Repository navigation
Durability comparison: celld 0.6.3 tip + publish honesty (ClawQL-only, keyed lost-ack) - #1455
Open
danielsmithdevelopment wants to merge 97 commits into
Open
danielsmithdevelopment wants to merge 97 commits into
danielsmithdevelopment wants to merge 97 commits into
Conversation
…ensity-fair-2725 Fair parked-agent density: sync latency vs probe; fdatasync/prealloc arms
…olumns Wire journal/plan entryMatchesCall into ProgramCellDO resume so a divergent entry fails closed (journal_divergence) instead of replaying. Scaffold columns.completed and columns.stepsRepeated on the fleet phased board and document progress on the durable-execution claim path — still not a publishable “durable execution” claim. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Since 06e0e6f memory_ingest is native-tool kind write (session IFC, no mandate park). Update fail-closed smoke to expect ok:true instead of the old park_destructive_write assertion. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Clear OSV leftover DoS via quadratic percent-decoding in mobile query-string graph; pin override at root and apps/mobile. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Full-stack smoke flakes when mcp-api-adapter briefly cannot reach clawql-mcp (502 upstream CallTool / fetch failed) while memory_* succeed. Retry a few times in callAdapterTool so AgentSessionDO adapter_search is race-tolerant. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
…ted-step-durability-2725 Checkpointed-step durability: ProgramCellDO digest hard-stop + board columns
Promote baseline-vs-resume step outcome equality into compareProgramStepStateEffect, wire Node durable-runner and kill-node stand-in goldens, and publish fleet columns.statePreserved scaffolding (measured: false until trials pair a no-crash baseline). Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Empty commit so PR workflows re-run without waiting on action_required approval of the autofix push. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
…h-state-preserved-2725 feat(program): claim-path state-preserved scaffolding
Add first-class plan tool sleep for claim-path §2: parse durationMs, journal sleep.<i>.wakeAt before waiting, land a replayable entry on complete, and re-arm remaining delay on resume. Model/approval-wait steps remain open. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Unblocks clawql-api DTS build after adding plan sleep steps. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
npm pack smoke failed on missing @aws-sdk/credential-provider-web-identity@^3.972.85 (transient registry gap); version is published now. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
…h-journaled-sleep-2725 feat(program): journal plan sleep with wakeAt re-arm
Add plan tool await_approval for claim-path §2: journal await_approval.<i>.executionId, park while the mandate is pending (observe-only), and land a settled entry on resume. Never consumes. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
…h-approval-wait-2725 feat(program): journal await_approval park/resume waits
Publish columns.waitApproval on the phased fleet board (measured: false), pointing at Node await_approval + density restartSurvival until fleet injects await trials. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Complete the §4 scaffolding: waitApproval object on the phased board, honesty string, and claim-path §4 status row. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
…h-wait-column-2725 feat(bench): scaffold claim-path waitApproval fleet column
Claim-path §2: first-class plan model_call lands the host response in the journal; resume replays without a second call. Mid-call crash after requestDigest with no entry re-requests once. Wire MCP host modelCall to the inference gateway and pendingStatus for await_approval. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
…h-model-journal-2725 feat(program): journal model_call with replay and one re-request
Sibling module under formal/tla/program: NoCompletedStepRerun and OutcomeUnknownHonest under MaxSteps≤3. Target must pass; weak WeakAllowCompletedRerun must counterexample. Wire CI + claim-path docs. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
…h-program-tla-2725 feat(formal): ProgramResume TLA+ dual-config for claim-path §5
Auto-wire R2 from CLOUDFLARE_ACCOUNT_ID + sync/access keys. Commit lab n=100 phased evidence with completed.rateOuterPhases=1 and writeStepsRerun=0 (dups=0). statePreserved/waitApproval remain unmeasured. Stock N=1000 safety claim unchanged. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Negative control on clawql.2 with GATE=0 and FENCED_ACK=0: every during_call hang-/health-then-kill trial lost intent. Harness reproduces the hole; fenced clawql.2 board with RUST_LOG=timing started next. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Paired efficacy: same hang-/health-then-kill during_call board. Unfenced control lost intent every trial; clawql.2 fence lost none. Durable-wait p50/p99 not publishable yet (logs truncated on restart); harness now appends celld logs when RUST_LOG includes timing. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Record callSent.sawInFlight so a gate that never releases egress cannot fake 0 lost intents. Fail those trials as call_never_sent (availability). Document TTL=3000 resume floor vs stock ~12–14s and residual-rate wording. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
A leftover FLEET_PHASED_EVIDENCE pointing at n40 caused the call-sent remeasure to overwrite clawql.1 evidence. Accept OUT first and log path. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Remeasure shows 100/100 pets inFlight before kill, 0 never-sent, 100/100 outcome_unknown with intent visible, 0 lost intents. Disclose CELLD_TTL_MS=3000 resume floor and residual-rate wording. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Add /.well-known/celld/alive that never hangs; time readiness Health RPCs to 503 instead of wedging TCP; hard-abort stuck #490 maintenance phases. Split Helm liveness to alivePath. Update Cost Plus correction copy for N=100 verified + v3 underway. Harness waitNode prefers /alive. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Omit removed CELLD_OUTPUT_GATE/FENCED_ACK envs for stock ≥0.6.3. Draft private Deno bug-report email; do not send until 0.6.3 wedge N=100 finishes. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Dual boards shared /tmp/fleet-phased-wedge-{A,B}.flag, so a wedge arm hung
a parallel stock board and startNode unlinkSync cleared active wedges.
Namespace by listen port (FLEET_PHASED_LOG_NS override). Update the private
ry@deno.com draft for DO-compat framing, tip 0.6.3 check, and format-patch
submission (PRs disabled).
Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Stock always-gate: 0/100 lost_midflight_intent, 100/100 callSent on the same wedged during_call board that lost 100/100 with 0.4.0 GATE=0. Draft leads with tip held; optional Fleet VerifyOwnership only. Do not email yet. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Wedged N=100 on 0.6.3 held (TTL=3000 disclosed). Replace bug-report posture with a short Fleet VerifyOwnership question. Restart v3 for binary alignment (claim must rest on shippable 0.6.3, not clawql.3); stale-owner probe still open. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Health-wedge alone does not stop node-lease renewal, so own.json never left the victim. Freeze the victim, wait for ownership flip, thaw, and fail on inFlight after both move and thaw. Tip N=40 board uses this path. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
…eover celld updates own.json lazily on access; freeze alone left ownership on the victim. Poll survivor program/status while frozen so the lease can move. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Stock celld 0.6.3 closed SIGSTOP→own.json→SIGCONT with 40/40 ownership moves, 0 staleSend, 0 lost intents; keep Ry draft as short question. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Product pin is denoland/celld v0.6.3. Remove clawql.* source and celld-vendor-release; install via attested release *.gz in Streams CI. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Pin CELLD_PINNED_VERSION to v0.6.3 with CI attestation on release *.gz; rewrite public/Cost Plus copy to credit tip; retarget latency to durable_wait; offer wedged+stale-owner boards upstream. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Add runtime/version, intentPresentOnResume, and callVerifiedSent columns; remeasure ClawQL arms; import tip fleet wedge into program-celld; mark 0.4.0 rows superseded; regenerate package manifest with N=50 expansion pilots. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
…shells Rerun ClawQL file-journal at N=50 to match peer pilot denominators. Split ProgramCellDO into during_call / wedged_owner / stale_owner rows. Peers share tip harnessCommit with null metrics (pending_adapter). Prefer N=50 pilots in the config-review package. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Add model-faithful golem-at-least-once-sim and pi-interrupt-sim arms (measured placeholders; real golem/pi-durable stay pending_adapter). Ship ClawQL lost-ack expansion with idempotent, lookup_by_reference, and neither modes. Refresh durable_wait partial from live v3 board. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
cursor
Bot
changed the title
cursor
Bot
changed the base branch from
cursor/celld-alive-ready-lease-sweep-2725
to
cursor/program-density-fair-2725
October 31, 2026 18:20
Pass programId ctx to host.execute; narrow flattenArmResults for tsc. Surface model-sim and lost_ack rows on the www durability table. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Streams celld evidence CI still expected ghcr.io/denoland/celld:v0.4.0 after the tip pin moved to stock 0.6.3. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Surface finalized comparison N=50 rows (sims, lost-ack, stale-owner), correct tip gate caveat, and tip artifact links. Tier C claim stays on published phased-evidence until stock-TTL v3 greens, then swap. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Durability section meta line and artifact footer cite celld 0.6.3 tip package identity plus wedge/stale/durable_wait evidence paths. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Pull peer sims out of the default comparison board and /benchmarks table (opt-in harness only). Remasure N=50 with lost-ack keyed resend so idempotency keys resolve to one unique effect instead of parking unknown. Add mid-flight claim caveat, phase availability notes, Executor/Panguard fairness credits, equal-arm headline, and tamper-evident WORM wording. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
cursor
Bot
changed the title
…ver_sent Live Golem/Pi N=50 existed on PR #1432 (DURABILITY_COMPARISON_LIVE_PEERS=1); archive that evidence and reference-pr1432 adapters for schema-v2 rewire. Restore Tier F resume-vs-journal (remeasured ≤500 steps, snapshots not needed). Document that during_call never-sent waits the full 8s inFlight poll — gate latency alone cannot explain the rise; leave v3 classification unchanged. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
Stop once failures exceeded the frozen 2% budget (44/500; abs ceiling 20/1000). Correctness stayed 0 on completed write-kill trials; await arm not run on a degrading fleet. Bucket shows 712 accumulated node registrations; never-sent trials never reached the durable-ack gate. v4 blocked on lease-sweep + alive-vs-ready soak under the same frozen criteria. Co-authored-by: Daniel Smith <danielsmithdevelopment@gmail.com>
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Tip schema v2 durability comparison on stock celld 0.6.3, publish honesty, and v3 N=1000 stock-TTL result: availability FAILED (stopped early).
v3 board (stopped)
docs/benchmarks/program-cell-fleet-kill/phased-n1000-v3-correctness-availability.json(status=failed)nodes/fleet_*.json(356/side); 68 rebuild eventsPublish honesty (prior)
/benchmarkstable; peers in reviewNext (not this PR’s claim path)
Test plan