Cross-Feature Intelligence

Aggregates historical pipeline data into a scoped brief for /cspec brainstorm. Spec: .correctless/specs/cross-feature-intelligence.md. Architecture: ABS-037.

What It Does

The Correctless pipeline generates rich data across features – deferred review findings, Devil’s Advocate themes, override patterns, lens recommendations, debug investigations, and phase effectiveness history – but each /cspec run previously started from a near-blank slate. This feature adds a deterministic aggregation script that synthesizes 6 data sources into a single JSON brief, filtered by file-scope overlap and recency-weighted. /cspec reads this brief during the Socratic brainstorm (Step 0a) to surface concerns from prior workflow runs – as context, not constraints.

How It Works

graph LR
    A["deferred-findings.json"] --> S["cross-feature-intel.sh"]
    B["devadv/report-*.md"] --> S
    C["overrides/*.json"] --> S
    D["lens-recommendations-*.json"] --> S
    E["debug-investigation-*.md"] --> S
    F["workflow-effectiveness.json"] --> S
    S -->|"JSON brief"| G["/cspec Step 0a<br/>(advisory read)"]
    S -->|"metadata"| H["/cstatus<br/>(health reporting)"]

The script (scripts/cross-feature-intel.sh, 876 lines) reads each source, extracts structured entries, applies recency filtering (90-day exclusion), scope filtering (by affected files), and caps the output at 30 entries with per-section minimums. The output is a single JSON artifact at .correctless/meta/cross-feature-intel.json.

Data Sources

Source	Location	Entries Extracted
Deferred findings	`.correctless/meta/deferred-findings.json`	Open findings (status = “open”)
Devil’s Advocate	`.correctless/artifacts/devadv/report-*.md`	`## DA-NNN:` headings with severity
Overrides	`.correctless/meta/overrides/*.json`	Override entries, collapsed by reason hash
Lens recommendations	`.correctless/artifacts/lens-recommendations-*.json`	Recommended lenses, collapsed by name
Debug investigations	`.correctless/artifacts/debug-investigation-*.md`	Root cause summaries with file refs
Phase effectiveness	`.correctless/meta/workflow-effectiveness.json`	Post-merge bugs by missed phase

Usage

The script runs automatically when /cspec starts a new feature. To run manually:

# Unfiltered (all entries)
bash .correctless/scripts/cross-feature-intel.sh --base .correctless

# Scoped to specific files
bash .correctless/scripts/cross-feature-intel.sh --base .correctless --scope "hooks/workflow-gate.sh,scripts/lib.sh"

Key Design Decisions

Script over LLM aggregation (DD-001): deterministic bash is testable and reproducible.
Single brief file (DD-002): one JSON read in /cspec instead of 6 fragile separate reads.
Anti-anchoring over UNTRUSTED fence (DD-004): the data is internal project history (human-reviewed at creation), not external untrusted content. The risk is cognitive anchoring, not prompt injection. Calibration examples tell the agent when to weight and when to dismiss.
Advisory only (PRH-001): the brief never gates any phase transition or blocks any skill.

Configuration

No configuration needed. The script derives all paths from the --base argument. The 90-day staleness threshold and 30-entry cap are constants in the script.

Consumers

/cspec: reads during Step 0a brainstorm. Presents 3-5 most recent entries. Framed as context, not constraints. Anti-anchoring directive with calibration examples prevents over-reliance on historical patterns.
/cstatus: shows intelligence health in 3 states (no data / stale / current). Dormant on pre-upgrade projects.

Known Limitations

File-scope filtering currently applies only to debug investigations (only source with file-scoped data). Other sources are included unconditionally as project-wide concerns.
Date arithmetic uses file mtime for some sources (devadv reports, debug investigations), which is approximate after git operations (ENV-003).
The brief is not fed to /creview or /creview-spec yet (DEFER-001 in the spec).