fix: added additional constraints around parallelization bugs in verification tasks

This commit is contained in:
Dark-Alex-17 committed 2026-09-16 00:36:33 -06:00
1 parent 4cc5563ba8
commit 9fb90be851
5 files changed
+34 -5

No files matched your search

+3 -1
View File
@@ -37,7 +37,9 @@ variables:
description: | description: |
JSON array of shell commands the CALLER declares for the deterministic JSON array of shell commands the CALLER declares for the deterministic
run_checks stage — e.g. '["cargo test --all"]'. Never inferred from the run_checks stage — e.g. '["cargo test --all"]'. Never inferred from the
prompt; empty array = none declared. prompt; empty array = none declared. Must not boot shared stacks or
services (compose projects, fixed ports) — those collide with the probe
lane running in parallel.
default: '[]' default: '[]'
settings: settings:
+4 -1
View File
@@ -342,7 +342,10 @@ instructions: |
here is executed with a shell — a trust boundary the prose must not cross. here is executed with a shell — a trust boundary the prose must not cross.
Budget: each declared command gets at most 15 minutes and the whole set 55 minutes — Budget: each declared command gets at most 15 minutes and the whole set 55 minutes —
declare targeted build/unit-test commands, not full integration suites; a command declare targeted build/unit-test commands, not full integration suites; a command
that exceeds the budget is recorded as a red run and fails conformance. that exceeds the budget is recorded as a red run and fails conformance. Never declare
commands that boot shared stacks or services (docker compose projects, dev servers,
fixed ports) — the probe lane owns end-to-end runs and WILL collide with them in
parallel; keep verification to generate/lint/build/unit/integration commands.
It returns `GAUNTLET: PASS/BLOCKED`. On BLOCKED with real review findings apply the same It returns `GAUNTLET: PASS/BLOCKED`. On BLOCKED with real review findings apply the same
fix-and-rerun rules as below (resume the SAME Sisyphus session with the findings verbatim, fix-and-rerun rules as below (resume the SAME Sisyphus session with the findings verbatim,
+4 -2
View File
@@ -115,8 +115,10 @@ nodes:
the change (absolute path preferred); empty string when the caller the change (absolute path preferred); empty string when the caller
named none. named none.
- `plan_context`: the plan / spec / acceptance-criteria text the - `plan_context`: the plan / spec / acceptance-criteria text the
implementation must satisfy, verbatim (or a file path the caller implementation must satisfy, copied through VERBATIM — never
pointed at). Empty string when the caller supplied none. summarize it and never replace pasted text with file paths or a
sentence saying it was included; the adversary lane fails closed on
whatever you drop. Empty string when the caller supplied none.
- `rigor`: poc | prototype | production. Default "production" when - `rigor`: poc | prototype | production. Default "production" when
not stated. not stated.
- `security_posture`: prototype | standard | hardened. Default - `security_posture`: prototype | standard | hardened. Default
@@ -38,6 +38,25 @@ def git(*args):
out = {"project_dir": proj} out = {"project_dir": proj}
# Deterministic plan_context recovery: the parse LLM proved a lossy channel
# for verbatim spec text (observed replacing ~4 KB of pasted criteria with
# file paths + a summary sentence). When the raw prompt carries the callers'
# labeled plan block, cut it out verbatim and override the extraction — but
# only when the recovered block is LONGER, so a good extraction is never
# degraded. The adversary lane fails closed on whatever goes missing here.
_prompt = state.get("initial_prompt")
if isinstance(_prompt, str) and _prompt:
_m = re.search(
r"Plan / acceptance criteria:\s*(.+?)"
r"(?=\n\s*(?:Local-run recipe|Forced lanes|Rigor:|Security posture:|Extra context)\b|\Z)",
_prompt,
re.S | re.I,
)
if _m:
_block = _m.group(1).strip()
if len(_block) > len((state.get("plan_context") or "").strip()):
out["plan_context"] = _block
# retry_gate measures its wall-clock budget from this stamp; it is set # retry_gate measures its wall-clock budget from this stamp; it is set
# before the git work so a failed diff still starts the clock. # before the git work so a failed diff still starts the clock.
if not state.get("gauntlet_started_at"): if not state.get("gauntlet_started_at"):
+4 -1
View File
@@ -324,7 +324,10 @@ instructions: |
executed with a shell — a trust boundary the prose must not cross. executed with a shell — a trust boundary the prose must not cross.
Budget: each declared command gets at most 15 minutes and the whole set 55 minutes — Budget: each declared command gets at most 15 minutes and the whole set 55 minutes —
declare targeted build/unit-test commands, not full integration suites; a command that declare targeted build/unit-test commands, not full integration suites; a command that
exceeds the budget is recorded as a red run and fails conformance. exceeds the budget is recorded as a red run and fails conformance. Never declare commands
that boot shared stacks or services (docker compose projects, dev servers, fixed ports) —
the probe lane owns end-to-end runs and WILL collide with them in parallel; keep
verification to generate/lint/build/unit/integration commands.
It computes diff signals deterministically, selects lanes (code-review always; adversary when a It computes diff signals deterministically, selects lanes (code-review always; adversary when a
spec/plan is present; security on attack-surface signals; probe on consumer-facing surface with a spec/plan is present; security on attack-surface signals; probe on consumer-facing surface with a