chore(routing): retire haiku from delivery-lifecycle roles (#680)

Haiku can't reliably emit the structured envelopes the lifecycle now
runs on — pass_review's per-AC criteria_verified, delegate's
covers_parent_criteria, the findings ledger. A haiku QA/PM claims, gets
validation-rejected, idles, respawns, and loops without advancing a task
(2026-07-24 live: fe-qa on haiku looped four awaiting_qa tasks to zero
progress). The per-token savings (~2x under the Sonnet-5 promo, 3x after)
are dwarfed by the cost of a review that never completes.

Three coordinated changes: ROLE_MODEL_MAP's qa/documenter defaults move
haiku -> sonnet (the actual source of the live incident); the cost_tiered
developer:low -> haiku seed retires to empty (the floor would upgrade it
anyway); and a structured-verb capability floor upgrades any below-floor
Anthropic assignment to sonnet at resolution — from a pin, a ROLE row,
or a future map edit — in both the assignment and legacy paths.
Non-Anthropic providers are untouched (an Anthropic-tier floor, not a
provider policy). pr_reviewer/auditor stay on opus.

Co-authored-by: Renn F <rennf93@users.noreply.github.com>
This commit is contained in:
Renzo F
2026-07-24 12:14:13 +02:00
committed by GitHub
co-authored by Renn F
parent e1ae58489c
commit a3f84f3165
5 changed files with 132 additions and 26 deletions
+7 -4
View File
@@ -119,13 +119,16 @@ MODEL_MAP: dict[str, str] = {
# Default model by role
ROLE_MODEL_MAP: dict[str, str] = {
"developer": "sonnet",
# QA — mechanical gate work (read diff, run the gate, pass/fail); its cost is
# cache-dominated, so the cheapest tier fits. Haiku ignores effort (fine here).
"qa": "haiku",
# QA — emits structured review envelopes (pass_review's per-AC
# criteria_verified, the findings ledger) the haiku tier can't reliably
# produce: a haiku QA claims, gets validation-rejected, idles, respawns,
# and loops without passing (2026-07-24 live incident). Sonnet is the
# floor for every structured-verb lifecycle role.
"qa": "sonnet",
# PR reviewer — reviews untrusted external/fork PRs and gates root→master;
# highest-stakes review, so opus rather than the sonnet review tier.
"pr_reviewer": "opus",
"documenter": "haiku",
"documenter": "sonnet",
"cell_pm": "sonnet",
# Main PM — cost is dominated by cache read/write of a large coordination
# context; Sonnet 5's cache-write is ~12x cheaper than Opus. Experiment: watch
+59 -13
View File
@@ -33,7 +33,7 @@ Self-hosted (LOCAL) provider support:
from __future__ import annotations
from dataclasses import dataclass
from dataclasses import dataclass, replace
from typing import TYPE_CHECKING, Any, ClassVar, Literal, cast
import httpx
@@ -72,17 +72,16 @@ if TYPE_CHECKING:
_OLLAMA_TAGS_TIMEOUT = 5.0 # seconds
_log = structlog.get_logger(__name__)
# Day-1 cost-tiered seed applied by apply_mode('cost_tiered'): (role,
# complexity, model_name). "haiku" is the catalog's cheap Anthropic tier
# (see MODEL_CATALOG_BY_NAME) — developer's LOW-complexity work is
# mechanical/cache-dominated the same way QA already runs on haiku
# (ROLE_MODEL_MAP). developer is the only entry: qa/documenter already
# default to haiku in ROLE_MODEL_MAP (no saving to seed), and cell_pm is
# deliberately excluded from complexity overrides entirely — a coordinator
# role, never offered a row (see _COMPLEXITY_OVERRIDE_ROLES in
# api/routes/provider.py). Extend this tuple to seed more role:complexity
# rows; nothing else needs editing.
_COST_TIERED_SEED: tuple[tuple[str, str, str], ...] = (("developer", "low", "haiku"),)
