Skip to content

Latest commit

 

History

History
77 lines (59 loc) · 4.8 KB

File metadata and controls

77 lines (59 loc) · 4.8 KB

squad (DeDuva fork) — Roadmap

This file is the repo's only status ledger. A PR that completes, starts, pauses, or supersedes a milestone updates this file in the same PR. Scope authority for the bench track is packages/duva-bench/PLAN.md — this file says where things are, not how the next piece gets built.

Mission

A fork of bradygaster/squad replatformed to be model- and harness-agnostic, hosting two research packages: squad-lab (goal-setting and cross-provider A/B experiments — one ADP intent, N verified runs) and duva-bench's squad track (in-process factorial experiments over coding-agent arms, twin toolsets, pre-registered statistics).

Where this fits

squad and ADP are each other's test fixture: squad wraps the gh CLI, ADP's locked success criterion is that unmodified gh works against it, so squad runs against ADP with zero code changes. Statistics come from adp-replay's adp_replay.stats.paired, pinned by commit. The Harbor track of duva-bench lives at github.com/DeDuva/duva-bench. The dependency map is ADP's docs/ecosystem.md; milestone reports and the portfolio audit live in the private github.com/DeDuva/duva-lab-tpm.

Milestone ledger

The fork itself

Item Status Evidence / detail
Replatform (model/harness-agnostic) complete 2026-08-07 The replatform-0.11 line is spent; dev was made default at its tip. PRs #74–#89 carry the lab work
Upstream sync paused Decision DEFER, taken against a stated rubric (duva-lab-tpm reports/m1-upstream-sync-decision.md). Resume condition: the rubric's terms change

squad-lab (packages/squad-lab)

Item Status Evidence / detail
M0–M14 complete Fifteen milestone reports in duva-lab-tpm reports/ — through the 80-run, ~$28 variance study (M14). ⚠️ M12's token figures are known wrong and annotated there
Next scope (M15) not started — awaiting a decision squad-lab is active, not complete (author, 2026-08-08). duva-lab-tpm reports/m15-next-scope-proposal.md collects every forward ask, marks four as shipped, and puts three options with a recommendation. It is a proposal, not a plan of record

duva-bench, squad track (packages/duva-bench)

Item Status Evidence / detail
S0–S7 complete 2026-08-08 One PR per milestone into dev. Gates SG1 and SG2 passed. S7 ran a live 24-trial pilot: $8.03, 24/24 runs verified, pre-registration printed unchanged
Gate SG3a (this track's own result) passed 2026-08-08 studies/a-tool-familiarity-pilot/FINDINGS.md — noise floor stated before any contrast, every included run verified
Study A proper not started — awaiting a decision The pilot's real finding: the pre-registered primary metric (hallucinatedCallRate) has a floor effect — 0.0000 on all 24 trials in every arm. Study A needs a re-registered metric, not an amendment; no design is written yet
Gate SG3b (cross-track memo) paused Precondition: the Harbor track reaching M8 with a shared task set. Harbor's own pause was lifted 2026-08-08 (an end-to-end probe ran there), so this is now a matter of that track doing the work. The expected result is registered in advance in studies/a-tool-familiarity-pilot/CROSS-TRACK.md

Now / Next / Later

  • Now: nothing in flight.
  • Next: two author decisions unblock everything queued here — squad-lab's M15 option, and Study A's replacement metric (then a new pre-registration and a pilot-scale re-run before committing to the full study).
  • Later: SG3b when the Harbor track reaches M8; multi-turn quiescence only if the cross-track memo shows this track earns it (PLAN.md §5).

Blockers and open decisions

  • squad-lab M15 — decision needed (author): pick among the three options in m15-next-scope-proposal.md, or decline them all. Proposal dated 2026-08-08.
  • Study A primary metric — decision needed (author): the three candidate designs are named in the pilot's FINDINGS (larger tool surface; tools withheld from the schema; tasks requiring an absent tool). Re-running at pilot scale costs real money (calibration: $8.03 for 24 trials).
  • No blockers otherwise. Verified 2026-08-08: keys, statistics venv, and the ADP stack all work; the statistics path is CI-run and commit-pinned (PR #104).

Plan documents

  • packages/duva-bench/PLAN.md — the squad track's plan of record: agent rules (§0, including the import boundary), the reuse map (§2), milestones S0–S7 with gates.
  • packages/duva-bench/studies/a-tool-familiarity-pilot/ — the executed pilot: pre-registration, findings, cross-track registration.
  • duva-lab-tpm (private) — squad-lab's milestone reports M0–M14, the M15 proposal, and the portfolio audit.