docs: record container-only retry and crash acceptance

Co-authored-by: GPT-6 <gpt-6@openai.com>
This commit is contained in:
ZeusinaandGPT-6 Bot committed 2026-10-11 10:04:05 +03:00
1 parent e9b70b00b6
commit 11f9197ba8
1 file changed
+5 -4
+5 -4
View File
@@ -4,9 +4,10 @@
## Текущий статус ## Текущий статус
- **2026-10-11 container-owned control (pending working-tree changes):** all technical child restart control belongs to the entrypoint/PID 1 inside the same container, not host systemd. Safe cleanup precedes each retry; three runtime starts per rolling six hours include the initial start, with cancellable 15s/30s backoff capped at 60s. Configuration/safety/DB/journal/rollback/cleanup failures and exhausted retry budget stop fail-closed. SIGTERM cancels and joins the monitor, restores safely, then stops target-local zapret and SS children. No additional dependency or host service installation is required. Compose keeps `restart: "no"` to avoid safety loops and now allows 240s stop grace: 140s sequential cleanup including 45s monitor join, plus room for a deferred 45s startup command/kill grace and overhead. A fresh-shell join watchdog is cancelled promptly on child completion; graceful shutdown never needs its full timeout. Container-platform lifecycle and host-boot policy remain deployment-operator responsibilities. - **2026-10-11 container-owned control, committed/deployed:** `5258ae2`, `6467108`, `fe23eec` replace external supervision with PID1-owned runtime-cycle retries (three starts per rolling six hours, cancellable 15s/30s backoff capped at 60s). Guard/config/DB/rollback/cleanup failures remain fail-closed. Reconciliation precedes activation; fresh-shell watchdogs avoid EXIT/TERM trap deadlock; private monitor sessions allow cleanup of orphaned scan workers before restart. SS uses private JSON, never exported password/argv. Graceful SIGTERM joins/restores before stopping owned daemons; Compose has no restart policy and 240s grace, monitor join 45s. Host/container-platform lifecycle belongs to the operator, not a generated host manager. Historical transient acceptance units were unloaded and the tracked systemd unit removed.
- **2026-10-11 strategy absence (pending working-tree changes):** genuine validated no-qualified-strategy outcomes in bootstrap/recovery persist a wait of at least two hours in SQLite. Both fresh bootstrap and periodic recovery use the greater of two hours and the configured recovery cooldown, preserving longer settings. The same container continues light probes without scans/apply during the wait; startup respects the persisted expiry, and normal threshold-driven recovery resumes after expiry. Other ordinary outcomes keep the default 30m cooldown. Malformed reports/runner/safety errors are not absence. The separate 90-minute whole-job watchdog plus 30-second restoration reserve remains an abnormal technical ceiling, not a normal no-strategy result or the absence wait. - **2026-10-11 strategy absence:** genuine validated no-qualified-strategy results persist `max(2h, configured cooldown)` in SQLite, preserve longer existing deadlines and other leases, and survive monitor/container startup. Lightweight probes continue; heavy saved-first/scan/apply waits until expiry. Other ordinary outcomes retain 30m cooldown. Malformed reports and explicit nfqws2 infrastructure failures are vetoed before both positive and negative selection; they are never disguised as absence. Qualification remains both TLS protocols 3/3. The separate 90-minute whole-job watchdog plus bounded restoration remains abnormal, not normal absence.
- Root operator docs now describe this contract, preserving immutable tagged/digest promotion, BuildKit/pinned builds, private `.env`, whole stopped-state backup/restore and exact legacy `managed-test` guards for a new isolated target only. No old target adoption or Xray changes. New runtime changes still need their owners' focused checks and isolated live acceptance (retry/backoff, SIGTERM, durable absence expiry/resume); this documentation update does not claim those checks were run. - **Verification:** full Go suite, targeted race tests and shell/static/Python fixtures pass. Negative-only checker fault injection demonstrated a persisted 7200-second pause, unchanged checker count 7→7 through internal monitor restart, same container/PID1, DockerRestartCount0, zero candidates and graceful exit0. Active-check crash fixture demonstrated old separate-PGID workers gone before replacement activation, unowned session peer preserved, direct unchanged PID1/container and exit0. Private reports: `/home/agent/container-absence-acceptance-roqcpxq1/actualreport.json`, `/home/agent/active-scan-crash-5z2a6xb_/actual-report.json`. These validate control flow, not new network qualification; real two-hour wall-clock expiry was not waited out (expiry covered by injected-clock tests). Numeric PID signalling retains the usual read/signal PID-reuse race; strict reuse-proof signalling would require pidfds.
- **Working target:** immutable image `sha256:e28b3340126891686e7f3cf1fce6a651b37259fc3bd51be225b8864124a7c103` (`rc-fe23eec`), private `.env` selector and password preservation verified. Generation35, active[1,18], both groups committed, Player API OK, queue300 only. Main container `e31475d0f2a8528c54089097d2e0129bc29a07f3c9a9e8d15ac6966a573166ce`/host PID518298/start time unchanged during real child SIGKILL; monitor PID184→546, DockerRestartCount0. Old container PID266922/start time unchanged; Xray untouched. Current README/operator/deployment guides describe only container-owned supervision and retry policy; this remains an isolated production candidate, not a blanket playback/production-readiness claim.
## История реализации (не текущие инструкции) ## История реализации (не текущие инструкции)
@@ -15,7 +16,7 @@ installation/restart/start-limit guidance and its acceptance tests are **superse
by container-owned control above; they impose no current deployment requirement. by container-owned control above; they impose no current deployment requirement.
Earlier “not deployed/not verified” snapshots do not override later recorded Earlier “not deployed/not verified” snapshots do not override later recorded
isolated-target acceptance. Production acceptance of the new supervision/absence isolated-target acceptance. Production acceptance of the new supervision/absence
policy remains pending. policy is recorded above; historical pending labels are superseded.
- **2026-10-11 production-candidate update (remote checks dated 2026-10-10 UTC):** `ed4298d` adds one 90-minute budget per managed bootstrap/recovery job, shared across phases. Watchdog expiration is abnormal/fatal, never ordinary nonpass/cooldown continuation; clean timeout permits technical exit 75 only after worker join and bounded 30-second restoration. Any guard/DB/rollback/cleanup or mixed unknown error vetoes restart (78); configuration errors use 64, graceful shutdown 0. Independent review reproduced and closed mixed-timeout, fatal-discovery and fatal-health/deadline races. Preferred frontend seed is now the layout-safe alternative verified in prior 60/60 requests; cached definitions are not rewritten. - **2026-10-11 production-candidate update (remote checks dated 2026-10-10 UTC):** `ed4298d` adds one 90-minute budget per managed bootstrap/recovery job, shared across phases. Watchdog expiration is abnormal/fatal, never ordinary nonpass/cooldown continuation; clean timeout permits technical exit 75 only after worker join and bounded 30-second restoration. Any guard/DB/rollback/cleanup or mixed unknown error vetoes restart (78); configuration errors use 64, graceful shutdown 0. Independent review reproduced and closed mixed-timeout, fatal-discovery and fatal-health/deadline races. Preferred frontend seed is now the layout-safe alternative verified in prior 60/60 requests; cached definitions are not rewritten.
- **2026-10-10 restart/startup hardening:** optional `deploy/managed-target/zapret-autotune.service` uses foreground Compose, no blanket Docker restart, systemd allowlisted 75 with bounded starts (3 per 6h); shutdown grace 120s / unit stop budget 180s. `61e6e00` adds guarded fixed-DB `managed reconcile` before initial activation; init/reconcile/start are bounded (45/45/30s plus 5s kill grace), all startup refusals map to 78. Unknown state is not sourced even by cleanup. Actual systemd tests: technical 75 retries bounded, 78 does not retry; killing monitor in final image caused one whole-target restart, preserving both groups and Player API OK. Permanent production service was not installed/enabled; transient acceptance units were stopped. - **2026-10-10 restart/startup hardening:** optional `deploy/managed-target/zapret-autotune.service` uses foreground Compose, no blanket Docker restart, systemd allowlisted 75 with bounded starts (3 per 6h); shutdown grace 120s / unit stop budget 180s. `61e6e00` adds guarded fixed-DB `managed reconcile` before initial activation; init/reconcile/start are bounded (45/45/30s plus 5s kill grace), all startup refusals map to 78. Unknown state is not sourced even by cleanup. Actual systemd tests: technical 75 retries bounded, 78 does not retry; killing monitor in final image caused one whole-target restart, preserving both groups and Player API OK. Permanent production service was not installed/enabled; transient acceptance units were stopped.