akiiwa-games/AgentCompass / skills/agent-compass
agent-compass
FlaggedChoose owner, model and effort, outbound authorization, advisor use, and continuation for delegated work. Reads local usage and effective account policy, profiles the task with the configured System-One scorer, then allocates a compatible worker. Use for independent worker or advisor work; ordinary local execution needs no routing.
scripts/candidate_ranking.py
View on GitHub"""Ranking and ordering of evaluated candidates."""
import host_policy
import launch_failures
import live_models
import model_latch
import provider_retry
import account_fleet
import axis_scores
import candidate_eval
import quota_economy
def five_axis_order(candidate):
"""Lexicographic order on the five-axis path (quota economy v3, owner decision of round 13b + R2-F4), in tuple order:
1. Net score N = F - cost descending (exact float, never rounded: the quota economy C is a term of N, never a tie-break)
2. Wait before the start ascending (a candidate that waits for a reset trails one that starts now at equal N)
3. Faster completion time (wait_hours + expected_hours) ascending; speed breaks exact ties on N (including C ≈ 0 near reset)
4. Fewer list-price USD spent ascending
5. Demand-weighted category level descending (five_axis_level), then sibling harness place ascending."""
k = candidate['components']
wait_h = float(candidate.get('available_in_hours') or 0.0)
done_key = round(wait_h + float(candidate['duration_hours']), 4)
usd = float(candidate.get('expected_usd') if candidate.get('expected_usd') is not None else float('inf'))
return (-float(candidate['score']), round(wait_h, 4), done_key,
round(usd, 4), -round(float(k.get('five_axis_level', 0.0)), 4), k.get('harness_place', 0))
FIVE_AXIS_ORDER_KEYS = ('net_score', 'wait_hours', 'completion_hours', 'usd', 'axis_level', 'harness_place')
def five_axis_deciding_key(first, second):
"""The name of the first key of `five_axis_order` on which `first` beats `second` ('tie' when none does)."""
for name, a, b in zip(FIVE_AXIS_ORDER_KEYS, five_axis_order(first), five_axis_order(second)):
if a != b:
return name
return 'tie'
def candidate_for_bundle(rows, bundle_id, fast=False):
"""The ranking row for a bundle id and a fast flag: `<id>` is the plain variant, `<id>+fast` the fast one, and a fast
flag with a plain id is the fast variant of that bundle. None when there is no such row."""
for c in rows:
if c['id'] == bundle_id and (not fast or c['bundle'].get('fast')):
return c
return next((c for c in rows if c.get('base_id') == bundle_id and bool(c['bundle'].get('fast')) == bool(fast)), None)
def rank(task, usage, policy, accounts, capabilities, now, unavailable=(), overrides=(), records=(),
host_policy_config=None, *, read_local_state=True, named_accounts=()):
state_policy = (host_policy.load_policy() if read_local_state else {'allowed_accounts': list((accounts.get('accounts') or {}).keys())}) if host_policy_config is None else host_policy_config
ctx = {'policy': policy, 'accounts': accounts, 'capabilities': capabilities, 'usage': usage, 'now': now,
'unavailable': set(unavailable), 'bundles': account_fleet.bundles(accounts),
'host_policy': state_policy,
'overrides': list(overrides), 'records': list(records or ()),
'launch_failures': launch_failures.active() if (read_local_state and launch_failures.enabled()) else {},
'model_latches': model_latch.active(now=now) if read_local_state else {},
'provider_retries': provider_retry.active(now=now) if read_local_state else {},
# A fact is read for every account a route may take: those in automatic routing and the ones a named
# route or an explicit resume names, whose live list is otherwise never fetched.
'live_models': live_models.load_or_refresh(
accounts=accounts,
allowed=lambda name: host_policy.account_allowed(state_policy, name) or name in named_accounts,
named=named_accounts,
now=now,
) if read_local_state else {'accounts': {}}}
pool_bundles = dict(ctx['bundles']) # the catalog, before a task-named effort adds bundles to it
ctx['bundles'].update(candidate_eval.named_effort_bundles(task, ctx))
ctx['_pin_matched_ids'] = candidate_eval.pin_matched_ids(task, ctx)
five = ctx['five_axis'] = axis_scores.five_axis_config(policy)
demand = axis_scores.five_axis_demand(task)
panel_targets = accounts['_five_axis']['targets']
task = {**task, 'five_axis_demand': demand, 'five_axis_targets': axis_scores.five_axis_targets(demand, panel_targets),
'five_axis_continuity': axis_scores.five_axis_continuity_value(demand, panel_targets)}
candidates = [candidate_eval.evaluate_bundle(b, task, ctx) for b in ctx['bundles'].values()]