# Cost-tiered seed applied by apply_mode('cost_tiered'): (role, complexity,
# model_name). RETIRED to empty (2026-07-24): the one entry seeded
# developer:low -> haiku, but haiku can't reliably emit the structured
# lifecycle envelopes (see _below_capability_floor), so the "cheap tier for
# mechanical work" premise no longer holds — the capability floor would
# upgrade the seeded row back to sonnet anyway, making the seed a confusing
# no-op. cost_tiered mode stays wired (empty seed = inert) so an operator can
# re-seed a genuinely-above-floor role:complexity tier here later; nothing
# else needs editing.
_COST_TIERED_SEED: tuple[tuple[str, str, str], ...] = ()
# derive_mode()'s single-GLOBAL-assignment lookup — a provider type maps to
# its "mode" label 1:1 for every mode `apply_mode` can set via a sole GLOBAL
@@ -173,6 +172,29 @@ INTERACTIVE_UNSUPPORTED_PROVIDERS: tuple[ModelProvider, ...] = (
)
# Capability floor: haiku-class models cannot reliably emit the structured
# review/coordination envelopes the lifecycle now requires (pass_review's
# criteria_verified, delegate's covers_parent_criteria, the findings ledger)
# — a haiku QA/PM/reviewer claims, gets validation-rejected, idles, respawns,
# and loops without ever advancing a task (2026-07-24 live: fe-qa on haiku
# looped 4 awaiting_qa tasks to zero progress). Every delivery-lifecycle role
# runs on structured verbs, so any Anthropic assignment below the floor is
# upgraded to the floor model at resolution, from whatever source (a pin, a
# ROLE row, or the retired cost_tiered seed). Non-Anthropic providers are
# untouched — this is an Anthropic-tier floor, not a provider policy.
_CAPABILITY_FLOOR_MODEL = "sonnet"
_BELOW_FLOOR_MODEL_MARKER = "haiku"
def _below_capability_floor(resolved: _ResolvedAssignment) -> bool:
"""True iff an Anthropic assignment resolves below the structured-verb
floor (a haiku-class model). Provider-agnostic names never match."""
return (
resolved.provider.type is ModelProvider.ANTHROPIC
and _BELOW_FLOOR_MODEL_MARKER in resolved.model_name.lower()
)
def _interactive_exempt(agent_slug: str, resolved: _ResolvedAssignment) -> bool:
"""True iff a GLOBAL/ROLE row lands an interactive agent on a
delivery-only provider — the resolver then keeps it on the legacy path.
@@ -209,6 +231,7 @@ class ModelRoutingService(BaseService):
"""
role = get_agent_role(agent_slug) or ""
resolved = await self._resolve_assignment(agent_slug, role, complexity)
resolved = self._floor_below_capability(agent_slug, resolved)
if resolved is not None and _interactive_exempt(agent_slug, resolved):
# A fleet-wide GLOBAL/ROLE row landed an interactive agent on a
# delivery-only provider (e.g. the one-click Codex/Gemini mode).
@@ -325,9 +348,32 @@ class ModelRoutingService(BaseService):
)
return None
def _floor_below_capability(
self, agent_slug: str, resolved: _ResolvedAssignment | None
) -> _ResolvedAssignment | None:
"""Upgrade a below-floor Anthropic assignment to the structured-verb
floor (see `_below_capability_floor`); pass everything else through."""
if resolved is None or not _below_capability_floor(resolved):
return resolved
self.log.info(
"Upgrading below-floor model to the structured-verb floor",
agent_slug=agent_slug,
from_model=resolved.model_name,
to_model=_CAPABILITY_FLOOR_MODEL,
)
return replace(resolved, model_name=_CAPABILITY_FLOOR_MODEL)
def _legacy_route(self, role: str) -> AgentRoute:
"""Legacy fallback: role-default short name through MODEL_MAP."""
"""Legacy fallback: role-default short name through MODEL_MAP.
The structured-verb capability floor applies here too: any role whose
map entry is a below-floor tier is upgraded, so a future map edit
(or a role not covered above) can't silently reintroduce a haiku
lifecycle agent.
