Harnesses

the forty-three it reads, and what each one supports

deja indexes forty-three coding-agent histories into one retroactive index. A fix found in one agent is recalled inside another.

✅ works · — possible, not built yet · ✕ the harness has no such mechanism · ⚠ waiting on the harness itself · ? not investigated

• aider mcp — aider has no MCP client. Not a design limit but an unimplemented feature: several open requests and an open PR adding it, so this becomes work the day they ship it. Until then recall reaches aider through the read: context file (https://github.com/Aider-AI/aider/pull/5539)
• aider skill — aider loads instruction files, it does not load them by name on demand: --read (and a CONVENTIONS.md by habit) puts a file in every message, which is what deja already writes as aider's context file. There is no catalogue an agent picks from, so a skill would only be the same always-on text under another name (aider 0.86.2 --help: --read/--file/--config, no skills flag)
• aider command — the slash command set is built in; adding custom commands is an open feature request upstream, not something a third party can wire today (https://github.com/Aider-AI/aider/issues/894)
• Continue auto — Continue's CLI carries a Claude-shaped hooks system — Claude's own event set, additionalContext, settings read from ~/.continue/settings.json — and nothing fires it in the published release: measured on @continuedev/cli 1.5.47, hooks written for every event loaded ("Hooks loaded: 8 handler(s)" in its own log) and no handler ran, while the same run's tool call went through. Becomes work the day a release fires them (https://github.com/continuedev/continue/tree/main/extensions/cli/src/hooks)
• Cherry Studio command — No slash-command surface for a third party: commands are the app's own and an MCP server cannot register one (#3644).
• Roo Code auto — Roo Code is archived: RooCodeInc/Roo-Code was archived on 2026-05-15, the hooks PRs (#10785, #11128, #11663) never merged, @roo-code/cli is gone from npm, and the last extension, 3.54.0, has no hook event in its bundle. Recall arrives when the model calls the tool. (https://github.com/RooCodeInc/Roo-Code)
• Zed auto — Zed's agent exposes no lifecycle hook: nothing runs before a prompt is sent, and the extension API covers language servers, slash commands, context servers and debuggers only, so there is no point at which recall could be injected. Becomes work the day Zed ships one (no hook API in Zed 1.22.0 (#52729 closed unmerged) or extension API 0.7)
• Devin CLI command — Devin CLI has no user-registered slash-command file: its plugin manifest and skills directories are the only extension surfaces the docs describe, so a `/deja` command has no file to live in. (bundled docs at share/devin/docs/extensibility (3000.11.3))
• JetBrains AI Assistant auto — AI Assistant 262.10968.170 runs no user hooks: no hook configuration or hook runner is in the plugin, so nothing can put recall in front of its chat. (https://plugins.jetbrains.com/plugin/22282-jetbrains-ai-assistant)
• JetBrains AI Assistant skill — AI Assistant 262.10968.170's own chat reads no SKILL.md. Its skill folders (IdeManagedSkillsLayout, ~/.claude/skills, ~/.codex/skills) are handed only to the ACP agents it hosts (AgentSkillsAdditionalRootsProvider: junie, claude-acp, codex), and the chat's skill picker opens only for a Claude or Codex agent (AgentSkillsContextPopupContributor). Those agents already get deja's skill from their own directories. (https://plugins.jetbrains.com/plugin/22282-jetbrains-ai-assistant)
• JetBrains AI Assistant command — AI Assistant 262.10968.170 reads custom chat commands only from a project's .aiassistant/commands (CustomUserCommandsService); there is no user-level directory to write one to. (https://plugins.jetbrains.com/plugin/22282-jetbrains-ai-assistant)
• Windsurf / Devin Desktop auto — Cascade runs user hooks from ~/.codeium/windsurf/hooks.json, but none of their output reaches the model: exit 2 only blocks the action, and show_output is shown in the UI and is not offered for pre_user_prompt. Recall reaches Cascade through the MCP tool, the global rule and the skill. (https://docs.devin.ai/desktop/cascade/hooks)
• MiniMax Code command — No slash-command surface for a third party: the V1 plugin package carries skills, agents, hooks and scripts, and the manifest has no field for a command, so a /deja command has nowhere to be written (MiniMax plugin reference, package structure and manifest fields).

Hook surfaces

What reaches the agent in each harness, surface by surface. A test installs every target into a temporary home and fails when a target wires a surface this table does not claim, or the other way round.

✅ works · ◐ partly, see the note · — not built yet · ✕ blocked by the harness itself. Hover a cell for how it was checked: on a live stand, or by a fixture test.

Each format is documented in the session format registry, with synthetic fixtures and a conformance test that catches upstream drift.

Adding a harness

A new parser is one file in internal/sources/ modeled on an existing one, plus wiring at the discovery, doctor, install and registry points. See the open harness issues on GitHub if you'd like to add yours.

Found this useful? Star deja-vu on GitHub.