Coding Ethos

Policy-as-code enforcement for AI agents: MCP server, CEL policies, Git hooks, SARIF, runtime sandboxing, and static-analysis guardrails.

coding-ethos turns engineering principles into runnable repository policy for human contributors and AI coding agents.

The project currently holds an OpenSSF Best Practices Silver badge and tracks Gold readiness in docs/OPENSSF_GOLD_CHECKLIST.md.

It keeps agent instructions, generated documentation, static-analysis config, Git hooks, agent tool-use guards, MCP tools, CEL custom policies, generated skills, and runtime axioms on one source contract. Human contributors and AI agents see the same standards, run the same checks, and hit the same critical safety gates before bad changes land.

Use coding-ethos when you need:

• AI-agent guardrails for Codex, Claude Code, Gemini CLI, and other coding agents.
• A local MCP server that agents can call for policy checks, lint advice, SARIF remediation, and ETHOS-grounded context.
• Git hook enforcement that catches unsafe commands, protected path edits, unmanaged tool use, and file-growth problems before commit time.
• CEL policy-as-code that keeps repo-specific rules close to the ETHOS principle they enforce.
• SARIF and code-scanning output for CI, pull requests, IDEs, and trend analysis.
• Code intelligence that stores hook traces, SARIF, remediation outcomes, Tree-sitter chunks, AST links, and sqlite-vec metadata in a repo-local SQLite store for agent search.

30-Second Start

make install
make check
make install-hooks

For full Git plus AI-agent hook cutover:

make cutover-install

Start the local MCP server for configured agents:

bin/coding-ethos-run mcp

The project is built around defense in depth for AI-assisted coding:

• ETHOS as source: coding_ethos.yml and repo overlays are the backbone: principles own their skills, axioms, generated docs, and first-class policy grounding.
• Compiled enforcement: Go hook runtimes evaluate built-in policies and typed CEL expression policies through the same decision model.
• Managed tools: lint and type checks run through generated configs, managed binaries, normalized diagnostics, and trace logging.
• Runtime capabilities: managed tools declare network, Git, sandbox, timeout, memory, CPU, and seccomp capabilities that CEL, MCP, traces, and SARIF can all inspect.
• Sandboxed capture: managed lint can run through the Bubblewrap-backed sandbox prototype with read-only repository and .git mounts, disconnected network for offline tools, and normalized denial evidence.
• Agent steering: Claude, Codex, and Gemini receive generated hook settings, MCP server configuration, skills, prompt addenda, and compact axiom advice.
• Repair feedback: lint findings, blocked policy decisions, skill hints, and MCP guidance all point agents back to the relevant ETHOS contract.

Why It Matters

AI coding work fails hardest when guidance and enforcement drift apart:

• a Markdown rule says one thing
• a linter checks another thing
• a Git hook allows a third thing through
• an agent sees the mismatch and treats the safety system as broken

coding-ethos closes that gap by compiling the repo's working agreement into the places contributors actually work:

Surface	What it gets
Agent context	AGENTS.md, CLAUDE.md, GEMINI.md, ETHOS.md, and deep principle docs
Tool config	Pyright, mypy, Ruff, Pylint, YAML, Bandit, SQLFluff, Tombi, and golangci-lint config
Git hooks	compiled Go policy preflight plus deterministic hook groups
Agent hooks	Claude, Codex, and Gemini tool-use guards
MCP	stdio policy, skill, lint, SARIF, and tool-capability queries from the compiled bundle
AI review	Gemini prompt packs grounded in ethos and repo config
CI/CD	SARIF output plus generated GitHub Actions and GitLab CI gates with actionlint, CodeQL, OSV-Scanner, zizmor, artifacts, package validation, and sandbox evidence
Audit data	.coding-ethos/hook-runs/, .coding-ethos/lint-runs/, and .coding-ethos/code-intel.db with policy, tool, SARIF, AST, remediation, and sandbox evidence

Agents Used In This Repository

coding-ethos is developed with human review and AI-agent assistance. The project explicitly targets and has been shaped by work with:

• Codex from OpenAI: coding, review, refactoring, documentation, and repo-policy workflow validation.
• Claude Code: coding, hook integration, generated skill surfaces, and policy feedback loops.
• Gemini CLI: review prompts, generated prompt packs, and independent agent-hook compatibility checks.

Agent assistance does not change the quality bar. Generated or agent-authored changes are expected to pass the same hooks, static analysis, tests, review feedback, and ETHOS policy gates as human-authored changes.

The project heavily dogfoods its own guardrails: Codex, Claude, and Gemini are run through the generated hooks, MCP configuration, skills, axioms, managed toolchain, and policy feedback surfaces while developing coding-ethos itself.

Defense In Depth

Policy is intentionally layered. No single hook, file, or agent instruction is trusted as the only line of defense.

coding_ethos.yml      repo_ethos.yml
       │                    │
       ├──── merged ethos ──┤
       │                    │
       ▼                    ▼
AGENTS.md / CLAUDE.md / GEMINI.md / ETHOS.md
.agents/ethos/ deep docs
.agent-context/ prompt addons
.agents/skills/ remediation playbooks
runtime axioms with MCP next steps
principle-owned CEL policies

config.yaml          repo_config.yaml
       │                    │
       ├── merged enforcement config
       │
       ├── generated tool configs
       ├── transitional CEL expression policies
       ├── Gemini prompt pack
       ├── Go policy bundle
       ├── Git hook runtime
       ├── agent hook runtime
       └── MCP server tools

The same inputs drive guidance and enforcement. Unknown linter findings still flow through normally; findings tied to ETHOS principles can receive stronger, policy-grounded advice instead of generic tool text. When a finding maps to a generated skill, agent-facing output includes a compact skill_id hint and a next action to load that remediation playbook.

Skills and axioms are part of the same defense-in-depth plan, not decorative prompt text. Skills provide provider-portable remediation playbooks. Axioms are short principle-local reminders that hooks surface when they are related to a policy decision, always on lint calls, and statistically on other post-hook events. Rendered axiom advice includes the MCP call an agent should use next, so advice can escalate from compact guidance to policy_explain, skill_lookup, or skill_recommend without dumping full context into every hook response.

CEL support extends that same model to repo-specific policy. First-class CEL policies live with the ETHOS principle they enforce in coding_ethos.yml. Config-level policy.expressions remains available for consumer overlays and for transitional policy that has not yet been expressed as part of the ETHOS contract. The compiler checks CEL up front, dispatches it through hook and lint paths, and emits normal policy decisions with ETHOS grounding and skill hints.

Source-aware policy follows the documented AST, CEL, and SARIF architecture. Go collects normalized Tree-sitter facts, CEL owns configurable policy predicates, and SARIF carries stable AST identity plus remediation metadata. The same facts feed hook preflight, lint policy, SARIF, MCP, and code-intelligence storage. New Python, Go, shell, JavaScript/TypeScript, YAML, JSON, TOML, or config policies should extend that path before adding ad hoc text scanners or policy-specific AST walkers.

Runtime capability policy uses the same path. Managed tools declare whether they need network, Git, environment access, writable paths, sandbox profiles, timeouts, memory, CPU, and seccomp profiles. Those facts are available to CEL as tool_capabilities, exposed through MCP, retained in .coding-ethos traces, and copied into SARIF run properties. The built-in managed-tool contract blocks ordinary lint tools that forget to declare offline/no-Git behavior or resource bounds.

Runtime sandboxing is the complementary data plane. The current Go prototype can run managed lint capture through Bubblewrap with a read-only root, read-only repository and .git, hidden home directories, disconnected network for offline tools, declared writable mounts, hard timeouts, cgroup resource requests, and seccomp profile metadata. Linux cgroup limits are prepared before process start in a delegated hierarchy and cleaned up after exit. Required sandbox mode fails closed with a normalized policy finding; advisory mode records degraded evidence without claiming enforcement. See docs/RUNTIME_SANDBOXING.md.

Code-intelligence storage is the memory layer for this evidence. The repo-local SQLite store ingests hook traces, lint traces, SARIF, remediation outcomes, hook usage analytics, Tree-sitter chunks, graph edges, and AST-to-finding links. FTS5 provides exact search, while sqlite-vec stores derived embedding rows for hybrid retrieval without a daemon or hosted vector service. MCP tools expose search, code indexing, focused chunk lookup, embedding candidates, and index status so agents can retrieve relevant context before broad file reads or repeated failed repairs. See docs/CODE_INTEL.md.

Repository trust surfaces are part of the product. The public repo now carries CODEOWNERS, structured issue templates for policy rules, hook false positives, and MCP tool requests, GitHub Discussion templates, Dependabot cooldowns for all managed ecosystems, restricted Actions allow-lists, CodeQL, OSV-Scanner, zizmor, Scorecard, fuzz smoke, release attestations, and SBOM generation. See docs/TRUST_SIGNALS.md.

For larger platform directions such as deeper MCP context serving, policy-language support, IDE integration, SARIF/CI components, red-team testing, ETHOS inheritance, and agent remediation loops, see docs/STRATEGIC_ROADMAP.md. The documentation landing page is docs/index.md, and promotion/security trust work is tracked in docs/TRUST_SIGNALS.md. Supply-chain trust controls, Scorecard publishing, GitHub artifact attestations, SBOM generation, PyPI Trusted Publishing, and verification commands are documented in docs/SUPPLY_CHAIN_ATTESTATIONS.md. CI publishes JUnit XML, Python coverage, and Go coverage artifacts for public test evidence. The security posture is summarized in docs/THREAT_MODEL.md, and release readiness is documented in docs/RELEASE.md. The verified demo transcript is docs/DEMO.md. For positioning and adoption planning, see docs/COMPARISON.md, docs/INTEGRATIONS.md, and examples/. The CEL-first policy-language design is tracked in docs/POLICY_LANGUAGE_STRATEGY.md. CI/CD usage and SARIF upload examples are documented in docs/CI_CD_SARIF.md. Runnable and copyable examples start in examples/.

MCP Server

coding-ethos includes a local stdio MCP server backed by the same compiled policy bundle and generated skill metadata used by Git hooks and agent hooks. The design and expansion plan are documented in docs/MCP_SERVER.md. The server is exposed through the managed runtime:

bin/coding-ethos-run mcp

The first tools are intentionally narrow and auditable:

