100 Commits

This Branch

This Branch

All Branches

Author	SHA1	Message	Date
jschoubben	9e11790bd7	Say a read-only mesh-cli call in no line of its own, and count them hourly (hq issue 762) ... A page polling plans every few seconds wrote about 14,000 journal lines a day. A line is judged a read by the generic command verb's reading forms (ADR 0266), on its every word, so plans go, conditions silence and settings set are still said; a refused or failed read is said too.	2026-10-15 00:28:17 +02:00
jschoubben	9f072c110c	Take no rebuild into a waiting walk, and ask what a tier left unasked (hq issue 754) ... A rebuild that landed while a walk waited for mesh-delivery's word was taken as the walk's build of that module; once let go, the walk asked a tier only when every module of it was unasked, so it never asked the rest and stood still. A walk now takes only the outcomes of builds it asked, and none while it waits; a rebuild does not join a waiting walk and says it runs beside it; and a tier asked in part is asked the rest, including a module held built that the walk never asked for, which is somebody else's build and not sent as the walk's.	2026-10-15 00:02:28 +02:00
jschoubben	e4acab32d6	End the check of a pull request's old head when a new head is checked ... A check judges one head; once the pull request has another, its verdict answers nobody and its run holds a build machine the new head waits behind. Asking a check of R#N at H now drops every waiting check of R#N at another head and kills every one in flight, saying each in the log. Announcements are unordered and a recheck asks a recorded head, so this happens only when the forge confirms H is the pull request's current head; otherwise nothing ends and that is said. A one-shot recheck ends them within the call. Tries, walk builds, group checks, other pull requests' checks and rechecks of the same head are left alone. novox/hq issue 726	2026-10-14 22:10:15 +02:00
jschoubben	b34055f580	Register the compared module after the fixture's own registration (review of #336)	2026-10-14 21:53:50 +02:00
jschoubben	480657d95b	Register the module a walk's store test compares with, as the mesh runs it (review of #336)	2026-10-14 21:43:20 +02:00
jschoubben	38c87a7e47	Compare a build with the one the mesh registered, not the newest, which may be one put back (review of #336, hq issue 721)	2026-10-14 21:30:09 +02:00
jschoubben	6acce60f8b	List the store test of a walk whose every build came out the same (hq issues 611, 721)	2026-10-14 21:23:46 +02:00
jschoubben	4390589c61	Leave out a dependent of a base that came out the same, and end a walk with every build the same as done (ADR 0267 rule 3; hq issues 607, 611, 721)	2026-10-14 21:19:14 +02:00
jschoubben	e6a28a2d28	Say why the walk moves every module when the checks' builds cannot be read (lint, hq issue 721)	2026-10-14 21:15:24 +02:00
jschoubben	32bd8b5d5f	Leave a module out of a walk when its check built it the same as what runs (hq issue 721) ... A one-sentence change in a package mesh-builder imports but never reaches moved build-agent. A merge check now compares each module it builds with the build the mesh holds, the whole image without file times, labels or creation time, and a walk is cut without a module every check of its pull requests found the same, while the mesh still holds that build. A comparison that fails moves the module as before.	2026-10-14 21:07:25 +02:00
jschoubben	34050dc6a3	Say a one-machine walk's report was its run: no gate reads it since ADR 0360 (hq issue 720)	2026-10-14 20:55:04 +02:00
jschoubben	3ec4473b29	Run the slow test packages in shards and stop compiling the standard library per judge test (hq issue 717) ... The controller's check spent 3-5 minutes in cmd/mesh-controller, whose few hundred store tests run one after another, while most packages end in seconds. merge-check.sh now compiles each slow package once and deals its tests round-robin into processes that run beside each other and beside the rest of the suite; every test still runs once, under the race detector, through its package's TestMain. The listed store tests run beside the suite instead of after it. internal/builder's judge tests ran a real go build with a fresh Go cache each, compiling the standard library every time: half the suite's CPU. The fake runtime fills an empty cache from one warmed per process. The lease tests run in parallel. The link test of slow work now sets its one-second window after the serving loop asserted the controller's thirty-second one, so it proves the heartbeat (it passed without it) in half the time.	2026-10-14 20:38:35 +02:00
jschoubben	bbee20adf9	Build a merge's modules of a tier at once, and beside the check's tests where the judge cache allows (hq issue 716) ... A check built its images one after another, after merge-check.sh. With a judge cache the builds run during the tests, out of their reach; a failed check lets go of them.	2026-10-14 20:20:49 +02:00
jschoubben	afed887217	Build a check's builds for its merge where the host's container engine sees them (hq issue 713) ... The check built in a temporary directory of the build holder's container, which the host's engine mounted empty: every check build failed and every walk built again.	2026-10-14 20:14:20 +02:00
jschoubben	704b21b3c5	Carry the new controller's user list to the bus when its walk ends, as the gate's pass did (hq ADR 0236 §34, ADR 0360)	2026-10-14 19:36:28 +02:00
jschoubben	5b455b8a24	Keep the first-machine gate behind a switch for walks begun before hq ADR 0360, and test both paths	2026-10-14 19:35:07 +02:00
jschoubben	0e97413875	Assert the backlog walk sends every machine at once and passes on the report (hq ADR 0360)	2026-10-14 19:21:42 +02:00
jschoubben	da8a5e9bc0	Send a merge's builds to every machine at once, with no first-machine gate (hq ADR 0360) ... The operator chose the leanest rollout: no gate, no one machine first, failures loud and fixed forward. A walk sends every rolled-out module to every machine running it in one step, built-by orders no tier, and a walk ends on its machines' reports. The backlog walk sends every machine at once and passes each on its applied report.	2026-10-14 19:19:49 +02:00
jschoubben	c707bf18d8	Say which check made a build the walk does not take, and test the merged heads ... A build of another pull request or of a head the forge did not merge is in no verdict the walk is handed; the walk now names the check that made it, so its "built:" line says which head it was (novox/hq issue 703).	2026-10-14 10:05:31 +02:00
jschoubben	060f2a9ca5	Build for the merge from a copy its tests never touch, and keep or take only the same head's build ... The gate and the repository's check mount the workspace read-write, so a build from that tree could hold what the tests wrote under a key that describes the merge. A check now clones the merge outside the workspace before anything of the change runs, and builds from that. A check keeps, and a walk takes, only a build an earlier check of the same pull request at the same head made: another pull request's build, or an earlier head's, is never registered. The walk reads the verdicts once (novox/hq issue 703, review of #328).	2026-10-14 09:56:23 +02:00
jschoubben	1666747a22	Keep a check build whose key is unchanged instead of building it again ... A fresh check asked when only the trunk moved rebuilt every image, though the reuse key said nothing it is built from changed: a controller check took about 12 minutes instead of 5.5. The check now keeps a build kept under its key when a verdict of the build seat names that key and every digest, names it as the build's own check so a walk takes it, and the cleanup keeps it while any check names it (novox/hq issue 703).	2026-10-14 09:41:39 +02:00
jschoubben	3ce5a0239b	Keep a try off the forge and bound how many run (hq issue 686) ... Review of #314: a try's start named its commit, which the forge's holder matches to an open pull request's head and marks pending. Its start now names no repository or commit. One try per commit and four in all, a try never ages ahead of deliveries or pull request checks, and its id is checked before it is a bus subject.	2026-10-14 02:29:28 +02:00
jschoubben	65508d9df9	Try a pushed branch on the build seat before its pull request (hq issue 686) ... Store tests skip without a store and no agent may start one, so each failure cost a 15-30 minute review round. The try verb asks the build seat for the repository's merge-check.sh on the ref merged into the trunk, as a check in the build queue, and reads its outcome from the build's log: no status on the forge, nothing registered, nothing built for a merge.	2026-10-14 02:20:37 +02:00
jschoubben	8123ca5654	Test that a branch is tried before its pull request (hq issue 686) ... A builder's store tests first ran in the pull request's merge check, a review round late. These fail on main: a try asks the repository's own check alone, its outcome sets no status and registers nothing, an unknown ref is refused before anything is asked, and the answer carries the failing lines.	2026-10-14 02:20:37 +02:00
jschoubben	b482122d66	Say a read error of a silent rest, and keep the node-engine one at a time in a kept step (hq ADR 0353) ... An empty answer on a read error would clear release-silent unseen.	2026-10-14 01:26:00 +02:00
jschoubben	2f5cc146fe	Keep a walk's sends to what was kept, and say a silent rest (hq ADR 0353 review) ... A release's own gates held its other sends back for ever; an adopted send lost what the machine ran before; a kept send could grow a build no gate passed; a take-over matched by commit alone; a rest that went silent ended done. The node-engine and the resolver still go one machine at a time, since a fault that cuts a machine off shows only as silence.	2026-10-14 01:18:51 +02:00