"""
short = ROLE_MODEL_MAP.get(role, "sonnet")
if _BELOW_FLOOR_MODEL_MARKER in short.lower():
short = _CAPABILITY_FLOOR_MODEL
return AgentRoute(
provider_id=None,
provider_type=ModelProvider.ANTHROPIC,
+6 -6
View File
@@ -956,8 +956,10 @@ async def test_apply_mode_cost_tiered_seeds_day1_rows(
"/api/providers/complexity-overrides", headers=_HDR_PM
)
rows = {(r["role"], r["complexity"]): r["model_name"] for r in listing.json()}
# cell_pm is deliberately excluded (a coordinator role) — only developer.
assert rows == {("developer", "low"): "haiku"}
# Seed retired (2026-07-24): its only entry was developer:low -> haiku,
# which the structured-verb capability floor would upgrade to sonnet
# anyway. cost_tiered mode stays wired but seeds nothing now.
assert rows == {}
@pytest.mark.asyncio
@@ -979,11 +981,9 @@ async def test_apply_mode_cost_tiered_is_additive_preserves_global(
assert response.status_code == HTTPStatus.OK
assignments = response.json()["assignments"]
scopes = {(a["scope"], a["scope_value"]) for a in assignments}
# The pre-existing GLOBAL row from 'ollama' mode survives untouched.
# The pre-existing GLOBAL row from 'ollama' mode survives untouched — the
# non-wiping contract holds even with the seed retired to empty.
assert ("global", None) in scopes
assert ("role", "developer:low") in scopes
# cell_pm is deliberately excluded from cost_tiered — a coordinator role.
assert ("role", "cell_pm:low") not in scopes
# =============================================================================
+54
View File
@@ -77,3 +77,57 @@ async def test_resolve_for_agent_no_assignment_is_silent_legacy_fallback() -> No
assert route.provider_type == ModelProvider.ANTHROPIC
svc.log.warning.assert_not_called()
@pytest.mark.asyncio
async def test_haiku_lifecycle_assignment_upgraded_to_the_floor() -> None:
"""The structured-verb capability floor: a haiku-class Anthropic
assignment on any lifecycle role resolves to sonnet instead, so a QA/PM
agent can produce the structured review envelopes (2026-07-24 retirement).
"""
provider = MagicMock(
enabled=True, id="prov-anthropic", type=ModelProvider.ANTHROPIC
)
resolved = _ResolvedAssignment(
provider=provider,
model_name="claude-haiku-4-5-20251001",
scope=AssignmentScope.ROLE,
)
svc = _svc()
captured: dict[str, str] = {}
async def _fake_route(res: _ResolvedAssignment, _slug: str) -> object:
captured["model"] = res.model_name
return object()
with (
patch.object(svc, "_resolve_assignment", AsyncMock(return_value=resolved)),
patch.object(svc, "_route_from_resolved", _fake_route),
):
await svc.resolve_for_agent("be-qa")
assert captured["model"] == "sonnet"
@pytest.mark.asyncio
async def test_non_anthropic_below_floor_name_is_untouched() -> None:
"""The floor is an Anthropic-tier policy — a non-Anthropic model whose
name happens to contain the marker is never upgraded."""
provider = MagicMock(enabled=True, id="prov-grok", type=ModelProvider.GROK)
resolved = _ResolvedAssignment(
provider=provider, model_name="grok-haiku-ish", scope=AssignmentScope.ROLE
)
svc = _svc()
captured: dict[str, str] = {}
async def _fake_route(res: _ResolvedAssignment, _slug: str) -> object:
captured["model"] = res.model_name
return object()
with (
patch.object(svc, "_resolve_assignment", AsyncMock(return_value=resolved)),
patch.object(svc, "_route_from_resolved", _fake_route),
):
await svc.resolve_for_agent("be-qa")
assert captured["model"] == "grok-haiku-ish"
+6 -3
View File
@@ -23,9 +23,12 @@ def test_sonnet_5_is_priced() -> None:
assert cost > 0.0
def test_qa_role_routes_to_haiku() -> None:
# Phase 2: QA is mechanical gate work → cheapest tier.
assert ROLE_MODEL_MAP["qa"] == "haiku"
def test_no_lifecycle_role_routes_below_the_structured_verb_floor() -> None:
# Retired haiku from lifecycle roles (2026-07-24): a haiku QA/documenter
# can't emit the structured review envelopes (criteria_verified, findings)
# and loops without passing. No role's default may be a haiku-class tier.
for role, model in ROLE_MODEL_MAP.items():
assert "haiku" not in model.lower(), f"{role} routes to a below-floor {model}"
def test_main_pm_role_routes_to_sonnet() -> None: