Builds the router part of novox/hq ADR 0308 (while a delivery walks, what it touches is held from the operator and told once, with the walk's result, when it ends) for issue 532, and fixes issue 531 (every restart of the tool runner popped the four silenced conditions up on the desk again, twice, one of them as "A mesh message was withheld").
Merge order: mesh-controller #247 (adr-0308-walk-touches, the walk events) first, then this. Merged alone, this holds nothing: no plan-moved carries phase yet, so no walk opens; the §6 fix works on its own.
What it reads (§1)
The controller's plan-moved (now declared in consumes). Only a body with phase is a walk's start or end; any other plan-moved is ignored. Fields read, as #247 says them: walk (else id), phase (started|ended), touches[{machine, modules, carried}] (modules is read; carried is listed there too), outcome (delivered|failed|stopped, required on the end), bound_seconds (missing or 0: 900), budget_class, started_at (on both, the bound runs from it), ended_at (else at, else updated). A body with a phase that cannot be read is refused by name, like an unreadable condition event. An event named walk with the same body is read the same way. Code: cmd/messenger/walks.go.
What it does
§2 Held: while a walk is open and within its bound, a message (raise, change, reminder, escalation, reopen, the sync's urgent-never-said, a notice) whose subject is a machine the walk touches, or a module it touches, is held instead of sent. It is kept on the walk with held-at and what was to be said, and operator-channel.history records it with outcome held. If the subject names its machine, a module is matched on that machine only. A message touched by overlapping walks is kept on each and told only when the last of them ends or passes its bound. operator-channel.open shows held_by_walks.
§3 Never held: an ask's own message, and a condition an open ask is about; kinds secret-given / secret-replaced; subjects no open walk touches; a walk's own failure (scopes plan, walk, delivery, build, release, and anything naming an open walk's id); a clearing or a silence edit.
§4 End: one message, sent alone (never folded into a digest, not held to the desk's 15-minute rhythm): Delivered|Delivery failed|Delivery stopped: <modules> on <machines>, It took N min., Still open after it: with one line per held message still open, and N problem(s) it raised resolved during it. Held messages it tells are marked said, so nothing more is sent of them. If the walk delivered with nothing held and nothing open, the message is one line (Delivered: … , in N min). It is sent nowhere on a kind listed in the new setting only-what-needs-me (a list of kinds, empty by default; no channel level existed, so I added this as a small design limit). The history says why.
§5 Bound: at started_at + bound_seconds, the walk's held messages still open are sent with The walk <id> is still running, since HH:MM. (the id is dropped if the content rule refuses it), and the cleared ones are counted. After that the walk holds nothing. A walk whose end never comes is forgotten a day after it started. A start heard after its bound has passed opens nothing. An end heard more than 10 min late with nothing open closes quietly.
Held state survives a router restart: the walks are kept in a new state bucket walks, one key per walk, read back by Load.
§6 / issue 531, cmd/messenger/restart.go. I found three causes:
Twice: the desk's standing gives the latest notifier start of any machine. Two machines' notifiers started a second apart, so two standings came and each made a pass at once. Now a loss is acted on once it has stood for 10 s (LostSettle), so the standings of one restart make one pass.
Again after the router's own restart: the record of which loss was handled was in memory (shows, holder.go:332-351). It is now kept in the state sent, key passes, and read back on start.
What it said: the pass edited every silenced message (restart.go:99-105), which showed it on the desk again; the withheld one came back as "A mesh message was withheld". Now the pass says again only what is open, asked there, never answered, not withheld for its content and not held by a walk. Everything else is left as it is. A failed pass is retried after a minute, not on every 5-second tick.
Tests (walks_test.go; ADR 0308 "How it is checked")
§2: TestDuringAWalkWhatItTouchesIsHeldAndTheHistorySaysSo (a module's condition, a machine's, and a notice: nothing sent, history held, kept open). TestAMessageTouchedByTwoOverlappingWalksIsHeldUntilTheSecondEnds.
§3: TestWhatAWalkNeverHoldsIsSentAtOnce (secret given and replaced on a touched machine, an untouched machine, a delivery's and the plan's own failure on a touched machine, an ask, and an ask about a held condition).
§4: TestAWalksEndIsToldOnceWithWhatIsStillOpen (outcome, reach, 7 min, two open lines, one cleared counted, nothing said again for 20 min). TestADeliveredWalkWithNothingHeldIsOneLineOrNothing (one line; nothing on an only-what-needs-me kind, with the history saying so).
§5: TestAHoldEndsAtTheWalksBound, table-driven from the controller's bound_seconds: 1200 gives 20 min, 600 gives 10 min, 360 gives 6 min, and none gives 15 min. It checks nothing is sent before the bound, the still-running line with since HH:MM, the cleared count, and that nothing is held after the bound.
Restart: TestHeldMessagesSurviveARouterRestart.
§6 / 531: TestARestartSaysNothingItHasSaid. Four silenced messages are shown on the desk, one withheld for a file path, plus a fifth open, asked and withheld. Two standings come a second apart with a tick between: nothing is sent, edited or replaced, and exactly one (answers lost) pass runs. The router is then restarted from the same state and the standing repeated: still one pass, nothing said.
The decoder: TestTheWalkEventsAreReadAsTheControllerSaysThem (#247's start and end bodies, carried, ended_at, an ordinary plan-moved, an end without an outcome refused, the default bound, a start older than its bound).
Changed: restart_test.go. Issue 384's tests asserted that the silenced message is edited again; ADR 0308 §6 reverses that. The router's restart is now simulated by reading the state back.
Mutations run: turning off holding, never-held, LostSettle or the withheld skip each fails a test.
Checks.go vet and go test -race pass (5 runs, plus -count=30). MESH_GATE=<controller built from main d854598> MESH_CHECK_MODULES=messenger sh merge-check.sh passes locally: words, agent-home, paths, every manifest through module check (69 long-running resources without health, as kept), and the module's tests. A trial merge with #178 (issue 383, which also touches messenger) is clean (git merge-tree).
Reviewers' checklist
The one choke point for new messages is deliver (plus the sync's urgent path and edit for reopen and silence-over). A clearing or a silence edit is never held, and nothing in the asks path is held (hold refuses asked, an ask's own record, and a record with an open ask).
subjectOf matching: for scopes machine, node and bus the machine is the first token. For module, core and seat the module is the first token and the machine the second. Any other scope and a notice match each of their names as either. Too broad a match only delays a message by at most the bound; a missed one is sent as today.
The bound can never silence the mesh: a walk past its bound holds nothing new, a walk whose end never comes is forgotten after a day, and a start heard after its bound opens nothing.
The end message's lines are each message's own title, made under the content rule (say). The whole message is checked again, and a refusal falls back to plain words.
§6: the pass leaves silenced, withheld (Refused), answered and held records alone. passes is persisted, and LostSettle (10 s) merges one restart's standings into one pass.
Manifest: consumes gains mesh-controller.plan-moved, state gains walks, and the settings default gains only-what-needs-me: []. No node name appears in the manifest.
Live check after both merges (ADR 0308): the next tool-runner delivery should give the operator one message at its end, and operator-channel.history should show held entries and their release, with one (answers lost) pass per notifier restart and no silenced edits.
By adr0308@g14. Not merged; waits for an independent review and for mesh-controller #247.
Builds the router part of novox/hq **ADR 0308** (while a delivery walks, what it touches is held from the operator and told once, with the walk's result, when it ends) for **issue 532**, and fixes **issue 531** (every restart of the tool runner popped the four silenced conditions up on the desk again, twice, one of them as "A mesh message was withheld").
**Merge order: mesh-controller #247 (`adr-0308-walk-touches`, the walk events) first, then this.** Merged alone, this holds nothing: no plan-moved carries `phase` yet, so no walk opens; the §6 fix works on its own.
## What it reads (§1)
The controller's `plan-moved` (now declared in `consumes`). Only a body with `phase` is a walk's start or end; any other plan-moved is ignored. Fields read, as #247 says them: `walk` (else `id`), `phase` (`started`|`ended`), `touches` `[{machine, modules, carried}]` (`modules` is read; `carried` is listed there too), `outcome` (`delivered`|`failed`|`stopped`, required on the end), `bound_seconds` (missing or 0: 900), `budget_class`, `started_at` (on both, the bound runs from it), `ended_at` (else `at`, else `updated`). A body with a `phase` that cannot be read is refused by name, like an unreadable condition event. An event named `walk` with the same body is read the same way. Code: `cmd/messenger/walks.go`.
## What it does
- **§2 Held:** while a walk is open and within its bound, a message (raise, change, reminder, escalation, reopen, the sync's urgent-never-said, a notice) whose subject is a machine the walk touches, or a module it touches, is held instead of sent. It is kept on the walk with held-at and what was to be said, and `operator-channel.history` records it with outcome `held`. If the subject names its machine, a module is matched on that machine only. A message touched by overlapping walks is kept on each and told only when the last of them ends or passes its bound. `operator-channel.open` shows `held_by_walks`.
- **§3 Never held:** an ask's own message, and a condition an open ask is about; kinds `secret-given` / `secret-replaced`; subjects no open walk touches; a walk's own failure (scopes plan, walk, delivery, build, release, and anything naming an open walk's id); a clearing or a silence edit.
- **§4 End:** one message, sent alone (never folded into a digest, not held to the desk's 15-minute rhythm): `Delivered|Delivery failed|Delivery stopped: <modules> on <machines>`, `It took N min.`, `Still open after it:` with one line per held message still open, and `N problem(s) it raised resolved during it.` Held messages it tells are marked said, so nothing more is sent of them. If the walk delivered with nothing held and nothing open, the message is one line (`Delivered: … , in N min`). It is sent nowhere on a kind listed in the new setting **`only-what-needs-me`** (a list of kinds, empty by default; no channel level existed, so I added this as a small design limit). The history says why.
- **§5 Bound:** at `started_at + bound_seconds`, the walk's held messages still open are sent with `The walk <id> is still running, since HH:MM.` (the id is dropped if the content rule refuses it), and the cleared ones are counted. After that the walk holds nothing. A walk whose end never comes is forgotten a day after it started. A start heard after its bound has passed opens nothing. An end heard more than 10 min late with nothing open closes quietly.
- **Held state survives a router restart:** the walks are kept in a new state bucket `walks`, one key per walk, read back by `Load`.
- **§6 / issue 531**, `cmd/messenger/restart.go`. I found three causes:
1. *Twice:* the desk's standing gives the latest notifier start of any machine. Two machines' notifiers started a second apart, so two standings came and each made a pass at once. Now a loss is acted on once it has stood for 10 s (`LostSettle`), so the standings of one restart make one pass.
2. *Again after the router's own restart:* the record of which loss was handled was in memory (`shows`, holder.go:332-351). It is now kept in the state `sent`, key `passes`, and read back on start.
3. *What it said:* the pass edited every silenced message (restart.go:99-105), which showed it on the desk again; the withheld one came back as "A mesh message was withheld". Now the pass says again only what is open, asked there, never answered, not withheld for its content and not held by a walk. Everything else is left as it is. A failed pass is retried after a minute, not on every 5-second tick.
## Tests (`walks_test.go`; ADR 0308 "How it is checked")
- §2: `TestDuringAWalkWhatItTouchesIsHeldAndTheHistorySaysSo` (a module's condition, a machine's, and a notice: nothing sent, history `held`, kept open). `TestAMessageTouchedByTwoOverlappingWalksIsHeldUntilTheSecondEnds`.
- §3: `TestWhatAWalkNeverHoldsIsSentAtOnce` (secret given and replaced on a touched machine, an untouched machine, a delivery's and the plan's own failure on a touched machine, an ask, and an ask about a held condition).
- §4: `TestAWalksEndIsToldOnceWithWhatIsStillOpen` (outcome, reach, 7 min, two open lines, one cleared counted, nothing said again for 20 min). `TestADeliveredWalkWithNothingHeldIsOneLineOrNothing` (one line; nothing on an `only-what-needs-me` kind, with the history saying so).
- §5: `TestAHoldEndsAtTheWalksBound`, table-driven from the controller's `bound_seconds`: 1200 gives 20 min, 600 gives 10 min, 360 gives 6 min, and none gives 15 min. It checks nothing is sent before the bound, the still-running line with since HH:MM, the cleared count, and that nothing is held after the bound.
- Restart: `TestHeldMessagesSurviveARouterRestart`.
- §6 / 531: `TestARestartSaysNothingItHasSaid`. Four silenced messages are shown on the desk, one withheld for a file path, plus a fifth open, asked and withheld. Two standings come a second apart with a tick between: nothing is sent, edited or replaced, and exactly one `(answers lost)` pass runs. The router is then restarted from the same state and the standing repeated: still one pass, nothing said.
- The decoder: `TestTheWalkEventsAreReadAsTheControllerSaysThem` (#247's start and end bodies, `carried`, `ended_at`, an ordinary plan-moved, an end without an outcome refused, the default bound, a start older than its bound).
- Changed: `restart_test.go`. Issue 384's tests asserted that the silenced message is edited again; ADR 0308 §6 reverses that. The router's restart is now simulated by reading the state back.
- Mutations run: turning off holding, never-held, `LostSettle` or the withheld skip each fails a test.
**Checks.** `go vet` and `go test -race` pass (5 runs, plus `-count=30`). `MESH_GATE=<controller built from main d854598> MESH_CHECK_MODULES=messenger sh merge-check.sh` passes locally: words, agent-home, paths, every manifest through `module check` (69 long-running resources without health, as kept), and the module's tests. A trial merge with #178 (issue 383, which also touches messenger) is clean (`git merge-tree`).
## Reviewers' checklist
- [ ] The one choke point for new messages is `deliver` (plus the sync's urgent path and `edit` for reopen and silence-over). A clearing or a silence edit is never held, and nothing in the asks path is held (`hold` refuses `asked`, an ask's own record, and a record with an open ask).
- [ ] `subjectOf` matching: for scopes machine, node and bus the machine is the first token. For module, core and seat the module is the first token and the machine the second. Any other scope and a notice match each of their names as either. Too broad a match only delays a message by at most the bound; a missed one is sent as today.
- [ ] The bound can never silence the mesh: a walk past its bound holds nothing new, a walk whose end never comes is forgotten after a day, and a start heard after its bound opens nothing.
- [ ] The end message's lines are each message's own title, made under the content rule (`say`). The whole message is checked again, and a refusal falls back to plain words.
- [ ] §6: the pass leaves silenced, withheld (`Refused`), answered and held records alone. `passes` is persisted, and `LostSettle` (10 s) merges one restart's standings into one pass.
- [ ] Manifest: `consumes` gains `mesh-controller.plan-moved`, `state` gains `walks`, and the settings default gains `only-what-needs-me: []`. No node name appears in the manifest.
- [ ] No conflict with #178 (issue 383).
**Live check after both merges** (ADR 0308): the next tool-runner delivery should give the operator one message at its end, and `operator-channel.history` should show `held` entries and their release, with one `(answers lost)` pass per notifier restart and no silenced edits.
By adr0308@g14. Not merged; waits for an independent review and for mesh-controller #247.
A delivery's own restarts reached the operator as notifications while it ran, and every restart of the
tool runner popped the four silenced messages up again on the desk, twice. The router now reads the
controller's walk start and end on plan-moved, holds what an open walk touches until its end or bound,
tells the end as one message, and makes the restart pass once per restart from its own state, leaving
silenced, withheld, answered and held messages as they are.
Deliverynovox/mesh-catalog@a92c08fd5b55 — delivered since 2026-10-12T11:25:44Z
Delivery plan — builds messenger → novox; no bus step
tier 0: messenger
novox: receives messenger
Machines
messenger on novox: passed — healthy 1 times over 1m26s
messenger on the rest running it: sent
Transitions
2026-10-12 10:08 (new) → proposed (announced): novox/mesh-catalog#230's head announced
2026-10-12 10:10 proposed → checked (checked): the verdict names this commit
2026-10-12 10:10 checked → ready (accepted): the gate passed or warned, and the repository's own check did not fail
2026-10-12 11:23 ready → published (merged): on the trunk its modules follow: merged there, and its walk opened — or nothing for a walk to move
2026-10-12 11:23 published → delivering (go): its walk started: let go by this owner in its group's order, by a person, or on the controller's own path — not a failed walk that carried a later commit for it, after which it waits for a walk of its own commit
2026-10-12 11:25 delivering → delivered (done): its walk is done, carrying its own commit: every machine of its deploy plan passed, or was left as its policy says
The commit's note under refs/notes/mesh-plan keeps every transition: git log --notes=mesh-plan.
<!-- mesh-delivery:view -->
**Delivery** `novox/mesh-catalog@a92c08fd5b55` — **delivered** since 2026-10-12T11:25:44Z
**Delivery plan** — builds messenger → novox; no bus step
- tier 0: messenger
- novox: receives messenger
**Machines**
- messenger on novox: passed — healthy 1 times over 1m26s
- messenger on the rest running it: sent
**Transitions**
- 2026-10-12 10:08 (new) → proposed (announced): novox/mesh-catalog#230's head announced
- 2026-10-12 10:10 proposed → checked (checked): the verdict names this commit
- 2026-10-12 10:10 checked → ready (accepted): the gate passed or warned, and the repository's own check did not fail
- 2026-10-12 11:23 ready → published (merged): on the trunk its modules follow: merged there, and its walk opened — or nothing for a walk to move
- 2026-10-12 11:23 published → delivering (go): its walk started: let go by this owner in its group's order, by a person, or on the controller's own path — not a failed walk that carried a later commit for it, after which it waits for a walk of its own commit
- 2026-10-12 11:25 delivering → delivered (done): its walk is done, carrying its own commit: every machine of its deploy plan passed, or was left as its policy says
The commit's note under `refs/notes/mesh-plan` keeps every transition: `git log --notes=mesh-plan`.
mesh/merge-gate: PASS — every machine composes with the change as it did without (4 of 4 compose) (modules: messenger)
mesh/repo-check: PASS — its merge-check.sh passed, each part in its toolchain: merge-check.sh (go), merge-check-typescript.sh (typescript)
Delivery plan — builds messenger → novox; no bus step
tier 0: messenger
novox: receives messenger
read by no module's build: modules/messenger/README.md, modules/messenger/cmd/messenger/holder_test.go, modules/messenger/cmd/messenger/manifest_test.go, modules/messenger/cmd/messenger/restart_test.go, modules/messenger/cmd/messenger/walks_test.go
Run by the build seat on novox as build-1791799044893033135 (builds --log build-1791799044893033135).
--- the gate: messenger
messenger: ok, 7 tool(s), invokes seat:mesh-controller.conditions, seat:mesh-controller.root-free, seat:issue-tracker.tracking, keeps state open, sent, asks, identities, walks; WARNING: settings ask(s) for a setting and do(es) not say whether it is trusted, so it counts as trusted: set at the terminal alone; say "trusted" false where nothing trusts it (novox/hq issue 339)
long-running resources without health: 0
files asking for a setting without saying whether it is trusted: 1
1 manifest(s) checked. Judged against the seats this binary carries; a claim on one of the mesh's own seats is judged fully at registration, and a seat declared by a module not given here reads as unknown. Identities judged on a 6-character machine name, against the bounds of the providers given here
judging 1 module(s) changed against 4 machine(s), as the snapshot of 2026-10-12T09:56:46Z says they run
composing every machine as the mesh is
the bus's user list leaves out 5 user(s) the mesh has minted no credential for: controller, node.f35, node.idt, node.ygcyr, node.yvbpsm. Each is a user that cannot connect until one is issued
the bus's user list leaves out 5 user(s) the mesh has minted no credential for: controller, node.f35, node.idt, node.ygcyr, node.yvbpsm. Each is a user that cannot connect until one is issued
composing every machine with the change
the bus's user list leaves out 5 user(s) the mesh has minted no credential for: controller, node.f35, node.idt, node.ygcyr, node.yvbpsm. Each is a user that cannot connect until one is issued
the bus's user list leaves out 5 user(s) the mesh has minted no credential for: controller, node.f35, node.idt, node.ygcyr, node.yvbpsm. Each is a user that cannot connect until one is issued
merge gate: PASS — every machine composes with the change as it did without (4 of 4 compose)
judged by controller development build against the facts of 2026-10-12T09:56:46Z
modules the change moves:
- messenger (changed)
machines:
- a machine (f35): composes
- holds mesh-dns-resolver (idt): composes
- holds git, holds mesh-artifact-store, holds mesh-broker, holds mesh-catalog, holds mesh-delivery, holds mesh-dns-resolver, holds mesh-store, holds mesh-vault, holds npm-package-registry, the control node, the hub (ygcyr): composes
- a machine (yvbpsm): composes
a merge would rebuild 1 module(s) in 1 tier(s): messenger
notes:
- modules/messenger/README.md, modules/messenger/cmd/messenger/holder_test.go, modules/messenger/cmd/messenger/manifest_test.go, modules/messenger/cmd/messenger/restart_test.go and modules/messenger/cmd/messenger/walks_test.go read by no module's build, and rebuild nothing (issue 280)
--- its merge-check.sh (go toolchain)
ok git.novox.be/novox/mesh-catalog/checks/words 0.003s
words: 12 retired words and 70 daemons' lines, none in the 42479 strings of 683 files
ok git.novox.be/novox/mesh-catalog/checks/agenthome 0.014s
agent-home: no archive and no file written into under the agent's home
ok git.novox.be/novox/mesh-catalog/checks/paths 0.008s
paths: no module names another's state on its machine (3 known, listed to go)
long-running resources without health: 69, as kept
testing modules/messenger
ok messenger/cmd/messenger 2.215s
--- its merge-check-typescript.sh (typescript toolchain)
**Merge check** at `991ffcfd`
- `mesh/merge-gate`: **PASS** — every machine composes with the change as it did without (4 of 4 compose) (modules: messenger)
- `mesh/repo-check`: **PASS** — its merge-check.sh passed, each part in its toolchain: merge-check.sh (go), merge-check-typescript.sh (typescript)
**Delivery plan** — builds messenger → novox; no bus step
- tier 0: messenger
- novox: receives messenger
- read by no module's build: modules/messenger/README.md, modules/messenger/cmd/messenger/holder_test.go, modules/messenger/cmd/messenger/manifest_test.go, modules/messenger/cmd/messenger/restart_test.go, modules/messenger/cmd/messenger/walks_test.go
Run by the build seat on novox as `build-1791799044893033135` (`builds --log build-1791799044893033135`).
```
--- the gate: messenger
messenger: ok, 7 tool(s), invokes seat:mesh-controller.conditions, seat:mesh-controller.root-free, seat:issue-tracker.tracking, keeps state open, sent, asks, identities, walks; WARNING: settings ask(s) for a setting and do(es) not say whether it is trusted, so it counts as trusted: set at the terminal alone; say "trusted" false where nothing trusts it (novox/hq issue 339)
long-running resources without health: 0
files asking for a setting without saying whether it is trusted: 1
1 manifest(s) checked. Judged against the seats this binary carries; a claim on one of the mesh's own seats is judged fully at registration, and a seat declared by a module not given here reads as unknown. Identities judged on a 6-character machine name, against the bounds of the providers given here
judging 1 module(s) changed against 4 machine(s), as the snapshot of 2026-10-12T09:56:46Z says they run
composing every machine as the mesh is
the bus's user list leaves out 5 user(s) the mesh has minted no credential for: controller, node.f35, node.idt, node.ygcyr, node.yvbpsm. Each is a user that cannot connect until one is issued
the bus's user list leaves out 5 user(s) the mesh has minted no credential for: controller, node.f35, node.idt, node.ygcyr, node.yvbpsm. Each is a user that cannot connect until one is issued
composing every machine with the change
the bus's user list leaves out 5 user(s) the mesh has minted no credential for: controller, node.f35, node.idt, node.ygcyr, node.yvbpsm. Each is a user that cannot connect until one is issued
the bus's user list leaves out 5 user(s) the mesh has minted no credential for: controller, node.f35, node.idt, node.ygcyr, node.yvbpsm. Each is a user that cannot connect until one is issued
merge gate: PASS — every machine composes with the change as it did without (4 of 4 compose)
judged by controller development build against the facts of 2026-10-12T09:56:46Z
modules the change moves:
- messenger (changed)
machines:
- a machine (f35): composes
- holds mesh-dns-resolver (idt): composes
- holds git, holds mesh-artifact-store, holds mesh-broker, holds mesh-catalog, holds mesh-delivery, holds mesh-dns-resolver, holds mesh-store, holds mesh-vault, holds npm-package-registry, the control node, the hub (ygcyr): composes
- a machine (yvbpsm): composes
a merge would rebuild 1 module(s) in 1 tier(s): messenger
notes:
- modules/messenger/README.md, modules/messenger/cmd/messenger/holder_test.go, modules/messenger/cmd/messenger/manifest_test.go, modules/messenger/cmd/messenger/restart_test.go and modules/messenger/cmd/messenger/walks_test.go read by no module's build, and rebuild nothing (issue 280)
--- its merge-check.sh (go toolchain)
ok git.novox.be/novox/mesh-catalog/checks/words 0.003s
words: 12 retired words and 70 daemons' lines, none in the 42479 strings of 683 files
ok git.novox.be/novox/mesh-catalog/checks/agenthome 0.014s
agent-home: no archive and no file written into under the agent's home
ok git.novox.be/novox/mesh-catalog/checks/paths 0.008s
paths: no module names another's state on its machine (3 known, listed to go)
long-running resources without health: 69, as kept
testing modules/messenger
ok messenger/cmd/messenger 2.215s
--- its merge-check-typescript.sh (typescript toolchain)
```
Independent review of head 991ffcf against ADR 0308, issue 531 and the walk events of mesh-controller #247.
Verdict: REQUEST_CHANGES. The forge refuses that state from a tool (ADR 0266), so it is posted as a comment. Three findings let a held message go untold, which is the worst fault this change can have. I reproduced each one with a throwaway test against the head; the probe file was deleted afterwards.
Blocking
B1. A held message is lost for good when no channel is ready at the walk's end or bound.walks.go:486-496 (tellWalk): when no channel is ready, kept is empty. The open records are not marked said, which is correct, but nothing sets r.Pending. The walk is already deleted (walks.go:314,325), or its Held is emptied at the bound (walks.go:445). Tick retries only r.Pending != "" (holder.go:771), and a reminder needs FirstSent (holder.go:784), so nothing ever says the record. Repro: a walk holds a warning on novox; desk and telegram are not ready at the walk's end and ready again a minute later. After 30 h of ticks, nothing was sent. The log line "what it held stays open here" says otherwise. Fix: when kept is empty, set r.Pending = whats[r.Key] (or raised) on each open record and persist it, as deliver's no-channel path does. Add a test.
B2. Overlapping walks release early when the second walk starts after the message was held.walks.go:337-348 (release) counts a message as held elsewhere only when another walk already has its key in Held. Scenario: walk A touches novox and holds machine.novox.network. Walk B then starts and also touches novox. A ends, and the message is told while B is still open. ADR 0308 §2 says a message touched by several open walks is held until the last of them ends. The existing test raises the message only while both walks are open. Repro: a probe with that order printed said while b still open: true. Fix: in release, also count it as elsewhere when another walk that is still holding touches(subjectOf(r)), and move the held entry (with its At/What) to that walk.
B3. A walk forgotten after a day drops what it held.walks.go:423-436 (boundWalks) checks the gone case (more than 24 h since start) before due. A walk whose bound tick never ran, for example a router that was down for a day and read the walk back from the walks bucket, is deleted with its Held never released. Repro: hold a message, advance the clock 25 h with no tick, then run ticks. The message was never said. Fix: release what it holds (as at the bound) before forgetting it, or test due before gone.
Not blocking
N1restart.go:162,170: a held record whose message has answers on a kind that lost its answers is left alone, and the pass is still stored as done in passes. On release, tellWalk sets LastSaid = now (walks.go:522), so that message is never said again with working answers; the end message carries only its title. Consider leaving such a loss open for those keys until the hold ends, or re-offering after release.
N2walks.go:131-141: a subject that names a touched machine is held whatever its module, for example a disk condition on novox during a leaf walk of one module there. This is a literal reading of §2 and is fine. It is worth one line in the README.
N3walks.go:468-470: the fallback message drops every open line but still marks those records told. say already sanitises each line, so this case should be rare. Consider not marking the records said on the fallback.
Checked and fine
The decoder matches #247's body (walkSaid embedded beside inventory.Plan: walk, phase, touches[{machine, modules, carried}], outcome, bound_seconds, budget_class, started_at, ended_at). None of these names collides with the Plan's own JSON fields, and an ordinary plan-moved has no phase.
The §3 never-held paths (walks.go:621-645, hold at 650-652): asks, conditions with an open ask, secret-given/secret-replaced, the plan/walk/delivery/build/release scopes, a walk's own id, clearings and silence edits.
§4: one message, sent alone (outbox.go flushChannel), past the desk's 15-minute rhythm (carriesWalk); the one-line form and the quiet setting only when nothing is held.
§5: the bound runs from started_at, defaults to 900 s, and gives the still-running line.
§6 / 531: LostSettle merges the passes, passes is kept in sent, and silenced, withheld, answered and held records are skipped.
The setting only-what-needs-me is declared in the settings file default beside time-zone and read strictly (main.go).
go test -race ./... and go vet pass in modules/messenger.
Required checks at review time: mesh/merge-gate pending (building), mesh/repo-check missing.
Independent review of head 991ffcf against ADR 0308, issue 531 and the walk events of mesh-controller #247.
**Verdict: REQUEST_CHANGES.** The forge refuses that state from a tool (ADR 0266), so it is posted as a comment. Three findings let a held message go untold, which is the worst fault this change can have. I reproduced each one with a throwaway test against the head; the probe file was deleted afterwards.
## Blocking
**B1. A held message is lost for good when no channel is ready at the walk's end or bound.** `walks.go:486-496` (tellWalk): when no channel is ready, `kept` is empty. The open records are not marked said, which is correct, but nothing sets `r.Pending`. The walk is already deleted (`walks.go:314,325`), or its `Held` is emptied at the bound (`walks.go:445`). Tick retries only `r.Pending != ""` (`holder.go:771`), and a reminder needs `FirstSent` (`holder.go:784`), so nothing ever says the record. Repro: a walk holds a warning on novox; desk and telegram are not ready at the walk's end and ready again a minute later. After 30 h of ticks, nothing was sent. The log line "what it held stays open here" says otherwise. Fix: when `kept` is empty, set `r.Pending = whats[r.Key]` (or `raised`) on each open record and persist it, as `deliver`'s no-channel path does. Add a test.
**B2. Overlapping walks release early when the second walk starts after the message was held.** `walks.go:337-348` (release) counts a message as held elsewhere only when another walk already has its key in `Held`. Scenario: walk A touches novox and holds `machine.novox.network`. Walk B then starts and also touches novox. A ends, and the message is told while B is still open. ADR 0308 §2 says a message touched by several open walks is held until the last of them ends. The existing test raises the message only while both walks are open. Repro: a probe with that order printed `said while b still open: true`. Fix: in release, also count it as elsewhere when another walk that is still holding `touches(subjectOf(r))`, and move the held entry (with its `At`/`What`) to that walk.
**B3. A walk forgotten after a day drops what it held.** `walks.go:423-436` (boundWalks) checks the `gone` case (more than 24 h since start) before `due`. A walk whose bound tick never ran, for example a router that was down for a day and read the walk back from the `walks` bucket, is deleted with its `Held` never released. Repro: hold a message, advance the clock 25 h with no tick, then run ticks. The message was never said. Fix: release what it holds (as at the bound) before forgetting it, or test `due` before `gone`.
## Not blocking
- **N1** `restart.go:162,170`: a held record whose message has answers on a kind that lost its answers is left alone, and the pass is still stored as done in `passes`. On release, `tellWalk` sets `LastSaid = now` (`walks.go:522`), so that message is never said again with working answers; the end message carries only its title. Consider leaving such a loss open for those keys until the hold ends, or re-offering after release.
- **N2** `walks.go:131-141`: a subject that names a touched machine is held whatever its module, for example a disk condition on novox during a leaf walk of one module there. This is a literal reading of §2 and is fine. It is worth one line in the README.
- **N3** `walks.go:468-470`: the fallback message drops every open line but still marks those records told. `say` already sanitises each line, so this case should be rare. Consider not marking the records said on the fallback.
## Checked and fine
- The decoder matches #247's body (`walkSaid` embedded beside `inventory.Plan`: `walk`, `phase`, `touches[{machine, modules, carried}]`, `outcome`, `bound_seconds`, `budget_class`, `started_at`, `ended_at`). None of these names collides with the Plan's own JSON fields, and an ordinary plan-moved has no `phase`.
- The §3 never-held paths (`walks.go:621-645`, `hold` at 650-652): asks, conditions with an open ask, `secret-given`/`secret-replaced`, the plan/walk/delivery/build/release scopes, a walk's own id, clearings and silence edits.
- §4: one message, sent alone (`outbox.go` flushChannel), past the desk's 15-minute rhythm (`carriesWalk`); the one-line form and the quiet setting only when nothing is held.
- §5: the bound runs from `started_at`, defaults to 900 s, and gives the still-running line.
- §6 / 531: `LostSettle` merges the passes, `passes` is kept in `sent`, and silenced, withheld, answered and held records are skipped.
- The setting `only-what-needs-me` is declared in the settings file default beside `time-zone` and read strictly (`main.go`).
- `go test -race ./...` and `go vet` pass in modules/messenger.
- Required checks at review time: mesh/merge-gate pending (building), mesh/repo-check missing.
- No channel ready at a walk's end or bound: what it held is left unsent and tried
again every minute, as any other message, instead of lost with the walk.
- At a walk's end or bound, a message another open walk touches is held on by
that walk, even one started after it was held (hq ADR 0308 §2).
- A walk read back past its forget is released at its bound first, then forgotten.
- A walk's end is kept a day: a start heard after it (same started_at, or not
later than the end) opens nothing, also after the router restarts; a later
start under the same id still opens.
- A walk's plain end, when its words cannot be said, leaves what it held to be
said on its own instead of marking it told.
- The restart pass no longer skips a held message: one with an open ask is never
held (ADR 0308 §3), so its answers are made to work again.
mesh/merge-gate: PASS — every machine composes with the change as it did without (4 of 4 compose) (modules: messenger)
mesh/repo-check: PASS — its merge-check.sh passed, each part in its toolchain: merge-check.sh (go), merge-check-typescript.sh (typescript)
Delivery plan — builds messenger → novox; no bus step
tier 0: messenger
novox: receives messenger
read by no module's build: modules/messenger/README.md, modules/messenger/cmd/messenger/holder_test.go, modules/messenger/cmd/messenger/manifest_test.go, modules/messenger/cmd/messenger/restart_test.go, modules/messenger/cmd/messenger/walks_test.go
Run by the build seat on ace as build-1791799691250601847 (builds --log build-1791799691250601847).
--- the gate: messenger
messenger: ok, 7 tool(s), invokes seat:mesh-controller.conditions, seat:mesh-controller.root-free, seat:issue-tracker.tracking, keeps state open, sent, asks, identities, walks; WARNING: settings ask(s) for a setting and do(es) not say whether it is trusted, so it counts as trusted: set at the terminal alone; say "trusted" false where nothing trusts it (novox/hq issue 339)
long-running resources without health: 0
files asking for a setting without saying whether it is trusted: 1
1 manifest(s) checked. Judged against the seats this binary carries; a claim on one of the mesh's own seats is judged fully at registration, and a seat declared by a module not given here reads as unknown. Identities judged on a 6-character machine name, against the bounds of the providers given here
judging 1 module(s) changed against 4 machine(s), as the snapshot of 2026-10-12T09:59:43Z says they run
composing every machine as the mesh is
the bus's user list leaves out 5 user(s) the mesh has minted no credential for: controller, node.f35, node.idt, node.ygcyr, node.yvbpsm. Each is a user that cannot connect until one is issued
the bus's user list leaves out 5 user(s) the mesh has minted no credential for: controller, node.f35, node.idt, node.ygcyr, node.yvbpsm. Each is a user that cannot connect until one is issued
composing every machine with the change
the bus's user list leaves out 5 user(s) the mesh has minted no credential for: controller, node.f35, node.idt, node.ygcyr, node.yvbpsm. Each is a user that cannot connect until one is issued
the bus's user list leaves out 5 user(s) the mesh has minted no credential for: controller, node.f35, node.idt, node.ygcyr, node.yvbpsm. Each is a user that cannot connect until one is issued
merge gate: PASS — every machine composes with the change as it did without (4 of 4 compose)
judged by controller development build against the facts of 2026-10-12T09:59:43Z
modules the change moves:
- messenger (changed)
machines:
- a machine (f35): composes
- holds mesh-dns-resolver (idt): composes
- holds git, holds mesh-artifact-store, holds mesh-broker, holds mesh-catalog, holds mesh-delivery, holds mesh-dns-resolver, holds mesh-store, holds mesh-vault, holds npm-package-registry, the control node, the hub (ygcyr): composes
- a machine (yvbpsm): composes
a merge would rebuild 1 module(s) in 1 tier(s): messenger
notes:
- modules/messenger/README.md, modules/messenger/cmd/messenger/holder_test.go, modules/messenger/cmd/messenger/manifest_test.go, modules/messenger/cmd/messenger/restart_test.go and modules/messenger/cmd/messenger/walks_test.go read by no module's build, and rebuild nothing (issue 280)
--- its merge-check.sh (go toolchain)
ok git.novox.be/novox/mesh-catalog/checks/words 0.003s
words: 12 retired words and 70 daemons' lines, none in the 42501 strings of 683 files
ok git.novox.be/novox/mesh-catalog/checks/agenthome 0.011s
agent-home: no archive and no file written into under the agent's home
ok git.novox.be/novox/mesh-catalog/checks/paths 0.006s
paths: no module names another's state on its machine (3 known, listed to go)
long-running resources without health: 69, as kept
testing modules/messenger
ok messenger/cmd/messenger 1.972s
--- its merge-check-typescript.sh (typescript toolchain)
**Merge check** at `a92c08fd`
- `mesh/merge-gate`: **PASS** — every machine composes with the change as it did without (4 of 4 compose) (modules: messenger)
- `mesh/repo-check`: **PASS** — its merge-check.sh passed, each part in its toolchain: merge-check.sh (go), merge-check-typescript.sh (typescript)
**Delivery plan** — builds messenger → novox; no bus step
- tier 0: messenger
- novox: receives messenger
- read by no module's build: modules/messenger/README.md, modules/messenger/cmd/messenger/holder_test.go, modules/messenger/cmd/messenger/manifest_test.go, modules/messenger/cmd/messenger/restart_test.go, modules/messenger/cmd/messenger/walks_test.go
Run by the build seat on ace as `build-1791799691250601847` (`builds --log build-1791799691250601847`).
```
--- the gate: messenger
messenger: ok, 7 tool(s), invokes seat:mesh-controller.conditions, seat:mesh-controller.root-free, seat:issue-tracker.tracking, keeps state open, sent, asks, identities, walks; WARNING: settings ask(s) for a setting and do(es) not say whether it is trusted, so it counts as trusted: set at the terminal alone; say "trusted" false where nothing trusts it (novox/hq issue 339)
long-running resources without health: 0
files asking for a setting without saying whether it is trusted: 1
1 manifest(s) checked. Judged against the seats this binary carries; a claim on one of the mesh's own seats is judged fully at registration, and a seat declared by a module not given here reads as unknown. Identities judged on a 6-character machine name, against the bounds of the providers given here
judging 1 module(s) changed against 4 machine(s), as the snapshot of 2026-10-12T09:59:43Z says they run
composing every machine as the mesh is
the bus's user list leaves out 5 user(s) the mesh has minted no credential for: controller, node.f35, node.idt, node.ygcyr, node.yvbpsm. Each is a user that cannot connect until one is issued
the bus's user list leaves out 5 user(s) the mesh has minted no credential for: controller, node.f35, node.idt, node.ygcyr, node.yvbpsm. Each is a user that cannot connect until one is issued
composing every machine with the change
the bus's user list leaves out 5 user(s) the mesh has minted no credential for: controller, node.f35, node.idt, node.ygcyr, node.yvbpsm. Each is a user that cannot connect until one is issued
the bus's user list leaves out 5 user(s) the mesh has minted no credential for: controller, node.f35, node.idt, node.ygcyr, node.yvbpsm. Each is a user that cannot connect until one is issued
merge gate: PASS — every machine composes with the change as it did without (4 of 4 compose)
judged by controller development build against the facts of 2026-10-12T09:59:43Z
modules the change moves:
- messenger (changed)
machines:
- a machine (f35): composes
- holds mesh-dns-resolver (idt): composes
- holds git, holds mesh-artifact-store, holds mesh-broker, holds mesh-catalog, holds mesh-delivery, holds mesh-dns-resolver, holds mesh-store, holds mesh-vault, holds npm-package-registry, the control node, the hub (ygcyr): composes
- a machine (yvbpsm): composes
a merge would rebuild 1 module(s) in 1 tier(s): messenger
notes:
- modules/messenger/README.md, modules/messenger/cmd/messenger/holder_test.go, modules/messenger/cmd/messenger/manifest_test.go, modules/messenger/cmd/messenger/restart_test.go and modules/messenger/cmd/messenger/walks_test.go read by no module's build, and rebuild nothing (issue 280)
--- its merge-check.sh (go toolchain)
ok git.novox.be/novox/mesh-catalog/checks/words 0.003s
words: 12 retired words and 70 daemons' lines, none in the 42501 strings of 683 files
ok git.novox.be/novox/mesh-catalog/checks/agenthome 0.011s
agent-home: no archive and no file written into under the agent's home
ok git.novox.be/novox/mesh-catalog/checks/paths 0.006s
paths: no module names another's state on its machine (3 known, listed to go)
long-running resources without health: 69, as kept
testing modules/messenger
ok messenger/cmd/messenger 1.972s
--- its merge-check-typescript.sh (typescript toolchain)
```
Independent re-review of the fixes for review 14, in a fresh clone. Both required checks are green on a92c08f: mesh/merge-gate passes (it builds messenger for novox, and 4 of 4 machines compose) and mesh/repo-check passes. go vet is clean, and go test -race -count=3 ./... in modules/messenger passes.
The fixes for review 14, each checked
I ran each new test on the old head 991ffcf, using the test file from a92c08f. All six fail there and pass here.
No channel ready at the end or at the bound.tellWalk (walks.go:570-585) marks each open or alone held record as Pending and persists it. Tick sends it again through deliver (holder.go:774-776), which can hold it again if another walk touches it. Test: TestWhatAWalkHeldIsSaidOnceAChannelCanSendAgain, end and bound. On 991ffcf it fails at walks_test.go:471 ("not left unsent").
