ganymede: implement tier logic — LIGHT skip, claim-shape detector, pre-merge promotion

- Claim-shape detector: if YAML has type: claim, force STANDARD minimum (Theseus)
- Random pre-merge promotion: 15% of LIGHT → STANDARD before eval (Rio)
- LIGHT_SKIP_LLM config flag: skip domain+Leo review for LIGHT (Rhea: env var rollback)
- Updated both_approve: domain_verdict=skipped is valid for LIGHT auto-approve
- Cost recording: only charge for reviews that actually ran
- SAMPLE_AUDIT_RATE bumped 0.10 → 0.15, audit model = Opus (Leo: different family from Haiku)

Multi-agent design review: Rio (gaming vectors, model diversity), Theseus (correlated
blindspots, claim-shape guard), Rhea (shadow mode, config flag, deployment), Leo (approval).

Pentagon-Agent: Ganymede <F99EBFA6-547B-4096-BEEA-1D59C3E4028A>
This commit is contained in:
m3taversal 2026-03-13 18:05:43 +00:00
parent 410cf32cfe
commit ffa718e834
2 changed files with 137 additions and 35 deletions

View file

@ -107,8 +107,17 @@ OPENROUTER_DAILY_BUDGET = 20.0 # USD
OPENROUTER_WARN_THRESHOLD = 0.8 # 80% of budget
# --- Quality ---
SAMPLE_AUDIT_RATE = 0.10 # 10% of LIGHT merges
SAMPLE_AUDIT_RATE = 0.15 # 15% of LIGHT merges get pre-merge promotion to STANDARD (Rio)
SAMPLE_AUDIT_DISAGREEMENT_THRESHOLD = 0.10 # 10% disagreement → tighten LIGHT criteria
SAMPLE_AUDIT_MODEL = MODEL_OPUS # Opus for audit — different family from Haiku triage (Leo)
# --- Tier logic ---
# LIGHT_SKIP_LLM: when True, LIGHT PRs skip domain+Leo review entirely (auto-approve on Tier 0 pass).
# Set False for shadow mode (domain review runs but logs only). Flip True after 24h validation (Rhea).
LIGHT_SKIP_LLM = os.environ.get("LIGHT_SKIP_LLM", "false").lower() == "true"
# Random pre-merge promotion: fraction of LIGHT PRs upgraded to STANDARD before eval (Rio).
# Makes gaming unpredictable — extraction agents can't know which LIGHT PRs get full review.
LIGHT_PROMOTION_RATE = float(os.environ.get("LIGHT_PROMOTION_RATE", "0.15"))
# --- Polling intervals (seconds) ---
INGEST_INTERVAL = 60

