mirror of
https://github.com/razzant/ouroboros.git
synced 2026-08-04 16:19:50 +00:00
ae53779ccd
Owner-approved big release (plan: большой_релиз_v6.26.0). Workstreams:
- WS-A lint gate: ruff F-rules step in CI quick-test; fixed the real F823
NameError class (supervisor/events.py utc_now_iso shadowing); make lint/health.
- WS-B memory integrity: atomic dialogue_blocks with corrupt-quarantine
(.corrupt-<ts>.bak + memory_store_corrupt event); honest scratchpad journal
(block_append_failed, corrupt storage renders as corruption not "(empty)");
full Pattern Register window (16K cap vs 3K cut); merge-aware scratchpad
consolidation under the sidecar lock; backlog writes through locked helpers;
chat omission notes ("[N older unconsolidated messages omitted]").
- WS-C provider SSOT: provider registry (prefixes/credentials/resolution) in
provider_models.py replacing 5 duplicated knowledge sites; credential-aware
consolidation model; pricing empty-fetch retry; direct-route correctness
(o-series max_completion_tokens, reasoning_effort, Anthropic error bodies,
per-request timeouts on cached clients); deep/plan review budgets reserve
output headroom inside the 1M window (min(SSOT, window − output − margin)).
- WS-D races/state: utils.update_json_locked (locked RMW, loud TimeoutError);
task_results merge under per-file lock (cancel-latch safe); update_state
migration for owner binding, evolution counters, budget_line, post-task
activation; queue-lock coverage (enqueue, timeouts, worker health, task_done,
snapshot); visible supervisor death (supervisor_error + owner notice + ghost
consciousness/chat-agent cleanup on re-init); chat.jsonl rotation via
os.replace under the append lock; unbounded _outbox removed; host-service
port probe + actual-port panic sweep; skill lifecycle lane deadline
(OUROBOROS_SKILL_LIFECYCLE_TIMEOUT_SEC, dedupe-leak-safe); consilium
force-plan structured flag (plan_review_aggregate from the FULL result);
atomic writers (update-intent, repo-manifest, metrics cache, post-task pair);
review_state.save_state raises on lock timeout (advisory ledger honesty);
apply_pending_request no-ops while a campaign is active;
api_update_apply kills workers only AFTER update validation, with respawn
on aborted checkout and "interrupted" terminal status.
- WS-E immune hardening: triad anti-refusal coverage contract (empty array
needs NO_FINDINGS sentinel or bare-[] body; refusal prose with [] is
parse_failure and never enters quorum); BIBLE P3 "Owner-chosen enforcement,
loud advisory" bound + CHECKLISTS sync; every advisory pass-through of a
blocking signal writes review_advisory_override + persistent
advisory_overrides (surfaced by review_status); ToolEntry.mutates_worktree
with dispatcher-level before/after worktree diff invalidation (covers error
paths; read-only runs no longer invalidate; redundant manual calls removed);
scope review fails closed without its checklist; advisory final write is a
locked re-read merge; synthesis severity defaults to critical with WARNING
fallbacks; find removed from SAFE_SHELL_COMMANDS; anti-thrashing state
survives advisory criticals; preflight env scrubs secret-class variables;
.git/index.lock age gate in worker startup checks.
- WS-F security: file-browser symlink containment on the RESOLVED path across
all endpoints (out-of-root symlink targets are listed but inert; tests
rewritten to the new contract deliberately); HMAC-signed session cookies
(server-side persisted key, 30-day TTL, Secure on TLS) replacing the
permanent password-derived cookie; password-class settings mask to a
constant placeholder; conservative SSRF guard for the MAIN agent (link-local
/cloud-metadata only, LAN stays reachable, per-request route re-validation);
ClawHub zip-slip hardening (":"/backslash segments rejected + post-join
containment) and lazy no-proxy OuroborosHub opener; single-execution
signature dispatch for extension handlers (no TypeError re-run after side
effects); onboarding postMessage origin checks; SHA256-pinned
python-standalone download with pipefail; payload-resident dependency
fingerprints only corroborate durable deps.json; skill payload re-hash
immediately before spawn (TOCTOU narrowing).