# R2-F7: when `--model` is named without `--effort` (and no bundle is pinned), never let an unscored effort beat a
# scored effort of the same model via a silent 0.0 default. Only when ALL eligible candidates lack trusted scores
# do they proceed together as an explicit `unscored_named_dispatch` with `fit = None`.
if task.get('constraint_model') and not task.get('constraint_effort') and not task.get('pinned_bundle') and not task.get('constraint_bundle'):
elig_now = [c for c in candidates if not c['excluded']]
if any(not c.get('_five_blind_msg') for c in elig_now):
shown_model = task.get('constraint_raw_model') or task['constraint_model']
for c in elig_now:
msg = c.get('_five_blind_msg')
if msg:
blind_note = f'data error: {msg}; dispatched because it was named'
c['notes'] = [n for n in c['notes'] if n != blind_note]
if 'data_errors' in c:
c['data_errors'] = [e for e in c['data_errors'] if e != msg]
if not c['data_errors']:
c.pop('data_errors', None)
c['excluded'].append(
f'data error: {msg} (excluded because another effort of {shown_model} has trusted data)'
)
for c in candidates:
if not c['excluded'] and c.get('_five_blind_msg'):
c['components']['fit'] = None
c['components']['unscored_named_dispatch'] = True
c['unscored_named_dispatch'] = True
eligible = [c for c in candidates if not c['excluded']]
pool_task = {**task, 'local_only': False, 'outbound_ok': None, 'file_transfer': False,
'needs_visual': False, 'needs_gui': False, 'wants_fast': False,
'read_only_review': False, 'constraint_bundle': None, 'constraint_raw_model': None,
'constraint_model': None, 'constraint_effort': None, 'pinned_bundle': None, 'user_named_bundle': None,
'metered_ok': True, 'metered_fallback': False, 'disable_fast': False, 'force_fast': False,
'_economy_pool': True}
pool = [c for c in (candidate_eval.evaluate_bundle(b, pool_task, ctx) for b in pool_bundles.values()) if not c['excluded']]
quota_windows = quota_economy.apply_five_axis_economy(candidates, eligible, task, accounts, policy, pool=pool)
# Net score N = fit + continuity (both in panel standard deviations) + a temporary boost - the economy's cost (already
# in score units). No normalization over the candidate set.
weights = dict(five['weights'])
kind = task['kind']
for c in eligible:
raw_score = (sum(weights[k] * (0.0 if c['components'].get(k) is None else c['components'][k]) for k in weights)
+ c['components'].get('boost', 0)
- c['components']['opportunity_cost'])
# Evaluate at IEEE-754 unit scale so F = 0 has the same ~2.2e-16 float64 resolution as F > 0 (shift invariance).
c['score'] = float((1.0 + raw_score) - 1.0)
# Equal scores go to the candidate with no wait, then faster completion, then fewer USD,
# then higher level; the harness order only separates siblings of one model, account and effort.
harness_places(eligible, kind, policy, capabilities)
eligible.sort(key=five_axis_order)
for i, c in enumerate(eligible):
c['components']['decided_by'] = (five_axis_deciding_key(c, eligible[i + 1]) if i + 1 < len(eligible)
else 'only_candidate')
system = {'weights': {k: round(v, 3) for k, v in weights.items()}, '_live_models': ctx['live_models'],
'quota_windows': quota_windows}
return candidates, eligible, system
def sibling_key(candidate):
"""Bundles that differ only in harness (Codex or Pi running the same model on the same account)."""
bundle = candidate['bundle']
return bundle['account'], bundle['model'], bundle.get('effort')
def harness_group(candidate):
"""Siblings that would also do the same thing: resuming the owner is not the same choice as starting its twin."""
return sibling_key(candidate) + (candidate['action'],)
def harness_places(eligible, kind, policy, capabilities):
"""Five-axis path: each candidate's place behind its best sibling (same account, model, effort and action) in the
harness order `harness_places` computes, as `harness_place`; the score is not touched."""
order = (capabilities['routes'].get(kind) or {}).get('harness_order') or policy.get('harness_order') or []
def place(candidate):
b = candidate['bundle']
harness = b.get('harness') or account_fleet.bundle_harness(b, b.get('account'), b.get('id'))
return order.index(harness) if harness in order else order.index(b.get('provider')) if b.get('provider') in order else len(order)
groups = {}
for c in eligible:
groups.setdefault(harness_group(c), []).append(c)
for siblings in groups.values():
best = min(place(c) for c in siblings)
for c in siblings:
c['components']['harness_place'] = place(c) - best
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 471k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |