You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
connector-host: align completion generation budget with Nexus-new loop (5) and track TEST_IMPACT revalidation (Nexus-new#1384) #511
Host-side follow-up for James3014/Nexus-new#1384 (completion liveness under unrelated main churn), on the nexus/connector-host-v1 line.
Gap 1: generation budget (fixed by the linked PR)
Nexus-new nexus/orchestrator/github_completion_loop.py sets MAX_INTEGRATION_GENERATIONS = DEFAULT_MAX_INTEGRATION_GENERATIONS = 5. The host caps it at 3 in four places:
src/github-completion-host.tsparseGitHubCompletionToolInput default 3 and max 3
CompletionHostPort.materialize and readRequiredCheckspositiveInt(generation, 3)
src/nexus-gateway-proxy.ts schema maxGenerations ... .max(3) and the "at most 3" description
Gap 2: revalidate_affected_dimension always fails closed (design question, not fixed)
The host returns passed:false, requires_fresh_candidate_acceptance:true for every dimension, so TEST_IMPACT always ends with FRESH_CANDIDATE_ACCEPTANCE_REQUIRED (#1384 AC4 is not live). Running TEST_IMPACT on the exact integration head needs contract decisions on the Nexus-new side first. See the PR body for the questions.
Separate open item, found during review of #512 and not fixed there.
Absolute vs. relative generation counting mismatch. The host port guards (parseCompletionGeneration, used in materialize_integration_head and read_required_checks) cap the absolute integration generation number (current_generation + 1) at MAX_COMPLETION_INTEGRATION_GENERATIONS (now 5). The Nexus-new loop (run_github_completion_loop) budgets relative to start_generation (current_generation - start_generation >= max_generations), so a resumed run can legitimately produce I6 and later.
Effect: when a PR is already at I3 and a new loop run starts, the host rejects I6 with generation_INVALID after two more generations, although the loop allows five. This failure is closed (a reject, not a wrong merge), but it limits liveness for resumed completions.
Possible fix (needs a decision): pass start_generation from the bridge and cap generation - start_generation, or drop the absolute cap and rely on the loop's relative budget. This stays open independently of the TEST_IMPACT revalidation design question.
Classification: KEEP_OPEN / source fix integrated, end-to-end consumer outcome unverified. Target for these changes is nexus/connector-host-v1, not DevSpace default main.
Verified merged lineage on the specialized branch:
#512 → cf62a6b0f627c754e00a1ebd608309d06ff910fe: per-invocation generation ceiling aligned to 5.
#513 → 8f1fe2545c72ca00bfeb1670a72a2c51208f002a: read_head_check_runs port for Nexus-new #2068.
#514 → 974e8876f7c8dc9818f99f67c8a29a8bd06443f9: macOS-only Smoke matrix; final GitHub CI run 38480566759 success. Actions pins remain immutable.
#515 → 532c0e4e64a9a9e10c2af0b4e668a124c5ad3e78 (exact tested PR head 4c4993014d0fc4dd0cd2d6817edd9f987e315192): correct absolute-vs-relative generation mismatch. RED macOS run 38492574701 fails exactly at I6/generation_INVALID; GREEN macOS run 38492817361 success, including typecheck/test/build/doctor. Independent read-only Agy Gemini review agt_8880d6d2 terminal RECOMMEND_MERGE, recorded in PR comment; not a Nexus v4 independent acceptance receipt. Port now allows positive safe absolute I6+ while maxGenerations remains capped at five per call.
Nexus-new #2068 is merged at 00ccc35cb7c26b334e0f12b23b02eb91ad038fe7, implementing the exact-head TEST_IMPACT path. No live #1384 AC11 pilot is inferred.
Remaining evidence gates, not passed or waived:
Runtime/host activation of the exact specialized DevSpace connector source and independently proving its current caller projection, not merely branch merge.
Specialized branch does not have .nexus-core/config.toml; Nexus Core check runs on fix(#511): allow absolute completion generations after I5 on resume #515 remain failure. Mac Smoke SUCCESS is not Core SUCCESS. No Core validation, governance approval or deployment should be inferred from the specialized branch lacking required checks.
Nexus-new #1384 AC11 requires a truly authorized live single-caller Completion Loop while multiple legitimate unrelated protected merges occur during CI, zero full Candidate re-acceptances, exact-head CAS merge and post-merge readback; NOT_RUN/NOT_PROVEN. The canonical Nexus Gateway is not at fresh current main. A fresh Gateway recovery authorization, if necessary, belongs to test(cutover): repair cutover-progression fixture and chain cutover tests into npm test (#517) #526's governed owner, not an ad-hoc restart.
Next safe gate: obtain a fresh authorization/readiness contract for the canonical Mac connector/Gateway generation, then bind a valid independently accepted pilot Candidate, re-run exact-head checks, and execute actual #1384 AC11. Keep this Issue OPEN until the live host-side outcome is verified. No force merge, bypass, synthetic concurrency or Gateway recovery effect is authorized by this comment.
Runtime readback 2026-10-15 (bundle-G-gateway, Mac mini, read-only)
Source alignment is complete on nexus/connector-host-v1 (head 532c0e4e, #512/#513/#515). The running public connector host has not picked any of it up:
Check
Observed
LaunchAgent
com.nexus.mcp.public.connector, PID started 2026-10-14 08:00 local
maxGenerations: z.number().int().min(1).max(3) (still 3, not 5)
Source checkout
.git points at ~/workspace/devspace/.git/worktrees/devspace-connector-host-v1, which no longer exists, so the build has no recoverable revision
Python bridge
NEXUS_CANONICAL_SOURCE_ROOT=~/workspace/Nexus-new, NEXUS_PYTHON_BIN=~/workspace/Nexus-new/.venv/bin/python (the shared main clone, which lags main; it also sets no NEXUS_SELF_HOSTED_CANONICAL_STATE_DIR, see Nexus-new#1539)
So remaining gate 1 ("runtime/host activation of the exact specialized connector source") is not met: activating it means rebuilding dist/ from 532c0e4e in a fresh tree and restarting the public OAuth connector. That is a live-service change outside an auto-deployer, so it waits for an Owner go-ahead; nothing was changed here. Nexus-new#1384 AC11 stays NOT_RUN.
Public connector redeployed on the Mac mini, 2026-10-16 07:10 +08 (Owner approved in chat 2026-10-16).
Before: build_id nexus-1.0.1-nexus.7-528dc505 (dist from 2026-09-25), /healthzok:false, manifest_status: drifted, pid 27167 since 10/14.
After: build_id nexus-1.0.1-nexus.7-532c0e4e built from nexus/connector-host-v1@532c0e4e (npm ci, typecheck, 14 test files, build all green), /healthzok:true, manifest_status: verified, 45 effective tools, proxy to gateway 127.0.0.1:8766.
Only com.nexus.mcp.public.connector was restarted (launchctl kickstart -k). First attempt rolled back automatically because generated/build-identity.json was not swapped, so /healthz kept reporting 528dc50; second attempt swapped dist/, generated/, node_modules/ and the package files together.
Backup of the previous dist, generated, node_modules, package files and plist: ~/.local/state/nexus-connector-backup-20261015-2354/.
So #512/#515 (maxGenerations 5, TEST_IMPACT revalidation) are now live on the public path. The AC11 standing grant for the real high-concurrency trial (Nexus-new#1384) is deliberately not issued yet; it will be requested when that trial is scheduled.
