Kill the connector at every ledger state, and prove it recovers by jorgemanrubia · Pull Request #743 · basecamp/basecamp-cli

Kill the connector at every ledger state, and recover

jorgemanrubia

committed

7ace7a4

Run the recovery harness against the real agents, opted into

jorgemanrubia

committed

7665f7a

Scale the prompt budget check by a measured tokenizer ratio

jorgemanrubia

committed

46f41e4

Follow the rebase: bind the dispatch with its context, reconcile a restart's sending intent

jorgemanrubia

committed

3719fff

Assert on the worker the ledger recorded, and join the follow-up before the crash

jorgemanrubia

committed

25bc026

Check the feed's checkpoint after the walk closed the loss

jorgemanrubia

committed

c3a543e

Put the import's tombstone entry first, so a crash at the first entry has something to leave behind

jorgemanrubia

committed

668e1cf

Watch every checkpoint the walk's run holds, not only its last

jorgemanrubia

committed

0cd2603

Satisfy the linter: the ledger's own context, and no needless conversion

jorgemanrubia

committed

eef8b35

Close the adversarial review: a real state directory, an ordered overflow check, a guard that fires

jorgemanrubia

committed

0a986f1

Keep the real-agent run off every Basecamp but a closed port, and expect a launching crash to be held

jorgemanrubia

committed

5d60ed9

Floor the checkpoint watcher on what stands long enough to be seen

jorgemanrubia

committed

359d86b

Order the overflow check by event, not by page: one page may serve both

jorgemanrubia

committed

5f4e697

Hold the harness to the one-owner rule: a surviving tree keeps its attempt, and no settled attempt's worker runs

jorgemanrubia

committed

e817578

Group the harness's imports

jorgemanrubia

committed

625f151

Close the second adversarial review: no line kill races the dispatcher, holds proved by work that goes on, groups with children

jorgemanrubia

committed

512b0ae

Answer Copilot: a portable zombie check, a comparator that cannot overflow, no dead reply lister

jorgemanrubia

committed

ac80714

Close the third adversarial review; rebase onto the outbox's start and the credential rule

jorgemanrubia

committed

547e3dc

Take the task token from the connector's socket, as the bridge does, and keep the harness's paths short enough for one

jorgemanrubia

committed

67c6c0c

Dial the token socket on the turn's context

jorgemanrubia

committed

99293fc

Follow the prompt's URL cap: the worst case is the longest URL it repeats, and a longer one is omitted

jorgemanrubia

committed

f5e6ba1

Give the opt-in real run a stored credential, since the bridge hands its server none, and keep every run's log

jorgemanrubia

committed

ac289c6

Group the real run's imports

jorgemanrubia

committed

7c48a0e

Let a loaded box take its time: a run fails on what the ledger says, not on the clock

jorgemanrubia

committed

25535bf

Close the fourth adversarial review: watch the session directory, hold the declaration to the socket, and keep the harness's deadline outside the run's

jorgemanrubia

committed

f56ab34

Make the credential check fail closed: what it could not read, and what it never had to check

jorgemanrubia

committed

2f77c5e

Require the connector to say what it decided, where the ledger cannot show it decided anything

jorgemanrubia

committed

bf4129f

Check every task token the connector minted, whoever its worker was

jorgemanrubia

committed

03856d5

Navigation Menu

Kill the connector at every ledger state, and prove it recovers - #743

Kill the connector at every ledger state, and prove it recovers#743
jorgemanrubia wants to merge 28 commits into
connect-operator-commandsfrom
connect-recovery-harness

Commits on Sep 17, 2026

Kill the connector at every ledger state, and recover

Run the recovery harness against the real agents, opted into

Scale the prompt budget check by a measured tokenizer ratio

Follow the rebase: bind the dispatch with its context, reconcile a restart's sending intent

Assert on the worker the ledger recorded, and join the follow-up before the crash

Check the feed's checkpoint after the walk closed the loss

Put the import's tombstone entry first, so a crash at the first entry has something to leave behind

Watch every checkpoint the walk's run holds, not only its last

Satisfy the linter: the ledger's own context, and no needless conversion

Close the adversarial review: a real state directory, an ordered overflow check, a guard that fires

Keep the real-agent run off every Basecamp but a closed port, and expect a launching crash to be held

Floor the checkpoint watcher on what stands long enough to be seen

Order the overflow check by event, not by page: one page may serve both

Hold the harness to the one-owner rule: a surviving tree keeps its attempt, and no settled attempt's worker runs

Group the harness's imports

Close the second adversarial review: no line kill races the dispatcher, holds proved by work that goes on, groups with children

Answer Copilot: a portable zombie check, a comparator that cannot overflow, no dead reply lister

Close the third adversarial review; rebase onto the outbox's start and the credential rule

Take the task token from the connector's socket, as the bridge does, and keep the harness's paths short enough for one

Dial the token socket on the turn's context

Follow the prompt's URL cap: the worst case is the longest URL it repeats, and a longer one is omitted

Give the opt-in real run a stored credential, since the bridge hands its server none, and keep every run's log

Group the real run's imports

Let a loaded box take its time: a run fails on what the ledger says, not on the clock

Close the fourth adversarial review: watch the session directory, hold the declaration to the socket, and keep the harness's deadline outside the run's

Make the credential check fail closed: what it could not read, and what it never had to check

Require the connector to say what it decided, where the ledger cannot show it decided anything

Check every task token the connector minted, whoever its worker was