dummyindex

The persistent context engine for a repo. A Claude Code and OpenAI Codex skill family that turns any codebase into a .context/ folder agents can navigate without broad grepping — deterministic AST extraction plus a multi-agent council (dev, architect, critics) that fills in the judgment.

pip install --user dummyindex          # or: uv tool install dummyindex

dummyindex install                     # flagless = universal + linked: one real
                                        # .agents/skills tree, .claude/skills
                                        # symlinked to it — discoverable by
                                        # Claude Code, Codex, Cursor, Gemini CLI

cd /path/to/your/repo
# Claude Code: /dummyindex .
# Codex:       $dummyindex .

The bootstrap is the only time you need the terminal. After that, use slash commands in Claude Code or $skill-name mentions in Codex; the Python CLI is the deterministic backbone those skills invoke.

After the first run, both hosts receive durable instructions to consult .context/ before searching source broadly. Claude uses dummyindex-managed hooks and .claude/CLAUDE.md; Codex uses the active project instruction file (AGENTS.override.md, AGENTS.md, or a configured fallback) and native skill discovery. Codex has its own hook surface, but dummyindex does not install into it today. For fallback names and the byte limit, dummyindex reads Codex's platform system config, then user config, then the root .codex/config.toml only when the persistent user config explicitly trusts that project via [projects."<absolute-root>"].trust_level. It cannot observe a session's selected profile, -c overrides, or nested launch-directory config layers.

What it is

dummyindex runs in two modes per repo. Setup mode (one-time) builds .context/ and writes host guidance. Ongoing mode plans, builds, remembers, audits, and cleans generated context through the installed skill family.

The workflow bodies are shared. Installed Codex copies include a compatibility header that maps Claude's Task/Agent vocabulary to Codex explorer, worker, and default subagents; maps skill calls to $name; and uses the active Codex model through --model current. The core index, plan, build, memory, audit, and GC workflows work without Claude state. Both hosts prepare and use the same capability bodies under .context/equipment/, registered in equipment.json. Claude can add thin native discovery references with equip apply --platform claude; default preparation creates no new Claude setup. Both hosts can delegate through native agents without a manifest and pass selected policy mandates to their workers.

On Claude, managed hooks add a PreToolUse document guard, SessionStart drift/memory/GC signals, and a Stop reconcile nudge. On Codex, durable active-instruction-file guidance and explicit $dummyindex* workflows provide the same core context lifecycle without claiming that Claude hook definitions are active.

Core principle: dummyindex stays the spine — deterministic commands manage context and state, while the active host's dispatched agents do judgment and production-code work.

Install

User-global (one-time):

pip install --user dummyindex        # or: uv tool install dummyindex
dummyindex install                   # flagless = universal + linked (default)

Betas are cut from development and are opt-in: pip install --user --pre dummyindex (or uv tool install --prerelease allow dummyindex).

A flagless install is universal by default: one real .agents/skills tree, plus one relative symlink per skill family on the Claude side — so a single install works in Claude Code, Codex, Cursor, and Gemini CLI with no further flags. CLAUDE.md and AGENTS.md are both written.

~/.agents/skills/dummyindex/          # real files — the portable rendering
~/.agents/skills/dummyindex-plan/     # ...and the other 6 sibling families
~/.claude/skills/dummyindex        -> ../../.agents/skills/dummyindex
~/.claude/skills/dummyindex-plan  -> ../../.agents/skills/dummyindex-plan

Links always point .claude → .agents, never the reverse: .agents is the portable rendering, and Claude Code is the only host that reads only .claude, so it's the one side that gets symlinked.

Narrow to one host with --platform claude or --platform agents (real trees on both sides, no linking — there's nothing to link to):

dummyindex install --platform claude # ~/.claude/skills/dummyindex*/
dummyindex install --platform agents # ~/.agents/skills/dummyindex*/

--platform agents is the platform-agnostic selector — the same .agents/skills family is discoverable by Codex, Cursor, Copilot CLI, OpenCode, Amp, Gemini CLI, Goose, Pi, Cline, and other Agent-Skills/AGENTS.md harnesses. --platform codex still works as a deprecated alias: it prints one deprecation warning and renders byte-identical skill trees.

Linking is a tri-state. The default (no flag) is AUTO: link when possible, otherwise fall back to writing real, duplicated trees with a warning. --copy opts out for one run — always writes real, duplicated trees under both .claude/skills and .agents/skills, exactly like every release before this one. --link is the strict form: it errors instead of silently falling back to copy when the Claude side can't be symlinked.

Updating dummyindex migrates duplicated repos automatically. A plain dummyindex install — the same command /dummyindex-update runs — detects an existing duplicated layout (real trees under both .claude/skills and .agents/skills) or a claude-only layout, and converts every proven Claude copy to a link in that same run, printing a claude skill migrated -> ... line per family the first time. This happens even when both copies already carry the same version stamp — repair alone would leave those alone, but migration forces the conversion. After that first run, the layout is already linked and reinstalls are quiet.

Windows: on a checkout with core.symlinks=false, or without Developer Mode enabled, AUTO falls back to copy mode with a warning — the install still succeeds and both hosts still work. A materialized-link placeholder (a plain file standing in for the symlink, the shape that checkout produces) is recovered by running git config core.symlinks true and re-checking out the repo; the next dummyindex install then replaces it with a real link.

Per-repo (no global state):

cd /path/to/your/repo
dummyindex install --scope project
# universal + linked, rooted at this repo instead of $HOME

dummyindex install --platform agents --scope project
# narrowed: writes only .agents/skills/dummyindex*/ and the active project
# instruction file

To remove:

dummyindex uninstall
# flagless = removes both hosts, mirroring a flagless install

dummyindex uninstall --platform agents
# add --scope project [--dir PATH], or use --platform both

Uninstall removes the whole named family directories it installed — the main skill plus its 7 siblings — under .claude/skills/ and/or .agents/skills/ at the target scope, so don't store your own files inside .claude/skills/dummyindex*/ or .agents/skills/dummyindex*/; they won't survive an uninstall. Narrowing removal with --platform agents on a linked layout also sweeps the now-dangling Claude-side links and prints the exact dummyindex install --platform claude command to restore that surface, since the Claude side was only ever links into the tree that command just removed.

Rerunning install also repairs stale copies an older dummyindex version left behind (stale preambles, stale managed guidance blocks), scoped to the platform(s) and scope you selected — everything else is reported with the exact command to fix it. A duplicate copy at both user and project scope is reported, never deleted, unless you pass --dedupe user|project; a copy stamped newer than the running package is left alone unless you pass --force-downgrade. A foreign symlinked family dir — one dummyindex didn't create — is always refused; a dummyindex-owned link (the linked layout itself) is recognized and left in place or healed, never mistaken for a stray duplicate. See docs/COMMANDS.md for the full contract.

Cursor already reads .claude/agents/ natively, so a repo's context equip-generated Claude agents work there with no extra rendering.

Quickstart

The same workflows have host-native spellings:

In Codex you can also open /skills and select any installed dummyindex skill. dummyindex installs Agent Skills rather than custom prompts, so $dummyindex-plan is the supported equivalent of the Claude slash invocation.

CLI — the agent's deterministic backbone (no LLM cost). The skill and council run these; you don't type them by hand. The only terminal commands a human runs are the install bootstrap above. Shown here for transparency:

dummyindex ingest . --platform agents # build .context/ + active Codex/Cursor/etc. project guidance
dummyindex context query "how does auth work"   # ranked feature shortlist
dummyindex context rebuild --changed .          # quick deterministic backbone refresh
dummyindex context reconcile .                  # what drifted since the last reconcile (commit-anchored)
dummyindex context hooks status .              # Claude integration hook health
dummyindex context --help            # full command list

Token usage reads Claude Code transcripts and is therefore Claude-only. The Claude install includes /tokens; a Codex-only install does not create it. Codex already provides /status for current context/session tokens and /usage for account usage.

dummyindex usage                     # current chat: context window + dedup session totals
dummyindex usage daily               # per-day totals across every project (also: session|monthly|blocks)

Full command reference: docs/COMMANDS.md — every slash command and CLI command in one place.

Docs

• docs/COMMANDS.md — every command (slash + CLI) in one page
• docs/README.md — docs index (guide + reference + internal)
• docs/guide/ — twelve conceptual docs: architecture, data model, CLI surface, lifecycle, retrieval model, and more

License

MIT — see LICENSE.