Skip to content

queue.md

Latest commit

 

History

History
100 lines (78 loc) · 5.52 KB

queue.md

File metadata and controls

100 lines (78 loc) · 5.52 KB
Raw

Queue

The one file the human maintains by hand, and the only thing they have to think about between slots: what I want done next, roughly in order.

Order is priority. There is no priority: field here — moving an entry up is the act of prioritising it. Nothing else about an entry needs to be true for it to sit in this file; the planner decides what is actually startable.

A batch is never composed here. pyauto-brain batch plan proposes the next one against a review-minute budget (PyAutoBrain/agents/conductors/batch/AGENTS.md), and the human approves or edits that proposal in their slot — they do not hand-pick members.

What the planner actually reads, and nothing else:

  1. this file — the kind: prompt entries, in file order;
  2. the draft/ backlog, graded through the sizing faculty;
  3. the review-queue depth, derived from active.md: a row naming an open PR (library-pr: / workspace-pr:) or whose status: says awaiting-merge / PR open. That is what "the open-PR state" means here — the ledger's own count, not a live one. Nothing in plan calls GitHub, and a shift stays plannable offline.

The order in this file is what the planner does with it: a queued prompt outranks an unqueued one, and among queued prompts the file's order wins. Below the queue it sorts cheapest-first, because that part of the list is read when there is a slot to fill and the question is what fits in it. An entry written against draft/… keeps its place after the prompt moves to active/.

An entry leaves the queue when its work reaches complete/, not when it is dispatched. A task in flight is still the thing the human wants.

Schema

## <n>. <label>
- kind: prompt | retired
- ref: draft/<work-type>/<target>/<name>.md    # kind: prompt
- note: <one line, optional>

Two kinds. prompt is a named prompt file — the live case, and the only one pyauto-brain batch plan reads. retired is an entry that has left the queue (shipped, shelved, or moved to the Cortex) but whose note is worth keeping, because what the human wanted and why it stopped being wanted is the queue's only history. Retired entries carry the pointer that replaced them (shipped-as: / shelved-as: / moved-to:) and live in the footer, not the ordered list.

The two non-prompt kinds — "the named epic's next phase" and "anything ready on this theme" — were removed on 2026-09-03. Both were written for science epics, which now live in the Cortex, and no code ever read them (_batch.py globs draft/**/*.md). The optional lane field went with them: where the work can run is a Cortex phase-header key now (PyAutoCortex/REFERENCE.md), and every Cortex phase is local-dev by construction.

1. numba solve vs JAX sparse operator

  • kind: prompt
  • ref: draft/research/autoarray/numba_solve_vs_jax_sparse_operator.md
  • note: review intake 2026-08-31 — same linear algebra? GPU-JAX amenable or CPU-sparsity-bound?

2. Witness campaign

  • kind: prompt
  • ref: draft/feature/pyautomind/witness_campaign.md
  • note: highest-leverage backlog item (151/153 judge); human-declared witnesses only; fill work

Retired

Entries that have left the queue. Kept for the record of what was wanted and what replaced it; batch plan ignores them.

Retire parked-at-ship under --auto

  • shipped-as: PyAutoBrain#365 → complete/2026/09/batch-no-park-at-ship.md (2026-09-07)
  • note: human 2026-08-31 — "the parked thing is an annoying middle ground which requires human time"; the AUTONOMY.md doctrine edit shipped: an effective-supervised --auto run now decides-and-flags at the ship checkpoint instead of parking, and the batch planner rejects only human-required

Carried members formalisation

  • shelved-as: complete/archive/shelved/batch_carry_forward.md (2026-09-03)
  • note: the Cortex half shipped (_batch.py carried_members, carried:/carried-from: on the Cortex batch record); the dev half retired with the two-slot-batching epic

Subhalo validation follow-up wave

  • moved-to: PyAutoCortex phases/subhalo_validation/ (2026-09-01)
  • note: review tweak 2026-08-31 — human wants to inspect the _adapt_split_fix result; no subhalo[2]; RectangularBilinear runs. The wave is now four Cortex phases (one per lens); record complete/2026/09/subhalo-followup-moved-to-cortex.md

Euclid docs + structure tidy (phase 3a addendum)

  • ref: euclid-dr1-prep
  • shipped-as: euclid#47 → PR#48, complete/2026/09/restore-pipeline-narrative-prose.md (2026-09-01)
  • note: review tweak 2026-08-31 was folded into the phase-3 prompt and shipped with it — README structural edits, the scripts/tools/ move and the AGENTS.md shortening all landed in PR #48; the start_here.py contradiction was ruled at plan time (kept as a documented shim over initial_lens_model.fit)

Batches strip on the dashboard

  • shipped-as: PyAutoBrain#341 → complete/2026/09/batch-status-box.md (2026-09-03)
  • note: shipped as the batch status box on both dashboards; the prompt was folded into that record (2b5af675) and ref: no longer resolves — reconciled 2026-09-03

batch collect

  • shipped-as: PyAutoBrain#332 → PR#333, complete/2026/09/batch-collect.md (2026-09-02)
  • note: pyauto-brain batch collect shipped; the phase's remaining verb is re-filed as draft/feature/pyautobrain/batch_slice.md

Euclid Lane: recut

  • shipped-as: superseded by cortex-birth phase 5 (2026-09-02)
  • note: complete/archive/shelved/batch_science_lane.md is marked superseded; leg A (the Lane: recut) has no surface left — every Cortex phase is Lane: local-dev by construction