Overlapping walks.release (walks.go:357-416) checks every other walk that is open and still within its bound. When such a walk touches the subject, it takes the held message over, keeping the original held-at, and history records it as held. Test: TestAMessageIsHeldOnByALaterWalkThatTouchesIt. On 991ffcf the first walk's end tells it (walks_test.go:496).
The bound is said before the one-day forget.boundWalks (walks.go:471-532) collects due independently of gone and forgets the gone walks in a defer after the bound messages are told. Test: TestAWalkReadBackADayLateSaysWhatItHeldBeforeItIsForgotten. On 991ffcf it fails with "forgotten unsaid".
An end heard before its start. The end keeps a tombstone in h.ended, persisted in the bucket walks under the walk's key (walks.go:338-344), and loadWalks reads it back (walks.go:850-852). A start whose at is at or before the end, or equal to the end's started_at, opens nothing (walks.go:295-303). A later end of an earlier run under the same id leaves the open run alone (walks.go:324-330). Tombstones are dropped after a day, unless the id is open again (walks.go:483-499). Test: TestAStartHeardAfterItsEndOpensNothing, which also covers the router's restart and a retry under the same id. On 991ffcf the start opens a walk (walks_test.go:551).
restart.go no longer skips held records (restart.go:168-172). The pass still says again only records whose ask is open (askKinds), and §3 never holds those. Test: TestARestartSaysAgainAnAskedMessageAWalkTookBeforeItsAsk. On 991ffcf nothing is replaced.
The plain fallback end message. When Check refuses the end message, the open records move to alone, are not marked told, and each is sent through deliver after the plain message (walks.go:542-545, 618-626). Test: TestWhatAWalksPlainEndCannotCarryIsSaidOnItsOwn. On 991ffcf the record is marked told and never said.
I checked against branch adr-0308-walk-touches at 44983d6. walkSaid (cmd/mesh-controller/walk_said.go:38-62) carries walk=plan id, phase, touches[{machine, modules, carried}], outcome (on the end only, from Plan.Outcome(): delivered, failed or stopped), bound_seconds, budget_class, started_at on both events and ended_at on the end. It is embedded beside inventory.Plan, and none of its JSON names collides with a field of Plan. DecodeWalk reads every one of these. A retry starts the walk again under the same id with a new started_at (inventory/walksaid.go:68-70). The tombstone check and the late-end check in walks.go handle that correctly.
Remaining findings (not blocking)
A. A held message's first telling can be lost across a router restart.tellWalk marks the open records said (LastSaid, FirstSent, Sent[ch]="", Pending="") and persists them when the walk's message is queued (walks.go:594-616), not when it is sent. Suppose the channel is ready but Send fails. The outbox then keeps the walk's message in memory only. If the router restarts before a retry succeeds, that message is gone, and the records already say they were told: Load does not mark them History, and the sync's urgent-never-said path skips them. I reproduced it with a scratch test, not committed: one urgent held record, the end sent while both channels fail, then a restart from the same state and 30 min of ticks. Nothing is said, and the history says "told in the message of walk plan-p". The first word comes only as the reminder, 1 h later for urgent and 12 h later for a warning. The window is small (a failing send plus a restart before the next retry), so I do not hold the merge for it. Suggested follow-up issue: mark the records told when the walk-end item is actually sent, for example by having the item carry its record keys, which is how a single item's record is marked in flush.
B. A takeover in release skips the filters that decide what is still open. walks.go:378-386 hands a message to a later walk as soon as the record is live. It does this before checking whether the record is silenced, was said after it was held (lastSaid(r).After(hm.At)), or has an open ask (§3). I reproduced the last case with a scratch test: a condition is held by plan-a, its ask is offered (and so said), plan-b starts on the same machine, and plan-a ends. heldBy then gives [plan-b], and history records a second held. Nothing is said twice, because plan-b's end filters the record out. But operator-channel.open shows held_by_walks and history shows held for a message that ADR 0308 §3 says is never held. The fix is to apply the same checks before the takeover. Doing that needs neverHeld without h.mu, or openAsk checked before the lock.
C. The order check compares a capped time. In Walked, at is capped at now. The tombstone check (walks.go:298) compares that capped at with end.Ended and end.Started. Suppose the controller's clock is ahead of the router's, and a short walk's end arrives before its start. The capped start can then come after end.Ended and fail to equal the uncapped end.Started, so the start opens a walk that has already ended, and that walk holds until its bound. Comparing ev.StartedAt (uncapped) with end.Started would close this. Both run on novox today, so this is theoretical.
Not merged by me.
Verdict: approve at a92c08f
Independent re-review of the fixes for review 14, in a fresh clone. Both required checks are green on a92c08f: mesh/merge-gate passes (it builds messenger for novox, and 4 of 4 machines compose) and mesh/repo-check passes. `go vet` is clean, and `go test -race -count=3 ./...` in modules/messenger passes.
## The fixes for review 14, each checked
I ran each new test on the old head 991ffcf, using the test file from a92c08f. **All six fail there and pass here.**
1. **No channel ready at the end or at the bound.** `tellWalk` (walks.go:570-585) marks each open or `alone` held record as `Pending` and persists it. `Tick` sends it again through `deliver` (holder.go:774-776), which can hold it again if another walk touches it. Test: `TestWhatAWalkHeldIsSaidOnceAChannelCanSendAgain`, end and bound. On 991ffcf it fails at walks_test.go:471 ("not left unsent").
2. **Overlapping walks.** `release` (walks.go:357-416) checks every other walk that is open and still within its bound. When such a walk touches the subject, it takes the held message over, keeping the original held-at, and history records it as `held`. Test: `TestAMessageIsHeldOnByALaterWalkThatTouchesIt`. On 991ffcf the first walk's end tells it (walks_test.go:496).
3. **The bound is said before the one-day forget.** `boundWalks` (walks.go:471-532) collects `due` independently of `gone` and forgets the gone walks in a `defer` after the bound messages are told. Test: `TestAWalkReadBackADayLateSaysWhatItHeldBeforeItIsForgotten`. On 991ffcf it fails with "forgotten unsaid".
4. **An end heard before its start.** The end keeps a tombstone in `h.ended`, persisted in the bucket `walks` under the walk's key (walks.go:338-344), and `loadWalks` reads it back (walks.go:850-852). A start whose `at` is at or before the end, or equal to the end's `started_at`, opens nothing (walks.go:295-303). A later end of an earlier run under the same id leaves the open run alone (walks.go:324-330). Tombstones are dropped after a day, unless the id is open again (walks.go:483-499). Test: `TestAStartHeardAfterItsEndOpensNothing`, which also covers the router's restart and a retry under the same id. On 991ffcf the start opens a walk (walks_test.go:551).
5. **restart.go no longer skips held records** (restart.go:168-172). The pass still says again only records whose ask is open (`askKinds`), and §3 never holds those. Test: `TestARestartSaysAgainAnAskedMessageAWalkTookBeforeItsAsk`. On 991ffcf nothing is replaced.
6. **The plain fallback end message.** When `Check` refuses the end message, the open records move to `alone`, are not marked told, and each is sent through `deliver` after the plain message (walks.go:542-545, 618-626). Test: `TestWhatAWalksPlainEndCannotCarryIsSaidOnItsOwn`. On 991ffcf the record is marked told and never said.
## The decoder against mesh-controller #247
I checked against branch adr-0308-walk-touches at 44983d6. `walkSaid` (cmd/mesh-controller/walk_said.go:38-62) carries `walk`=plan id, `phase`, `touches[{machine, modules, carried}]`, `outcome` (on the end only, from `Plan.Outcome()`: delivered, failed or stopped), `bound_seconds`, `budget_class`, `started_at` on both events and `ended_at` on the end. It is embedded beside `inventory.Plan`, and none of its JSON names collides with a field of Plan. `DecodeWalk` reads every one of these. A retry starts the walk again under the same id with a new `started_at` (inventory/walksaid.go:68-70). The tombstone check and the late-end check in walks.go handle that correctly.
## Remaining findings (not blocking)
A. **A held message's first telling can be lost across a router restart.** `tellWalk` marks the open records said (`LastSaid`, `FirstSent`, `Sent[ch]=""`, `Pending=""`) and persists them when the walk's message is *queued* (walks.go:594-616), not when it is sent. Suppose the channel is ready but `Send` fails. The outbox then keeps the walk's message in memory only. If the router restarts before a retry succeeds, that message is gone, and the records already say they were told: `Load` does not mark them `History`, and the sync's urgent-never-said path skips them. I reproduced it with a scratch test, not committed: one urgent held record, the end sent while both channels fail, then a restart from the same state and 30 min of ticks. Nothing is said, and the history says "told in the message of walk plan-p". The first word comes only as the reminder, 1 h later for urgent and 12 h later for a warning. The window is small (a failing send plus a restart before the next retry), so I do not hold the merge for it. Suggested follow-up issue: mark the records told when the `walk-end` item is actually sent, for example by having the item carry its record keys, which is how a single item's record is marked in `flush`.
