rounds: a stock seat gets /ccc-work as its prompt's slash command; the acts read ccc-agent list/send too (ruling A16) #42

Open
caoer115 wants to merge 0 commits from agent/379d2e5a-stock-skill-slash into main
Owner

Ruling A16 (leader 21b7d2e1). Stacked on #41 (branch off its head 99b81cd); merging this lands #41 with it.

r271's and r272's stock-seat-acts seats skipped a prose load /ccc-work. r272's then listed and sent through ccc-agent, which the readers did not read, so the row fell to fail although every act landed.

  • Skill load, deterministic: 65 (stock-acts) and 66 (stock-e2e) prompts open with /ccc-work <task>, the way an official-binary user loads it. transcript.py skill-load reads what the binary recorded. Probed on official 2.1.301 TUI on nyc-2:
    • a loaded skill is the <command-name>/ccc-work</command-name> user entry, then an isMeta entry Base directory for this skill: <dir>/ccc-work with the body, and no Skill call;
    • an unknown name leaves only a system Unknown command: /<name>, and the model is never called;
    • a Skill call answered also counts as loaded.
  • Verdict (acts_decide): every act landed and the skill loaded → pass; asked in words and skipped → missed; the slash command loaded nothing → fail.
  • Acts readers: they also read the seat's own live entry of ccc-agent list and the result of ccc-agent send <peer>. A real daemon failure still reads fail.

Gate (~/.local/share/ucc/work/379d2e5a/slash-replay/{replay.sh,replay.out} on devstation; probe transcripts in …/slash-probe/):

  • r270, r271, r272 acts seats: new verdict missed (recorded pass / missed / fail). r272 sees-self now reads yes off ccc-agent list. r270's harness predates the load check, and its seat made no load either.
  • Probe: /ccc-work loaded → pass; a slash entry with no body → fail (noload); an unknown name → unknown: Unknown command: /ccc-nonexistent-xyz.
  • replay-acts.sh on r272's seat: 23 mutations × 4 decisions, all resolve.
  • bash -n on acts.sh, 65, 66, k4.sh; py_compile on transcript.py: ok.
Ruling A16 (leader 21b7d2e1). Stacked on #41 (branch off its head `99b81cd`); merging this lands #41 with it. r271's and r272's `stock-seat-acts` seats skipped a prose `load /ccc-work`. r272's then listed and sent through `ccc-agent`, which the readers did not read, so the row fell to `fail` although every act landed. - **Skill load, deterministic:** 65 (stock-acts) and 66 (stock-e2e) prompts open with `/ccc-work <task>`, the way an official-binary user loads it. `transcript.py skill-load` reads what the binary recorded. Probed on official 2.1.301 TUI on nyc-2: - a loaded skill is the `<command-name>/ccc-work</command-name>` user entry, then an `isMeta` entry `Base directory for this skill: <dir>/ccc-work` with the body, and no Skill call; - an unknown name leaves only a system `Unknown command: /<name>`, and the model is never called; - a Skill call answered also counts as loaded. - **Verdict (`acts_decide`):** every act landed and the skill loaded → pass; asked in words and skipped → missed; the slash command loaded nothing → fail. - **Acts readers:** they also read the seat's own live entry of `ccc-agent list` and the result of `ccc-agent send <peer>`. A real daemon failure still reads fail. Gate (`~/.local/share/ucc/work/379d2e5a/slash-replay/{replay.sh,replay.out}` on devstation; probe transcripts in `…/slash-probe/`): - r270, r271, r272 acts seats: new verdict `missed` (recorded pass / missed / fail). r272 sees-self now reads yes off `ccc-agent list`. r270's harness predates the load check, and its seat made no load either. - Probe: `/ccc-work` loaded → `pass`; a slash entry with no body → `fail` (noload); an unknown name → `unknown: Unknown command: /ccc-nonexistent-xyz`. - `replay-acts.sh` on r272's seat: 23 mutations × 4 decisions, all resolve. - `bash -n` on acts.sh, 65, 66, k4.sh; `py_compile` on transcript.py: ok.
r271 K4 claude-patched W4: the seat answered "The subagent is running in the
background; I'll print its answer when it returns, then
MARK-K4-claude-patched-W4" while its background Agent ran; wait_text matched
the mark in that planned line and k4_kill ended the seat before the W4 put, so
K4-actor/grade/principal read fail on a write that never ran.

k4_wait_act ends a subagent's write (TUI W4, -p W3) when the record exists in
the user's root, or when the subagent's answer is in the seat's transcript:
transcript.py sub-answer reads the Agent call's result (never a background
launch's async_launched notice) or the task-notification naming the call's
tool-use-id. A seat's own write (W1-W3) keeps its mark wait.
r271's and r272's stock-seat-acts seats skipped a prose "load /ccc-work" on the
same binary and model whose r270 seat followed it; r272's then listed and sent
through ccc-agent, which the readers did not read, so the row fell to fail
while every act had landed.

- The skill reaches the seat as the prompt's slash command (65 stock-acts, 66
  stock-e2e), the way an official-binary user loads it. transcript.py
  skill-load reads the load off what the binary recorded: the
  <command-name>/ccc-work</command-name> entry and the isMeta "Base directory
  for this skill: <dir>/ccc-work" body (official 2.1.301, probed on nyc-2), or a
  Skill call answered; an unknown name leaves only "Unknown command: /<name>".
- acts_decide: every act landed and the skill loaded -> pass; asked in words and
  skipped -> missed; the slash command loaded nothing -> fail (the binary's).
- acts_read also reads the seat's own live entry of `ccc-agent list` and the
  result of `ccc-agent send <peer>`; a real daemon failure still reads fail.

Replay: r270, r271, r272 read missed (never pass); the probe's slash load
decides pass, a slash entry with no body fail; replay-acts.sh mutations move no
decision.
This branch is already included in the target branch. There is nothing to merge.
View command line instructions

Checkout

From your project repository, check out a new branch and test the changes.
git fetch -u origin agent/379d2e5a-stock-skill-slash:agent/379d2e5a-stock-skill-slash
git switch agent/379d2e5a-stock-skill-slash

Merge

Merge the changes and update on Forgejo.

Warning: The "Autodetect manual merge" setting is not enabled for this repository, you will have to mark this pull request as manually merged afterwards.

git switch main
git merge --no-ff agent/379d2e5a-stock-skill-slash
git switch agent/379d2e5a-stock-skill-slash
git rebase main
git switch main
git merge --ff-only agent/379d2e5a-stock-skill-slash
git switch agent/379d2e5a-stock-skill-slash
git rebase main
git switch main
git merge --no-ff agent/379d2e5a-stock-skill-slash
git switch main
git merge --squash agent/379d2e5a-stock-skill-slash
git switch main
git merge --ff-only agent/379d2e5a-stock-skill-slash
git switch main
git merge agent/379d2e5a-stock-skill-slash
git push origin main
Sign in to join this conversation.
No reviewers
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
caoer115/ucc!42
No description provided.