jschoubben	c2ea9b4c51	Send the rest at once after the first passes, and resume a stored gate (hq ADR 0353) ... A release now sends every other machine in one step once its first machine passed, each judged by its own gate; a failing one is put back without stopping the others, a silent one is left. A send is kept before it is made, and a machine already sent what a gate is for is not sent again: the gate counts from that send and takes over another walk's readings of the same build there.	2026-10-14 01:06:49 +02:00
jschoubben	b03d257f81	Test that the rest go at once after the first passes and a gate survives a new controller (hq ADR 0353) ... The release walked four machines one after another in eleven minutes, and a new controller re-sent and re-judged a build its machine already ran.	2026-10-14 01:06:49 +02:00
jschoubben	873b19566a	Read the delivery group's order again, now the catalogue and hq left it	2026-10-14 00:08:44 +02:00
jschoubben	b0753e2aac	Let a drain lapse, bound the gate's wait on the controller's clock, keep a busy wait out of apply time ... Review of #283: a drain whose replacement was withdrawn held the agent for ever with its health green; the gate trusted the engine's until; a machine of the rest waiting at the bound read as silent and its apply time held the wait; an unknown mesh-builder word exited 0; plan said nothing of a busy an older engine is not sent; and a node with a report and no send broke the reading of every last report (the store test's NULL).	2026-10-14 00:07:31 +02:00
jschoubben	a69c0e50dc	Merge origin/main into feat/335, and pin mesh-host at #98's new head	2026-10-14 00:01:21 +02:00
jschoubben	5c606fd1d4	Take only a check build its verdict names, and keep a build another check still names ... The artifact store takes writes from anything on its network, so a record planted under another merge's key could have been registered: the walk now takes a scratch build only when the build seat's own verdict for one of its pull requests names its key and digests. Deleting by digest also took builds another check shared (review of #311; novox/hq ADR 0350).	2026-10-13 23:16:55 +02:00
jschoubben	675466a8c3	Hold the store tests' modules as the settle time did, and fault a dip after any healthy reading ... Two tier tests relied on the two-minute settle to keep a module unjudged; the witness window and the gate's spacing now hold it. A healthy reading taken before the controller's floor counted no pass, so a dip after it did not take the two-minute pass away (novox/hq ADR 0352 decision 3).	2026-10-13 23:07:15 +02:00
jschoubben	cd49b50c40	Build a change once: the check builds what its merge moves, the walk takes it by key ... A merge built every moved module a second time although its check had judged the same tree (issue 649). The check now builds into the scratch namespace under a reuse key (source fingerprint plus the builder's own digest), and the walk of the trunk commit copies a build of the same key by digest instead of building, saying per module why when it cannot (novox/hq ADR 0350).	2026-10-13 23:04:40 +02:00
jschoubben	41eefc2f8c	Test that a healthy node-engine passes at two minutes on one reading	2026-10-13 22:58:37 +02:00
jschoubben	c3946e5d72	Pass the node-engine and resolver on their third reading, at most two minutes ... The operator wants these gates as fast as they can be: counting readings only after two minutes added about 80 s to every node-engine rollout for no safety. Count from the send, pass on the third healthy reading, and pass a module healthy at every reading at two minutes on what it has (novox/hq ADR 0352).	2026-10-13 22:55:46 +02:00
jschoubben	524295926f	Pass the controller's first-node gate on one reading after the witness window ... The controller runs on one machine, so three readings over two minutes guarded no other machine and held every walk behind it. Judge each module by its own rule: one reading for the controller and the tool runner once the node-engine's witness window has ended, one for mesh-delivery, three for the node-engine and the resolver with the first two minutes after the send (novox/hq ADR 0351). The gate's verdict names the rule it applied.	2026-10-13 22:39:32 +02:00
jschoubben	b1546ed28a	Hold a walk's first-machine gate while its busy container finishes its build ... A merge into mesh-controller replaced every build agent and threw away the checks in flight (issue 594). The manifest's container busy is checked and sent only to an engine that reads it; a report's busy waits are kept, the gate starts at the report that says the container replaced, and the wait is a phase of the delivery time. mesh-builder busy drains the agent and answers while a build is in hand (novox/hq ADR 0335).	2026-10-13 00:53:10 +02:00
jschoubben	ee1d7e5cae	Mark a rerun check with a short (run N) after its verdict, and word a re-hand-out neutrally ... The forge clips a status to 140 characters, so the long sentence pushed the verdict's own words out; and a paused holder's hand-back or a bus delay also count as a delivery, so the cause is not claimed (novox/hq issue 594).	2026-10-13 00:31:57 +02:00
jschoubben	43cab56452	Say a build handed out again restarted, in its log, the queue and the check's verdict ... A build agent replaced mid-build leaves its ask to be handed out again, and the second run looked like a first: the queue and the pull request showed nothing of the 5-8 minutes lost (novox/hq issue 594).	2026-10-13 00:26:08 +02:00
jschoubben	7f3ff32361	Give a fast-renewing test lease a serving controller's time to renew, and prove the minted hash's cost ... A renewal timed out at the renew interval, so the lease test's 1 s renewals could lose the lease on a slow build machine (novox/hq issue 591, review of #279).	2026-10-13 00:08:48 +02:00
jschoubben	2819915e8e	Stop the controller's tests waiting out fixed intervals ... Reading the build events waited two seconds on every queue read, the race detector's exit sleep cost a second per command line a test runs, and the lease and plan-wait tests waited out production intervals (novox/hq issue 591).	2026-10-12 23:53:06 +02:00
jschoubben	7892019d9e	Copy each store test's database from a template migrated once, and hash test bus passwords at bcrypt's least cost ... Running ~90 migrations into a fresh database per store test, and bcrypt at the default cost under the race detector, were most of the controller's check (novox/hq issue 591).	2026-10-12 23:53:06 +02:00
jschoubben	e727589e48	End a walk's apply phase at each machine's own report of its send (hq issue 582) ... The walk records its send to the rest only after every machine was sent, which takes 20-28 s, so the machine sent first fell outside the 2 s slack, its report was missed and the phase ended at the next walk's send to it. Keep each machine's own send time with what it was sent, match its apply duration by that send and digest, and show each machine's report in the walk's phases (ADR 0282 decision 6).	2026-10-12 23:18:28 +02:00
jschoubben	fc573f951a	Remove only a move's proven second copy, and never delete a handed-out original (hq issue 558) ... Second review of #276: a dead letter delivered again beside its original shares its request id and was removed as a copy, losing the ask; and a publish that outlived the pause could delete an original a holder had taken. A copy now names the message it copies, removals are said, the publish waits inside the pause's margin, and an original handed out meanwhile keeps its place while the copy goes.	2026-10-12 21:46:49 +02:00
jschoubben	1ded7dff34	Make a queue move lossless and run its passes on the acting controller alone (hq issue 558) ... Review of #276: a move deleted before it published, so a failed publish lost an ask; a verb's own process ran a pass beside the serving one, so one could resume the queue while the other moved asks and a build ran twice. The copy is now published first and second copies removed by id, asks only signal the acting controller's loop, moves stop with 5 s of the pause left, a cancel finds a moved ask by id, and an unreadable ask keeps its age.	2026-10-12 21:38:11 +02:00
jschoubben	50c55a773e	Mark that the build agent takes work, so its health check can see it ready (novox/hq issue 544) ... The agent listens on nothing, so the node-engine could judge it only alive and held it starting for the whole grace on every machine a walk moved it to. It now writes a mark once it is bound to its worker, after its seat verbs are served, and 'mesh-builder ready' passes only while the mark is there.	2026-10-12 21:33:12 +02:00
jschoubben	c6e234455f	Hand a delivery build out before pull-request checks (hq issue 558) ... A merged change's walk waited minutes on its build behind checks in a first-come queue. The queue now hands deliveries out first, a check that has waited ten minutes before deliveries asked after it, and says each ask's kind; the order is kept in the work queue itself.	2026-10-12 21:25:46 +02:00
jschoubben	875ebc093e	Format PullUpdated, and keep the announced base head on a rechecked member ... A head checked again through mesh-delivery reached the direct verdict with no base head, so a head no build judges was answered with no trunk commit and the merge tool would refuse it (novox/hq ADR 0323).	2026-10-12 17:52:12 +02:00
jschoubben	bc623bc1f9	Name the announced trunk commit on a verdict no build judges ... A change touching no module, in a repository with no merge-check.sh, was answered by the controller with no trunk commit, so the merge tool refused it, asked again and refused it for ever (novox/mesh-tools-go; novox/hq ADR 0323).	2026-10-12 17:44:18 +02:00
jschoubben	ce9dc8e157	Run no hook and no signature setting when a check makes its merge ... A build machine's or a clone's git configuration must not run code or change the merge a check judges (novox/hq ADR 0323).	2026-10-12 17:39:59 +02:00
jschoubben	25deb499bf	Say why a merge's changed files read as none on an error	2026-10-12 17:31:36 +02:00
jschoubben	83a9cf6e3a	Judge a pull request's head merged into its trunk, and name the trunk commit ... A check judged the head alone, so two pull requests that each passed could break the trunk together (node-engine #66 on 2026-10-09), and a recheck after the trunk moved judged the same head again. The build seat now merges the head into the trunk's head in its own clone, never pushed, fails a head that does not merge cleanly naming the files, and the verdict names the head, the trunk commit and the commits read beside it, for the forge and mesh-delivery to weigh against the trunk's next move (novox/hq ADR 0323, issue 341).	2026-10-12 17:27:01 +02:00
jschoubben	7b050abdcb	Stop composing ADR 0312's group grant, so deliveries to the media machine pass again ... Since #256 delivered, the node-engine judges every read-write access by the path's group alone and states six media modules unhealthy on ace, though they write their folders as the owner; every delivery there fails its gate (hq issue 567). ADR 0325 replaces ADR 0312; until it is built, the controller sends no grant, as before #256. This reverts commit 3cfc121670.	2026-10-12 17:25:46 +02:00
jschoubben	0cd8bc0bfb	Answer the build queue as data too, so a check can say how far back it waits ... The queue verb answered only a table, so nothing outside the controller could say an ask's place in the queue. It now asks its command for JSON: each ask's state, kind, repository, commit, machine, since and position among the waiting, with the table carried beside it and answered as the verb's output, both written from one read. A repository's password, if one were ever written into it, is left out of what is shown. novox/hq issue 455	2026-10-12 14:22:22 +02:00
jschoubben	7385d9c092	Ask a node's engine which of its directories the walk would refuse, read-only (issue 358) ... Before the node-engine gives a mesh directory its mode and owner only through a walk that follows no link another account could make, the operator is to see which directories that walk would refuse on every node, the control-node included, where no tool reads a path. `node judge-directories <node>` signs the ask as a hand-over is, under its own context, and prints the engine's one line per directory; it changes nothing, so it is a seat verb an agent may call. The node is granted the subject, the controller alone its publish.	2026-10-12 14:17:01 +02:00
jschoubben	a23b190747	Merge remote-tracking branch 'origin/main' into fix/451-a-recheck-is-the-step	2026-10-12 13:56:52 +02:00
jschoubben	74bd90df7e	Pin the node-engine at its merge of the access grant, and capture it beside (hq ADR 0312, issue 537)	2026-10-12 13:47:28 +02:00
jschoubben	30599c1858	Pin the node-engine at its reviewed grant, and hold the mount match to the path itself (hq ADR 0312, issue 537 review) ... A path beginning with the same letters is another path, and a read-only mount of a read-write access is still granted: the grant follows what mounts it.	2026-10-12 13:47:17 +02:00
jschoubben	10d4712f82	Name who a read-write access grants its path's group to, so modules write shared data through the group the operator gave it (hq ADR 0312, issue 537) ... Each read-write access is composed with a grant: the module's containers that mount its path and the account its tools run as (the operator's, or the module's own under ADR 0259 §8); that runtime restarts on the access, so it holds a group the engine put its account in. A read access grants nothing. Sent only to an engine whose statement says GrantContract (4), since an older one refuses the field. Pins the node-engine at the branch that reads it.	2026-10-12 13:47:17 +02:00
jschoubben	e737e13bbb	Advise a recheck for an unanswered check on a ready or rejected delivery (issue 451) ... That check answered once, so the build queue is the wrong place to look. mesh-delivery names the delivery's state as data (held); the controller reads it and never the words. A line without it keeps the advice of 438.	2026-10-12 13:47:11 +02:00
jschoubben	a3181e356b	Ask only once the bus enforces the controller's own grant, judged for real and failing closed (review of hq issue 353) ... change held asking back, a list sent was taken for a list the bus enforces, a plan that could not be worked out read as not behind, and no test held the real judgement. Now the asker asks only when the composition grants the controller itself to publish its ask, and the bus's own module says that grant is in force on the running server (nats_user_can's in_force, mesh-catalog); each fact not known is not held, said with why. brokerBehind takes a list nobody could compose as behind.	2026-10-12 13:21:32 +02:00