B. **A takeover in `release` skips the filters that decide what is still open.** walks.go:378-386 hands a message to a later walk as soon as the record is live. It does this before checking whether the record is silenced, was said after it was held (`lastSaid(r).After(hm.At)`), or has an open ask (§3). I reproduced the last case with a scratch test: a condition is held by plan-a, its ask is offered (and so said), plan-b starts on the same machine, and plan-a ends. `heldBy` then gives `[plan-b]`, and history records a second `held`. Nothing is said twice, because plan-b's end filters the record out. But `operator-channel.open` shows `held_by_walks` and history shows `held` for a message that ADR 0308 §3 says is never held. The fix is to apply the same checks before the takeover. Doing that needs `neverHeld` without `h.mu`, or `openAsk` checked before the lock.
C. **The order check compares a capped time.** In `Walked`, `at` is capped at `now`. The tombstone check (walks.go:298) compares that capped `at` with `end.Ended` and `end.Started`. Suppose the controller's clock is ahead of the router's, and a short walk's end arrives before its start. The capped start can then come after `end.Ended` and fail to equal the uncapped `end.Started`, so the start opens a walk that has already ended, and that walk holds until its bound. Comparing `ev.StartedAt` (uncapped) with `end.Started` would close this. Both run on novox today, so this is theoretical.
Not merged by me.
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Builds the router part of novox/hq ADR 0308 (while a delivery walks, what it touches is held from the operator and told once, with the walk's result, when it ends) for issue 532, and fixes issue 531 (every restart of the tool runner popped the four silenced conditions up on the desk again, twice, one of them as "A mesh message was withheld").
Merge order: mesh-controller #247 (
adr-0308-walk-touches, the walk events) first, then this. Merged alone, this holds nothing: no plan-moved carriesphaseyet, so no walk opens; the §6 fix works on its own.What it reads (§1)
The controller's
plan-moved(now declared inconsumes). Only a body withphaseis a walk's start or end; any other plan-moved is ignored. Fields read, as #247 says them:walk(elseid),phase(started|ended),touches[{machine, modules, carried}](modulesis read;carriedis listed there too),outcome(delivered|failed|stopped, required on the end),bound_seconds(missing or 0: 900),budget_class,started_at(on both, the bound runs from it),ended_at(elseat, elseupdated). A body with aphasethat cannot be read is refused by name, like an unreadable condition event. An event namedwalkwith the same body is read the same way. Code:cmd/messenger/walks.go.What it does
operator-channel.historyrecords it with outcomeheld. If the subject names its machine, a module is matched on that machine only. A message touched by overlapping walks is kept on each and told only when the last of them ends or passes its bound.operator-channel.openshowsheld_by_walks.secret-given/secret-replaced; subjects no open walk touches; a walk's own failure (scopes plan, walk, delivery, build, release, and anything naming an open walk's id); a clearing or a silence edit.Delivered|Delivery failed|Delivery stopped: <modules> on <machines>,It took N min.,Still open after it:with one line per held message still open, andN problem(s) it raised resolved during it.Held messages it tells are marked said, so nothing more is sent of them. If the walk delivered with nothing held and nothing open, the message is one line (Delivered: … , in N min). It is sent nowhere on a kind listed in the new settingonly-what-needs-me(a list of kinds, empty by default; no channel level existed, so I added this as a small design limit). The history says why.started_at + bound_seconds, the walk's held messages still open are sent withThe walk <id> is still running, since HH:MM.(the id is dropped if the content rule refuses it), and the cleared ones are counted. After that the walk holds nothing. A walk whose end never comes is forgotten a day after it started. A start heard after its bound has passed opens nothing. An end heard more than 10 min late with nothing open closes quietly.walks, one key per walk, read back byLoad.cmd/messenger/restart.go. I found three causes:LostSettle), so the standings of one restart make one pass.shows, holder.go:332-351). It is now kept in the statesent, keypasses, and read back on start.Tests (
walks_test.go; ADR 0308 "How it is checked")TestDuringAWalkWhatItTouchesIsHeldAndTheHistorySaysSo(a module's condition, a machine's, and a notice: nothing sent, historyheld, kept open).TestAMessageTouchedByTwoOverlappingWalksIsHeldUntilTheSecondEnds.TestWhatAWalkNeverHoldsIsSentAtOnce(secret given and replaced on a touched machine, an untouched machine, a delivery's and the plan's own failure on a touched machine, an ask, and an ask about a held condition).TestAWalksEndIsToldOnceWithWhatIsStillOpen(outcome, reach, 7 min, two open lines, one cleared counted, nothing said again for 20 min).TestADeliveredWalkWithNothingHeldIsOneLineOrNothing(one line; nothing on anonly-what-needs-mekind, with the history saying so).TestAHoldEndsAtTheWalksBound, table-driven from the controller'sbound_seconds: 1200 gives 20 min, 600 gives 10 min, 360 gives 6 min, and none gives 15 min. It checks nothing is sent before the bound, the still-running line with since HH:MM, the cleared count, and that nothing is held after the bound.TestHeldMessagesSurviveARouterRestart.TestARestartSaysNothingItHasSaid. Four silenced messages are shown on the desk, one withheld for a file path, plus a fifth open, asked and withheld. Two standings come a second apart with a tick between: nothing is sent, edited or replaced, and exactly one(answers lost)pass runs. The router is then restarted from the same state and the standing repeated: still one pass, nothing said.TestTheWalkEventsAreReadAsTheControllerSaysThem(#247's start and end bodies,carried,ended_at, an ordinary plan-moved, an end without an outcome refused, the default bound, a start older than its bound).restart_test.go. Issue 384's tests asserted that the silenced message is edited again; ADR 0308 §6 reverses that. The router's restart is now simulated by reading the state back.LostSettleor the withheld skip each fails a test.Checks.
go vetandgo test -racepass (5 runs, plus-count=30).MESH_GATE=<controller built from main d854598> MESH_CHECK_MODULES=messenger sh merge-check.shpasses locally: words, agent-home, paths, every manifest throughmodule check(69 long-running resources without health, as kept), and the module's tests. A trial merge with #178 (issue 383, which also touches messenger) is clean (git merge-tree).Reviewers' checklist
deliver(plus the sync's urgent path andeditfor reopen and silence-over). A clearing or a silence edit is never held, and nothing in the asks path is held (holdrefusesasked, an ask's own record, and a record with an open ask).subjectOfmatching: for scopes machine, node and bus the machine is the first token. For module, core and seat the module is the first token and the machine the second. Any other scope and a notice match each of their names as either. Too broad a match only delays a message by at most the bound; a missed one is sent as today.say). The whole message is checked again, and a refusal falls back to plain words.Refused), answered and held records alone.passesis persisted, andLostSettle(10 s) merges one restart's standings into one pass.consumesgainsmesh-controller.plan-moved,stategainswalks, and the settings default gainsonly-what-needs-me: []. No node name appears in the manifest.Live check after both merges (ADR 0308): the next tool-runner delivery should give the operator one message at its end, and
operator-channel.historyshould showheldentries and their release, with one(answers lost)pass per notifier restart and no silenced edits.By adr0308@g14. Not merged; waits for an independent review and for mesh-controller #247.
Delivery
novox/mesh-catalog@a92c08fd5b55— delivered since 2026-10-12T11:25:44ZDelivery plan — builds messenger → novox; no bus step
Machines
Transitions
The commit's note under
refs/notes/mesh-plankeeps every transition:git log --notes=mesh-plan.Merge check at
991ffcfdmesh/merge-gate: PASS — every machine composes with the change as it did without (4 of 4 compose) (modules: messenger)mesh/repo-check: PASS — its merge-check.sh passed, each part in its toolchain: merge-check.sh (go), merge-check-typescript.sh (typescript)Delivery plan — builds messenger → novox; no bus step
Run by the build seat on novox as
build-1791799044893033135(builds --log build-1791799044893033135).Independent review of head
991ffcfagainst ADR 0308, issue 531 and the walk events of mesh-controller #247.Verdict: REQUEST_CHANGES. The forge refuses that state from a tool (ADR 0266), so it is posted as a comment. Three findings let a held message go untold, which is the worst fault this change can have. I reproduced each one with a throwaway test against the head; the probe file was deleted afterwards.