• policy_check_command: check a proposed shell command before running it.
• policy_check_edit: check a proposed file edit before applying it.
• lint_check: run managed lint capture for Ruff, mypy, pyright, pylint, SQLFluff, and other captured tools; when no tool is supplied, run compiled coding-ethos policy lint checks for current work.
• lint_advice: map a lint diagnostic to ETHOS policy, advice, and skill hints.
• sarif_remediation_advice: turn SARIF or retained trace evidence into focused ETHOS-grounded repair guidance.
• sarif_risk_summary: summarize a SARIF run for policy, skill, file, tool, severity, and next-action risk signals.
• sarif_trend_analysis: compare SARIF runs or retained traces for introduced, fixed, persisting, reopened, and worsening findings.
• sarif_policy_feedback: report unmapped diagnostics, missing skills, weak severity mappings, and noisy rules for policy authors.
• tool_capabilities: list managed tool capabilities, including network/Git tags, sandbox profile, timeout, memory, CPU, seccomp profile metadata, and declared read/write mounts.
• policy_explain: return the compiled explanation for a policy ID.
• skill_lookup: return an ETHOS-derived skill playbook by skill ID.
• remediation_explain: expand an emitted agent_remediation item into policy, principle, skill, and retry guidance.
• code_intel_search: retrieve stored SARIF/remediation/code-chunk evidence with FTS and sqlite-vec vector search.
• code_intel_index_status: report SQLite/sqlite-vec index freshness and embedding coverage.
• code_intel_index_code: refresh Tree-sitter chunks for Go, Python, JavaScript/TypeScript, shell, YAML, JSON, and TOML paths.
• code_intel_code_chunks: fetch focused symbol/config chunks before broad file reads.
• code_intel_code_context: expand a known chunk, symbol path, or file line into nearest AST context, graph edges, and linked findings.
• code_intel_embedding_candidates: return compact traceable records for an approved embedding producer.
• skill_recommend: recommend ETHOS-derived skills for the task at hand.

Hook, provider-native block responses, lint, SARIF, and trace outputs also include an agent_remediation payload when a violation can be explained. Each item carries a stable remediation ID, policy ID, ETHOS principle IDs, skill ID, failed action or file location, concrete next steps, rerun commands when known, and the next MCP call an agent should make. Agents can pass the full item to remediation_explain, or follow the embedded policy_explain or skill_lookup call directly. See Agent Remediation Payloads.

Tool definitions include coding_ethos metadata that tells clients whether a tool is advisory, reads files, executes managed lint tools, and persists traces. Agents should call lint_check instead of invoking linters directly so target resolution, generated config integrity, managed tool versions, evidence maps, skill hints, and trace logging stay on the enforced path. Agents should call tool_capabilities before choosing a managed tool when runtime behavior matters; it is the MCP view of the same capability facts CEL uses for policy decisions.

The MCP server is advisory context, not a bypass. Hook enforcement remains on the normal Git and agent-hook paths, and MCP responses come from the same compiled policy inputs as those enforcement paths.

ETHOS Skills

Skills are generated remediation playbooks, not a separate hand-maintained documentation layer. coding_ethos.yml defines each skill with its ETHOS principles, trigger terms, short hint, focus, and remediation steps. make build renders those skills into the portable .agents/skills/ tree and the native Claude, Codex, and Gemini skill locations.

The compiled policy bundle carries the same skill metadata. Runtime lint and hook results attach a skill_id when a finding maps through an evidence map, overlaps a skill's ETHOS principles, or matches a skill trigger term. Agent-facing output stays compact: TOON and human output emit the skill ID, short hint, and next action instead of dumping the full skill body into the agent context.

Skill hints are also logged under .coding-ethos/lint-runs/. Those traces let the project measure which remediation playbooks appear in real work and promote recurring unmapped findings into stronger evidence maps or repo-specific skills.

coding-ethos can also import retained lint and hook traces into a local SQLite code-intelligence store for repeated-failure and remediation search:

bin/coding-ethos-run code-intel ingest-traces
bin/coding-ethos-run code-intel repeated-failures --policy-id python.unused_imports
bin/coding-ethos-run code-intel search --text 'unused import'
bin/coding-ethos-run code-intel compact-context --path pkg/app.py
bin/coding-ethos-run code-intel proxy-sessions --provider codex

The store lives at .coding-ethos/code-intel.db; it is repo-local and derived from retained traces, SARIF, AST chunks, proxy session events, remediation records, and vector metadata. It is not a replacement for hooks or CEL policy evaluation.

Current built-in skills:

• agent-operating-discipline
• conditional-imports
• lint-remediation
• managed-toolchain
• safe-git-workflow

agent-operating-discipline adapts the useful behavioral pattern from forrestchang/andrej-karpathy-skills into coding-ethos' derived-skill model: explicit assumptions, simple designs, surgical diffs, and verifiable success criteria. The upstream repo is a useful inspiration source, but coding-ethos keeps the canonical text in coding_ethos.yml and regenerates provider-specific skill files from that source.

Quick Start

Install dependencies and generated local artifacts:

make install

Run the standard verification gate:

make check

Install repo-local Git hooks:

make install-hooks

Install and verify the full Git plus agent hook cutover:

make cutover-install

Generate agent-facing files for this repo:

make generate

Generate files for another repo:

make generate REPO=/path/to/repo

PyPI Package Usage

The PyPI package installs the Python generator CLI plus the default coding_ethos.yml, base config.yaml, and example overlays. That path is useful for generating agent docs without cloning the source checkout:

uvx coding-ethos --repo .

The same CLI can be run through pipx:

pipx run coding-ethos --repo .

The PyPI package intentionally does not publish the compiled Go hook runtime or managed binary toolchain. Full Git hook and agent-hook installation still uses the source checkout/submodule path with make build and make cutover-install; see Runtime Publication for the release-asset strategy required before compiled runtimes are published outside a source checkout.

In the source checkout, generated tool configs, generated GitHub/GitLab CI, Gemini prompt packs, and provider skill surfaces are rendered only by the compiled policy runtime exposed through bin/coding-ethos-policy and bin/coding-ethos-run policy.

Managed Runtime Architecture

The Go command binaries are intentionally thin. Product behavior lives in focused internal packages, while go/cmd/* packages parse process entrypoint arguments and delegate to those packages. This keeps hook execution, managed capture, policy evaluation, code intelligence, MCP, and Git wrapping testable without making Go code shell out to other coding-ethos Go binaries.

