Skip to content

review: assess bounded local-footprint admission under Janitor pressure #2458

Closed

Description

@jerryfane

Assignment: Jarvis messages376223/376245; consumer owner gitmoot, producer-contract owner workspace-janitor (376230/376239). This is separate from #2457 (queued remote CI deadlock) and #2456 (message policy). Current deployed/base commit e6f39cb checks Janitor before every new review allocates its read-only worktree, including explicit remote E2B. Remote execution still has local preparation and retained data.

Assess the existing producer paths and reservation primitives against /root/fleet-tools/state/disk-containment-20261031/review-admission-contract-376224.json. Required proposal: fresh matching policy/device/mount evidence and immediate available bytes; full enforced incremental physical-allocation bound B; conservatively charged outstanding budgets C; atomic cross-process measurement/admission/durable reservation before any allocation; A=min(snapshot,immediate)>canonical reserve+C+B on every filesystem. Unknown/unbounded producers and external commitments are not zero. Crash/restart retains charges until producer termination and retained-artifact accounting are proved. Include DB/WAL, git objects, staging/snapshots, runtime/toolchain caches, logs and artifacts.

Deliver source-bound inventory, enforced limits versus unknowns, existing primitives versus missing guarantees, and one smallest supported activation proposal with exact policy/security risks for any required owner decision. No live policy, reserve, global latch, security guard or spend changes are authorized by this assessment. Preserve existing ordinary/unbounded gates and queued-job identities. No numeric review budget is established by instantaneous free space.

Before any authorized activation: concurrent admissions; stale/policy/device/mount mismatch refusal; boundary arithmetic; failed allocation cleanup; crash/restart retention; all local-stage/runtime physical caps; retained-artifact transfer; actual bounded review end-to-end. PR67 jerryfane/geelark-agent head c379adbbdd55e3b3069a1c61a1d50e01fb7d26ad is a reported prequeue-blocked consumer, not authorization to dispatch or bypass its hold.

Activity

  1. jerryfane commented on Oct 31, 2026

    @jerryfane
    CollaboratorAuthor

    Additional reported affected exact-head consumers (Enyo376264, not independently re-read GitHub): jerryfane/data-core#27 head21233f58ab741a754bab51f6a6a222e5583a7368; #28 head7607df8e9c1815292a8c3205b5fd701c59214360. No new jobs, paid fallback or guard changes initiated. These are the same prequeue resource boundary, separate from #2457 queued-review CI classification.

  2. jerryfane commented on Oct 31, 2026

    @jerryfane
    CollaboratorAuthor

    Jarvis376287 adds reported affected review gitmoot/sandboxd#79 exact446ac8326b18a521a1c7cd9795168313c0150350: level3 dispatch refused before job creation; blocks Forge#4 real Firecracker/HerdrUp acceptance. Sandboxd retains review/integration ownership. Janitor376289 explicitly limits this assignment to finishing the local-write inventory and proposing smallest genuinely enforceable versus explicitly risk-accepted policy; do not implement a broad byte-ledger/storage architecture or silently weaken the invariant. Current holds remain unchanged.

  3. jerryfane commented on Oct 31, 2026

    @jerryfane
    CollaboratorAuthor

    Source assessment at e6f39cb: explicit remote review still allocates a full detached local checkout before enqueue, can fetch shared Git objects, creates temporary history.git during SyncIn, and retains local transcripts/results/SQLite WAL. Archives stream through io.Pipe; normal review does not invoke changeset Collect (implement-only), and E2B OMP skips Local-only toolcache staging. Those avoided allocations are not zero total local footprint.

    No existing atomic physical-byte reservation or enforced whole-route bound B was found. Existing execbackend_attempts uses guarded SQL for provider dollars/counts, not bytes. store.go:70-76 explicitly says64MiB journal limit does not bound peak WAL with pinned readers. Cleanup/schema-v2 budgets authorize deletion, not creation. Cross-producer outstanding commitments C are unknown. Full private inventories: /root/fleet-tools/state/gitmoot-review-admission-deadlock-20261031/bounded-review-local-writes.txt and bounded-review-reservations.txt.

    Disposition: current supported guaranteed route remains ordinary admission after producer clearance. A reserve-aware remote-only exception would be best-effort risk acceptance, not satisfaction of the proposed physical B/C guarantee; unbounded shared writers can cross reserve between measurements. No numeric allowance, safe activation, or new quota architecture is claimed. Janitor376294 and Jarvis376295 received exact gaps; awaiting whether an existing producer-side quota mechanism can close them or an explicit changed-guarantee decision is needed. Existing gates remain unchanged. No dispatch/spend.

  4. jerryfane commented on Oct 31, 2026

    @jerryfane
    CollaboratorAuthor

    Owner decision relayed by Jarvis376302 from completed interactive question: ALLOW MONITORED REVIEWS. Explicitly best-effort, not a hard physical-byte guarantee; one-at-a-time review-only exception with fresh disk checks and monitoring, no new paid capacity. Implement supported route before use; preserve canonical reserve, cleanup latch, ordinary/unbounded gate, exact-head review/security checks and queued identities. Monitor failures/low reserve must close admission and cancel only the owned review; retain slot through termination/cleanup. No blind automatic retry. Activation after applicable checks/review is within this assignment; no further phase approval. Janitor376307 owns matching monitor contract. No broad reservation framework.

    Owner also asks to move required bulk E2B preparation to Storage Box. New worktrees already use verified Storage Box; assess/move remaining history staging while leaving SQLite/WAL local. Placement is not a guarantee against concurrent shared writes.

    Consumer metadata update (Enyo376296): EFM PR264 now headfe49c0cd2e146f30beb64bfecfceb4f362baeb12 supersedes6df5e869. Prior-head checks/reviews do not transfer.

  5. jerryfane commented on Oct 31, 2026

    @jerryfane
    CollaboratorAuthor

    Superseding direct owner instruction: reject the proposed one-review-at-a-time restriction; fix the local blocker for remote reviews. Direction: choose E2B before local preparation, prepare exact head/base remotely through existing scoped access, and do not apply the host cleanup-target gate to that remote path. Retain existing provider concurrency/budget, exact-head/security checks, and ordinary local-job admission. No new single-slot restriction or general reservation system. Previous one-at-a-time proposal is superseded, not activated. No new spending authorized.

  6. jerryfane commented on Oct 31, 2026

    @jerryfane
    CollaboratorAuthor

    Tracked as steps1–3,5–6 of highest-priority owner epic #2459. One-at-a-time/monitored-slot proposal is superseded. Remove local preparation/admission dependency for actual E2B reviews while retaining existing provider concurrency/spend and exact-head/security controls.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions