Code Intel Pipeline

Local orchestration for the current code-understanding stack:

• rg: precise text and file inventory.
• repowise: long-lived semantic index, wiki, and multi-repo workspace memory.
• Understand Anything: Claude skill that produces .understand-anything/knowledge-graph.json.
• sentrux: architectural rules and structural regression gate.

This is a thin pipeline. It does not vendor or wrap the tools internally yet. The goal is to prove the workflow before merging capabilities into a future CodeNexus-style tool.

Run

Install or verify this machine first:

D:\projects\_tools\code-intel-pipeline\install-code-intel-pipeline.ps1 -RepoPath C:\path\to\repo -CheckProvider

On a teammate machine, add -InstallMissing to install missing command-line tools when a supported installer is available:

D:\projects\_tools\code-intel-pipeline\install-code-intel-pipeline.ps1 -RepoPath C:\path\to\repo -CheckProvider -RepairSkillLinks -InstallMissing

The installer checks rg, git, python, repowise, sentrux, the shared Codex/Claude skill links, Understand Anything, config, repo doctor, and optional provider access. It never writes API keys. JSON output includes installActions so another agent can tell whether each tool was already present, installed, needs a new shell, or failed.

Use -AuditInstallPlan to print the supply-chain notes for each tool before installing anything:

D:\projects\_tools\code-intel-pipeline\install-code-intel-pipeline.ps1 -RepoPath C:\path\to\repo -AuditInstallPlan

Always start with the doctor:

D:\projects\_tools\code-intel-pipeline\check-code-intel-tools.ps1 -RepoPath C:\path\to\repo

D:\projects\_tools\code-intel-pipeline\run-code-intel.ps1 -RepoPath C:\path\to\repo -Mode normal

With aliases and defaults:

D:\projects\_tools\code-intel-pipeline\run-code-intel.ps1 -Repo my-repo -Mode normal

Team default:

D:\projects\_tools\code-intel-pipeline\check-code-intel-tools.ps1 -RepoPath C:\path\to\repo
D:\projects\_tools\code-intel-pipeline\run-code-intel.ps1 -RepoPath C:\path\to\repo -Mode normal

Stable wrapper:

D:\projects\_tools\code-intel-pipeline\invoke-code-intel.ps1 -RepoPath C:\path\to\repo -Mode normal

Batch wrappers:

D:\projects\_tools\code-intel-pipeline\invoke-code-intel.ps1 -Config D:\projects\_tools\code-intel-pipeline\pipeline.config.json -Repos k-atana,glyph-arts -Mode normal
D:\projects\_tools\code-intel-pipeline\invoke-code-intel.ps1 -Config D:\projects\_tools\code-intel-pipeline\pipeline.config.json -All -Mode lite

For a first baseline:

D:\projects\_tools\code-intel-pipeline\run-code-intel.ps1 -RepoPath C:\path\to\repo -Mode normal -SaveSentruxBaseline

For a full architecture refresh, run the Claude skill first:

/understand C:\path\to\repo --language zh --full

Then run the pipeline again.

MiniMax Anthropic-compatible provider config is expected at user level:

ANTHROPIC_BASE_URL=https://api.minimaxi.com/anthropic
REPOWISE_PROVIDER=anthropic
ANTHROPIC_API_KEY=<secret>
ANTHROPIC_AUTH_TOKEN=<secret>

Do not write secrets into repo files. Repo-local config may store provider/model names only.

Agent Skill

The shared skill lives at:

C:\Users\Administrator\.agents\skills\code-intel-pipeline

The distributable copy is included in this repo at:

skill\SKILL.md
skill\agents\openai.yaml

Claude and Codex both point at that same directory:

C:\Users\Administrator\.claude\skills\code-intel-pipeline
C:\Users\Administrator\.codex\skills\code-intel-pipeline

The OpenAI/Codex metadata is in agents/openai.yaml and is validated with the skill-creator quick validator.

Modes

• lite: inventory and status checks only.
• normal: update or initialize repowise, check Understand graph, run Sentrux gate.
• full: same as normal, but the emitted Understand command includes --full.

Add -RepowiseDocs to make the pipeline run scoped wiki generation instead of index-only repowise refresh for repos that define scoped repowise settings.

When -RepowiseDocs is set, the pipeline runs a provider preflight first. If quota or rate limits are hit, docs generation is disabled for that run and the summary records the provider failure category.

When repowiseScopePaths or repowiseRootFiles is configured for a repo, the pipeline uses a sparse git worktree under %LOCALAPPDATA%\code-intel\repowise\<repo> by default and runs repowise there instead of at the noisy repo root.

For k-atana, that shadow worktree is the supported team path. It indexes current backend working-tree contents plus selected root metadata files without crawling nested external repos under tools/.

Example:

D:\projects\_tools\code-intel-pipeline\run-code-intel.ps1 -RepoPath C:\path\to\repo -Mode normal -RepowiseDocs

Smoke test:

D:\projects\_tools\code-intel-pipeline\test-code-intel-pipeline.ps1 -RepoPath C:\path\to\repo

Provider preflight:

D:\projects\_tools\code-intel-pipeline\test-code-intel-provider.ps1 -Json

Install check:

D:\projects\_tools\code-intel-pipeline\install-code-intel-pipeline.ps1 -RepoPath C:\path\to\repo

Use -RepairSkillLinks when the shared code-intel-pipeline skill should be installed or repaired for Codex and Claude. If .agents\skills\code-intel-pipeline is missing, the installer seeds it from this repo's bundled skill\ directory first, then creates the .codex and .claude junctions. Use -InstallMissing only on machines where you want the script to install missing CLI tools. It uses winget for Git/Python/ripgrep, pip for repowise, and cargo install sentrux --locked only if Cargo is present. Unsupported installs are reported as manual fixes instead of hidden magic. The installer never writes API keys.

Artifact index:

D:\projects\_tools\code-intel-pipeline\update-code-intel-index.ps1

The index is written to:

%LOCALAPPDATA%\code-intel\artifacts\index.md

Set CODE_INTEL_ARTIFACT_ROOT or pass -ArtifactRoot when the team wants artifacts on a shared project disk.

Artifacts

Each pipeline run writes:

• summary.md: skim-first operator summary.
• report.json: machine-readable run state and failure categories.
• understanding.md: Karpathy-style understanding report for handoff and human inspection.
• files.txt: exact rg --files inventory.

Read understanding.md when handing work to another teammate or agent. It lists assumptions, verified facts, unverified areas, failure categories, and the single next action.

Templates

Reusable process templates live under templates\:

• idea-file.md: write the idea and success criteria before a nontrivial pipeline or tool change.
• understanding-report.md: standalone understanding-first handoff format.
• dependency-audit.md: pre-install dependency review format.

Rules

• Keep generated tool state local at first.
• Commit only intentional governance files such as .sentrux/rules.toml.
• Do not merge tool internals into the future unified tool until this pipeline has survived real repo work.
• Treat Understand Anything as the architecture snapshot, repowise as memory, and sentrux as the gate.
• Treat understanding.md as the team handoff layer: do not ship a run you cannot explain.
• Audit new install surfaces before adding them to -InstallMissing.
• Use sentruxPath for legacy-heavy repos where vendored or research-copy code would make a full-repo gate noisy.
• For k-atana, the pipeline now uses scoped repowise: sparse worktree plus live file sync for backend and a few root metadata files. That avoids nested tool repos while preserving current working-tree changes in the indexed scope.
• MiniMax-backed repowise index-only mode is stable.
• Scoped wiki generation now runs through Run-ScopedRepowiseDocs.py with a small default budget (coverage_pct=0.02) so the team can smoke-test docs without asking the provider to eat the whole repo in one shot.
• If the provider is rate-limited, the helper leaves docs_enabled=false plus docs_skip_reason=no pages generated; likely provider quota or rate limit. That means the local toolchain is fine and the blocker is upstream quota, not local indexing.

Future Merge Shape

When the pipeline is stable, merge by interface, not by copying whole projects:

• SearchProvider: backed by repowise.
• ArchitectureGraphProvider: backed by Understand Anything graph artifacts.
• StructureGateProvider: backed by sentrux.
• ExactSearchProvider: backed by rg.

That keeps the unified tool small instead of becoming a museum of other people's abstractions.