Conversation
added 29 commits
September 30, 2026 23:53
…eal only while an update is proven kit.json ruvnetBrain:true made autoUpdateOptOut stand down forever, but agentic-kit schedules nothing on the owner's Mac, so the knowledge base aged by hand only. Ownership is now honoured only while a successful refresh or CURRENT verdict is proven within 36h; otherwise SessionStart self-heals and the knowledge line names the Brain's own update, never --enable-nightly. Also de-flakes the 35-day currency test that compared a real-clock age to a fixed NOW.
…t puts the gold store planSourceRoute is the routing block of searchAllPrimary moved verbatim into an exported function (git diff -w shows only the move). scripts/route-gold-rank.mjs runs it with no model and reports gold-in-1/3/5, declines and stores opened with Wilson intervals.
…ole, tolerate a self-updating host CLI - consoleRestartState prunes a receipt whose pid is dead (ESRCH) and whose port does not answer /api/runtime with the same identity (scripts/console-instances.mjs); --doctor re-reads live receipts instead of trusting the recorded pending-console-restart snapshot. - syncHostsAfterUpdate replaces an owned Console still serving the previous runtime by running the activated runtime's own launcher (--serve, never --open) in that scope and port, waits for the receipt to name the new runtime and the old process to exit, and records exactly why when it cannot. - claude/codex invocations go through scripts/host-cli.mjs runHostCli: a CLI seen this run (or whose PATH link dangles) that is missing is retried 3 times over ~20s, then reported as one line; a CLI that was never installed still fails at once.
…the update path From the customer-state-matrix agent's measured patches (2026-09-30), unchanged except where noted: 1 update-apply: a machine with no host and no active spine has nothing to converge (exit 0, not 1). 2 refresh-run: physicalPath(); lock inheritance compares directories, not spellings (symlinked cache). 3 update-storage-transaction: a kill mid candidate-build quarantines the unsealed candidate instead of RECOVERY_REQUIRED forever; the quarantine is released one cycle later only if bytes are unchanged. 6 install --update: recovers an interrupted storage transaction before looking for the updater; no fresh reinstall fallback on retained-copy refusal or a transient manifest failure (403/408/429/5xx/offline). 7 corpus-canary: opt-in private-overlay and older-runtime cases (default stays clean) + the customer-state-matrix harness (registered in wired-check as a human-run harness). Also repairs corpus-canary's BREAK-IT test, red on origin/main: its literal no longer matched carryLiveNodeModules()' guard line.
…data per KB build metadataSourceRoute broke a top-overlap tie by repository name and kept one store. It now keeps up to METADATA_ROUTE_TIES (3) stores tied at the best overlap, ordered by how many of their entries reach it. Route-only on the 206-need set (kb 4.3.37): gold in top 3 42/206 -> 83/206. The metadata scan reads an in-process inverted index keyed by the KB build identity (manifest stat + generated stamp) and the store file's dev/inode/mtime/size, so an update never serves the previous build's index even when a rewritten store keeps its mtime and size.
…s into the brain Matrix patch 6 (revised): run recoverIncompleteStorageTransactions whenever the transaction ledger exists, not only when forge-update.mjs is missing. The preflight re-stamps RUNTIME-IDENTITY.json, after which recovery can never prove live equals the identity sealed at LOCKED, so a pre-activation kill wedged every later update at RECOVERY_REQUIRED.
…e gist store alone An unhyphenated "rUv" was read as provenance intent, so "which rUv library runs vector search in the browser?" searched ruv-gists only (5 of 6 product probes on the 4.3.37 corpus). It now signals provenance only next to something rUv authored or said (publish/write/post/announce, tutorial, spec, post ...). After: those 5 route to agentdb, ruvector, ruv-fann, ruview, agentic-flow; all 7 provenance probes still route to ruv-gists.
…with store identity stripped One described need per store (180+ stores) so a routing change is validated on far more than the three repositories the novice need set covers.