jschoubben	71b3aeaf2c	Keep a group member of the other kind's batch from being overtaken (hq issue 525) ... A merge's group member counted as merged when its merge waited in any batch, but kinds never share a batch: the later member got a batch of its own, cut at once, and walked before the earlier member's batch was cut, breaking the group's order (ADR 0276 §5). Only a member kept in a batch of the merge's own kind counts as merged now; one kept in the other kind's keeps the merge waiting, and its batch names that member, so the two are cut in the order their windows close (review of #250).	2026-10-12 12:54:11 +02:00
jschoubben	6e5a1df575	Judge the installer's user list against mesh-host with the groups grant (hq issue 525) ... The controller now asks mesh-delivery.groups, and mesh-host a76baa4b carries that grant in the installer's first user list (mesh-host #85). The captured copy is moved to that commit so the genesis test judges the list that ships.	2026-10-12 12:31:50 +02:00
jschoubben	4d203b5a5d	Ask mesh-delivery whether a merge's group still assembles (hq issue 525) ... A merge from any head waited for the window because the controller could not see a group's members, so the window skip changed nothing in production and unrelated merges walked one after another. The controller now asks mesh-delivery's groups, within 3 s and before the plans are held: a merge waits only when its head's group has another member whose pull request is open and whose merge it does not keep yet (ADR 0306 §2). No answer means it waits, and its batch says why. A headed merge in no waiting group and related to no batch is cut at once, as one from no head is (ADR 0304 §1, ADR 0306 §3). The controller's grant names mesh-delivery.groups. The installer's first user list in mesh-host must carry it too before the genesis test can pass.	2026-10-12 12:14:46 +02:00
jschoubben	6196a4815e	Let a merge that shares nothing skip the merge window (hq issue 525) ... A merge waited the whole window even when nothing assembling shared anything with it. It now joins only a batch it is related to (ADR 0304 §1, read by ADR 0306 §3: repository, a module the cut would move, a graph edge, a group head), folding the newer into the oldest when it relates to two. A merge related to none starts a batch whose window is closed at once (PlanBatch.CutAtOnce), unless its delivery group may still have a member to merge (ADR 0306 §2). The controller cannot see a group's unmerged members (that is mesh-delivery's record), so a merge from any head still waits, in a batch of its own: the safe error is to wait. Walks stay one at a time; a batch cut at once queues behind an open walk. A stop's merges walked again join a related batch, or their own.	2026-10-12 12:14:46 +02:00
jschoubben	ac5d34f613	Give the unheld log test its own writer and memory, so it passes on a repeat (novox/hq issue 486) ... logUnheld kept what it had logged in package state and the test swapped os.Stderr, so a second run in one process logged nothing and failed.	2026-10-12 11:41:41 +02:00
jschoubben	b03ec3b64e	Offer no hand start for a walk that is only waiting its turn (issue 529) ... A hand start there would skip its group's order; point at it only for a walk nothing holds, and say why nothing holds it.	2026-10-12 11:34:54 +02:00
jschoubben	4aacd98f8d	Cap the forge wait as a whole, and wait out only a forge not served (novox/hq issue 486) ... The registration holds the controller's one receive loop. Checking the bound only between asks let an ask in flight hold it near two minutes, the deaf-controller limit. The whole wait now runs under a 60 s deadline, still says it waited when that deadline ends it, and a tool that did not answer in time is not asked again: it has already held the loop its whole wait.	2026-10-12 11:32:14 +02:00
jschoubben	ed037784e2	Name the hold a waiting walk waits on in plans (issue 529) ... A walk whose delivery was held for a person said only that it waits for mesh-delivery's word and that plans go starts it by hand, pointing the reader at a hand start instead of the person's decision. Ask mesh-delivery what holds each waiting walk and say the held delivery, its reason and mesh-delivery.release, or the delivery not at its turn.	2026-10-12 11:27:05 +02:00
jschoubben	47205403d3	Wait out a forge not served for a moment when registering a build (novox/hq issue 486) ... A registration asked the forge while node-tools re-served after a tool-runner rollout, heard "nothing serves gitea.gitea_api", and failed a good build's plan. The bus's not-served answer is now link.ErrNothingServes (and a timeout link.ErrNoAnswer), and the registration asks again for up to 90 s on those alone, never on a refusal; past the bound it fails saying how long it waited.	2026-10-12 11:23:03 +02:00
