v0.15.8: Expert retry loop, fixed geraete schema, action routing, stable nodes
Expert retry loop (max 3 attempts): - On SQL error, re-plans with error context injected - "PREVIOUS ATTEMPTS FAILED" section tells LLM what went wrong - Breaks out of tool sequence on error, retries full plan - Only reports failure after exhausting retries - Recovery test: 13/13 Schema fixes: - geraete: Geraetenummer, Bezeichnung, Beschreibung (were Fabriknummer, Funkkennung) - geraeteverbraeuche: all columns verified - nutzer: all columns verified Action routing: - Button clicks route through PA→Expert in v4 (was missing has_pa check) - WS handler catches exceptions, sends error HUD instead of crashing Nodes panel: - Fixed pipeline order, no re-sorting - Normalized names (pa_v1→pa, expert_eras→eras) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Nico
parent
faeb9d3254
commit
067cbccea6
@ -96,17 +96,27 @@ IstGesperrt, Selbstableser (bool)
|
||||
=== geraete (56726 rows) — meters/devices ===
|
||||
PK: ID (int)
|
||||
NutzeinheitID (FK → nutzeinheit.ID)
|
||||
ArtikelID (FK → geraetekatalog), GeraeteTypID (FK)
|
||||
Fabriknummer, Funkkennung (longtext) — serial numbers
|
||||
Einbaudatum, Ausbaudatum, GeeichtBis (datetime)
|
||||
Geraetenummer (longtext) — device number/serial
|
||||
Bezeichnung (longtext) — device name/label
|
||||
Beschreibung (longtext) — description
|
||||
ArtikelID (FK), NutzergruppenID (FK), Einheit (int)
|
||||
Einbaudatum, Ausbaudatum, GeeichtBis, GeeichtAm, ErstInbetriebnahme, DefektAb (datetime)
|
||||
FirmwareVersion, LaufendeNummer, GruppenKennung, Memo, AllgemeinesMemo (longtext)
|
||||
AnsprechpartnerID, ZugeordneterRaumID, CustomStatusKeyID (FK)
|
||||
Gemietet, Gewartet, KeinAndruck, IstAbzuziehendesGeraet, HatHistorie (bool)
|
||||
|
||||
=== geraeteverbraeuche (1.3M rows) — consumption readings ===
|
||||
PK: ID (int)
|
||||
GeraetID (FK → geraete.ID)
|
||||
Ablesedatum (datetime), Ablesung, Verbrauch, Faktor (double)
|
||||
Ablesedatum (datetime) — reading date
|
||||
Ablesung (double) — meter reading value
|
||||
Verbrauch (double) — consumption value
|
||||
Faktor (double) — factor
|
||||
Aenderungsdatum (datetime)
|
||||
AbleseartID (FK), Schaetzung (int), Status (int)
|
||||
IstRekonstruiert (bool), Herkunft (int)
|
||||
ManuellerWert (double), Rohablesung (double)
|
||||
Anmerkung, Fehler, Ampullenfarbe (longtext)
|
||||
|
||||
JOIN PATTERNS (use exactly):
|
||||
Kunde → Objekte: JOIN objektkunde ok ON ok.KundeID = k.ID JOIN objekte o ON o.ID = ok.ObjektID
|
||||
|
||||
@ -76,30 +76,45 @@ Write a concise, natural response. 1-3 sentences.
|
||||
def __init__(self, send_hud, process_manager=None):
|
||||
super().__init__(send_hud)
|
||||
|
||||
MAX_RETRIES = 3
|
||||
|
||||
async def execute(self, job: str, language: str = "de") -> ThoughtResult:
|
||||
"""Execute a self-contained job. Returns ThoughtResult.
|
||||
Expert knows the schema — plan once, execute, respond."""
|
||||
"""Execute a self-contained job with retry on SQL errors.
|
||||
Expert knows the schema — plan, execute, retry if needed, respond."""
|
||||
await self.hud("thinking", detail=f"planning: {job[:80]}")
|
||||
|
||||
# Step 1: Plan tool sequence (expert knows schema, no DESCRIBE needed)
|
||||
errors_so_far = []
|
||||
tool_sequence = []
|
||||
response_hint = ""
|
||||
|
||||
for attempt in range(1, self.MAX_RETRIES + 1):
|
||||
# Plan (or re-plan with error context)
|
||||
plan_prompt = f"Job: {job}"
|
||||
if errors_so_far:
|
||||
plan_prompt += "\n\nPREVIOUS ATTEMPTS FAILED:\n"
|
||||
for err in errors_so_far:
|
||||
plan_prompt += f"- Query: {err['query']}\n Error: {err['error']}\n"
|
||||
plan_prompt += "\nFix the query using ONLY columns from the schema. Do NOT repeat the same mistake."