…rain home; test --doctor's live re-read Matrix patch 4 (kb/brain-profile.mjs): published names with dots (dspy.ts, ruv.io) are valid store names (still no leading dot, '..', separator or .big/.rvf suffix), and stores in the release's PUBLIC-RVF-GENERATIONS.json (concepts, ruv-gists) are release-managed unless fenced private or opted out. Matrix patch 5 (kb/lifecycle-evidence-retention.mjs): the brain home (the evidence roots' parent) may be a symlink to another disk; references are compared by physical path. An evidence ROOT that is itself a link is still refused — new TEETH test through a real and a linked brain home. --doctor: the live re-read of Console receipts is now withLiveConsoleState(), exported and tested.
…STER #26) canonical-qa gains a full-suite job: scripts/full-suite-gate.mjs runs every file vitest.config.mjs includes and fails on any red not in the reviewed tests/known-red.json (class, evidence, owner), on a quarantined test that now passes, on file-level errors, and on any test file vitest silently never runs. Before this, 552 of 589 test files ran in no automatic gate. Also: ruflo-daemon-autostart reads the version of the binary it executes, not a hard-coded $HOME path; drop a dangling QE lane path and fix a wrong executed-by comment in corpus-seed.yml.
…uth green) install-release-fallback: the rate-limit reset time is built from a Date and asserted with toContain. customer-state-matrix: N-k expectations are read from the fixture release list, not typed versions. no-restated-truth was red on main for the first and on this branch for the second (carried with patch 7).
…me each known-red owner lesson-migrate-agentdb asserts the developer's own AgentDB stores through the real ruflo CLI; it was red in every environment but the owner's. It now lives with the other diagnostics under tests/diagnostics/vitest.config.mjs and is an explicit excludedFiles entry. tests/known-red.json is transitional: every entry names its owning branch, and a release ships with it empty.
…--update Matrix D8 / patch 8: an installed updater run directly (cd kb && node forge-update.mjs --apply) can predate the current candidate layout and fail the guard (measured on 4.3.28 vs 4.3.39); the npx door places the current updater first. Changed: the SessionStart BEHIND hint (behavioural test through heartbeat()), kb/forge-update.mjs --check's 'A newer build exists' line (asserted in the real --check run in forge-update-apply-rollback), kb/forge-currency.mjs's brain line, docs/ARCHITECTURE-MAP.md.
…orts FLAKY Measured: a full run at load 100-150 turned four budget/latency tests red that pass alone (restore-semantics 3745ms vs 1000ms, decision-gate 2000ms budget, ux-qe probe, issue-83). A test red in the full run but green in one isolated retry is listed as FLAKY in the verdict, never hidden; red in both runs, or absent from the retry, stays RED. Sabotage-tested.
…s final) Local paths are replaced by placeholders (<kb>, <scratch>, <worktree>, ...). The final run's off-topic and held-out outputs were byte-identical to the baseline's; one copy of each is kept.
…f 4.3.40 Reapplies, without conflicts, the hook commit reverted out of release/4.3.40 (cd385f3, saved as .claude/worktrees/hook-fixes.patch) and the adversarial review's work-in-progress corrections (hook-review-fixes-wip.patch): - grounding-stamp.sh: success markers and refusals read from tool_response only (a model-written query can no longer forge a stamp); refusals checked before any marker; fast-lane card and host-replaced oversize results recognised; oversize file read only under $HOME/.claude/projects/*/tool-results/ - grounding-turn-evidence.mjs: brainAnswered() recognises the same shapes - set -u: initialise the read variable in design-wall, ground-before-write, grounding-stamp, protect-brain-state, route-dispatch - ground-ruvnet.sh: 2>/dev/null before the file redirect (no shell error text) - codex-hooks.json: SessionEnd launcher 2500ms / 3s (the host's hard cap) - scripts/hook-qualify*.mjs + fixtures + tests: the hook qualification matrix (its path-leak guard assembles the owner's username instead of carrying it) Follow-up commit finishes the review corrections and the Stop false alarm.
…on a topic; finish review corrections THE FALSE ALARM. Gate 1 arms on any prompt naming the rUv stack, and in this repository nearly every prompt does, so the Stop gate said "no successful search_ruvnet call is in this turn's transcript" on release-status, git/CI, disk/backup and memory-write answers. grounding-turn-gate now demands a search only when the final answer asserts what a rUv product does/can/cannot/requires /says (ruvCapabilityClaims, deterministic, no model). Measured on hand-labelled real Stop points (kept outside the repository) through the real decide(), origin/main vs this commit: 183 real deliveries of the correction: false positives 158/176 -> 0/177 held-out set (70 Stop points, never tuned on): FP 68/68 -> 0/68, FN 0/2 -> 2/2 (both implicit shapes) tests/fixtures/grounding-turn-stop-points.json is a MINIMAL public-safe fixture: 45 deciding excerpts (7 must-fire claims, 6 pinned known misses, 3 grounded claims, 29 hard negatives across release/git-CI/disk/memory), no paths, usernames or sensitive strings, guarded by a test that refuses them. Review corrections finished: - grounding-turn-gate: a long turn whose tail cannot see the turn start falls back to the stamp evidence (turn = null) instead of returning null (silent pass) - brainAnswered(): oversize file read only under $HOME/.claude/projects/*/ tool-results/, never a symlink, never with a refusal before the answer - stamp forgery tests per case: symlink, `..` path, case-forged markers - Codex SessionEnd: wrapper budget 2200ms (inside the 2500ms launcher, 3s host cap) handed down; session-snapshot plans inside it, captures the NEW snapshot first and defers outbox replay under 4000ms (deferred, never dropped) - set -u: learn-capture.sh and kling-preflight.sh (red on macOS bash 3.2); the regression test discovers every set -u + timed-read hook and spawns bash through resolveBash(), never a literal path - grounding-turn-replay.mjs reports gate1WouldFire under the new rule plus gate1ArmedUnsearched (the old count); CONTRIBUTING and hooks.json describe it - wired-check: register scripts/hook-qualify.mjs (human-run matrix)
…n update identifierScan keyed its warm-process cache by directory path, so after an update swapped kb/ under the same path the MCP worker kept answering with the previous build's stores. It now keys on the same KB build identity the router metadata index uses (extracted to kb/kb-build-identity.mjs) and drops the old build's scans.
# Conflicts: # data/convergence-manifest.json # scripts/wired-check.mjs
# Conflicts: # data/convergence-manifest.json
…back, not export The real-AgentDB progression test ended with a `ruflo memory export` probe that returned count 0. Root cause is the probe, not the product: export has no --path and on ruflo 3.49.0 ignores CLAUDE_FLOW_DB_PATH/MEMORY_PATH, reading a cwd-derived store (0 rows from the project root, 1 from <root>/.swarm). The same two product-written rows read back byte-identical through `memory list --path` and exact `memory retrieve --path` on the canonical store (the --path surfaces ruflo wires per scripts/smoke-memory-db-path.mjs, #2105). The probe now reads every row back through the CLI with --path and compares canonical digests. Red when the product writes outside the documented namespace (capture/restore stay self-consistent; the CLI readback fails), green on the product as shipped. Drops its tests/known-red.json entry.
# Conflicts: # data/convergence-manifest.json # tests/known-red.json
….swarm store is created ruflo creates <cwd>/.swarm on every invocation even when --path names the store. The progression store ran it from inside <root>/.swarm, leaving an unused nested store in every project (measured on ruflo 3.49.0). When the store is the default <root>/.swarm/memory.db, run from the project root so that directory is the store's own. Proven: the real-AgentDB test fails on the old cwd (nested store present) and passes now; a unit assertion pins the cwd.