jschoubben	cde0dc287b	Say a restart count runs from when the node-engine began watching it ... The engine reports its own count, kept across recreates and starts by hand, not systemd's NRestarts or the runtime's RestartCount (novox/hq issue 458).	2026-10-12 10:46:38 +02:00
jschoubben	63079e2c67	Say a restart count is automatic restarts since the last start or creation ... NRestarts and RestartCount reset on a start by hand or a recreation, so "whole life" was as wrong as "since" (novox/hq issue 458).	2026-10-12 10:45:40 +02:00
jschoubben	0d08d5c730	Say a unit's restart count is its whole life, not since its state began ... The count comes from systemd's NRestarts or the runtime's restart count; beside "since" it read as thousands of restarts tonight (novox/hq issue 458).	2026-10-12 10:42:19 +02:00
jschoubben	9460689fff	Relate a batch whose moves are not known to every change ... The review found that a merge kept without its moves, judged again at the cut, gave its batch no modules, so it could read as unrelated to a walk it will build beside. An unknown change is never unrelated.	2026-10-12 02:06:54 +02:00
jschoubben	96005a8bfd	Say what relates two changes, so unrelated ones can walk beside each other ... The helper decides, by repository, module (widened to dependents), delivery group head or module graph edge, whether two changes are related, as novox/hq ADR 0304 decision 1 reads in ADR 0306 decision 3. Nothing calls it yet; it is the first step of issue 525's parallel walks.	2026-10-12 02:04:38 +02:00
jschoubben	e6bd8f1dee	Say a leaf delivery's budget as three minutes in the times verb's description	2026-10-12 01:46:26 +02:00
jschoubben	5b9c81bf26	Hold a leaf delivery to three minutes (hq ADR 0304 decision 4, issue 525) ... ADR 0304 lowered a leaf module's delivery budget from five minutes to three, the time its measured phases support once a leaf is judged on one reading and waits only for what it shares.	2026-10-12 01:44:18 +02:00
jschoubben	d8ec8baa22	Judge a send on no machine as the core, and test a send carrying the core with the real classes ... The review found that a send whose first machine runs none of its modules would pass at once with no reading, and that every mixed send in the tests judged its leaf as the core.	2026-10-12 01:38:09 +02:00
jschoubben	4ee131e990	Let the controller ask an issued credential's tool in its own name, and seal it again over a remade pair (hq issue 526, ADR 0305) ... The bus left the controller out of the caller-named grants, so issued could never reach plex's tool; the controller is now granted exactly the tools of offers whose software issues the credential, composed from the records, so the installer's first user list is unchanged. A pair remade on a key change or accepted by hand kept the issued stamp and was never sealed again; those writes now clear it.	2026-10-12 01:21:20 +02:00
jschoubben	faf8b76b84	Cite ADR 0305's clauses by their accepted-for-review numbers (issue 526)	2026-10-12 01:09:26 +02:00
jschoubben	ad3b4f185b	Judge a leaf module on one healthy reading, a core module on three (hq issue 524) ... ADR 0282 decided readings by class and ADR 0304 took a leaf module to one, but the gate never learned it: every build, leaf or core, needed three readings over at least two minutes, so a container module that was healthy at once still waited about two and a half minutes on its first machine. The number of readings and the settle time now follow the module's class: three after two minutes for the controller's own path and the resolver, one for every other module. Tests that judge a gate's other rules keep leaf modules on three readings.	2026-10-12 01:08:14 +02:00
jschoubben	db86dd52c8	Take a provider's software-issued credential from the provider, sealed (hq issue 526, ADR 0305) ... A token a media server got from its vendor reached its consumers only as a value the mesh minted, which the server refused, or as a copy typed by hand. An offer may now say its software issues the credential; the new issued verb carries no value and asks the provider's tool, in the controller's own name, for the value sealed to one call, then seals it to every pair that does not hold it yet.	2026-10-12 01:07:16 +02:00
jschoubben	e25fdf776f	Make the store-tests step fail on a malformed line, a failed go test and a shortened list ... The step let tests escape it: a one-field line ran nothing, go test's exit status was ignored, a pull request could drop entries main lists, and a result written mid-line by go test -v read as NOT RUN, hiding a FAIL (novox/hq issue 515).	2026-10-12 00:28:32 +02:00
jschoubben	6d2138217f	Let several modules declare one package, and refuse only present beside absent (hq ADR 0303) ... Assigning the media-importer beside youtube was refused because both declare ffmpeg. The node-engine installs a missing package and never removes one a declaration stopped naming, so a second declaration changes nothing on the machine; what two modules can contradict is one declaring a package present and another absent, which stays refused.	2026-10-11 21:55:20 +02:00
