perf(f1): cache-guard schema introspect + F1 toolbox in skill (~-5 min per session)

Transcript analysis (session 2026-07-06-175012, GLM 5.2) showed F1 at 567s:
182s wasted on a useless `tht schema introspect` (re-introspecting the remote
DWH although physical.yaml was already materialized) plus ~220s of model
thinking inflated by ~7 exploratory turns (--help/find/cat). The actual
searches cost ~15s; reviewer gates (~145s, untouched) are the quality contract.

- schema_cmd.py: introspect now exits 0 with "OK (cache)" in ~1s when
  physical.yaml exists; --refresh forces the real re-introspection.
  Deterministic cross-model guarantee, verified live on psd (163 tables, 1.2s).
- SKILL.md: F1 toolbox (only `tht search find` + `tht schema render`; no
  introspect/--help/filesystem browsing; batch all searches in one turn);
  F4 step 1 is render-only with a one-shot introspect fallback.
- tht-gate.js: `tht schema introspect ... --refresh` added to FORBIDDEN
  (maintenance stays shell-only, never in-session).
- tests: 4 new pytest cases (cache hit placement proven with fake credentials,
  refresh bypass, corrupt-catalog fall-through, render fallback message) and
  2 gate anti-bypass JS cases.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
2026-07-07 16:02:45 +02:00
co-authored by Claude Fable 5
parent 0cad04a8d8
commit 34aeda006e
5 changed files with 147 additions and 5 deletions
@@ -0,0 +1,27 @@
const test = require("node:test");
const assert = require("node:assert");
const { createFakePi } = require("./fake_pi_runtime.js");
const installGatePromise = import("../../tht-gate.js").then((m) => m.default);
test("tht schema introspect --refresh e' bloccato in sessione (manutenzione)", async () => {
const installGate = await installGatePromise;
const { pi } = createFakePi();
installGate(pi);
const res = await pi.emit("tool_call", {
toolName: "bash",
input: { command: "tht schema introspect -c workspaces/psd.yaml --refresh" },
});
assert.equal(res?.block, true);
});
test("tht schema introspect senza --refresh passa (cache hit innocuo)", async () => {
const installGate = await installGatePromise;
const { pi } = createFakePi();
installGate(pi);
const res = await pi.emit("tool_call", {
toolName: "bash",
input: { command: "tht schema introspect -c workspaces/psd.yaml" },
});
assert.equal(res, undefined);
});
+2
View File
@@ -109,6 +109,8 @@ const FORBIDDEN = [
/\btht\s+cte\s+plan\b/,
// La promozione in memoria passa dal reviewer (reviewer_memory_promote), mai da shell.
/\btht\s+memory\s+(promote|save-one)\b/,
// La re-introspezione del DWH (~3 min) è manutenzione fuori sessione, mai in workflow.
/\btht\s+schema\s+introspect\b[^\n]*--refresh/,
];
const PROTECTED_FILES =
/(review_decisions\.jsonl|session_manifest\.yaml|cte_plan\.json)/;
+15 -2
View File
@@ -148,10 +148,21 @@ it is complete. (The backend already refuses resume for finalized/archived sessi
Prerequisite: you must already be in Phase 1.
**F1 toolbox.** The only commands you need here are `tht search find` and `tht schema
render` — both fast, read-only lookups over workspace artifacts already on disk.
Evidence lives in `<workspace>/evidence/**` and is what `tht search find --kind evidence`
returns — do not browse it with `find`/`cat`. Do NOT run `tht schema introspect`: it is
a maintenance command that re-reads the remote DWH (~3 minutes); the catalog
`artifacts/mschema/physical.yaml` is already in the workspace. Do NOT explore with
`--help` or ad-hoc shell commands — every command you need is named in this skill.
1. Explore the DWH and knowledge base: `tht search find "<term>"` (evidence + schema, LSH
over real values) and `tht search find --kind evidence "<term>"`. The LSH exposes
EVERY column where a value appears — it does not collapse to a single best match,
so a value like "ablazione" may anchor on multiple columns.
First list ALL the ambiguous terms in the question, then run the `tht search find`
calls for every term in ONE batch (a single message with multiple shell invocations)
— not one lookup per turn.
2. For each ambiguity (clinical term, population, time window, outcome), present the
candidate interpretations (`recommended:true` on the best) + "Altro". Pick the widget
by the question's shape:
@@ -230,8 +241,10 @@ Phase 3 (CLI exit 5).
Prerequisite: Phase 3 closed.
1. `tht schema introspect` + `tht schema render --format mschema-text` for the schema
context. Copy table/column names EXACTLY from it — never invent objects.
1. `tht schema render --format mschema-text` for the schema context (the catalog
`artifacts/mschema/physical.yaml` is already in the workspace; only if render fails
with `physical.yaml non trovato`, run `tht schema introspect` once, then render).