…ling Review finding S4: any test that failed in the full run and passed alone was waved through, unbounded and unrecorded, so a value-assertion red caused by real shared state could hide. Now a test is FLAKY only when its full-run failure message is a timing/budget failure (vitest test/hook timeout, or an explicit budget/ceiling/latency/'took Nms'/'exceeded Nms' assertion) AND it passed the isolated retry. Every FLAKY is recorded with its failure in the JSON verdict and the GitHub job summary; more than 3 in one run fails the gate. Each rule is sabotage-tested. full-suite job cap 75 -> 120 min.
… the supplied KB; restore aliases without a public repo-aliases.json S5: the private-overlay canary test compared the verdict with all-checks-ok, a tautology; it now names every check green and private-store-preserved byte-identical. Making it real exposed a product defect it had hidden: restorePrivateOverlayState read the candidate's repo-aliases.json unconditionally, so a public bundle without one failed every private-overlay update (ENOENT). It now starts from an empty map. S6: with --installed-kb, extra cases reused and mutated the supplied KB. suppliedKbInstaller gives each extra case its own copy of the pre-clean snapshot, and runCanary refuses an extra case handed the clean case's KB.
# Conflicts: # data/convergence-manifest.json
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
# Conflicts: # data/convergence-manifest.json
…a index cache (n=69)
# Conflicts: # data/convergence-manifest.json
… Brain home Re-review blocker: ruflo writes snapshot content (hnsw.metadata.json) into <cwd>/.swarm and loads any metadata it finds there. One shared /tmp/ruvnet-brain-ruflo-cwd-<uid> pooled every project's text and, on a shared /tmp, could be pre-created or symlinked by another user. rufloCwdFor() now returns <RUVNET_BRAIN_HOME or ~/.cache/ruvnet-brain>/ruflo-cwd/<sha256(store path)>, and every directory it creates is verified by ensurePrivateDir(): a real directory (a symlink is refused), owned by this uid (else refused), mode 0700 (repaired). The test suite points the root at a private temp dir (RUVNET_RUFLO_CWD_ROOT in vitest.config.mjs) so no test touches the real Brain home.
…ps the user's keys and proxy S-B: a live pid whose port is silent (a busy Console) or answers another identity (pid reuse) is now a reported failure with the receipt kept — it is never deleted, so a live stale Console cannot be orphaned; only dead-pid + silent-port receipts are pruned (readConsoleReceipts). Still reported at once, with no launch and no 20s wait. S-C: consoleEnv is a denylist of this run's flags (RUVNET_NIGHTLY*, test/import switches, the refresh token, NODE_OPTIONS, npm_*/VITEST*), so provider API keys, HTTP(S)_PROXY/NO_PROXY and NODE_EXTRA_CA_CERTS the Console reads reach it.
…Y timing class - pre-commit-convergence-heal ran 'git config user.name/email' inside a linked worktree of the real repo; linked worktrees share the main .git/config, so every run rewrote the owner's identity to Fixture (143 commits since 2026-09-27). Identity is now passed per command with -c. A guard test runs the heal test in a separate process and asserts the real repo's --local user.name/user.email are unchanged (it restores them before failing): red with the old lines reinstated, green now. No other test configures a worktree of the real repo; the rest configure throwaway repos. - full-suite-gate: FLAKY timing class is now only a vitest test/hook timeout or a numeric bound comparison carrying a duration cue. Budget/ceiling/latency words in a value assertion stay RED. Proven with real failure messages; red when the old loose rule is restored.
…ink clone, else full copy) NIT 7: copies already use COPYFILE_FICLONE; the doc now says what that costs without reflink (about 3x the KB at peak) and why hardlinks are not used (an in-place rewrite would reach the caller's brain).