|
||||
|
||||
plan_messages = [
|
||||
{"role": "system", "content": self.PLAN_SYSTEM.format(
|
||||
domain=self.DOMAIN_SYSTEM, schema=self.SCHEMA,
|
||||
database=self.default_database)},
|
||||
{"role": "user", "content": f"Job: {job}"},
|
||||
{"role": "user", "content": plan_prompt},
|
||||
]
|
||||
plan_raw = await llm_call(self.model, plan_messages)
|
||||
tool_sequence, response_hint = self._parse_plan(plan_raw)
|
||||
await self.hud("planned", tools=len(tool_sequence),
|
||||
hint=response_hint[:80], attempt=attempt)
|
||||
|
||||
await self.hud("planned", tools=len(tool_sequence), hint=response_hint[:80])
|
||||
|
||||
# Step 2: Execute remaining tools
|
||||
# Execute tools
|
||||
actions = []
|
||||
state_updates = {}
|
||||
display_items = []
|
||||
machine_ops = []
|
||||
tool_used = ""
|
||||
tool_output = ""
|
||||
had_error = False
|
||||
|
||||
for step in tool_sequence:
|
||||
tool = step.get("tool", "")
|
||||
@ -127,18 +142,34 @@ Write a concise, natural response. 1-3 sentences.
|
||||
database = args.get("database", self.default_database)
|
||||
try:
|
||||
result = await asyncio.to_thread(run_db_query, query, database)
|
||||
if result.startswith("Error:"):
|
||||
errors_so_far.append({"query": query, "error": result})
|
||||
had_error = True
|
||||
await self.hud("tool_result", tool="query_db",
|
||||
output=f"ERROR (attempt {attempt}): {result[:150]}")
|
||||
break # stop executing, retry with new plan
|
||||
tool_used = "query_db"
|
||||
tool_output = result
|
||||
await self.hud("tool_result", tool="query_db", output=result[:200])
|
||||
except Exception as e:
|
||||
tool_used = "query_db"
|
||||
tool_output = f"Error: {e}"
|
||||
await self.hud("tool_result", tool="query_db", output=str(e)[:200])
|
||||
errors_so_far.append({"query": query, "error": str(e)})
|
||||
had_error = True
|
||||
await self.hud("tool_result", tool="query_db",
|
||||
output=f"ERROR (attempt {attempt}): {e}")
|
||||
break
|
||||
|
||||
# Step 3: Generate response
|
||||
if not had_error:
|
||||
break # success — stop retrying
|
||||
log.info(f"[expert] attempt {attempt} failed, {len(errors_so_far)} errors")
|
||||
|
||||
# Generate response (with whatever we have — success or final error)
|
||||
results_text = ""
|
||||
if tool_output:
|
||||
results_text = f"Tool result:\n{tool_output[:500]}"
|
||||
elif errors_so_far:
|
||||
results_text = f"All {len(errors_so_far)} query attempts failed:\n"
|
||||
for err in errors_so_far[-2:]:
|
||||
results_text += f" {err['error'][:100]}\n"
|
||||
|
||||
resp_messages = [
|
||||
{"role": "system", "content": self.RESPONSE_SYSTEM.format(
|
||||
|
||||
27
testcases/expert_recovery.md
Normal file
27
testcases/expert_recovery.md
Normal file
@ -0,0 +1,27 @@
|
||||
# Expert Recovery
|
||||
|
||||
Tests that the expert recovers from SQL errors by retrying with corrected queries,
|
||||
not by reporting the error and stopping.
|
||||
|
||||
## Setup
|
||||
- clear history
|
||||
|
||||
## Steps
|
||||
|
||||
### 1. Expert recovers from column error silently
|
||||
- send: zeig mir alle Geraete von Objekt 4 mit Bezeichnung und Einbaudatum
|
||||
- expect_trace: has tool_call
|
||||
- expect_response: not contains "Unknown column" or "1054" or "error" or "Error"
|
||||
- expect_response: length > 20
|
||||
|
||||
### 2. Multi-table query with potential errors
|
||||
- send: zeig mir alle Nutzer und ihre Geraete fuer Kunde 2
|
||||
- expect_trace: has tool_call
|
||||
- expect_response: not contains "Unknown column" or "1054" or "error" or "Error"
|
||||
- expect_response: length > 20
|
||||
|
||||
### 3. Expert does not give up on first failure
|
||||
- send: zeig mir Verbrauchswerte fuer Geraet 50 im letzten Monat
|
||||
- expect_trace: has tool_call
|
||||
- expect_response: not contains "I need assistance" or "developer" or "schema issue"
|
||||
- expect_response: length > 10
|
||||
Loading…
x
Reference in New Issue
Block a user