Run 202's findings: a revise's dependencies, the watchdog's redo, joined additions, one README #559

Closed
opened 2026-11-19 15:56:34 -05:00 by cmoriarty · 1 comment
Owner

What happens (run 202, braid#557)

  1. A revise's change to a group's Depends on: line is never seen. The scheduler reads .osf/budget.json, which is written once, after propose.
  2. When groups run side by side, each of the first ones writes README.md, so the second to merge conflicts on it.
  3. Two groups that each add a block at the same point of one file (the browser spec, the stylesheet) conflict, and an agent spends minutes resolving what is only two additions.
  4. The stall watchdog redoes a step that is waiting on its own shell command at 1200 s, before the command's own 1800 s limit. The redo starts the step from scratch: 72 min of Pairs' work was lost.
  5. The note for groups side by side says "write only your own suite's command", and a group whose suite is already in the joined command stops to work out what that means.

Fix (decided with the owner on 2026-11-19)

  1. Write .osf/budget.json again from the final tasks.md before apply starts.
  2. Side by side, only group 1 writes the README.
  3. Braid joins a conflicting part whose base side is empty, putting the run's lines first, then the group's. Only in files the base has; never lock files; JSON must still parse. Any other conflict still goes to the group.
  4. (a) A step is spared until its command's own limit plus a minute. (b) The shell wrapper stops a command that has printed more than about 20 MB. (c) The log names the rule that fired. (d) A stall redo of a step that has ticked a task keeps its work (the resume path) instead of resetting.
  5. The lane-command line becomes a condition: only when your work adds a suite the lane does not run yet.

Done when: each fix has a test that fails without it, 1 and 3 are shown by a simulated run, and braid#536's standard run starts on the build.

**What happens** (run 202, braid#557) 1. A revise's change to a group's `Depends on:` line is never seen. The scheduler reads `.osf/budget.json`, which is written once, after propose. 2. When groups run side by side, each of the first ones writes `README.md`, so the second to merge conflicts on it. 3. Two groups that each add a block at the same point of one file (the browser spec, the stylesheet) conflict, and an agent spends minutes resolving what is only two additions. 4. The stall watchdog redoes a step that is waiting on its own shell command at 1200 s, before the command's own 1800 s limit. The redo starts the step from scratch: 72 min of Pairs' work was lost. 5. The note for groups side by side says "write only your own suite's command", and a group whose suite is already in the joined command stops to work out what that means. **Fix** (decided with the owner on 2026-11-19) 1. Write `.osf/budget.json` again from the final tasks.md before apply starts. 2. Side by side, only group 1 writes the README. 3. Braid joins a conflicting part whose base side is empty, putting the run's lines first, then the group's. Only in files the base has; never lock files; JSON must still parse. Any other conflict still goes to the group. 4. (a) A step is spared until its command's own limit plus a minute. (b) The shell wrapper stops a command that has printed more than about 20 MB. (c) The log names the rule that fired. (d) A stall redo of a step that has ticked a task keeps its work (the resume path) instead of resetting. 5. The lane-command line becomes a condition: only when your work adds a suite the lane does not run yet. **Done when:** each fix has a test that fails without it, 1 and 3 are shown by a simulated run, and braid#536's standard run starts on the build.
Author
Owner

Shipped in PR 560 (merge 07bea0c), live on production at 21:39 UTC, about 3 minutes after the merge. Archived in PR 562 ([skip ci]).

  • The estimate follows each revision. The four revision commits run the estimate before they commit. sim/revision-moves-a-dependency fails with the estimate taken out of the operator's commit (group 2 starts beside group 1, and its edit of group 1's code misses) and passes with it.
  • A merge joins two additions at one point, the run's lines first, then the group's. It never joins a file both sides made, a lock file, a binary, or JSON that would not parse. On run 202's four real conflicts it gives what the agents wrote: three byte for byte, and one with its CSS blocks the other way round. sim/parallel-additions-joined fails with the join off and passes with it, and parallel-conflict still conflicts.
  • One README side by side: only the plan's first group is asked to write it.
  • The watchdog spares a step waiting on its own shell command until the command's own limit and a minute. A stalled step that has ticked a task keeps its work and continues from Braid's note, and the redo's reason names the call.
  • The wrapper stops a command that has printed more than 20 MB. It also reaps the shell now while it stops its group: on Linux every stop had waited about 4 s for nothing (4.45 s, against 0.14 s now).
  • The note asks for a suite's command only when the group's work adds one.

Tests: ./tools/test.sh is green (900 backend, 204 UI unit, 61 browser), each new test fails without its fix, and 21 simulated runs pass. One CI attempt hung on an unrelated scheduler test, filed as braid#561.

Next: braid#536's standard run on this build.

**Shipped** in PR 560 (merge `07bea0c`), live on production at 21:39 UTC, about 3 minutes after the merge. Archived in PR 562 (`[skip ci]`). - **The estimate follows each revision.** The four revision commits run the estimate before they commit. `sim/revision-moves-a-dependency` fails with the estimate taken out of the operator's commit (group 2 starts beside group 1, and its edit of group 1's code misses) and passes with it. - **A merge joins two additions at one point**, the run's lines first, then the group's. It never joins a file both sides made, a lock file, a binary, or JSON that would not parse. On run 202's four real conflicts it gives what the agents wrote: three byte for byte, and one with its CSS blocks the other way round. `sim/parallel-additions-joined` fails with the join off and passes with it, and `parallel-conflict` still conflicts. - **One README side by side:** only the plan's first group is asked to write it. - **The watchdog** spares a step waiting on its own shell command until the command's own limit and a minute. A stalled step that has ticked a task keeps its work and continues from Braid's note, and the redo's reason names the call. - **The wrapper** stops a command that has printed more than 20 MB. It also reaps the shell now while it stops its group: on Linux every stop had waited about 4 s for nothing (4.45 s, against 0.14 s now). - **The note** asks for a suite's command only when the group's work adds one. **Tests:** `./tools/test.sh` is green (900 backend, 204 UI unit, 61 browser), each new test fails without its fix, and 21 simulated runs pass. One CI attempt hung on an unrelated scheduler test, filed as braid#561. **Next:** braid#536's standard run on this build.
Sign in to join this conversation.
No labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
cmoriarty/braid#559
No description provided.