refactor(do): remove fable from the model selection policy (#896)

Owner directive: "our document for what agent to use should remove fable
entirely" — removed, not demoted.

- skills/meta/do/SKILL.md: drop the Fable-5 variant row, the manual-pick
  set and ordering, the cross-provider escalation target, and the
  felt-quality clause comparing fable to sol; reword the sentences that
  depended on them.
- scripts/build-dispatch.py: drop fable from VALID_MODELS,
  ANTHROPIC_MODELS, and the manual-override gate (now sonnet alone).
- hooks/routing-decision-recorder.py: drop fable from the valid-model set.
- scripts/validate_component_models.py: ALLOWED_MODELS is {sonnet, opus}.
- docs/PHILOSOPHY.md, agent-frontmatter-template.md,
  component-audit-checklists.md: same removal in live instructions.
- Tests updated to match, including SUPPLIED_CLAUDE_POINTS.

Opus-4.8 and Sonnet-5 prior-measurement rows are kept as manual-only:
they are the Anthropic lane's only DeepSWE data (Opus 5 is unbenchmarked)
and cover cost, latency, context-window, and fan-out constraints.

Dated historical records keep their fable references
(docs/what-didnt-work.md, scripts/routing-ab-results/) — editing them
would falsify experiment records; neither is a live routing instruction.

Author: notque

agents/system-upgrade-engineer/references/component-audit-checklists.md

@@ -33,7 +33,7 @@ For each agent in scope, check these fields in order:
 
 ```bash
 # 1. Check model field is current
-grep "^model:" agents/*.md | grep -v "sonnet\|opus\|fable"
+grep "^model:" agents/*.md | grep -v "sonnet\|opus"
 
 # 2. Check allowed-tools for deprecated tools
 grep -A 10 "allowed-tools:" agents/*.md | grep "deprecated_tool_name"
@@ -207,7 +207,7 @@ Do instead: Scope audit to the component types identified in the Change Manifest
 
 ```bash
 # Agents: stale model name
-grep "^model:" agents/*.md | grep -v "sonnet\|opus\|fable"
+grep "^model:" agents/*.md | grep -v "sonnet\|opus"
 
 # Hooks: syntax errors
 for f in hooks/*.py; do python3 -m py_compile "$f" 2>&1 | grep -v "^$" && echo "ERROR: $f"; done

docs/PHILOSOPHY.md

@@ -344,7 +344,7 @@ The biggest risk: rationalization. "Already done" (assumption). "Code looks corr
 
 Owner model-selection policy (ADR `model-selection-policy`; operational table in `skills/meta/do/SKILL.md`, Model Selection):
 
-**Harness-native routing:** each harness defaults to its own provider's model lane. Cross-provider dispatch is manual-only, never automatic. Start low, escalate on miss — high tiers cost 3-6x per Pass@1 point where measured. Plan budget ($200/month per provider) makes cost a first-class constraint. Three decision axes: the current session model (the harness runs Opus 5, the owner-directed Anthropic-lane default at every task class), DeepSWE Pass@1 (agentic completion) where a model has been measured, and owner-observed felt quality (fable > sol noticeable, opus > gpt-5.5 marginal); ties resolve in favor of felt quality.
+**Harness-native routing:** each harness defaults to its own provider's model lane. Cross-provider dispatch is manual-only, never automatic. Start low, escalate on miss — high tiers cost 3-6x per Pass@1 point where measured. Plan budget ($200/month per provider) makes cost a first-class constraint. Three decision axes: the current session model (the harness runs Opus 5, the owner-directed Anthropic-lane default at every task class), DeepSWE Pass@1 (agentic completion) where a model has been measured, and owner-observed felt quality (opus > gpt-5.5 marginal); ties resolve in favor of felt quality.
 
 | Task class | Anthropic lane (Claude Code) | OpenAI lane (Codex CLI) |
 |---|---|---|
@@ -354,7 +354,7 @@ Owner model-selection policy (ADR `model-selection-policy`; operational table in
 | High-risk | `opus` / `high` (unmeasured) | `gpt-5.6-sol` / `xhigh` (71 / $4.70) |
 | Max-power | `opus` / `xhigh` (unmeasured) | `gpt-5.6-sol` / `max` (73 / $8.39) |
 
