agentic-resources/workplans/AGENTIC-WP-0005-detect-hardening.md

---
id: AGENTIC-WP-0005
type: workplan
title: "Coding Session Memory — Detect Hardening (quality filter + infra signals)"
domain: helix_forge
repo: agentic-resources
status: finished
owner: codex
topic_slug: helix-forge
created: "2026-06-07"
updated: "2026-06-07"
state_hub_workstream_id: "d8b7b8d1-1d85-4d2a-8ccd-7b0366a9442d"
---

# Coding Session Memory — Detect Hardening

A focused hardening pass (call it Phase 1.5) so the Detect output is trustworthy
enough to drive an **infrastructure assessment**. Triggered by ad-hoc analysis of
the live store after Phase 2:

- Of **72 captured sessions, only 31 are real coding sessions**; the rest are
  health-checks / smoke-tests / interrupted runs (mostly `llm-connect` *"Say hello
  in one word"*). The `abandoned` outcome heuristic mislabels these, and Phase 2
  cataloged a **false-positive** "cross-flavor abandoned" pattern as
  `approved`/`distribution_ready`.
- All 31 real sessions read as `success`, so the current signal set
  (outcome + markers + cost) surfaces almost no genuine friction.
- The already-captured `tool_histogram` tells the real story: **~17% of tool
  activity in real sessions is State Hub MCP + task plumbing + `ToolSearch`
  schema-loading**, concentrated to 40–70% in some sessions — but `signals.py`
  never looks at it.

No new capture is needed — this is analysis the data already supports.

## Session-Quality Filter

```task
id: AGENTIC-WP-0005-T01
status: done
priority: high
state_hub_task_id: "9f8b4304-0a37-4f66-ad34-d93e12fba0d8"
```

Add `detect/quality.py` with `is_real_coding_session(digest)` that filters out
health-checks, smoke-tests, interrupted, and trivially-short sessions (event-count
floor, repo present, substantive edit/tool activity, not a single hello/interrupt
prompt). Wire it into the detect pipeline so signals/clusters only form over real
sessions — fixing the `abandoned` false-positive. Knobs under `[detect]` in
`config.toml`. Unit-tested on synthetic trivial-vs-real digests.

## Infra-Overhead + Thrash Signals

```task
id: AGENTIC-WP-0005-T02
status: done
priority: high
state_hub_task_id: "10d57b05-a731-4ece-bf45-f6a98ac77555"
```

Add `tool_histogram`-based extractors to `detect/signals.py`: a shared tool-bucket
helper (`shell` / `edit` / `read` / `statehub_mcp` / `task_mgmt` / `schema_load` /
`other`); `sig_infra_overhead` (PROBLEM when the statehub+task+schema share of tool
calls exceeds a threshold; magnitude = share; locus `infra_overhead`);
`sig_schema_thrash` (`ToolSearch` count over threshold; locus `schema_load`);
`sig_tool_thrash` (extreme single-tool repetition). Pure functions over digests;
thresholds configurable. Unit-tested.

## Re-run Live, Purge False Positives, Ranked Friction Report

```task
id: AGENTIC-WP-0005-T03
status: done
priority: high
state_hub_task_id: "8b9d029a-60d0-4caf-af62-4fcc9c9a645c"
```

Re-run `ingest → detect` over the real local sessions with the filter + new
signals. Purge the false-positive catalog entries seeded in Phase 2 (the
health-check `abandoned` pattern) and re-curate so the catalog reflects real
friction. Produce a ranked **friction assessment** (`docs/ASSESSMENT-infra-friction.md`)
of the major infrastructure problems — quantified per repo/flavor, infra-overhead
share, schema-thrash — with recommendations (incl. the State Hub / MCP skill
hypothesis). After workplan file updates, notify the operator to run from
`~/state-hub`:

```bash
make fix-consistency REPO=agentic-resources
```