…detector, degraded, node-path, windowsHide nits 4.4.0 re-review. S-A. Ordering was not guaranteed: a boundary counted only outbox rows as older work, so with an older boundary queued and a worker holding the lock, a new boundary at 1900 ms produced+captured inline and at 30000 ms replayed and captured inline, and the worker later committed the OLDER snapshot after the NEWER one. The replay lock is now the right to commit in order: every boundary takes it first; if a live worker holds it, or captures are queued (any budget), or a short budget has outbox debt, the boundary queues itself behind that work and hands the lock to a detached worker; otherwise it works inline while holding it. The lock carries an owner token: only the owner refreshes (heartbeat between every step) or releases it; a worker whose lock was taken over stops at once; a lock not refreshed for 120 s is taken over (rename-aside, then exclusive create). Any boundary therefore drains a queue a dead worker stranded. Per-step budget 45 s (< half the stale time). NIT 3. "will not" counts only with a capability verb, "now" only with a capability verb, and a same-sentence pronoun only after "X is a/an/the ..." or "X has/provides/ships ...": the reviewer's four false positives are silent, the S2 recalls still fire. Real labelled sets unchanged: 183 firings FP 0/177 FN 0/6; held-out 94 FP 0/89 FN 2/5; held-out 70 FP 0/68 FN 2/2. NIT 4. The degraded paragraph is matched to its fixed closing sentence, so a repo error spanning lines no longer fails a real answer closed. NIT 5. hook-shim passes its own node as RUVNET_NODE_BIN; grounding-stamp.sh prefers it over PATH (and no longer needs dirname). NIT 6. The detached worker is spawned with windowsHide.
# Conflicts: # data/convergence-manifest.json
# Conflicts: # data/convergence-manifest.json
… authorship words Review found product questions that still reached ruv-gists alone through words like share, threads, specs, writes, writing and post. ruvAuthorshipIntent now requires rUv as the subject of an act of saying or publishing, or rUv's written artifact. Six review questions are negative tests (red under the previous rule); eight provenance controls reach the gist store only through this rule (red without it).
…ver a literal path entrypoint-guard-safety refuses a literal /bin/bash spawn; the new hook-shim node-path test's PATH precondition used one.
…scan cache is bounded The private overlay, on-demand ingest and forge-refresh rewrite .passages.jsonl without touching manifest.json, so a build-identity key alone served stale scans after them. The key now adds a fingerprint of every passage sidecar (name, dev, inode, mtime, size); the build identity still covers a rewrite that keeps all four. Scans are LRU-bounded at SCAN_CACHE_MAX (64).
# Conflicts: # data/convergence-manifest.json
Linux probe: readSourceIdentity digested untracked files from git ls-files --others --exclude-standard, which includes .swarm/memory.db (and -wal/-shm, outbox, queue) whenever the customer's project does not gitignore .swarm. Every capture wrote memory.db, so every later boundary looked changed, the no-op path was never taken, and memory.db grew at every boundary. It passed on the owner's Mac only because the global gitignore lists .swarm/. The tracked, untracked and dirty digests now exclude .swarm, .claude-flow and ruvector.db by pathspec, independent of any gitignore. Tests run git with an empty GIT_CONFIG_GLOBAL and GIT_CONFIG_NOSYSTEM=1.
…ncy and memory harnesses One token dictionary per KB directory build and three typed arrays per store (CSR) replace a Map<token, Uint32Array> per store. scripts/route-index-memory.mjs on the 4.3.37 corpus (199 stores, --expose-gc, two runs each): retained 162.5 MB -> 43.9 MB; cold build 2.2-2.5 s both; warm call 2-3 ms both. Indexes for at most two KB directories are kept (LRU), so alternating directories does not rebuild. scripts/route-latency-warm.mjs is the paired warm latency harness behind the routing 4.4 numbers; --summarize over the committed rows reproduces latency-warm-abc/summary-paired.json.
… the 4.3.40 reproduction
… 488 questions; index 162.5 -> 43.9 MB)
# Conflicts: # data/convergence-manifest.json
# Conflicts: # data/convergence-manifest.json
# Conflicts: # data/convergence-manifest.json
Owner
Author
|
Probe complete: full-suite green on Linux (runs 4 and 5). Closing without merge; 4.4.0 ships from release/4.4.0. |
This branch was successfully 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
1 participant
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.
Throwaway PR to measure the new required full-suite job on a GitHub ubuntu runner before the 4.4.0 release PR. Will be closed without merging.
🤖 Generated with Claude Code