-The `/do` table is canonical and records the full DeepSWE Pass@1 / cost / tokens / steps data, including the prior Fable-5 / Opus-4.8 / Sonnet-5 measurements kept for manual picks. Opus 5 has no DeepSWE run yet, so its pts/USD is uncomputable and the selection rests on the session-model and owner-directive axes; effort still follows start-low-escalate-on-miss. Max-power requires `manual_model_override=true` in both lanes. Fable/sonnet are manual-only. Legacy `gpt-5.5` and non-default GPT-5.6 points are manual-only. Haiku is retired (routing was Haiku pre-#777; self-route since — `scripts/routing-ab-results/self-route-v1/VERDICT.md`). Defaults, not limits: escalate when cheaper output misses the bar; for anything that ships, intelligence > taste > cost, with cost a tie-breaker only. Fan-out uses the lane's low-risk point; one synthesis agent may run one tier higher.
+The `/do` table is canonical and records the full DeepSWE Pass@1 / cost / tokens / steps data, including the prior Opus-4.8 / Sonnet-5 measurements kept for manual picks. Opus 5 has no DeepSWE run yet, so its pts/USD is uncomputable and the selection rests on the session-model and owner-directive axes; effort still follows start-low-escalate-on-miss. Max-power requires `manual_model_override=true` in both lanes. Sonnet is manual-only. Legacy `gpt-5.5` and non-default GPT-5.6 points are manual-only. Haiku is retired (routing was Haiku pre-#777; self-route since — `scripts/routing-ab-results/self-route-v1/VERDICT.md`). Defaults, not limits: escalate when cheaper output misses the bar; for anything that ships, intelligence > taste > cost, with cost a tie-breaker only. Fan-out uses the lane's low-risk point; one synthesis agent may run one tier higher.
 
 **Coordinator model.** The main-thread coordinator routes and evaluates, never executes — its cost is input-dominated and Pass@1 measures execution it never does. Anthropic harness → opus (Opus 5, the session model); OpenAI harness → gpt-5.6-terra/high; downgrade the anthropic coordinator to sonnet only as a deliberate plan-limit measure, set via harness config (`/model`), not per-turn. Full rule: `skills/meta/do/SKILL.md`, Model Selection.
 

hooks/routing-decision-recorder.py

@@ -117,7 +117,6 @@
     {
         "sonnet",
         "opus",
-        "fable",
         "codex",
         "gpt-5.5",
         "gpt-5.6-sol",

hooks/tests/test_routing_decision_recorder.py

@@ -445,26 +445,26 @@ def test_gpt_56_model_effort_is_persisted_in_route_evidence(self, db_env, monkey
         assert context["recent"][0]["model"] == "gpt-5.6-sol@xhigh"
 
     def test_claude_model_effort_is_persisted_in_route_evidence(self, db_env, monkeypatch):
-        """Claude model@effort (fable@high) survives the marker round-trip (advisory effort)."""
+        """Claude model@effort (sonnet@high) survives the marker round-trip (advisory effort)."""
         a = _load(A_PATH, "rdr_claude_effort")
         monkeypatch.setattr(a, "append_pending_outcome", lambda *_a, **_k: None)
         monkeypatch.setattr(a, "claim_dispatch", lambda *_a, **_k: True)
         event = _agent_event(skill="go-patterns", session="claude-effort")
         event["tool_input"]["prompt"] = (
             "[do-route] agent=python-general-engineer skill=go-patterns complexity=complex "
-            "model=fable effort=high health=-\nReview the implementation."
+            "model=sonnet effort=high health=-\nReview the implementation."
         )
         with patch("sys.exit"), patch("sys.stdin.read", return_value=json.dumps(event)):
             a.main()
 
         decisions = [event for event in _read_events(db_env) if event["type"] == "decision"]
-        assert decisions[0]["model"] == "fable@high"
+        assert decisions[0]["model"] == "sonnet@high"
 
         sys.path.insert(0, str(LIB_DIR))
         import learning_db_v2 as ldb
 
         context = ldb.get_evidence_route_context("python-general-engineer:go-patterns")
-        assert context["recent"][0]["model"] == "fable@high"
+        assert context["recent"][0]["model"] == "sonnet@high"
 
     def test_no_telemetry_row_when_marker_absent(self, db_env, tmp_path, monkeypatch):
         # No [do-route] marker => no decision row AND no envelope row.

