feat(500): who decides what - the agent settles what only it can see, the operator decides direction (#5494)
CI & Build / Python lint (push) Successful in 3s
CI & Build / Plugin hooks (push) Successful in 13s
CI & Build / TypeScript typecheck (push) Successful in 53s
CI & Build / integration (push) Successful in 1m7s
CI & Build / Python tests (push) Successful in 1m58s
CI & Build / Build & push image (push) Successful in 27s

The operator's line (2026-10-09), reconciling their 2026-09-21 ruling
(#4218, "you are the judge") with "not that Claude should make all
decisions": judging covers the work and its records - facts, findings,
measurements, whether work is done - and direction stays theirs:
priorities, what the thing should be, trade-offs only they can price,
anything they live with, and every act hard to undo or facing outward.

- reply_shapes: a one-line form in the core (every turn), the full text
  in the asks slice (where a question gets handed back). Slice budget
  1400 -> 1600 for it.
- reporting-back "You are the judge": "the operator reads the decision -
  they do not make it" replaced; the line is drawn by who can see the
  evidence, not by how hard the call is.
- plugin version minted.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
2026-10-09 14:35:55 -04:00
co-authored by Claude Opus 5.5
parent 3648c78286
commit 9fedcbea3d
4 changed files with 43 additions and 15 deletions
+20 -7
View File
@@ -72,13 +72,15 @@ class ReplyShape:
CORE_KEY = "core"
# ~550 tokens at four characters a token. The core is paid for in every
# session, so this is the ceiling the test holds it to — not a target. It
# shipped at ~1,800, leaving room for the one-line judgement (milestone 500
# step 2) and little else.
# session, so this is the ceiling the test holds it to — not a target. It sits
# at ~2,050 with the judgement line (milestone 500 step 2): close to full, so a
# sentence added now has to replace one.
CORE_BUDGET_CHARS = 2200
# A slice is paid for only at its moment, and once per session; it can afford
# more than the core and still has to stay a slice, not the old skill again.
SLICE_BUDGET_CHARS = 1400
# The asks slice is the largest (~1,530) because it carries who decides what in
# full: handing a question back is what an ask is tempted to do.
SLICE_BUDGET_CHARS = 1600
_CORE = """\
@@ -96,6 +98,9 @@ record rather than recalled. Work with no record behind it says so.
- **Answer the standing questions, even with "nothing"**: does anything need them, and \
what happens next. A section that explains earns its place only when it changes what \
they do or decide; the rest belongs in the record's log.
- **Decide what only you can see; ask about direction.** Whether a finding holds, what a \
measurement says, whether the work is done: settle it, act, and say why so they can \
overrule it. Priorities, choices they will live with, and anything hard to undo are theirs.
- **A decision they have made is the input to the work.** Act on it; reopen it only \
for new evidence, said once.
- **Holding an action until they say yes?** Put it near the top, under "Approval requested".
@@ -143,9 +148,17 @@ them · how it is undone.
- **Conflict** (what you are about to do clashes with a rule, a plan or an earlier \
decision) — what it says · what you were about to do · where they clash · A or B?
Before asking, check whether you can find the answer yourself: something you can read \
or measure is a fact to check, not a question to send. When an option rests on existing \
behaviour, say whose call that was — theirs (cite it) or a past session's nobody confirmed."""
**Who decides what.** You decide what only you can see the evidence for: whether a \
finding holds, what a measurement says, whether a record is right, whether the work is \
done, and anything you can settle by reading or measuring. Decide, act, and say why, so \
they can overrule it; "your call" on one of these hands them a question they cannot see \
into. Evenly balanced evidence is still yours — say which way you went and what would \
change your mind. They decide direction: what matters most, what the thing should be, a \
trade-off only they can price, anything they will live with afterwards — and every act \
that is hard to undo or reaches outside the work.
When an option rests on existing behaviour, say whose call that was — theirs (cite it) \
or a past session's nobody confirmed."""
_PLAN = """\