Copy table/column names EXACTLY from it — never invent objects.
Also run `tht memory solved-search "<question>" --json`: similar already-solved
questions show which tables comparable questions used. Cite relevant precedents
(session id + tables) to the reviewer as CONTEXT — they are reference material,
@@ -0,0 +1,70 @@
from datetime import datetime
from typer.testing import CliRunner
from tht.cli import app
from tht.mschema.models import ColumnPhysical, PhysicalSchema, TablePhysical
def _write_catalog(tmp_path):
phys = PhysicalSchema(
database="d", schema="s", introspected_at=datetime(2026, 1, 1),
tables={
"dim_patient": TablePhysical(
comment="Anagrafica",
columns={
"cod_paz": ColumnPhysical(type="bigint", pk=True, comment="Codice paziente"),
},
)
},
)
out = tmp_path / "artifacts" / "mschema" / "physical.yaml"
phys.to_yaml(out)
return out
def _write_config(tmp_path):
cfg = tmp_path / "workspace.yaml"
cfg.write_text(
"database: {database: d, schema: s, user: u, password: p, transport: direct}\n"
f"paths: {{artifacts: {tmp_path/'artifacts'}, indexes: {tmp_path/'i'}, sessions: {tmp_path/'s'}}}\n"
)
return cfg
def test_introspect_cache_hit_skips_dwh(tmp_path):
# Le credenziali sono fasulle: se la guardia non scattasse PRIMA del branch
# transport, il comando tenterebbe la connessione e fallirebbe.
catalog = _write_catalog(tmp_path)
before = catalog.read_bytes()
cfg = _write_config(tmp_path)
res = CliRunner().invoke(app, ["schema", "introspect", "-c", str(cfg)])
assert res.exit_code == 0, res.output
assert "OK (cache)" in res.output
assert "1 tabelle" in res.output
assert catalog.read_bytes() == before
def test_introspect_refresh_bypasses_cache(tmp_path):
_write_catalog(tmp_path)
cfg = _write_config(tmp_path)
res = CliRunner().invoke(app, ["schema", "introspect", "-c", str(cfg), "--refresh"])
assert res.exit_code != 0
assert "OK (cache)" not in res.output
def test_introspect_corrupt_catalog_falls_through(tmp_path):
catalog = tmp_path / "artifacts" / "mschema" / "physical.yaml"
catalog.parent.mkdir(parents=True)
catalog.write_text("{not: [valid")
cfg = _write_config(tmp_path)
res = CliRunner().invoke(app, ["schema", "introspect", "-c", str(cfg)])
assert res.exit_code != 0
assert "OK (cache)" not in res.output
def test_render_without_catalog_guides_fallback(tmp_path):
cfg = _write_config(tmp_path)
res = CliRunner().invoke(app, ["schema", "render", "-c", str(cfg)])
assert res.exit_code == 1
assert "Esegui prima" in res.output
+33 -3
View File
@@ -30,9 +30,40 @@ def annotations_path(cfg) -> Path:
@schema_app.command("introspect")
def introspect_cmd(config: Path = CONFIG_OPT) -> None:
"""Introspeziona lo schema target e genera artifacts/mschema/physical.yaml."""
def introspect_cmd(
config: Path = CONFIG_OPT,
refresh: bool = typer.Option(
False,
"--refresh",
help="Forza la re-introspezione del DWH anche se physical.yaml esiste già.",
),
) -> None:
"""Introspeziona lo schema target e genera artifacts/mschema/physical.yaml.
Se physical.yaml esiste già, esce subito (cache); usa --refresh per rigenerarlo.
"""
cfg = _load_config_or_exit(config)
out = physical_path(cfg)
if out.exists() and not refresh:
from datetime import UTC, datetime
from tht.mschema.models import PhysicalSchema
try:
cached = PhysicalSchema.from_yaml(out)
except Exception:
pass # catalogo illeggibile: procedi con la re-introspezione
else:
ts = cached.introspected_at
if ts.tzinfo is None:
ts = ts.replace(tzinfo=UTC)
age_days = (datetime.now(UTC) - ts).days
typer.secho(
f"OK (cache): {out} esistente ({len(cached.tables)} tabelle, "
f"età {age_days}g). Re-introspezione solo con --refresh (manutenzione).",
fg=typer.colors.GREEN,
)
return
if cfg.database.transport == "rest":
from tht.db.introspect import introspect_rest
from tht.db.sampling import add_examples_rest
@@ -59,7 +90,6 @@ def introspect_cmd(config: Path = CONFIG_OPT) -> None:
except (OperationalError, IntrospectionError) as e:
typer.secho(f"ERRORE: {e}", fg=typer.colors.RED, err=True)
raise typer.Exit(code=1)
out = physical_path(cfg)
phys.to_yaml(out)
n_cols = sum(len(t.columns) for t in phys.tables.values())
n_ignored = sum(