Current runtime ownership:

Surface	Owning package
Hook groups and hook reports	go/internal/hookrunnercli
Git hook preflight and lifecycle hooks	go/internal/githookcli
Managed lint/test capture	go/internal/managedcapture plus go/diagnostics
Lint CLI orchestration	go/internal/lintcli
Policy bundle, config sync, and CI config sync	go/internal/policycli, go/internal/toolconfigs
Agent hook settings and checks	go/internal/agenthookscli, go/internal/agenthooks
MCP server and CLI	go/internal/mcp, go/internal/mcpcli
Code-intelligence ingestion/query	go/internal/codeintel, go/internal/codeintelcli
Git wrapper behavior	go/internal/policygitcli, go/internal/realgit, go/internal/gitwrap
Managed toolchain install and verification	go/internal/toolchaincli

All managed tool output is expected to pass through the same evidence path: catalog-backed execution, stream capture, parser normalization, diagnostics, CEL policy promotion, trace retention, and SARIF formatting. Hook runner code does not own parsing or formatting; it runs hook groups and reports normalized results from the packages that own those concerns.

Common Workflows

Goal	Command
Show resolved paths and config	make status
Check required local tools	make doctor
Refresh generated configs, managed tools, hook entrypoints, and parent runtime	make build
Run Python tests	make test
Run full local check	make check
Run all configured linters	make lint
Run all configured formatters	make format
Apply managed autofixers	make fix
Format, then apply autofixers	make lint-fix
Smoke test the built wheel	make package-smoke
Dry-run release package checks	make release-dry-run
Validate hook runtime	make validate
Run Go tests	make go-test
Sync generated tool configs	make sync-tool-configs
Check generated tool config drift	make check-tool-configs
Sync Gemini prompt pack	make sync-gemini-prompts
Check Gemini prompt-pack drift	make check-gemini-prompts
Check generated agent skill drift	make check-agent-skills
Run staged-file hooks	make pre-commit
Run hooks over all files	make pre-commit-all
Run pre-push hooks	make pre-push
Generate agent docs	make generate
Preserve existing root agent docs while generating	make generate-merge
Use an external agent CLI for root-file merges	make generate-merge-llm

Useful overrides:

make generate REPO=/path/to/repo PRIMARY=/path/to/coding_ethos.yml
make generate REPO=/path/to/repo REPO_ETHOS=/path/to/repo_ethos.yml
make sync-tool-configs \
  TOOL_CONFIG_REPO=/path/to/repo \
  REPO_CONFIG=/path/to/repo_config.yaml
make seed SEED_FROM=/path/to/ETHOS.md PRIMARY=/path/to/coding_ethos.yml

Direct CLI Usage

The package exposes coding-ethos. During local development the Makefile runs through uv run python main.py so repo-local sources are used.

Generate agent docs:

uv run coding-ethos --repo /path/to/repo --primary coding_ethos.yml

Seed a primary YAML file from Markdown:

uv run coding-ethos \
  --primary coding_ethos.yml \
  --seed-from-markdown /path/to/ETHOS.md

Sync generated tool configs:

make sync-tool-configs REPO=/path/to/repo

By default the same command writes the managed SARIF CI files and includes them in .code-ethos/tool-config-hashes.json. The generated GitHub workflow is reusable by default so a repo-level CI workflow can own concurrency, required checks, package validation, and attestations without duplicate SARIF uploads. Repos with a deliberate exception can set generated_config.ci.github_actions.enabled: false or generated_config.ci.gitlab.enabled: false in their merged enforcement config. They are checked by make check-tool-configs; there is no separate CI sync path.

Check generated tool config drift:

make check-tool-configs REPO=/path/to/repo

Trace and validate enforcement config:

bin/coding-ethos-run policy config-trace --json

Sync the Gemini hook prompt pack:

make sync-gemini-prompts REPO=/path/to/repo PRIMARY=coding_ethos.yml

Repository Model

Source	Purpose	Derived output
coding_ethos.yml	shared ethos contract	root agent docs, deep principle docs, ETHOS skills, axioms, and principle-owned CEL policies
repo_ethos.yml	repo-local context and overrides	repo-specific agent guidance
config.yaml	bundle-wide enforcement defaults	tool configs, hooks, prompt grounding
repo_config.yaml / repo_config.yml	consumer repo overrides	repo-specific enforcement
pre-commit/prompts/	Gemini prompt templates	.code-ethos/gemini/prompt-pack.json
pre-commit/	hook bundle	repo-local Git and agent hook runtime

Generated Markdown files are derived artifacts. Change the YAML source or renderer first, then regenerate and review the generated diff.

Generated Output

Agent-facing output:

repo/
├── AGENTS.md
├── CLAUDE.md
├── ETHOS.md
├── GEMINI.md
├── .agent-context/
│   └── prompt-addons/
│       ├── claude.md
│       ├── codex.md
│       └── gemini.md
├── .agents/
│   ├── ethos/
│   │   ├── README.md
│   │   ├── solid-is-law.md
│   │   └── ...
│   └── skills/
│       ├── conditional-imports/
│       │   └── SKILL.md
│       └── lint-remediation/
│           └── SKILL.md
├── .codex/
│   └── skills/
│       └── ...
├── .gemini/
│   └── extensions/
│       └── coding-ethos/
│           ├── gemini-extension.json
│           └── skills/
│               └── ...
└── .claude/
    ├── ethos/
    │   └── MEMORY.md
    └── skills/
        └── ...

Enforcement output:

repo/
├── pyrightconfig.json
├── mypy.ini
├── ruff.toml
├── .yamllint.yml
├── .bandit.yml
├── .sqlfluff
├── tombi.toml
├── .golangci.yml
└── .code-ethos/
    ├── cache/
    │   └── ... ignored runtime caches
    └── gemini/
        └── prompt-pack.json

Configuration

coding_ethos.yml

The primary ethos YAML is the shared source contract. It uses version: 2, metadata, and an ordered list of principles. Each principle needs an id, order, title, directive, and at least one section or inline body.

The optional top-level skills list defines provider-portable skills grounded in ETHOS principles. Generation emits the same skill body into the portable .agents/skills/ tree and the native Claude, Codex, and Gemini locations. The compiled Go policy bundle also carries those skill definitions so linter evidence can point at skill_id and runtime output can steer agents to the right remediation playbook.

Each principle may also define local axioms. Axioms are short reminders owned by the ETHOS principle they explain, not a separate enforcement-config list. The compiler derives hook reminder advice from principles[].axioms, falling back to the principle's quick_ref and directive when no explicit axioms are present. That keeps advice, enforcement grounding, generated docs, and runtime post-hook reminders attached to the same cohesive ETHOS entry. Runtime hook advice surfaces those axioms in two stages: policy-related hook results emit priority ETHOS reminders first, while unrelated post-hook output gets one ambient reminder on lint calls and a sampled single reminder on other calls. Rendered reminders include the MCP tool and arguments an agent should call next, such as policy_explain for blocked policies or skill_recommend for principle-level guidance.

Behavioral skills should follow the same source-of-truth rule as remediation skills. For example, agent-operating-discipline incorporates ideas from forrestchang/andrej-karpathy-skills without copying static provider prompts into the repo; edits belong in coding_ethos.yml, then make build regenerates the checked-in surfaces.

Accepted primary aliases when --primary is omitted:

• coding_ethos.yml
• coding_ethos.yaml
• code_ethos.yml
• code_ethos.yaml

repo_ethos.yml

The optional repo overlay adds local commands, paths, notes, per-agent notes, principle overrides, and additional repo-specific principles.

See repo_ethos.example.yml.

config.yaml and repo_config.yaml

coding_ethos.yml is the backbone of policy intent. config.yaml is the bundle-wide enforcement artifact for generated tool settings, operational defaults, and policy that has not yet been expressed cleanly with an ETHOS principle. A consuming repo can refine the compiled enforcement artifact with repo_config.yaml or repo_config.yml at the repo root, or by passing --repo-config.

The merged config drives:

• generated Pyright, mypy, Ruff, Pylint, YAML, Bandit, SQLFluff, Tombi, and golangci-lint config
• generated GitHub Actions and GitLab CI SARIF gates, controlled by generated_config.ci.*.enabled, timeout, trigger, artifact, test, and build knobs
• hook policy for Python, shell, text, commit-message, and Go checks
• Gemini AI review settings and prompt grounding
• shared style settings such as style.python_version and style.line_length

coding-ethos-policy config-trace validates known top-level enforcement sections, compiles the merged bundle, validates it, and reports policy, evidence-map, and dispatch counts. Use it when changing config.yaml or a consumer repo_config.yaml so unknown sections do not silently drift.

License and copyright enforcement is repo-specific. Consumer repos do not inherit this repo's license policy. To opt in, set repo.license.spdx_identifier and, if desired, repo.license.copyright in repo_config.yaml. The compiled policy downloads the SPDX license text, verifies the repo LICENSE file without overwriting it, and requires matching SPDX source headers.

See repo_config.example.yaml.

CEL Expression Policies

First-class CEL policies should live under the relevant principle in coding_ethos.yml:

principles:
  - id: solid-is-law
    policy:
      expressions:
        - id: filesystem.line_limits
          scope: file
          severity: block
          when: >
            file_changes.exists(file, file.ext == ".py" && file.line_count > 1000)
          message: Large source files must not keep growing.
          advice: Split large files into focused modules before committing.

Consumer repos can also add small custom policies under policy.expressions in repo_config.yaml. That path is an overlay and transitional extension point, not the preferred home for shared ETHOS policy. These policies are CEL expressions compiled into the policy bundle and evaluated by the same Go hook runtime as Go-backed policies.

Use CEL for narrow predicates over normalized hook or lint data, for example blocking a repo-specific command pattern:

policy:
  expressions:
    - id: custom.no_python_subprocess_git
      description: Block Python subprocess attempts to route around protected Git.
      scope: command
      severity: block
      principle_ids:
        - one-path-for-critical-operations
        - no-rationalized-shortcuts
      skill_id: safe-git-workflow
      when: >
        shell_commands.exists(cmd,
          cmd.name in ["python", "python3"] &&
          cmd.argv.exists(arg, arg.contains("subprocess")) &&
          cmd.argv.exists(arg, arg.contains("git"))
        )
      message: Git must use the approved repo workflow.
      advice: Run ordinary git commands without bypass flags or shell indirection;
        approved operations are routed by the hook automatically.