scripts/build-dispatch.py

@@ -124,9 +124,9 @@
 GPT_56_MODELS = ("gpt-5.6-sol", "gpt-5.6-terra", "gpt-5.6-luna")
 GPT_56_EFFORTS = ("low", "medium", "high", "xhigh", "max")
 LEGACY_GPT_55 = "gpt-5.5"
-VALID_MODELS = ("sonnet", "opus", "fable", "codex", LEGACY_GPT_55, *GPT_56_MODELS)
+VALID_MODELS = ("sonnet", "opus", "codex", LEGACY_GPT_55, *GPT_56_MODELS)
 VALID_PROVIDERS = ("anthropic", "openai", "other")
-ANTHROPIC_MODELS = ("fable", "opus", "sonnet")
+ANTHROPIC_MODELS = ("opus", "sonnet")
 
 # DeepSWE Pass@1 / cost benchmark defaults per provider lane.
 # `deterministic` deliberately has no model: use scripts.
@@ -220,8 +220,8 @@ def resolve_model_selection(decision: dict, provider: str = "anthropic") -> tupl
 
     Harness-aware: ``provider`` selects the automatic policy table.
     Anthropic lane defaults select Opus 5 at every task class (owner
-    directive + current session model); fable and sonnet are manual-only,
-    kept for felt-quality, context-window, and latency constraints.
+    directive + current session model); sonnet is manual-only, kept for
+    cost, context-window, and latency constraints.
     OpenAI lane defaults select GPT-5.6 Sol/Terra.  Effort is recorded in
     the marker for all models; for Claude lanes it is advisory (the harness
     Agent tool does not accept per-call effort).
@@ -286,10 +286,10 @@ def resolve_model_selection(decision: dict, provider: str = "anthropic") -> tupl
             raise InputError("legacy gpt-5.5 requires manual_model_override=true")
         return model, effort
 
-    # Claude models (fable, opus, sonnet) and codex wrapper.
+    # Claude models (opus, sonnet) and codex wrapper.
     # Effort is optional and advisory for Claude lanes — recorded in the
     # marker (model@effort) for telemetry but not passed to the Agent tool.
-    if model in ("fable", "sonnet"):
+    if model == "sonnet":
         if not manual:
             raise InputError(
                 f"'{model}' requires manual_model_override=true "

scripts/tests/test_build_dispatch.py

@@ -549,11 +549,6 @@ def test_non_default_openai_model_choices_require_manual_override(overrides):
 # ---------------------------------------------------------------------------
 
 SUPPLIED_CLAUDE_POINTS = {
-    ("fable", "max"): (70, 21.63, 119_000, 88),
-    ("fable", "xhigh"): (70, 13.41, 80_000, 68),
-    ("fable", "high"): (69, 9.18, 57_000, 59),
-    ("fable", "medium"): (65, 6.09, 40_000, 48),
-    ("fable", "low"): (60, 3.76, 25_000, 38),
     ("opus-4.8", "max"): (59, 13.22, 135_000, 120),
     ("opus-4.8", "xhigh"): (54, 8.01, 86_000, 95),
     ("opus-4.8", "high"): (52, 4.28, 50_000, 73),
@@ -599,12 +594,12 @@ def test_anthropic_policy_points_are_unmeasured_and_effort_rises_with_risk():
         previous = order.index(effort)
 
 
-def test_fable_max_still_costs_more_than_fable_xhigh():
-    """Recorded prior measurement: fable[max] matches fable[xhigh] Pass@1 at higher cost."""
-    fmax = SUPPLIED_CLAUDE_POINTS[("fable", "max")]
-    fxhigh = SUPPLIED_CLAUDE_POINTS[("fable", "xhigh")]
-    assert fmax[0] == fxhigh[0], "precondition: same Pass@1"
-    assert fmax[1] > fxhigh[1], "precondition: max costs more"
+def test_sonnet_max_still_costs_more_per_point_than_sonnet_xhigh():
+    """Recorded prior measurement: sonnet[max] buys 4 Pass@1 points for 2.2x the cost."""
+    smax = SUPPLIED_CLAUDE_POINTS[("sonnet", "max")]
+    sxhigh = SUPPLIED_CLAUDE_POINTS[("sonnet", "xhigh")]
+    assert smax[0] > sxhigh[0], "precondition: max scores higher"
+    assert smax[1] / smax[0] > sxhigh[1] / sxhigh[0], "precondition: max costs more per point"
 