View file

@ -1,23 +1,26 @@
"""Evaluate stage — PR lifecycle orchestration.
Ported from eval-worker.sh. Key architectural change: domain-first, Leo-last.
Sonnet (domain review) filters before Opus (Leo review) to maximize value per
scarce Opus call.
Tier-based review routing. Model diversity: GPT-4o (domain) + Sonnet (Leo STANDARD)
+ Opus (Leo DEEP) = two model families, no correlated blind spots.
Flow per PR:
1. Triage Haiku (OpenRouter) DEEP / STANDARD / LIGHT
2. Domain review Sonnet (Claude Max, overflow: OpenRouter GPT-4o)
3. Leo review Opus (Claude Max, overflow: queue) skipped for LIGHT
4. DEEP cross-family GPT-4o (OpenRouter) only if domain + Leo both approve
2. Tier overrides:
a. Claim-shape detector: type: claim in YAML STANDARD min (Theseus)
b. Random pre-merge promotion: 15% of LIGHT STANDARD (Rio)
3. Domain review GPT-4o (OpenRouter) skipped for LIGHT when LIGHT_SKIP_LLM=True
4. Leo review Opus DEEP / Sonnet STANDARD (OpenRouter) skipped for LIGHT
5. Post reviews, submit formal Forgejo approvals, update SQLite
6. If both approve status = 'approved' (merge module picks it up)
7. Retry budget: 3 attempts max, disposition on attempt 2+
Design reviewed by Ganymede, Rhea, Vida, Theseus.
Design reviewed by Ganymede, Rio, Theseus, Rhea, Leo.
LLM transport and prompts extracted to lib/llm.py (Phase 3c).
"""
import json
import logging
import random
import re
from datetime import datetime, timezone
@ -80,6 +83,25 @@ def _is_musings_only(diff: str) -> bool:
return has_musings and not has_other
# ─── Tier overrides ───────────────────────────────────────────────────────
def _diff_contains_claim_type(diff: str) -> bool:
"""Claim-shape detector: check if any file in diff has type: claim in frontmatter.
Mechanical check ($0). If YAML declares type: claim, this is a factual claim
not an entity update or formatting fix. Must be classified STANDARD minimum
regardless of Haiku triage. Catches factual claims disguised as LIGHT content.
(Theseus: converts semantic problem to mechanical check)
"""
for line in diff.split("\n"):
if line.startswith("+") and not line.startswith("+++"):
stripped = line[1:].strip()
if stripped in ("type: claim", 'type: "claim"', "type: 'claim'"):
return True
return False
# ─── Verdict parsing ──────────────────────────────────────────────────────
@ -133,10 +155,12 @@ async def _terminate_pr(conn, pr_number: int, reason: str):
await forgejo_api(
"POST",
repo_path(f"issues/{pr_number}/comments"),
{"body": f"**Closed by eval pipeline** — {reason}.\n\n"
f"This PR has been evaluated {config.MAX_EVAL_ATTEMPTS} times without passing. "
f"Source material will be re-queued for extraction with review feedback attached.\n\n"
f"See eval_issues for specific problems."},
{
"body": f"**Closed by eval pipeline** — {reason}.\n\n"
f"This PR has been evaluated {config.MAX_EVAL_ATTEMPTS} times without passing. "
f"Source material will be re-queued for extraction with review feedback attached.\n\n"
f"See eval_issues for specific problems."
},
)
await forgejo_api(
"PATCH",
@ -160,9 +184,17 @@ async def _terminate_pr(conn, pr_number: int, reason: str):
if cursor.rowcount == 0:
logger.warning("PR #%d: no source_path linked — source not requeued for re-extraction", pr_number)
db.audit(conn, "evaluate", "pr_terminated", json.dumps({
"pr": pr_number, "reason": reason,
}))
db.audit(
conn,
"evaluate",
"pr_terminated",
json.dumps(
{
"pr": pr_number,
"reason": reason,
}
),
)
logger.info("PR #%d: TERMINATED — %s", pr_number, reason)
@ -205,18 +237,34 @@ async def _dispose_rejected_pr(conn, pr_number: int, eval_attempts: int, all_iss
# Future: auto-fix module will push fixes here.
logger.info(
"PR #%d: attempt %d, mechanical issues only (%s) — keeping open for fix attempt",
pr_number, eval_attempts, all_issues,
pr_number,
eval_attempts,
all_issues,
)
db.audit(
conn,
"evaluate",
"mechanical_retry",
json.dumps(
{
"pr": pr_number,
"attempt": eval_attempts,
"issues": all_issues,
}
),
)
db.audit(conn, "evaluate", "mechanical_retry", json.dumps({
"pr": pr_number, "attempt": eval_attempts, "issues": all_issues,
}))
else:
# Substantive, mixed, or unknown — close and requeue
logger.info(
"PR #%d: attempt %d, %s issues (%s) — closing and requeuing source",
pr_number, eval_attempts, classification, all_issues,
pr_number,
eval_attempts,
classification,
all_issues,
)
await _terminate_pr(
conn, pr_number, f"substantive issues after {eval_attempts} attempts: {', '.join(all_issues)}"
)
await _terminate_pr(conn, pr_number, f"substantive issues after {eval_attempts} attempts: {', '.join(all_issues)}")
# ─── Single PR evaluation ─────────────────────────────────────────────────
@ -225,9 +273,7 @@ async def _dispose_rejected_pr(conn, pr_number: int, eval_attempts: int, all_iss
async def evaluate_pr(conn, pr_number: int, tier: str = None) -> dict:
"""Evaluate a single PR. Returns result dict."""
# Check eval attempt budget before claiming
row = conn.execute(
"SELECT eval_attempts FROM prs WHERE number = ?", (pr_number,)
).fetchone()
row = conn.execute("SELECT eval_attempts FROM prs WHERE number = ?", (pr_number,)).fetchone()
eval_attempts = (row["eval_attempts"] or 0) if row else 0
if eval_attempts >= config.MAX_EVAL_ATTEMPTS:
# Terminal — hard cap reached. Close PR, tag source.
@ -294,6 +340,26 @@ async def evaluate_pr(conn, pr_number: int, tier: str = None) -> dict:
# Step 1: Triage (if not already triaged)
if tier is None:
tier = await triage_pr(diff)
# Tier overrides (claim-shape detector + random promotion)
# Order matters: claim-shape catches obvious cases, random promotion catches the rest.
# Claim-shape detector: type: claim in YAML → STANDARD minimum (Theseus)
if tier == "LIGHT" and _diff_contains_claim_type(diff):
tier = "STANDARD"
logger.info("PR #%d: claim-shape detector upgraded LIGHT → STANDARD (type: claim found)", pr_number)
db.audit(
conn, "evaluate", "claim_shape_upgrade", json.dumps({"pr": pr_number, "from": "LIGHT", "to": "STANDARD"})
)
# Random pre-merge promotion: 15% of LIGHT → STANDARD (Rio)
if tier == "LIGHT" and random.random() < config.LIGHT_PROMOTION_RATE:
tier = "STANDARD"
logger.info(
"PR #%d: random promotion LIGHT → STANDARD (%.0f%% rate)", pr_number, config.LIGHT_PROMOTION_RATE * 100
)
db.audit(conn, "evaluate", "random_promotion", json.dumps({"pr": pr_number, "from": "LIGHT", "to": "STANDARD"}))
conn.execute("UPDATE prs SET tier = ? WHERE number = ?", (tier, pr_number))
# Update last_attempt timestamp (status already set to 'reviewing' by atomic claim above)
@ -307,10 +373,18 @@ async def evaluate_pr(conn, pr_number: int, tier: str = None) -> dict:
existing_domain_verdict = existing["domain_verdict"] if existing else "pending"
_existing_leo_verdict = existing["leo_verdict"] if existing else "pending"
# Step 2: Domain review FIRST (Sonnet — high volume filter)
# Step 2: Domain review (GPT-4o via OpenRouter)
# LIGHT tier: skip entirely when LIGHT_SKIP_LLM enabled (Rhea: config flag rollback)
# Skip if already completed from a previous attempt
domain_review = None # Initialize — used later for feedback extraction (Ganymede #12)
if existing_domain_verdict not in ("pending", None):
if tier == "LIGHT" and config.LIGHT_SKIP_LLM:
domain_verdict = "skipped"
logger.info("PR #%d: LIGHT tier — skipping domain review (LIGHT_SKIP_LLM=True)", pr_number)
conn.execute(
"UPDATE prs SET domain_verdict = 'skipped', domain_model = 'none' WHERE number = ?",
(pr_number,),
)
elif existing_domain_verdict not in ("pending", None):
domain_verdict = existing_domain_verdict
logger.info("PR #%d: domain review already done (%s), skipping to Leo", pr_number, domain_verdict)
else:
@ -349,12 +423,19 @@ async def evaluate_pr(conn, pr_number: int, tier: str = None) -> dict:
WHERE number = ?""",
(json.dumps(domain_issues), pr_number),
)
db.audit(conn, "evaluate", "domain_rejected", json.dumps({"pr": pr_number, "agent": agent, "issues": domain_issues}))
db.audit(
conn, "evaluate", "domain_rejected", json.dumps({"pr": pr_number, "agent": agent, "issues": domain_issues})
)
# Disposition: check if this PR should be terminated or kept open
await _dispose_rejected_pr(conn, pr_number, eval_attempts, domain_issues)
return {"pr": pr_number, "domain_verdict": domain_verdict, "leo_verdict": "skipped", "eval_attempts": eval_attempts}
return {
"pr": pr_number,
"domain_verdict": domain_verdict,
"leo_verdict": "skipped",
"eval_attempts": eval_attempts,
}
# Step 3: Leo review (Opus — only if domain passes, skipped for LIGHT)
leo_verdict = "skipped"
@ -385,7 +466,8 @@ async def evaluate_pr(conn, pr_number: int, tier: str = None) -> dict:
conn.execute("UPDATE prs SET leo_verdict = 'skipped' WHERE number = ?", (pr_number,))
# Step 4: Determine final verdict
both_approve = (leo_verdict == "approve" or leo_verdict == "skipped") and domain_verdict == "approve"
# "skipped" counts as approve (LIGHT skips both reviews deliberately)
both_approve = leo_verdict in ("approve", "skipped") and domain_verdict in ("approve", "skipped")
if both_approve:
# Get PR author for formal approvals
@ -431,18 +513,27 @@ async def evaluate_pr(conn, pr_number: int, tier: str = None) -> dict:
conn,
"evaluate",
"changes_requested",
json.dumps({"pr": pr_number, "tier": tier, "leo": leo_verdict, "domain": domain_verdict, "issues": all_issues}),
json.dumps(
{"pr": pr_number, "tier": tier, "leo": leo_verdict, "domain": domain_verdict, "issues": all_issues}
),
)
logger.info(
"PR #%d: CHANGES REQUESTED (leo=%s, domain=%s, issues=%s)",
pr_number,
leo_verdict,
domain_verdict,
all_issues,
)
logger.info("PR #%d: CHANGES REQUESTED (leo=%s, domain=%s, issues=%s)", pr_number, leo_verdict, domain_verdict, all_issues)
# Disposition: check if this PR should be terminated or kept open
await _dispose_rejected_pr(conn, pr_number, eval_attempts, all_issues)
# Record cost (domain review on OpenRouter, Leo depends on tier)
# Record cost (only for reviews that actually ran)
from . import costs
costs.record_usage(conn, config.EVAL_DOMAIN_MODEL, "eval_domain", backend="openrouter")
if tier != "LIGHT":
if domain_verdict != "skipped":
costs.record_usage(conn, config.EVAL_DOMAIN_MODEL, "eval_domain", backend="openrouter")
if leo_verdict not in ("skipped",):
if tier == "DEEP":
costs.record_usage(conn, config.EVAL_LEO_MODEL, "eval_leo", backend="max")
else:
@ -535,7 +626,9 @@ async def evaluate_cycle(conn, max_workers=None) -> tuple[int, int]:
).fetchall()
if stagger_limit and rows:
logger.info("Post-migration stagger: limiting eval batch to %d (migrated PRs: %d)", stagger_limit, migrated_count)
logger.info(
"Post-migration stagger: limiting eval batch to %d (migrated PRs: %d)", stagger_limit, migrated_count
)
if not rows:
return 0, 0