jschoubben	0fb7de7ef8	Merge remote-tracking branch 'origin/main' into feat/459-a-stop-re-walks-the-others ... # Conflicts: # testdata/store-tests	2026-10-11 21:06:47 +02:00
jschoubben	bc4662d656	End the line a walk's ask leaves open as each ADR 0299 store test ends, so its pass is read (issue 459) ... A walk's ask prints " tier N: " and the build ask the rest; the tests' fake ask prints nothing, so go test -v put three tests' --- PASS after it and the store-tests step read them as not run.	2026-10-11 21:06:44 +02:00
jschoubben	8d73f705af	Wait for the build's log lines to be stored before counting them (hq issue 507) ... A line is said with a plain publish, so a live subscriber can hold both lines while the stream has stored one or none yet; the test read the stream once and failed under load. It now waits up to 5 s for both.	2026-10-11 21:05:04 +02:00
jschoubben	0d4d94ff24	Stop tracking built binaries at the repository root, and ignore them (hq issue 470) ... Built artifacts belong in build/; three binaries (51 MB) were committed by a local build and rode in every clone and build context.	2026-10-11 20:52:45 +02:00
jschoubben	e2801773c3	Take an ordered engine's report as a send's answer only by its epoch and sequence (novox/hq issue 485) ... A periodic report about an identical earlier send names the same digest and can arrive after a later send the machine never received, so the digest and the time read it as applied. The digest-and-time rule stays for engines that report no order. Tests now hold the digest condition and the sequence one.	2026-10-11 20:44:54 +02:00
jschoubben	fae6bd0a87	Say each machine's last send in status and node show, and as fields in status --json (novox/hq issue 485) ... The send record of issue 234 is now readable only through the controller's verbs (issue 506), and only plan said it. status says per machine when it was last sent a declaration, by whom, from which generation and what it answered; node show says it in full with the refusal's words. A machine's answer is read from its report only when that report is about the send and came after it.	2026-10-11 20:40:48 +02:00
jschoubben	201bd28278	Keep a walk's withheld modules on its record, name every machine a stopped send reached, and list the stop's store tests (review of #232)	2026-10-11 20:25:20 +02:00
jschoubben	5446be1abb	Merge remote-tracking branch 'origin/feat/459-store-tests-are-proven' into feat/459-a-stop-re-walks-the-others	2026-10-11 20:23:14 +02:00
jschoubben	932203df20	Leave the store tests' proof to the check of its own (#233)	2026-10-11 20:23:14 +02:00
jschoubben	949990f5a8	Merge remote-tracking branch 'origin/main' into feat/459-a-stop-re-walks-the-others	2026-10-11 20:23:12 +02:00
jschoubben	731143f5b9	Fail the repo check when a listed store test did not run against the store (hq issue 459) ... A test that needs the store skips without one, and a skip passes unseen, so a change's store tests could pass the check without ever running. Tests named in testdata/store-tests must each run and pass, said by name.	2026-10-11 20:22:56 +02:00
jschoubben	309475bf6f	Walk a stopped walk's merges again in the open batch, in one act, and say what stays on the machines (review of #232, hq issue 459) ... A new batch beside the open one walked an older commit after a newer one, and a stop that failed partway could leave a merge on two plans. A module shared with the stopped merge, built at its branch, would deliver it; it is left out or withheld. A batch not yet cut can drop a merge, so mesh-delivery can stop a merge walked again. The stopped walk names what it left on which machines.	2026-10-11 20:12:27 +02:00
jschoubben	fbfdded74b	Let delivery-stop leave merges out and walk the others again, so a group stop ends only its own change (hq issue 459, ADR 0299) ... One walk carries every merge of its batch, so ending it for one group member ended every other delivery on it. delivery-stop now takes without: the walk ends, every other merge it answered is batched again, and each later merge of a left-out merge's repository is left out too, since it contains it.	2026-10-11 19:54:12 +02:00
jschoubben	eb8233ac7c	Give the grants step's restic fixture a check beside its tool, as the module rules require (repo-check of #230) ... A tool check alone is refused, so the store test's c2 never registered; the fixtures are now parsed by a test that needs no store.	2026-10-11 19:43:50 +02:00
jschoubben	325575b292	Keep the builds of the grants step's machine alone, not of every machine its composition resolves (repo-check of #230) ... Composing the bus's machine resolves the others on the same context, and a machine never recorded as sent refused the whole composition.	2026-10-11 19:30:10 +02:00