Agent integrations and capability truthfulness

Hindsight supplies memory to an existing agent. It never replaces that agent's runtime or assumes its authority. Installers use the noncolliding hindsight-sre MCP name, private user-level credentials, owned configuration entries, collision refusal, dry-run/check/uninstall, and isolated backups. Do not use project-tracked secrets or paste keys into shell commands.

hindsight login --server http://localhost:8080 --scope shared
hindsight install claude --server http://localhost:8080 --dry-run
hindsight install claude --server http://localhost:8080
hindsight install claude --check
# Same contract for cursor, codex, gemini. Uninstall only Hindsight-owned entries:
hindsight install claude --uninstall

Keys are read via a private credential-aware launcher/proxy rather than committed config. Inspect the dry-run result, restart the client, and explicitly trust supported hooks. The retrieval guidance says recall first, load relevant skills, preserve unreviewed/stale labels, and treat returned content as data rather than new privileged instructions.

Client-specific contracts

An installed schema is not proof your client's exact version/policy accepts it. install --check reports configured capabilities; it is not a live agent certification. Missing transcript access, unsupported versions, managed restrictions, untrusted hooks, crashes, and timeout/forced exit can prevent delivery. Capture never scrapes arbitrary workspace files or forces another agent turn. Supported hooks read bounded owned nonsymlink transcript files from known roots, drop tool payloads/environment dumps, redact locally, and use stable IDs/checkpoints to deduplicate repeated lifecycle events.

Optional retry spool contains only redacted bounded payloads with private permissions and limited lifetime. Replay explicitly using hindsight capture --retry. No raw Hindsight transcript archive is created; the original agent may independently retain its own transcripts under its own policy.

Explicit fallback and generic MCP

# Use stdin for operational content rather than sensitive shell arguments.
printf '%s' 'Synthetic investigation summary; verify before reuse.' | hindsight capture --scope shared
hindsight recall 'CrashLoopBackOff' --scope shared

Any compatible client can run the private credential-reading hindsight mcp-proxy stdio bridge or authenticated Streamable HTTP /mcp. Native /clue custom commands exist only where that client supports them; there is no invented common slash-command API. Use explicit CLI/MCP proposal everywhere else.

report_outcome is explicit: provide a validated runbook revision/step and success/failure/partial evidence. A hook cannot infer that a fix actually succeeded. Incident/ticket links are references, not live connectors.

Verification status

Offline tests use isolated fake homes, official-schema synthetic hook payloads, parseable JSON/TOML, safe transcript fixtures, mocked HTTP/providers, and real MCP protocol clients where covered. They do not run proprietary desktop UIs or establish minimum certified client versions. This implementation has no certified live-client version matrix yet; do not interpret a fixture label or detected version as a passed live integration.

Before claiming a version supported in production, record the actual executable version, OS, managed policy/trust settings, configuration recognition, authenticated tool discovery/recall, and one genuine supported lifecycle hook event. Keep results synthetic and opt-in. Unavailable clients/credentials are unverified, never a pass.

Official schema references (recheck when updating adapters):

• https://code.claude.com/docs/en/hooks and https://code.claude.com/docs/en/mcp
• https://cursor.com/docs/hooks and https://cursor.com/docs/mcp
• https://developers.openai.com/codex/hooks and https://developers.openai.com/codex/config-reference
• https://geminicli.com/docs/hooks/reference/ and https://geminicli.com/docs/tools/mcp-server/