Skip to content

Re-run the dispatch brake on recurrence#1

Open
Nanako0129 wants to merge 1 commit into
cablate:mainfrom
Nanako0129:recurrence-brake
Open

Re-run the dispatch brake on recurrence#1
Nanako0129 wants to merge 1 commit into
cablate:mainfrom
Nanako0129:recurrence-brake

Conversation

@Nanako0129

@Nanako0129 Nanako0129 commented Jul 19, 2026

Copy link
Copy Markdown

Problem

The dispatch brake evaluates one decision at a time. That leaves a blind spot for accumulation: small same-shape tasks that arrive one at a time — review findings with a known remedy, templated documents, single-file follow-up fixes — each pass the Direct-work alternative test individually. No single decision is wrong, and the main agent absorbs them serially for hours.

Observed

From a time/token attribution of two long orchestration sessions (75h + 26h wall-clock; full report in Nanako0129/pilotfish#16, with the matching policy amendments specced in Nanako0129/pilotfish#15, numbers cross-checkable against public TokenBar PRs):

Observation Value
Direct edits vs delegations in the 26h session 1,267 vs 12
Milestone fan-out (isolated worktrees, 3 concurrent) worked correctly — the brake did not over-block
Shape of the absorbed work review-finding fixes with known remedies (~700 source edits), templated writing (~610 doc edits: PR bodies, release accounting, plan documents)
Main-agent share of the session's output tokens ~92%

The primitive table already has the right row ("Repeated homogeneous items → Batch or workflow") — it just never fires, because at each decision point the shape on the table is one item, not the accumulated stream.

Change

One ### subsection under Start with the dispatch brake: after completing the same shape of small task directly about three times, if more instances are still arriving, re-enter the primitive table with the accumulated shape rather than the next single item — batch the remainder to a worker, keep per-item acceptance in the main agent. Capability-neutral wording, per the skill's own conventions; no version bump (left to your release flow).

🤖 Generated with Claude Code

The brake evaluates one decision at a time, which leaves a blind spot:
same-shape small tasks that arrive one at a time each pass the
direct-work test individually, and the main agent absorbs them serially
for hours without any single decision being wrong. Observed in a 26-hour
field session: 1,267 direct edits alongside 12 delegations, with
milestone fan-out working correctly the whole time.

After ~three direct repeats of the same shape with more arriving,
re-enter the primitive table with the accumulated shape: the remainder
are repeated homogeneous items — batch them, keep per-item acceptance.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant