
access = _fact(proof, "access", row, agent, at)
sources = capabilities.evidence(table, row["model"], row.get("effort"), needs)
try:
qualification = capabilities.assess(table, row["model"], row.get("effort"), needs)
except capabilities.InadequateCapability:
qualification = "inadequate"
if qualification == "adequate" and not all(
_fresh(source["dated"] + "T00:00:00Z", at, FACT_MAX_AGE) for source in sources):
qualification = "stale"
placement = successors.placement(table, agent, role, tier["round"], name, row, needs, at, worker_kind) if at else None
inherited = placement is not None and placement["status"] == "provisional" and override is None
rejected = []
if name not in allowed or name in JUDGMENT_ROUNDS:
rejected.append("role_ineligible")
for kind, status, positive in (("launch", launch["status"], "supported"),
("access", access["status"], "accessible"),
("capacity", capacity_status, "available")):
if status != positive:
rejected.append(kind + "_" + status)
if qualification != "adequate" and not inherited:
rejected.append("qualification_" + qualification)
if placement and placement["status"] in {"withdrawn", "retired", "inadequate", "future"}:
rejected.append("placement_" + placement["status"])
candidates.append({"tier_row": name, "model": row["model"], "effort": row.get("effort"),
"resource_weight": effective_multiplier(row), "launch": launch, "access": access,
"qualification": {"status": qualification, "sources": sources},
"placement": placement,
"capacity": capacity, "rejected": rejected})
chosen = tier
if mode == "minimum_adequate" and override is None:
eligible = [row for row in candidates if not row["rejected"]]
if not eligible:
raise UsageError("No qualified, accessible supported pair has current capacity for {} on {}; refresh only the cited candidate facts or explicitly pin the configured row.".format(role, agent.name),
{"role": role, "agent": agent.name, "routing_candidates": candidates})
selected = min(eligible, key=lambda row: (row["resource_weight"], EFFORT_ORDER[row["effort"]], row["model"], row["tier_row"]))
chosen = {**agent.tiers[selected["tier_row"]], "tier_row": selected["tier_row"],
**{key: tier[key] for key in ("round", "kind", "pressure_headroom", "de_escalated")}}
chosen.update(billing_window=billing_window(chosen), effective_multiplier=effective_multiplier(chosen))
return chosen, {"mode": mode, "override": override, "ordering": "declared_resource_weight_then_effort",
"monetary_cost": "unknown", "candidates": candidates}.tessl-plugin
hooks
rules
skills
adopt-fork-pr
herdr-foreman
classify
foreman
references
specialists
templates
tests
herdr-standup
migrate-to-plugin
onboard-repo
release
references
tests