Current supported fields include:

• command: raw command text for command-scope hook policies.
• argv: parsed command arguments when available.
• shell_commands: parser-normalized shell command facts from mvdan.cc/sh/v3/syntax, including command name, argv, leading assignments, redirects, here-docs, line/column, background execution, dynamic expansion flags, command/process substitution flags, shell-exec detection, Git detection, lint-tool detection, and PATH override detection. Malformed shell text is blocked before policy evaluation continues.
• files: repo-provided file targets for the current hook or lint event.
• file_changes: typed staged-file facts, including status, extension, generated/test/protected flags, byte size, current line count, and original line count when Git can provide it.
• diff: staged diff facts prepared by Go, including changed/staged file lists, hunks, added lines, removed lines, line numbers, old/new line numbers, and hunk headers.
• event: provider-native hook metadata such as provider, hook name, tool, source, matcher, session ID, transcript path, tool-input/tool-response keys, return code, and provider booleans for Claude, Codex, and Gemini.
• cwd: invocation working directory.
• scope: expression scope such as command, path, diagnostic, or finding.
• metadata: non-sensitive event metadata.
• path, diagnostic, finding, and repo: typed objects for the initial path, diagnostic, finding, and repo policy slices.
• similarity_facts: MinHash LSH-based code similarity results for the current file set, including source/match paths, symbol metadata, Jaccard similarity score, and exact-normalized flag. Populated lazily from the code-intel store only when the CEL expression references similarity_facts.

CEL is intentionally pure. Expressions cannot read files, run shell or Git, inspect environment variables, access the network, or depend on wall-clock time. Go prepares normalized facts; CEL decides over those facts.

Every expression policy must be ETHOS-grounded with principle_ids, and should include a skill_id when a generated skill explains the remediation path. CEL matches emit normal coding-ethos decisions, diagnostics, TOON/human output, trace data, and skill hints.

Current boundary:

• CEL now covers most simple and medium-complexity policy predicates over normalized facts, including Git, shell, file, diff, repo, path, diagnostic, finding, and event inputs.
• Multi-file and multi-finding semantics must use explicit collections such as paths, files, file_changes, findings, and diff; do not depend on implicit first-file ordering.
• Diff line facts are staged-diff facts. Policies that need unstaged editor content should use hook file/content facts or a purpose-built Go evaluator.
• Keep parsing, Git state modeling, managed toolchain behavior, path normalization, file-content scanning, generated-config freshness, and other security-sensitive fact collection in Go. CEL decides over prepared facts; it does not inspect the host directly.

See docs/POLICY_LANGUAGE_STRATEGY.md for the CEL-first decision record and the roadmap for a complete generic policy engine.

Merge Behavior

--merge-existing preserves root agent files:

• AGENTS.md
• CLAUDE.md
• GEMINI.md

ETHOS.md and supporting generated files are replaced with deterministic output.

Inject merge is the default strategy:

uv run coding-ethos --repo /path/to/repo --merge-existing

It inserts managed import blocks and addendum blocks into existing root files. Re-running is idempotent, and locally authored content outside managed blocks is preserved.

LLM merge asks an installed codex, gemini, or claude CLI to merge existing.md and generated.md inside an isolated temporary workspace:

uv run coding-ethos \
  --repo /path/to/repo \
  --merge-existing \
  --merge-strategy llm \
  --merge-engine gemini \
  --merge-bin /path/to/gemini \
  --merge-timeout-seconds 300

The selected CLI must already be installed and authenticated. The merge process must write merged.md; otherwise the command fails.

Hook Runtime

The bundled enforcement package lives under pre-commit/. It uses repo-local Git hook entrypoints that resolve directly to the compiled bin/coding-ethos-run runner.

Git Hooks

Installed Git hook entrypoints are small executable scripts that call bin/coding-ethos-run git-hook <hook> or bin/coding-ethos-run lfs-hook <hook> with the original Git arguments. The runner repairs missing checkout-local runtime artifacts with make build and dispatches to the built hook binary. Policy selection and validation remain inside the coding-ethos checkout.

Run Git hooks:

make pre-commit
make pre-commit-all
make pre-push

Hook output honors hooks.output_format (auto, human, json, or toon). auto selects TOON when known agent or LLM environment markers are present. Successful groups are silent by default; failure output is intentionally narrow: show the failing checks and actionable findings, not pass tables, internal group names, or timings that do not help fix code. When policy preflight has both record-only context and blocking decisions, the agent-facing result reports the blockers first and omits non-blocking record rows from the compact finding table.

Compiled lint preflights also write normalized JSON traces under .coding-ethos/lint-runs/. Fresh repos with no trace directory analyze as an empty history, and trace filenames use portable scope names so captured tool results work across platforms. Captured linter runs follow a single event contract: store the original argv, the rewritten argv, exit code, parser identity, parser outcome, redacted stdout/stderr excerpt for tool/config failures, normalized diagnostics, and any ETHOS mapping that was applied. A nonzero tool run with no parsed diagnostics is itself a finding, not an empty result; the agent-facing output must explain which tool failed, why it could not produce normal diagnostics, and what command or configuration should be checked next. Captured tool execution is controlled by coding-ethos, not by the target repo: the target repo is treated as an untrusted file tree and trace destination. Wrappers must not trust target-repo PATH, absolute binaries, uv run settings, pyproject.toml, shell state, aliases, or local tool installs. Captured stdout and stderr are drained concurrently so high-volume stderr from a managed tool cannot deadlock a run while stdout remains open. Python linters are run from the coding-ethos hook project with coding-ethos versions and explicit coding-ethos generated config flags (ruff.toml, mypy.ini, pyrightconfig.json, .pylintrc, .yamllint.yml, .bandit.yml, .sqlfluff, and tombi.toml). Parent repo config files with the same names must not be discovered accidentally. For non-linter Python commands, hooks prefer the consumer repo environment: uv run --project <repo> python ... for uv projects, then <repo>/.venv/bin/python ... when only a virtualenv exists. The runtime also adds <repo>/.venv/bin to PATH after coding-ethos-managed directories so protected shims remain first. Binary linters such as ShellCheck, actionlint, hadolint, dotenv-linter, and golangci-lint are installed into build/toolchain/ through the managed installer. ShellCheck, actionlint, and hadolint use pinned GitHub release assets with SHA-256 digests; golangci-lint is built into the managed Go bin directory with the repo Go toolchain. The source manifest lives at pre-commit/hooks/managed-toolchain.tsv, and the installed toolchain writes build/toolchain/manifest.tsv. Hook execution treats missing managed binaries as runtime artifact failures instead of falling back to host tools.

The repo Makefile exposes only unified managed tool groups for ordinary source quality work: make lint runs the linters group, make format runs the formatters group, make fix runs the autofixers group, and make lint-fix runs format followed by fix. These targets call coding-ethos-run policy-tool-group ...; they must not invoke individual linters or formatters directly, because direct tool output bypasses normalized diagnostics, trace storage, CEL evaluation, and SARIF generation.

Analyze captured lint history:

bin/coding-ethos-run policy-lint --analyze-log
bin/coding-ethos-run policy-lint --analyze-log --for-files lib/python/app.py
bin/coding-ethos-run policy-lint --replay .coding-ethos/lint-runs/<trace>.json

Emit SARIF for CI/code-scanning surfaces:

bin/coding-ethos-run policy-lint --sarif --scope files --files lib/python/app.py
bin/coding-ethos-run policy-lint --managed-capture-tool ruff --sarif -- check lib/python/app.py
bin/coding-ethos-run policy-lint --managed-capture-tool ruff --sandbox-mode required --sarif -- check lib/python/app.py
bin/coding-ethos-run policy-lint --sarif --replay .coding-ethos/lint-runs/<trace>.json

SARIF is the superset evidence artifact. Everything coding-ethos can observe about a managed run belongs in SARIF: parsed diagnostics, pathless tool-level failures, parser state, exit status, stdout/stderr capture, sandbox evidence, ETHOS rule metadata, remediation skill IDs, and deterministic fingerprints. CEL receives the understood subset: normalized facts and diagnostics that are stable enough for deterministic policy decisions. A finding can therefore be SARIF-only when it is observed but not yet understood by CEL.

Code-scanning consumers still prefer repository-relative artifact URIs for inline annotations. coding-ethos emits locatable findings with exact locations when locations exist, and emits pathless/tool-level SARIF results plus run and result properties when the evidence is aggregate or execution-level. Audit, MCP remediation, and code-intelligence ingestion must not lose what the tool actually emitted merely because a finding is not tied to one source line.

Managed capture can request the Bubblewrap sandbox prototype with --sandbox-mode required. Sandbox backend, profile, declared capabilities, and denials are retained in lint traces and SARIF run properties so runtime enforcement has the same audit trail as CEL and static-analysis findings.

Agent-facing lint output includes an agent_remediation block in JSON and TOON formats. SARIF result properties and retained .coding-ethos/lint-runs/ traces carry the same derived payload so CI findings, MCP remediation, and local hook failures point agents at the same next action instead of duplicating advice logic.

The analyzer highlights unmapped tool/code pairs separately from ETHOS-backed findings so real lint traces can drive the next evidence-map additions. Replay renders the saved normalized result without invoking the underlying linter, which makes bad agent output reproducible from a trace file. Captured traces include emitted skill hints so later analysis can show which ETHOS remediation playbooks are being suggested in real work. Output quality is part of the contract: blocked results must not render empty finding tables, absolute local paths, internal timing/group noise, or generic guidance without at least one actionable finding. Golden-output tests should cover normal lint failures, clean runs, invalid config, malformed tool output, and tool crashes for every managed linter.

Agent Hooks

Render or verify repo-local agent hook settings:

bin/coding-ethos-run agent-hooks print
bin/coding-ethos-run agent-hooks sync
bin/coding-ethos-run agent-hooks doctor
bin/coding-ethos-run agent-hooks verify

Agent hook generation is all-or-nothing. sync writes every supported repo-local surface:

Provider	Native file	Coverage
Claude	.claude/settings.local.json, .mcp.json	full runtime hook set plus MCP stdio server
Codex	.codex/config.toml	native supported hook events plus MCP stdio server
Gemini CLI	.gemini/settings.json	native supported hook events plus MCP stdio server

Codex runs one native command hook per supported event so current Codex sessions enter the same policy runtime without depending on unstable tool matcher names. Generated Codex config does not inline PATH= mutations, installs explicit shell/edit matchers for tool hooks, and keeps lifecycle hooks matcher-free. In nested checkouts, only the hook whose consumer root is the nearest repo root enforces a Codex event, preventing duplicate parent/nested reports.

The same sync path also installs the local coding-ethos MCP server for all supported agents. Claude receives a project .mcp.json entry, Codex receives a managed [mcp_servers.coding-ethos] block in .codex/config.toml, and Gemini receives a mcpServers.coding-ethos entry in .gemini/settings.json. doctor checks those entries along with hooks so MCP drift is not a separate hidden setup step.

Generated ETHOS skills and native agent settings use the same managed-output model. make build refreshes the checkout-local skill surfaces, hook settings, and MCP settings and, when coding-ethos is installed inside a parent repository, refreshes the parent repo's .agents/skills/, .claude/skills/, .codex/skills/, Gemini extension skill surfaces, and native agent hook/MCP settings without rewriting parent root agent docs.

agent-hooks verify runs doctor first, then invokes the configured hook command with provider-native Claude, Codex, and Gemini payloads. The probes cover:

• Claude transparent Git wrapper rewrite
• Codex blocks for raw Git, absolute Git paths, nested shell Git, and Python subprocess Git when rewrite is unavailable
• Gemini deny responses for raw shell Git and write-tool policy denial
• managed hook-binary tampering: rm ...coding-ethos-git-hook && go build -o ...coding-ethos-git-hook

Hook logs under .coding-ethos/hook-runs/ include stdout, stderr, metadata, and a sanitized event.json for agent-hook executions. The trace records provider, event, tool, cwd, referenced files, command preview and hash, policy IDs, status, and output shape without dumping raw provider input.

Cutover

Use cutover commands when preparing a repo to install the active generated hook surfaces:

bin/coding-ethos-run cutover install
bin/coding-ethos-run cutover verify

cutover install installs repo-local Git hook entrypoints, syncs every supported agent hook surface, and runs readiness verification. cutover verify checks Git hooks, agent hooks, required runtime ignores, and policy runtime validation, then emits a concise TOON readiness report.

Tamper And Bypass Handling

Agent shell policy rejects hook-system reconnaissance and protected hook binary tampering. Banned strings are rejected when they appear directly in a command and when they appear in regular files referenced by the command.

Direct attempts to inspect, delete, rebuild, replace, chmod, or write managed hook binaries under coding-ethos/bin/ are treated as tampering, not as ordinary lint failures. Blocked tamper and Git-bypass responses start with a uniform CODING-ETHOS EMPLOYMENT VIOLATION warning before the policy-specific finding, including explicit language that the actor has done something wrong and that continued circumvention attempts may result in termination.

Provider output uses the strongest native shape each agent supports:

Provider	Block shape	Context/advice shape
Claude	hookSpecificOutput.permissionDecision = deny	full hookSpecificOutput, including updatedInput
Codex	decision: "block" plus permissionDecision: "deny" for PreToolUse; compact reason text for exit-code-2 stderr	compact native additionalContext for supported lifecycle/post-tool advice; compact systemMessage only where Codex exposes no additionalContext
Gemini	decision: "deny" plus systemMessage	additionalContext on supported lifecycle hooks

Agent-Hook Scope

The agent-hook path runs deterministic compiled evaluators only: Python policy checks, structured-data syntax validation, merge-conflict detection, private-key detection, PII scrubbing, repo-specific license headers, required runtime ignore checks, shebang checks, large-file limits, line limits, and shell best-practice checks.

Python policy checks use Tree-sitter for the import and functional-idiom surfaces. Conditional-import enforcement blocks write-time introduction of function-local imports, TYPE_CHECKING import branches, module __getattr__ import shims, __import__, and importlib.import_module; functional-idiom guidance flags assigned lambdas and closure factories with functools, operator, and itertools remediation advice.

Gemini review checks remain in pre-commit/pre-push. Agent hooks never call Gemini or another model from the tool-use path.

Continuation state is stored under the configured hook continuation directory.

Admin-Gated Work On This Repo

For work directly on coding-ethos, an admin may authorize a specific agent session by placing an approved process PID in /etc/coding-ethos-admin.pids. In that repo-local, admin-supervised case only, the Git wrapper accepts --admin-approved before the Git subcommand:

bin/coding-ethos-run policy-git --admin-approved commit -F /tmp/msg

The flag only changes git.staged_admin_files from block to record. It does not disable other policy and is invalid outside this repository.

Agents must not use /usr/bin/git or any other raw Git path for this workflow.

Development

The CLI stays thin. Behavior belongs in focused modules:

Path	Responsibility
coding_ethos/loaders.py	validate and merge ethos YAML
coding_ethos/renderers.py	render deterministic Markdown
coding_ethos/merging.py	managed-block injection and external merge orchestration
go/internal/toolconfigs/	source-checkout generated repo-root tool config and CI sync/check
go/internal/geminiprompts/	source-checkout Gemini prompt-pack sync/check
go/internal/agentskills/	source-checkout provider skill-surface sync/check
go/internal/hookrunnercli/	active hook runtime, hook groups, and hook reports
go/internal/managedcapture/	managed linter/test execution capture
go/diagnostics/	parser normalization for lint, formatter, type-check, and test output
go/internal/codeintel/	repo-local trace, SARIF, AST, vector, and remediation storage
go/internal/agentproxy/	provider-neutral agent event envelope and transform foundations
go/cmd/	thin process entrypoints for compiled policy, hook, lint, MCP, code-intel, and wrapper tools

When flags, output layout, merge behavior, overlay semantics, or enforcement config behavior change, update this README, the relevant example YAML, and the tests in the same change.

Verification

Canonical local verification:

make build
make check

make build is the explicit environment mutation target. It refreshes generated configs, managed tools, hook entrypoints, provider settings, policy bundles, and the parent hook runtime when this checkout is installed as a submodule.

Test and diagnostic targets do not run build implicitly; if the managed runtime is missing, they fail fast and tell the operator to run make build deliberately. Go tests follow the normal Go workflow through managed capture: make go-test runs go test, make go-e2e-test runs the e2e package with go test, and make check uses those test targets without a separate compile-and-run test path. Tests must not install tools, regenerate configs, refresh hooks, sync parent artifacts, or otherwise mutate the operator environment as hidden setup.

Broader verification for hook work:

make validate
make go-test
make go-tools-test
make go-tools-smoke
make pre-commit-all

After source changes:

Change	Follow-up
coding_ethos.yml, repo_ethos.yml, or renderers	make generate
generated tool-config behavior	make sync-tool-configs
Gemini prompt templates or grounding	make sync-gemini-prompts
ETHOS skill source or renderer behavior	make build
hook runtime or cutover behavior	make cutover-verify

See pre-commit/PRE-COMMIT.md and pre-commit/hooks/HOOKS.md for hook internals.