19 KiB
Row 265 / #1770 — project rooms
Current result: verified under sage-3942; slot returned in goldie-0034. Full bus 957/957, sandbox 10/10, parser 15/15, rooms 18/18, twelve mutants killed, restored files green. Review/push packet below includes mos-1567 closure scope.
Goldie, 2026-10-24. Branch goldie/row265, base ec86f2c847. Brief docs/plans/2026-10-20_case2-rows.md § C2-09b rooms at blob3a5ed91a4fb3462df67e6939ae6f18290ac9b453; design agents/gosalyn/work/case2/DESIGN.md §6. Sage-3915 transfers ownership from Rocko to Goldie and review from Dewey to Boyd. No old test edit authorized. All tests/mutants wait for Sage CLEAR; queued behind Poe, Bentina and Sage's326 gate.
Design and clarified readings
A room is a bus project. Members are all loaded business instances whose definitions are in THREAD_DEFINITIONS, sorted by key. Canonical room threads have those keys as names and belong to the requesting human. open checks all name/instance/owner collisions before creating anything; it delegates missing inserts to existing thread.create dispatch, including default session metadata. send/read require the complete canonical set. No schema or migration change.
Sage-3921 settles two brief readings: the single captured timestamp applies to the room.send MESSAGE rows; existing route/delivery/event helpers keep their own timestamps in the same dispatch transaction. They are unchanged. A final EOF-terminated @name is a tag, so bare @continuity is invalid-request rather than a broadcast. Leading whitespace is accepted. Pure parseAddress returns {tags,body}, where body is the unaddressed remainder; room.send stores the ORIGINAL input unchanged. Export it from the bus index and expose a separate reusable test table for C2-10c parity.
send resolves tags against current #named results and member instance keys, deduplicates recipients, and refuses multiple member matches. It checks each target's current default-session fence and messaging rule, then inserts one normal INFO message, route and human.input per target under the existing request transaction. All message rows share one at. No tag broadcasts; tags after the first non-tag token remain plain text.
read includes all sessions of the caller's canonical threads, globally ordered by messages.seq. Each entry retains first-message fields plus thread/session/instance and adds ids (every grouped message id) and to (each instance once). Only consecutive human rows with identical at/body/class group, and a repeated instance starts a new entry, so same-millisecond identical sends remain separate. Paging limit counts merged entries (THREAD_PAGE defaults/max); clients continue after the last ids element. Cursor must belong to a room thread. Bounded raw scan limit*memberCount+1 completes requested groups without reading unbounded history. Sage-3921 accepts these routine choices as proposed in goldie-0030.
Scope / steps
- Read current thread.create/send/messages, transaction wrapper, naming and messaging helpers; audit source readers/pins. Claim265 via queue after next. Done/read, claim recorded at rev 3814 with op goldie-0265-start-1 (same-op replay confirmed).
- Add parser + index export, only three room cases in broker, README verb rows and new bus tests/fixtures. Do not modify existing thread/messaging/ACL/helper behavior.
- New tests cover all brief cases, atomicity across message/route/input rows, ownership/projects, permissions, paging/grouping and session fences. No suites or test runs before CLEAR.
- Record exact refusal→test→mutant matrix here before requesting the slot; one targeted killed mutant per named refusal, restoration byte-checks. New files whole then bus suite under CLEAR, serial timeout-k10, fresh31-byte mktemp TMPDIR plus UTC/load/busy header.
- Before in-review, grep packages/*/tests for broker.mjs/TESTER_VERBS and changed-file pins. Push own branch and hand tip/base/manifest directly to Sage/Boyd. Sage owns request posting/landing.
No Console/Discord/CLI room surface or cross-agent reply visibility added. No tester verb, role, schema or secret change. No host action or credential access.
New tests and refusal / mutant matrix
Only new test files and a new reusable vector fixture are added. No pre-existing test changes. Parser table: address.test.mjs consumes address-cases.mjs, including final EOF tags, leading whitespace, tags after the first word, repetition and blank/tag-only remainder. Broker cases: rooms.test.mjs covers sorted/reused canonical threads, other-instance and other-owner collisions, all three verbs' human/tester gates, missing projects/member set, partial/foreign/wrong-instance rooms, key and current-name addressing, ambiguity/same-member name+key, unknown/non-member/old-name tags, tags-only bodies, one/many/broadcast targets, original bytes, shared message timestamp, routes/deliveries/human.input, denied last recipient, clearing last session, all-session merged reads, agent reply metadata, merged paging/default/max, foreign/noncanonical/absent cursors, and same-at repeated sends. Synthetic historical rows append valid rows; they do not rewrite evidence or clocks. The same-at batches repeat the first recipient at each boundary, proving identical sends remain distinct.
Each named refusal has one mutation of its enforcement, run against the entire new rooms.test.mjs file. The driver mutants.py stores the exact original→mutant diff in logs/Mxx-*.diff, logs UTC/load/busy/TMPDIR, accepts only Node exit 1 with the named target test failing, and restores/verifies full broker bytes and SHA256 in finally. No mutation is retained. Generic pre-existing request/schema/argument policies are exercised through existing code; the matrix names the refusals added or deliberately inherited by these room cases.
| Mutant | Refusal | Guard relaxed | Required failing new test |
|---|---|---|---|
| M01-human | human-required | remove common room human check | room verbs require the human |
| M02-tester | tester-refused | add the three room verbs to TESTER_VERBS | room verbs require the human (explicit absence pin) |
| M03-project | project-not-found | remove room project lookup check | room verbs refuse missing projects |
| M04-members | thread-instance-missing | allow empty member set | room verbs refuse missing projects |
| M05-collision | room-thread-name-taken | remove canonical owner/instance collision check | room.open preflights instance collisions |
| M06-open | room-not-open | remove missing canonical thread check | send/read refuse a missing canonical member |
| M07-body | invalid-request | allow empty parsed remainder | unknown or non-member addressees and tags-only bodies |
| M08-unknown | unknown-addressee | allow zero member matches | unknown or non-member addressees and tags-only bodies |
| M09-ambiguous | addressee-ambiguous | allow multiple member matches | a legacy name colliding with another member key |
| M10-messaging | message-sender-refused | omit room target messaging check | messaging denial on the last target |
| M11-cursor | message-not-found | unknown room cursor falls back to seq zero | room.read groups one send, pages |
| M12-fence | session-clearing | omit room target session fence | a clearing last member fences the whole send |
For M02, the existing human check remains a second barrier: the mutant deliberately relaxes only the tester allowlist; the new explicit absence assertion kills it without loosening or running the old pin. Existing helpers keep their full session-state behavior. No helper or schema mutation ships.
Reader / hash audit
rg -n 'broker\.mjs|TESTER_VERBS' packages/*/tests --glob '!fixtures/*' produced reader-grep.txt (133 hits at audit time). These are imports except the concrete source read at bus/tests/limit-checks.test.mjs:323-326 (forbidden limit-check/token-limit imports) and the tester set assertions at bus/tests/tester-writes.test.mjs:18-32. The latter stays byte-for-byte unchanged. bus/tests/viewer.test.mjs:116 builds its coverage set from exported verb lists; rooms do not widen those lists. No old test needs an edit.
Pin check: recursively inspect packages//tests plus scripts/test-.sh for changed-file mentions (broker.mjs, address.mjs, packages/bus/README.md, packages/bus/src/index.mjs) with literal 64-hex values within eight lines: pin-context-grep.txt records no hits. A direct same-line hash/digest/sha256 grep finds registrySha256 imports only, which hash host registry data rather than broker bytes. The independent channel-proof.test.mjs baselines cover proof/transport files, none changed here. Positive control: pin-positive-control.txt records actual schema.sql / SCHEMA_SHA256 / createHash references in bus/tests/schema.test.mjs and bus/src/store.mjs. Schema bytes are unchanged.
Requested serialized slot plan (not run)
At each run: fresh TMPDIR=$(mktemp -d /mnt/storage/scratch/tmp/XXXXXX), date -u, uptime, and bash /mnt/storage/scratch/sage-busy-now.sh 2>/dev/null | tail -1 in the log. No concurrent test runs. Ignored root node_modules symlink targets the canonical checkout; it will not be committed.
timeout -k 10 120 node --test packages/bus/tests/address.test.mjswhole.timeout -k 10 120 node --test packages/bus/tests/rooms.test.mjswhole.timeout -k 10 600 node --test --test-concurrency=1 packages/bus/tests/*.test.mjsbus suite whole.python3 agents/goldie/work/queue-265/mutants.py: M01 through M12 serially, each runstimeout -k 10 120 node --test packages/bus/tests/rooms.test.mjswhole, fresh TMPDIR/header; stop on a live mutant or wrong failure. Record each diff and restoration digest.- Restored address.test.mjs whole and rooms.test.mjs whole separately, same commands as 1/2 and fresh directories/header each.
Estimate 5–10 minutes; bus timeout ceiling 10 minutes, individual new/mutant ceilings 2 minutes each. On first new-test red, record/report it. On any old-test red or need for old-test change, stop and ask Sage. Static node --check for new modules/tests and git diff --check passed; no tests or mutants run yet.
First slot STOP — sage-3932
2026-10-24T04:27Z: stopped on an old-test red. No old test changed or rerun, no mutant executed, no product adjustment after the failure, no commit/push claiming verification.
| Step | Exit | Counts / disposition | Log |
|---|---|---|---|
| 1 address whole | 0 | 14 pass, 0 fail/skip/cancel; 58.559 ms | logs/01-address.log |
| 2 rooms whole | 0 | 18 pass, 0 fail/skip/cancel; 1067.524 ms | logs/02-rooms.log |
| 3 bus whole | 1 (stopped) | partial 619 top-level ok, 1 not ok; not a completed-suite result | logs/03-bus.log |
| 4 twelve mutants | not run | stop rule | none |
| 5 restored files | not run | stop rule; no mutations were applied | none |
All three run headers use UTC/date, uptime, helper CLEAR and distinct 31-byte scratch TMPDIRs. Bus began04:25:22Z at load2.80/2.91/2.61, TMPDIR /mnt/storage/scratch/tmp/LaZAaK. Parser used /mnt/storage/scratch/tmp/rqcTH7; rooms has its own header. Explicit TAP reporter was selected for rooms/bus and the mutant driver so failure detection/counts are stable across Node26's default spec reporter; this changes reporting only.
The old red is project-runtime.test.mjs:12, assertion:14, test620, “Broker keeps the sandbox client import closure: only the listed imports, no lazy project import”. Actual imports include the new import { parseAddress } from './address.mjs';; the pinned array omits it. ERR_ASSERTION deepStrictEqual; individual duration2.469251ms. The full expected/actual/stack is in 03-bus.log. No other red was observed before stopping.
Reader-audit correction: reader-grep.txt:100 already contains project-runtime.test.mjs:13, a direct broker source read. The earlier prose incorrectly described the import-closure assertion as a normal import; I missed this old source pin during inspection. This is not a hash pin. Follow-up static trace finds seat/src/sandbox.mjs:50-60 RUNNER_FILES includes broker and landing but not address; seat/tests/sandbox.test.mjs:45-53 compares this mount list to the recursively imported source closure. A parser dependency therefore also needs the one new RUNNER_FILES path. Both an old-test pin change and this out-of-owned-files product change require Sage/Mos approval before edits. Recommended bounded change: add only the parser import to the project-runtime expected imports, add only packages/bus/src/address.mjs beside broker in RUNNER_FILES, and leave the generic sandbox closure test unchanged.
Broker bytes remain SHA256 71b32ee4afa322793aad81231441ee7badc3e5aff65ff008c6cc0f7f63b72781. git diff --check still passes. Slot returned in goldie-0032, pending ruling/new CLEAR; no reruns.
mos-1567 approved closure correction / second slot request
Sage-3940 relays Mos's mos-1567 and widened row scope (sage-3939). Applied exactly the approved pre-existing-file edits: project-runtime.test.mjs adds two lines at17-18, one Row265 comment and the parser import in the pinned array, no other line changed (numstat2 added/0 deleted); seat/src/sandbox.mjs adds only packages/bus/src/address.mjs immediately after broker in RUNNER_FILES. seat/tests/sandbox.test.mjs stays byte-for-byte unchanged. The parser remains a separate pure module. A NEW block in address.test.mjs asserts doesNotMatch(/\bimport\b/) on its source, with a node:fs import in that new test file; no existing assertion is moved or weakened.
Review packet must explicitly call out the RUNNER_FILES mount-list entry for Boyd's review. This is a read-only source-file mount needed by the broker's static dependency, with no new module dependency or role/verb authority. The first-run import pin failure remains recorded above and in logs/03-bus.log; new logs will have rerun-prefixed names, preserving first-slot bytes.
Requested rerun order, awaiting CLEAR after Bentina: first timeout -k 10 180 node --test --test-reporter=tap packages/seat/tests/sandbox.test.mjs whole, then the original five steps again from step1: address whole (now15 cases), rooms whole, bus whole with concurrency1 (includes project-runtime whole), M01–M12 whole-room-file mutants, then restored address and rooms files separately. Explicit TAP reporter is used throughout for stable logs/kill matching; timeouts otherwise120 per new/mutant run and600 for bus. Each run still gets fresh short mktemp TMPDIR, date/uptime/busy header, serialized with the same stop rule. Estimate5–10minutes. Only static node --check and git diff --check run since the STOP; all pass. No new test run yet.
Verified result — sage-3942 CLEAR, slot returned goldie-0034
All rerun positives exit 0 with zero failures, skips or cancellations. Full bus suite passed 957/957, including the approved project-runtime import pin. The unchanged sandbox test passed 10/10, including its recursive mount/import-closure assertion. New parser cases passed 15/15, rooms 18/18, and both passed again after restoration. No tests run after slot return.
| Run | Exit | Pass / fail | Duration ms | UTC / load / helper | Log |
|---|---|---|---|---|---|
| rerun-00-sandbox | 0 | 10 / 0 | 5526.027699 | Sat Oct 24 04:31:35 UTC 2026; 23:31:35 up 10 days, 14:51, 2 users, load average: 2.83, 2.91, 2.73; CLEAR | log |
| rerun-01-address | 0 | 15 / 0 | 54.634817 | Sat Oct 24 04:31:45 UTC 2026; 23:31:45 up 10 days, 14:51, 2 users, load average: 2.54, 2.84, 2.71; CLEAR | log |
| rerun-02-rooms | 0 | 18 / 0 | 921.788987 | Sat Oct 24 04:31:52 UTC 2026; 23:31:52 up 10 days, 14:51, 2 users, load average: 2.42, 2.81, 2.70; CLEAR | log |
| rerun-03-bus | 0 | 957 / 0 | 118397.418425 | Sat Oct 24 04:32:01 UTC 2026; 23:32:01 up 10 days, 14:51, 2 users, load average: 2.28, 2.77, 2.68; CLEAR | log |
| rerun-05-restored-address | 0 | 15 / 0 | 54.587943 | Sat Oct 24 04:35:02 UTC 2026; 23:35:02 up 10 days, 14:54, 2 users, load average: 3.10, 3.07, 2.83; CLEAR | log |
| rerun-05-restored-rooms | 0 | 18 / 0 | 991.877991 | Sat Oct 24 04:35:11 UTC 2026; 23:35:11 up 10 days, 14:55, 2 users, load average: 3.17, 3.09, 2.84; CLEAR | log |
Mutant driver exit 0; all twelve target tests killed their guard relaxation with Node exit 1. Every mutant log records exact bytes restored to broker SHA256 71b32ee4afa322793aad81231441ee7badc3e5aff65ff008c6cc0f7f63b72781. Driver receipt: rerun-04-mutants-driver.log.
| Mutant | Node exit | Pass / fail | Exact diff | Raw log |
|---|---|---|---|---|
| M01-human | 1, killed | 16 / 2 | diff | log |
| M02-tester | 1, killed | 17 / 1 | diff | log |
| M03-project | 1, killed | 17 / 1 | diff | log |
| M04-members | 1, killed | 17 / 1 | diff | log |
| M05-collision | 1, killed | 14 / 4 | diff | log |
| M06-open | 1, killed | 17 / 1 | diff | log |
| M07-body | 1, killed | 17 / 1 | diff | log |
| M08-unknown | 1, killed | 16 / 2 | diff | log |
| M09-ambiguous | 1, killed | 17 / 1 | diff | log |
| M10-messaging | 1, killed | 17 / 1 | diff | log |
| M11-cursor | 1, killed | 17 / 1 | diff | log |
| M12-fence | 1, killed | 17 / 1 | diff | log |
Final review scope and packet
Boyd reviews the three room cases, pure parser/export, README rows, new parser/room tests, and the added read-only RUNNER_FILES mount for packages/bus/src/address.mjs. mos-1567 authorizes that one sandbox source line and the sole old-test delta: project-runtime.test.mjs:17-18 adds exactly the Row265 comment and parser import pin (numstat 2 added / 0 deleted). Byte comparison against base after removing those additions proves all other old-test bytes unchanged. Removing the single RUNNER_FILES addition likewise restores sandbox.mjs to base. sandbox.test.mjs has no byte change. New address.test.mjs block proves the mounted parser source has no import word.
The widened changed-file hash-context audit includes sandbox.mjs: pin-context-with-sandbox.txt, zero 64-hex-near-reference hits. The original reader and positive-control receipts remain alongside it; the missed source pin and first STOP remain recorded, not concealed. No role/schema/tester allowlist or helper change ships. BUILD-LOG and SESSIONS only append tracking entries.
Base: ec86f2c847f6154a0677c6ed23c87dabc2c665d0. Branch: goldie/row265. MANIFEST.sha256 lists every changed source/test/tracking/evidence file except itself, with repository-relative paths, and is verified before commit/push. Raw first-red logs are retained separately from green reruns.
Evidence packaging: logs are stored losslessly as .log.gz; decompress with gzip -cd. Decompression was byte-compared with each raw run log before packing. This preserves the original failure/Node whitespace while keeping the committed diff whitespace-clean. The run driver recreates raw .log files when used. Mutant busy-check headers identify only their own Python driver PID3686656 as BUSY; positive runs report CLEAR. The driver held the assigned slot throughout, with serialized child tests.