Skip to content

feat(insight-kit): agents-system-v2 — Phase 1-6 restructure + T17 harness + B1 eval/omp extraction - #17

Merged
nyashkn merged 1 commit into
mainfrom
feat/agents-system-v2
Jun 9, 2026
Merged

nyashkn merged 1 commit into
mainfrom
feat/agents-system-v2

Conversation

@nyashkn

@nyashkn nyashkn commented Jun 9, 2026

Copy link
Copy Markdown
Owner

Summary

Bundles 60 commits of agents-system-v2 feature work into main. Three major threads:

1. Monorepo restructure (Phases 1-6)

Scaffolds the platform/ libs/ integrations/ surfaces/ deploy/ group layout (Phase 1), then moves each module group into its canonical location (Phases 2-5), and drops all shims (Phase 6).

2. T-series gate + harness work

  • T8/T9/T10/T11 — Layer-A guards + input-provenance check + coverage-warning gate + selection params
  • T16 — Layer-D render audit (token-join + prose lint)
  • T17 — eval harness: semantic field-diff vs audited-truth golden (V11, C10, C11)
  • T18 — L3 pi extension via uv-run seam
  • T19/T20 — purity lint + env capture
  • T21/T22/T23 — intervention reconciliation gate + research knowledge records + post-hoc utility verdict
  • T25 — adapter/cli/agents cutover onto the gate

3. B1 — eval-harness extraction from growth_insights (last commit, cdbec4b)

Extracts the runtime-side score gate + omp session parser out of growth_insights and into insight-kit so they're reusable across consumer repos.

  • platform/eval/score.py — 9-check structural + semantic score gate
  • integrations/omp/session_parser.py — omp 15.2.4 session JSONL parser
  • 37 new tests, all passing

Version bump: 0.1.0a3 → 0.1.4.

Test plan

  • pytest — 583/583 pass
  • B1 wired into growth_insights via editable path source — pytest tests/test_score.py tests/test_session_parser.py 38/38
  • Real T17 run scores 0.8889 via the extracted code (canonical implementation)

Companion

  • growth_insights PR — feat/consume-insight-kit-eval-omp (consumes B1)

…nsights

B1 of the eval-harness extraction. Two pure modules and their tests move
into insight-kit so any analyst/critic loop can reuse them without copying
files. growth_insights becomes the reference consumer.

## New modules

- `platform/eval/score.py` — 9-check structural + semantic score gate.
  Pure logic, no nairomarket coupling. CLI: `python -m insight_kit.platform.eval.score <run_dir>`
- `integrations/omp/session_parser.py` — omp 15.2.4 session JSONL parser.
  Reverse-engineers the hidden `.tmp` dotfile layout into typed ParsedTurn /
  ParsedToolCall. Includes the persona-boundary regex
  (`run_analyst_critic.py.*--critic-phase`).

Both expose clean public surfaces via `__init__.py` re-exports.

## Tests

Mirrored from growth_insights, adapted to native package imports (no importlib
shim). 37 new tests, all passing alongside the existing suite (583 total).

- `tests/platform/eval/test_score.py` — 20 tests covering regex variants,
  citation cross-check (multi-source, .py extractors, empty paths, bare
  tokens), length prior boundaries, integration scenarios.
- `tests/integrations/omp/test_session_parser.py` — 17 tests covering
  parsing, persona boundary, edge cases (empty/missing/truncated/no-usage),
  tool calls, find_session_file selection.
- `tests/integrations/omp/fixtures/omp_session_sample.jsonl` — anonymised
  real omp v15.2.4 session line.

## Version

0.1.0a3 → 0.1.4 (drop alpha, normalize patch).

## Next (B1 cont. on growth_insights side)

Replace growth_insights/deploy/eval/{score,omp_session_parser}.py with
thin re-export shims pointing at the new insight-kit modules. Add
insight-kit as a Python dep in growth_insights pyproject.toml (editable
path install). Verify growth_insights tests still pass via the shims.
@nyashkn
nyashkn merged commit c356ba3 into main Jun 9, 2026
@nyashkn
nyashkn deleted the feat/agents-system-v2 branch June 9, 2026 00:21
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