 
 def test_opus_max_requires_manual_override():
@@ -614,15 +609,14 @@ def test_opus_max_requires_manual_override():
 
 
 def test_prior_measurements_retained_for_manual_picks():
-    """Historical Fable-5 / Opus-4.8 / Sonnet-5 points stay recorded, not deleted."""
-    assert SUPPLIED_CLAUDE_POINTS[("fable", "low")] == (60, 3.76, 25_000, 38)
+    """Historical Opus-4.8 / Sonnet-5 points stay recorded, not deleted."""
     assert SUPPLIED_CLAUDE_POINTS[("opus-4.8", "max")] == (59, 13.22, 135_000, 120)
     assert SUPPLIED_CLAUDE_POINTS[("sonnet", "high")] == (48, 7.43, 87_000, 147)
 
 
-@pytest.mark.parametrize("model", ("fable", "sonnet"))
+@pytest.mark.parametrize("model", ("sonnet",))
 def test_off_policy_claude_models_require_manual_override(model):
-    """Opus 5 is the default; fable and sonnet are the manual-only picks."""
+    """Opus 5 is the default; sonnet is the manual-only pick."""
     with pytest.raises(bd.InputError, match="manual_model_override"):
         bd.build_marker(_decision(model=model))
     # With manual_override they work fine

scripts/tests/test_validate_component_models.py

@@ -52,14 +52,14 @@ def test_do_skill_is_exempt(tmp_path: Path) -> None:
     assert violations == []
 
 
-def test_sonnet_opus_and_fable_are_allowed(tmp_path: Path) -> None:
+def test_sonnet_and_opus_are_allowed(tmp_path: Path) -> None:
     repo = tmp_path / "repo"
     agents_dir = repo / "agents"
     agents_dir.mkdir(parents=True)
     (agents_dir / "demo-agent.md").write_text("---\nmodel: opus\n---\n", encoding="utf-8")
     skill_dir = repo / "skills" / "demo-skill"
     skill_dir.mkdir(parents=True)
-    (skill_dir / "SKILL.md").write_text("---\nmodel: fable\n---\n", encoding="utf-8")
+    (skill_dir / "SKILL.md").write_text("---\nmodel: sonnet\n---\n", encoding="utf-8")
     workflow_dir = repo / "skills" / "workflow" / "references"
     workflow_dir.mkdir(parents=True)
     (workflow_dir / "demo-pipeline.md").write_text("---\nmodel: sonnet\n---\n", encoding="utf-8")

scripts/validate_component_models.py

@@ -11,7 +11,7 @@
 
 REPO_ROOT = Path(__file__).resolve().parent.parent
 MODEL_RE = re.compile(r"^model:\s*(.+)$", re.MULTILINE)
-ALLOWED_MODELS = {"sonnet", "opus", "fable"}
+ALLOWED_MODELS = {"sonnet", "opus"}
 EXEMPT_COMPONENTS: set[tuple[str, str]] = {
     ("skill", "do"),
 }
@@ -58,7 +58,7 @@ def validate_models() -> list[Violation]:
                 component_type=component_type,
                 component=component,
                 file=str(path.relative_to(REPO_ROOT)),
-                issue=f"model '{model}' is not allowed; use 'sonnet', 'opus', or 'fable' (except /do)",
+                issue=f"model '{model}' is not allowed; use 'sonnet' or 'opus' (except /do)",
             )
         )
     return violations

skills/meta/agent-creator/references/agent-frontmatter-template.md

@@ -80,7 +80,7 @@ user_invocable: false
 | `name` | yes | string | lowercase, hyphens, numbers |
 | `description` | yes | quoted string | 60–120 chars; no "Use when:" prefix |
 | `color` | no | string | blue, green, yellow, red, purple, orange |
-| `model` | no | string | sonnet, opus, fable |
+| `model` | no | string | sonnet, opus |
 | `routing.triggers` | yes | list | 3–6 natural-speech phrases |
 | `routing.pairs_with` | no | list | agent names that exist on disk |
 | `routing.complexity` | yes | enum | Low, Medium, High (case-sensitive) |