Skip to content

v0.15.0 — :investigator cohort + per-cohort default-model axis

Choose a tag to compare

View all tags
@overdrive-dev overdrive-dev released this 17 Apr 02:35
· 431 commits to main since this release
v0.15.0
5bda4c3
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

v0.15.0 — 5 cohorts + per-cohort default-model axis (cost-drop release)

One feature (feature/cohort-taxonomy-v3, ADR-M010-A M010.1 amendment) grows the cohort taxonomy from 4 → 5 and binds each cohort to a default model — so presets now tune both axes (model and effort) per role in one flag. Ships a measurable cost reduction with zero reasoning regression on :planner / :adversarial paths.

Headline

  • New :investigator cohort — debugger, debug-session-manager, user-profiler split out of :verifier. Hypothesis-driven investigation of runtime/behavioral state is a distinct reasoning shape from static artifact verification.
  • Per-cohort default-model field — each cohort now declares a default model in pkg/.aihaus/skills/aih-calibrate/annexes/cohorts.md, informed by Anthropic's models overview:
Cohort Default model Default effort Rationale
:planner opus high (xhigh for binding-4) "Complex reasoning + agentic coding"
:doer sonnet high "Speed-intelligence balance" (daily driver)
:verifier haiku high "Near-frontier intelligence, fastest" — boolean artifact checks
:investigator sonnet high Hypothesis-driven reasoning, not artifact checks
:adversarial opus high/max Binding review; preset-immune per ADR-M008-C/M010-A
  • Shipped distribution: v0.14.0 32 opus / 11 sonnet / 0 haiku → v0.15.0 16 opus / 21 sonnet / 6 haiku. Roughly 30-40% cost reduction on an average multi-milestone run.

What changed

  • 16 agents moved from opus to sonnet or haiku per new cohort defaults (8 :doer opus→sonnet; 4 :verifier opus→haiku; 2 :verifier opus→sonnet overrides — nyquist-auditor + ui-auditor; 2 :investigator opus→sonnet — debugger, debug-session-manager).
  • 2 additional :verifier agents moved sonnet→haiku (doc-verifier, ui-checker).
  • Preset map rewritten (annexes/presets.md) — each preset now expresses 5 cohort tuples + bounded ### Overrides block. balanced preset on clean v0.15.0 install = zero-diff no-op (AC-07 relative to v0.15.0). cost-optimized pushes :verifier + :investigator to medium effort for maximum cost reduction.
  • Two :verifier overrides retained on sonnet:
    • nyquist-auditor (writes + iteratively debugs minimal behavioral tests — code-gen surface)
    • ui-auditor (6-pillar visual audit — needs vision + aesthetic judgment)
  • Smoke-test Check 28 updated to exercise :investigator cohort row in the v2 sidecar round-trip.

Binding decisions

  • ADR-M010-A M010.1 amendment (new, appended to .aihaus/decisions.md) — formalizes 4 → 5 cohorts + per-cohort default-model field. Additive extension, not supersession. Preserved:
    • ADR-M008-A amendment — cohort-scoped model: edits; per-agent dual-axis escape hatch.
    • ADR-M008-C / M010 Decision (d) — :adversarial cohort preset-immunity, 4 members unchanged (plan-checker, contrarian, reviewer, code-reviewer).
    • ADR-M009-A — sidecar ownership (user-owned, never-committed, absolute values, schema-gated). Schema v2 additive; new :investigator cohort name accepted; older readers warn+skip unknown rows.

Usage

New cohort CLI

# Downgrade :verifier cohort to haiku explicitly (matches new default)
/aih-calibrate --cohort :verifier --model haiku --effort high

# Tune :investigator specifically
/aih-calibrate --cohort :investigator --model sonnet --effort medium

# Apply the new cost-optimized preset (haiku verifiers + sonnet investigators)
/aih-calibrate --preset cost-optimized

Read-only audit

/aih-calibrate --inspect

Phase-1 report now includes cohort membership for all 43 agents plus the default-model column (GFM pipe table, no Unicode box-drawing — fix from v0.14.0).

Cohort membership

  • :planner (17) — analyst, architect, planner, product-manager, roadmapper, eval-planner, domain-researcher, phase-researcher, project-researcher, ai-researcher, framework-selector, ux-designer, ui-researcher, advisor-researcher, assumptions-analyzer, brainstorm-synthesizer, research-synthesizer
  • :doer (11) — implementer, frontend-dev, executor, code-fixer, doc-writer, intel-updater, test-writer, notion-sync, project-analyst, codebase-mapper, pattern-mapper
  • :verifier (8) — verifier, integration-checker, security-auditor, doc-verifier, eval-auditor, nyquist-auditor, ui-auditor, ui-checker
  • :investigator (3, new in v0.15.0) — debugger, debug-session-manager, user-profiler
  • :adversarial (4, preset-immune) — plan-checker, contrarian, reviewer, code-reviewer

Full (cohort, model, effort) per agent: pkg/.aihaus/skills/aih-calibrate/annexes/cohorts.md.

Upgrade notes

v0.14.0 → v0.15.0 — distribution shift, not regression

  • First /aih-calibrate --preset <any> after upgrading will mutate frontmatter for the 16 agents whose cohort defaults changed. This is intentional — the v0.15.0 cost-drop is the point of this release.
  • If you were running --preset balanced on v0.14.0 and you want the v0.15.0 balanced distribution, simply re-run /aih-calibrate --preset balanced after update.
  • If you want to preserve v0.14.0's opus-heavy distribution on v0.15.0, apply --preset quality-first or use explicit per-agent overrides. :adversarial members never change.

Schema v2 additive — v1 sidecars still restore

  • v0.13.0 schema=1 sidecars continue to restore byte-identically on v0.15.0 (unchanged from v0.14.0).
  • v0.14.0 schema=2 sidecars without cohort.investigator.* rows parse fine — v0.15.0 reader treats missing cohort rows as "use cohort defaults".
  • First /aih-calibrate run on v0.15.0 writes a v2 sidecar that includes cohort.investigator.* rows when relevant.

Known implications

  • Agents that were running at effort: xhigh and model: opus pre-v0.15.0 may respond slightly faster at model: sonnet but with less deliberation. The :planner cohort (17 members including all planning + research agents) stays on opus — the reasoning-critical path is untouched.
  • :verifier agents at haiku may produce slightly terser findings but catch the same failures; multiple M010 smoke tests verify this.

Claude Code requirements (unchanged)

  • effort: xhigh requires Claude Code ≥ v2.1.111. Older releases silently fall back to effort: high.
  • permissions.defaultMode: auto requires Claude Code ≥ v2.1.83. Older releases ignore the field.

Deferred / follow-up

  • R11 (carried from M009) — pre-existing update_aihaus_dir crash window. Out of scope; dedicated plan if prioritized.
  • R12 (carried from M009) — pre-v0.13.0 hand-edited frontmatter migration. One-time user action; documented in state-file.md.
  • Per-cohort effort-only CLI (--cohort :<name> --effort X without --model) — currently rejected by D-2 (both axes required). Could relax in a future minor bump if user feedback warrants it.
  • haiku-doers preset — proposed during M010 planning, not shipped. Candidate for v0.16.0 if haiku-tier :doer performance proves acceptable in real use.

Full changelog

Commit: 14f2072 feat(calibrate): split :investigator cohort + add per-cohort default-model axis (v0.15.0) on feature/cohort-taxonomy-v3.

See ADR-M010-A M010.1 amendment (local / gitignored — on-disk file in the dogfood install) for the formal decision record.

Assets 2
Loading