Activity
github-actions commentedon Oct 13, 2026
Authority:
RESEARCH_OBSERVATION_ONLY / NO_ENGINEERING_AUTHORITYHybrid replication capture receipt
{ "body_gzip_base64": "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", "capture_sha256": "833ba22e272919db9a123c9eda823ff85271428e3980ff8142ee29b159d963c6", "captured_at": "2026-10-13T11:36:37Z", "contract_sha256": "bb2ba35d72688d873fcca33c17fd44e0c982f6a2d1d8e9301856f0d27afa80fa", "created_at": "2026-10-13T11:36:19Z", "default_branch": "main", "issue_number": 511, "issue_updated_at": "2026-10-13T11:36:19Z", "pre_implementation_revision": "f7cfc4777fea59601ab546399fc86c47ad705d26", "repository": "James3014/devspace", "schema": "nexus.hybrid_replication.capture.v1", "source_event_id": "actions:38391375602:1", "title": "connector-host: align completion generation budget with Nexus-new loop (5) and track TEST_IMPACT revalidation (Nexus-new#1384)" }github-actions commentedon Oct 13, 2026
Authority:
RESEARCH_OBSERVATION_ONLY / NO_ENGINEERING_AUTHORITY{ "activation_boundary": "2026-10-13T04:05:51Z", "activation_state": "AUTOMATIC_CAPTURE_READY", "admitted_at": "2026-10-13T11:36:37Z", "capture_sha256": "833ba22e272919db9a123c9eda823ff85271428e3980ff8142ee29b159d963c6", "disposition": "CROSS_REPO_SCOPE_GAP", "exclusion_set_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855", "implementation_pr_numbers": [], "issue_state_at_admission": "open", "receipt_sha256": "17f638604265a57751ee875636be123c81fa50b131753568fc2fea87cec0ba12", "schema": "nexus.hybrid_replication.admission.v1", "task_key": "James3014/devspace#511", "tracked_parent_created_at": null, "tracked_parent_issue_number": null }docs(#511): completion generation cap is 5
James3014 commentedon Oct 13, 2026
Separate open item, found during review of #512 and not fixed there.
Absolute vs. relative generation counting mismatch. The host port guards (
parseCompletionGeneration, used inmaterialize_integration_headandread_required_checks) cap the absolute integration generation number (current_generation + 1) atMAX_COMPLETION_INTEGRATION_GENERATIONS(now 5). The Nexus-new loop (run_github_completion_loop) budgets relative tostart_generation(current_generation - start_generation >= max_generations), so a resumed run can legitimately produce I6 and later.Effect: when a PR is already at I3 and a new loop run starts, the host rejects I6 with
generation_INVALIDafter two more generations, although the loop allows five. This failure is closed (a reject, not a wrong merge), but it limits liveness for resumed completions.Possible fix (needs a decision): pass
start_generationfrom the bridge and capgeneration - start_generation, or drop the absolute cap and rely on the loop's relative budget. This stays open independently of the TEST_IMPACT revalidation design question.fix(#511): raise completion generation cap to 5 to match Nexus-new lo…
ci: pin connector-host smoke Actions to immutable SHAs (#511)
ci: run connector-host smoke tests under bash on Windows (#511)
ci: scope connector-host smoke matrix to macOS only (#511)
ci: pin connector-host branch Actions to full SHAs (#511) (#514)
fix(#511): bound completion runs relatively, not absolute I number
fix(#511): allow absolute completion generations after I5 on resume (#…
James3014 commentedon Oct 14, 2026
#511 macOS-only post-merge reconciliation — 2026-10-14
Classification: KEEP_OPEN / source fix integrated, end-to-end consumer outcome unverified. Target for these changes is
nexus/connector-host-v1, not DevSpace defaultmain.Verified merged lineage on the specialized branch:
