Ecosystem Borrow Audit

Ecosystem Borrow Audit

All gptengage calls follow ../../references/gptengage-invocation.md.

Intent and applicability

Produce a repository-evidenced backlog for the supplied ecosystem scope. Generic inferred ecosystem/borrowing analysis defaults to evidence-only. A bare explicit /ecosystem-borrow-audit preserves the full historical workflow, including 0.25, 0.5, 1, 1.5 ideation, unless the user selects evidence-only. Explicit external ideation follows the shared invocation authority/data contract.

Non-goals and must not

Do not widen a named-repository scope to the whole workspace, install tooling, invoke a second paid call for formatting, or present model ideas as source facts. Required investigation and local report preparation remain autonomous. Resolve only material scope/provider/data/spend decisions left unsettled; reuse grants.

Inputs

Extract from user request:

• Scope root (default: ${WORKSPACE_ROOT:-$HOME/Documents})
• Listings file path (default: $WORKSPACE_ROOT/listings.txt)
• Repo depth (default: depth 1)
• Ideate sigma list (default: 0.25, 0.5, 1, 1.5)
• Ideate CLI preference (codex, claude, or gemini; default: claude)

Workflow

1. Normalize paths and scope.

• Use supplied paths and validate them. A missing optional listings file does not block repository inspection. Do not silently repair an ambiguous path.
• Enumerate repositories within the selected scope/depth. Use git -C PATH rev-parse --show-toplevel and compare resolved roots to recognize ordinary repositories and linked worktrees without counting arbitrary nested folders.
• Build unified catalog: listings entries + git repos. Listings entries that are not git repos are contextual only: never scanned, scored, or ranked.

2. Run tier-1 scan across the unified catalog.

• For each repo/path, capture lightweight signals:
  • language/toolchain markers
  • recent commit date and remotes (for git repos)
  • README/docs architecture keywords (worker, policy, queue, memory, gate, retry, metrics, etc.)
• Score each project:
  • relevance (0-5)
  • borrowability (0-5)
  • integration cost (1-5)
  • novelty (0-5)

3. Run tier-2 deep reviews.

• Always include:
  • the target project repository
  • directly-related repositories discovered from docs, configs, and imports
  • any repository with high relevance and high borrowability from tier-1 scoring
• Add highest-scoring repos from tier-1.
• Extract concrete borrow candidates and file-backed evidence.

4. Reconcile missed opportunities.

• Compare findings with project roadmap/state docs (for target repo):
  • VISION.md, IMPLEMENTATION_PLAN.md, PROMPT.md
• Label findings:
  • already implemented
  • partially implemented
  • missed and high-value
  • low-value/not applicable

5. Produce ranked action backlog.

• Rank by impact, effort, risk, and dependency ordering.
• Include explicit next actions and validation checks.

6. Only in selected external-ideation mode, run sweeps independently per sigma.

• Resolve the existing gptengage command and verify relevant help. Choose a unique private artifact directory before execution. Do not use a shared fixed /tmp/ecosystem_audit directory or embed arbitrary seed text in shell code.
• For each distinct selected sigma, invoke once with --output json, the selected CLI/depth and per-invocation timeout. Capture stdout, stderr and exit status separately. For example, bind reviewed values before executing:

if "$GPTENGAGE" ideate "$SEED" --sigma "$SIG" --depth "$DEPTH" \
    --output json --cli "$CLI" --timeout "$CALL_TIMEOUT" \
    > "$RUN_DIR/$INDEX.json" 2> "$RUN_DIR/$INDEX.stderr"; then
  status=0
else
  status=$?
fi
printf '%s\n' "$status" > "$RUN_DIR/$INDEX.exit"

• Increment a safe numeric INDEX per selected sigma and record the sigma/file mapping. Validate successful stdout as one JSON document. Render that saved result locally into readable Markdown/text; never invoke another model just to obtain the same run's alternate format. Preserve failed outputs as failure evidence, not valid ideas. Even exit zero may contain partial expansions: inspect stderr/coverage and report missing branches.
• A per-call timeout is not a total sweep deadline. Account for sequential levels, concurrent waves and cleanup when setting an outer supervisor deadline; omit an arbitrary shorter outer timeout if no reliable total bound is known. If an explicit total deadline is required, use supervision that terminates children and records interruption. Do not silently substitute a backend after failure.

7. Merge ideation with repo-grounded findings.

• Deduplicate themes.
• Label each final idea:
  • repo-grounded
  • hybrid
  • ideate-only
• Prioritize repo-grounded and hybrid items.

Output Format

Return:

1. Coverage summary (counts + scope)
2. Ranked backlog (impact/effort/risk)
3. Missed previously section
4. Ideate summary by sigma when selected, otherwise evidence-only mode
5. Artifact paths to raw scans and ideate outputs

Guardrails

• Keep analysis evidence-backed with file paths and commands.
• Verify scoped repository identity with Git, including .git files used by worktrees; report nonrepository listings as contextual and all skipped coverage.
• If gptengage fails, report error and suggest:

~/.local/bin/gptengage status

• If preferred ideate CLI stalls/timeouts repeatedly, report the issue and run a fallback CLI only with user approval.