- WS-G process custody: ouroboros/process_custody.py — spawn_supervised
chokepoint + durable data/state/process_ledger.jsonl (pid, pgid,
fingerprint{start_time, cmd_sha256}, purpose, scope task|session|daemon,
owner_task, session_id); platform_layer.process_start_time primitive;
startup + periodic reaper killing ONLY strict-fingerprint matches from dead
generations/tasks (never by command-line class); migrations: services
(the orphan hole), workspace executor + local model + extension companions
(ledger write-through), worker_pids (write-through; legacy path retained);
parent lifelines (ppid watchdog, group-suicide only as group leader) in
worker_main, extension runner, Claude readonly child; conformance test
pinning the Popen allowlist; ARCHITECTURE/DEVELOPMENT/CHECKLISTS entries.
- WS-H native multimodal chat: supports_vision capability map (static
prefixes + OpenRouter /models input_modalities overlay); web chat uploads
ride the WS frame as structured attachments (additive ChatInbound field)
and image uploads become NATIVE image blocks via the existing Path B;
browser screenshots inject natively for vision models via the multipart
user-merge (tool result stays a string; file persisted under
data/uploads/screenshots for re-view); K=3 newest-image eviction with
caption placeholders carrying the vlm_query re-view path; image-aware token
estimates (fixed ~1.1K-token equivalent instead of base64 length, fixing
permanent emergency-compaction wedges); compaction renders images as
captions (no base64 into the summarizer); GigaChat/local lanes emit explicit
"[image omitted: model has no vision]"; internal _caption/_source_path
metadata stripped from provider payloads.
- WS-I housekeeping (partial): SETTLED_STATUSES SSOT (+ cycle-safe mirror pin);
owner_inject.py renamed to owner_mailbox.py; version-neutral envelope
wording; files.py import-block cleanup. Remaining WS-I/WS-J/WS-K items are
deferred with the owner's context-budget priority on review+release.
Review notes: triad+scope ran via scripts/run_external_review.py on the core
pack across 3 rounds to convergence (scope responded=PASS each round; round-2
criticals fixed: update_apply kill-order + respawn, lifecycle dedupe leak on
lane timeout, OUROBOROS_MAX_ROUNDS hot-reload + docs, toggle_evolution
NameError, async Anthropic timeout forwarding, supports_vision local check,
budget-update lock visibility, ChatInbound additive attachment contract,
file-browser doc sync). The FULL combined diff exceeds every triad model's
context window (~1.59M tokens > 1.05M) — reviewed in packs; remaining
cross-pack findings were verified as slicing artifacts. Adversarial critics
(GPT; Gemini/Opus rounds) ran on the working tree. Deliberate tradeoffs:
metadata-based eviction captions (no light-LLM call in the hot path);
worker ledger records use live-cmdline fingerprints with a synthetic-arg
fallback only where the OS offers no cmdline.
|
||
|---|---|---|
| .. | ||
| common | ||
| osworld | ||
| programbench | ||
| swe_bench | ||
| swe_bench_pro | ||
| terminal_bench | ||
| __init__.py | ||
| evolve_smoke.py | ||
| README.md | ||
Benchmark Devtools
This directory contains thin adapters around official benchmark harnesses. The adapters prepare Ouroboros tasks, capture artifacts, and preserve traces; they do not replace official scoring.
Supported surfaces:
- ProgramBench: official
programbench eval/infoand cleanroom submission layout. - Terminal-Bench: Harbor installed-agent adapter.
- SWE-bench and SWE-bench Verified: standard predictions JSONL for the official SWE-bench harness.
- SWE-bench Pro: official Pro eval wrapper and patch-capture methodology. The prediction/capture path is ledgered; the grader wrapper stays official-output-only.
- OSWorld: stop-closed skeleton and logs normalizer only.
Shared sidecar schemas:
- Run manifests record non-secret provenance: requested task ids where the
benchmark runner exposes them before execution, requested counts/selection
slots for deterministic first-N runs such as Terminal-Bench, exact argv,
official command shape, output paths, model slots, source commit, dirty-state
counts, and hashes. Defaults are adapter-specific (
run_manifest.json,<predictions>.run_manifest.json, orosworld_preflight.run_manifest.json). - Result ledgers are denominator-preserving Ouroboros JSONL files. They record
every requested instance, including setup failures, timeouts, and empty
patches, even when the official benchmark prediction/submission format only
accepts successful rows. Defaults are adapter-specific (
result_index.jsonl,<predictions>.ledger.jsonl, orosworld_preflight.ledger.jsonl).
These sidecars are audit artifacts, not replacement scoring. Official benchmark harnesses and official result files remain the scoring authority.