Blocking
B1. A held message is lost for good when no channel is ready at the walk's end or bound.
walks.go:486-496(tellWalk): when no channel is ready,keptis empty. The open records are not marked said, which is correct, but nothing setsr.Pending. The walk is already deleted (walks.go:314,325), or itsHeldis emptied at the bound (walks.go:445). Tick retries onlyr.Pending != ""(holder.go:771), and a reminder needsFirstSent(holder.go:784), so nothing ever says the record. Repro: a walk holds a warning on novox; desk and telegram are not ready at the walk's end and ready again a minute later. After 30 h of ticks, nothing was sent. The log line "what it held stays open here" says otherwise. Fix: whenkeptis empty, setr.Pending = whats[r.Key](orraised) on each open record and persist it, asdeliver's no-channel path does. Add a test.B2. Overlapping walks release early when the second walk starts after the message was held.
walks.go:337-348(release) counts a message as held elsewhere only when another walk already has its key inHeld. Scenario: walk A touches novox and holdsmachine.novox.network. Walk B then starts and also touches novox. A ends, and the message is told while B is still open. ADR 0308 §2 says a message touched by several open walks is held until the last of them ends. The existing test raises the message only while both walks are open. Repro: a probe with that order printedsaid while b still open: true. Fix: in release, also count it as elsewhere when another walk that is still holdingtouches(subjectOf(r)), and move the held entry (with itsAt/What) to that walk.B3. A walk forgotten after a day drops what it held.
walks.go:423-436(boundWalks) checks thegonecase (more than 24 h since start) beforedue. A walk whose bound tick never ran, for example a router that was down for a day and read the walk back from thewalksbucket, is deleted with itsHeldnever released. Repro: hold a message, advance the clock 25 h with no tick, then run ticks. The message was never said. Fix: release what it holds (as at the bound) before forgetting it, or testduebeforegone.Not blocking
restart.go:162,170: a held record whose message has answers on a kind that lost its answers is left alone, and the pass is still stored as done inpasses. On release,tellWalksetsLastSaid = now(walks.go:522), so that message is never said again with working answers; the end message carries only its title. Consider leaving such a loss open for those keys until the hold ends, or re-offering after release.walks.go:131-141: a subject that names a touched machine is held whatever its module, for example a disk condition on novox during a leaf walk of one module there. This is a literal reading of §2 and is fine. It is worth one line in the README.walks.go:468-470: the fallback message drops every open line but still marks those records told.sayalready sanitises each line, so this case should be rare. Consider not marking the records said on the fallback.Checked and fine
walkSaidembedded besideinventory.Plan:walk,phase,touches[{machine, modules, carried}],outcome,bound_seconds,budget_class,started_at,ended_at). None of these names collides with the Plan's own JSON fields, and an ordinary plan-moved has nophase.walks.go:621-645,holdat 650-652): asks, conditions with an open ask,secret-given/secret-replaced, the plan/walk/delivery/build/release scopes, a walk's own id, clearings and silence edits.outbox.goflushChannel), past the desk's 15-minute rhythm (carriesWalk); the one-line form and the quiet setting only when nothing is held.started_at, defaults to 900 s, and gives the still-running line.LostSettlemerges the passes,passesis kept insent, and silenced, withheld, answered and held records are skipped.only-what-needs-meis declared in the settings file default besidetime-zoneand read strictly (main.go).go test -race ./...andgo vetpass in modules/messenger.Merge check at
a92c08fdmesh/merge-gate: PASS — every machine composes with the change as it did without (4 of 4 compose) (modules: messenger)mesh/repo-check: PASS — its merge-check.sh passed, each part in its toolchain: merge-check.sh (go), merge-check-typescript.sh (typescript)Delivery plan — builds messenger → novox; no bus step
Run by the build seat on ace as
build-1791799691250601847(builds --log build-1791799691250601847).Verdict: approve at
a92c08fIndependent re-review of the fixes for review 14, in a fresh clone. Both required checks are green on
a92c08f: mesh/merge-gate passes (it builds messenger for novox, and 4 of 4 machines compose) and mesh/repo-check passes.go vetis clean, andgo test -race -count=3 ./...in modules/messenger passes.The fixes for review 14, each checked
I ran each new test on the old head
991ffcf, using the test file froma92c08f. All six fail there and pass here.tellWalk(walks.go:570-585) marks each open oraloneheld record asPendingand persists it.Ticksends it again throughdeliver(holder.go:774-776), which can hold it again if another walk touches it. Test:TestWhatAWalkHeldIsSaidOnceAChannelCanSendAgain, end and bound. On991ffcfit fails at walks_test.go:471 ("not left unsent").release(walks.go:357-416) checks every other walk that is open and still within its bound. When such a walk touches the subject, it takes the held message over, keeping the original held-at, and history records it asheld. Test:TestAMessageIsHeldOnByALaterWalkThatTouchesIt. On991ffcfthe first walk's end tells it (walks_test.go:496).boundWalks(walks.go:471-532) collectsdueindependently ofgoneand forgets the gone walks in adeferafter the bound messages are told. Test:TestAWalkReadBackADayLateSaysWhatItHeldBeforeItIsForgotten. On991ffcfit fails with "forgotten unsaid".h.ended, persisted in the bucketwalksunder the walk's key (walks.go:338-344), andloadWalksreads it back (walks.go:850-852). A start whoseatis at or before the end, or equal to the end'sstarted_at, opens nothing (walks.go:295-303). A later end of an earlier run under the same id leaves the open run alone (walks.go:324-330). Tombstones are dropped after a day, unless the id is open again (walks.go:483-499). Test:TestAStartHeardAfterItsEndOpensNothing, which also covers the router's restart and a retry under the same id. On991ffcfthe start opens a walk (walks_test.go:551).askKinds), and §3 never holds those. Test:TestARestartSaysAgainAnAskedMessageAWalkTookBeforeItsAsk. On991ffcfnothing is replaced.Checkrefuses the end message, the open records move toalone, are not marked told, and each is sent throughdeliverafter the plain message (walks.go:542-545, 618-626). Test:TestWhatAWalksPlainEndCannotCarryIsSaidOnItsOwn. On991ffcfthe record is marked told and never said.The decoder against mesh-controller #247
I checked against branch adr-0308-walk-touches at 44983d6.
walkSaid(cmd/mesh-controller/walk_said.go:38-62) carrieswalk=plan id,phase,touches[{machine, modules, carried}],outcome(on the end only, fromPlan.Outcome(): delivered, failed or stopped),bound_seconds,budget_class,started_aton both events andended_aton the end. It is embedded besideinventory.Plan, and none of its JSON names collides with a field of Plan.DecodeWalkreads every one of these. A retry starts the walk again under the same id with a newstarted_at(inventory/walksaid.go:68-70). The tombstone check and the late-end check in walks.go handle that correctly.Remaining findings (not blocking)
A. A held message's first telling can be lost across a router restart.
tellWalkmarks the open records said (LastSaid,FirstSent,Sent[ch]="",Pending="") and persists them when the walk's message is queued (walks.go:594-616), not when it is sent. Suppose the channel is ready butSendfails. The outbox then keeps the walk's message in memory only. If the router restarts before a retry succeeds, that message is gone, and the records already say they were told:Loaddoes not mark themHistory, and the sync's urgent-never-said path skips them. I reproduced it with a scratch test, not committed: one urgent held record, the end sent while both channels fail, then a restart from the same state and 30 min of ticks. Nothing is said, and the history says "told in the message of walk plan-p". The first word comes only as the reminder, 1 h later for urgent and 12 h later for a warning. The window is small (a failing send plus a restart before the next retry), so I do not hold the merge for it. Suggested follow-up issue: mark the records told when thewalk-enditem is actually sent, for example by having the item carry its record keys, which is how a single item's record is marked inflush.B. A takeover in
releaseskips the filters that decide what is still open. walks.go:378-386 hands a message to a later walk as soon as the record is live. It does this before checking whether the record is silenced, was said after it was held (lastSaid(r).After(hm.At)), or has an open ask (§3). I reproduced the last case with a scratch test: a condition is held by plan-a, its ask is offered (and so said), plan-b starts on the same machine, and plan-a ends.heldBythen gives[plan-b], and history records a secondheld. Nothing is said twice, because plan-b's end filters the record out. Butoperator-channel.openshowsheld_by_walksand history showsheldfor a message that ADR 0308 §3 says is never held. The fix is to apply the same checks before the takeover. Doing that needsneverHeldwithouth.mu, oropenAskchecked before the lock.C. The order check compares a capped time. In
Walked,atis capped atnow. The tombstone check (walks.go:298) compares that cappedatwithend.Endedandend.Started. Suppose the controller's clock is ahead of the router's, and a short walk's end arrives before its start. The capped start can then come afterend.Endedand fail to equal the uncappedend.Started, so the start opens a walk that has already ended, and that walk holds until its bound. Comparingev.StartedAt(uncapped) withend.Startedwould close this. Both run on novox today, so this is theoretical.Not merged by me.