Feature Parity: Kiro CLI vs Claude Code

Claude-mem on Kiro CLI keeps the full pipeline — capture, AI compression, storage, search, injection — with a handful of host-imposed differences documented here.

Hook mapping

Claude Code	Kiro CLI	Status
`SessionStart` → context injection	`agentSpawn` (stdout injected into context)	✅ Full — raw text instead of JSON envelope
`UserPromptSubmit` → session init + semantic inject	`userPromptSubmit` (stdout injected)	✅ Full
`PostToolUse` → observation capture	`postToolUse`, matcher `*`	✅ Full — Kiro tool names mapped (`fs_write`→`Write`, `fs_read`→`Read`, `execute_bash`→`Bash`)
`Stop` → session summary	`stop` (`assistant_response` in payload)	✅ Full — better: no transcript parsing needed
`Setup` → version check	—	❌ No Kiro equivalent — install and repair commands perform runtime setup explicitly
`PreToolUse: Read` → per-file context	—	❌ Blocked — Kiro does not inject `preToolUse` stdout
`SessionEnd` / `PreCompact` / `SubagentStop`	—	❌ No Kiro equivalent

Full parity (no user-visible difference)

Observation capture and compression, session summaries, session-start memory injection, per-prompt semantic injection, MCP search tools (search, timeline, get_observations, corpus tools, smart_* tree-sitter tools), the web viewer + SSE live feed, Chroma semantic search, modes, project exclusion, privacy tags, Telegram notifications, and all worker-side features — these are host-independent and work unchanged.

Compression provider: Kiro installs default to the Kiro provider — compression runs on your own Kiro subscription via headless kiro-cli chat --no-interactive through a dedicated hook-less, tool-less claude-mem-observer agent (the hook-less agent is the recursion guard: it prevents claude-mem from observing its own compression chats). No API key and no co-installed Claude Code needed. claude, gemini, and openrouter remain available via CLAUDE_MEM_PROVIDER.

Skills: Kiro implements the same Agent Skills standard (SKILL.md), so all skills install verbatim and appear as slash commands. Skills that orchestrate Claude Code subagents (do, make-plan, babysit, standup) load fine but their subagent steps depend on Kiro's delegate/subagent capabilities; version-bump is Claude-Code-plugin-specific.

Degraded, with compensation

Feature	Gap	Compensation
Per-file context on `Read`	`preToolUse` stdout is not injected; exit 2 would block the tool, which claude-mem never does	Per-prompt semantic injection is enabled by default on Kiro installs — relevant observations (including file-scoped ones) arrive with each prompt
Post-`/compact` re-injection	No compaction lifecycle event	Same: the next `userPromptSubmit` re-injects semantically relevant context
Subagent attribution	Kiro hook payloads carry no `agent_id`/`agent_type`	Observations from delegated sessions are recorded unattributed; revisit when Kiro ships its official subagents tool
Terminal hints (`systemMessage`)	Hooks have no user-visible output channel	Viewer URL is printed at install time and by `npx claude-mem status`
Transcript watcher	Kiro provides no `transcript_path`	Not needed — `stop` delivers `assistant_response` directly
Auto-update	No plugin/marketplace system in Kiro CLI (kirodotdev/Kiro#8578)	`npx claude-mem update`, then re-run the Kiro install (idempotent)

Safety invariants (why the port is shaped this way)

Raw stdout, not JSON: Kiro injects hook stdout verbatim. The Kiro adapter emits plain text for context and nothing at all otherwise — a JSON envelope would pollute your agent's context.
Silent stop hook: Kiro parses stop stdout for a {"decision": "block"} override that forces the conversation to continue. Claude-mem's stop hook is unit-tested to emit zero bytes.
Never exit 2: on Kiro, a preToolUse hook exiting 2 blocks your tool call. Claude-mem hooks soft-fail with exit 0 when the worker is unreachable.
Explicit timeout_ms: Kiro's 30s default is too short for a cold worker start; the installed hooks carry 60–120s timeouts (300s for the silent self-heal hook that restores plugin dependencies), with a dedicated warm-up hook first.

Verified against

kiro.dev CLI docs plus a live payload capture on kiro-cli 2.11.0 (July 2026) — see tests/fixtures/kiro/README.md for the captured contract, including two deviations from the docs the port handles: the session UUID arrives as the KIRO_SESSION_ID environment variable (no session_id payload field), and fs_read batches reads in an operations array. If behaviour changes after a Kiro upgrade, re-run the probe in tests/fixtures/kiro/probe/ and file an issue with the captured payloads.