cf62a6b0f627c754e00a1ebd608309d06ff910fe: per-invocation generation ceiling aligned to 5.8f1fe2545c72ca00bfeb1670a72a2c51208f002a:read_head_check_runsport for Nexus-new #2068.974e8876f7c8dc9818f99f67c8a29a8bd06443f9: macOS-only Smoke matrix; final GitHub CI run38480566759success. Actions pins remain immutable.532c0e4e64a9a9e10c2af0b4e668a124c5ad3e78(exact tested PR head4c4993014d0fc4dd0cd2d6817edd9f987e315192): correct absolute-vs-relative generation mismatch. RED macOS run38492574701fails exactly at I6/generation_INVALID; GREEN macOS run38492817361success, including typecheck/test/build/doctor. Independent read-only Agy Gemini reviewagt_8880d6d2terminalRECOMMEND_MERGE, recorded in PR comment; not a Nexus v4 independent acceptance receipt. Port now allows positive safe absolute I6+ while maxGenerations remains capped at five per call.00ccc35cb7c26b334e0f12b23b02eb91ad038fe7, implementing the exact-head TEST_IMPACT path. No live #1384 AC11 pilot is inferred.Remaining evidence gates, not passed or waived:
.nexus-core/config.toml; Nexus Core check runs on fix(#511): allow absolute completion generations after I5 on resume #515 remain failure. Mac Smoke SUCCESS is not Core SUCCESS. No Core validation, governance approval or deployment should be inferred from the specialized branch lacking required checks.Next safe gate: obtain a fresh authorization/readiness contract for the canonical Mac connector/Gateway generation, then bind a valid independently accepted pilot Candidate, re-run exact-head checks, and execute actual #1384 AC11. Keep this Issue OPEN until the live host-side outcome is verified. No force merge, bypass, synthetic concurrency or Gateway recovery effect is authorized by this comment.
James3014 commentedon Oct 15, 2026
Runtime readback 2026-10-15 (bundle-G-gateway, Mac mini, read-only)
Source alignment is complete on
nexus/connector-host-v1(head532c0e4e, #512/#513/#515). The running public connector host has not picked any of it up:com.nexus.mcp.public.connector, PID started 2026-10-14 08:00 localnode ~/.local/share/devspace-connector-host-v1/dist/cli.js serve, package1.0.1-nexus.7dist/build timedist/nexus-gateway-proxy.js:475maxGenerations: z.number().int().min(1).max(3)(still 3, not 5).gitpoints at~/workspace/devspace/.git/worktrees/devspace-connector-host-v1, which no longer exists, so the build has no recoverable revisionNEXUS_CANONICAL_SOURCE_ROOT=~/workspace/Nexus-new,NEXUS_PYTHON_BIN=~/workspace/Nexus-new/.venv/bin/python(the shared main clone, which lags main; it also sets noNEXUS_SELF_HOSTED_CANONICAL_STATE_DIR, see Nexus-new#1539)So remaining gate 1 ("runtime/host activation of the exact specialized connector source") is not met: activating it means rebuilding
dist/from532c0e4ein a fresh tree and restarting the public OAuth connector. That is a live-service change outside an auto-deployer, so it waits for an Owner go-ahead; nothing was changed here. Nexus-new#1384 AC11 stays NOT_RUN.james3014-nexus-agent commentedon Oct 15, 2026
Public connector redeployed on the Mac mini, 2026-10-16 07:10 +08 (Owner approved in chat 2026-10-16).
build_id nexus-1.0.1-nexus.7-528dc505(dist from 2026-09-25),/healthzok:false,manifest_status: drifted, pid 27167 since 10/14.build_id nexus-1.0.1-nexus.7-532c0e4ebuilt fromnexus/connector-host-v1@532c0e4e(npm ci, typecheck, 14 test files, build all green),/healthzok:true,manifest_status: verified, 45 effective tools, proxy to gateway 127.0.0.1:8766.com.nexus.mcp.public.connectorwas restarted (launchctl kickstart -k). First attempt rolled back automatically becausegenerated/build-identity.jsonwas not swapped, so/healthzkept reporting 528dc50; second attempt swappeddist/,generated/,node_modules/and the package files together.~/.local/state/nexus-connector-backup-20261015-2354/.So #512/#515 (maxGenerations 5, TEST_IMPACT revalidation) are now live on the public path. The AC11 standing grant for the real high-concurrency trial (Nexus-new#1384) is deliberately not issued yet; it will be requested when that trial is scheduled.