Solution for long term memory for agent coding CLIs and to facilitate handoff between different agent vendors
Rust
Merge remote-tracking branch 'origin/main' into merge/main-wrapper-into-3.0
AkitaOnRailsac35bf25a4a5
diff --cc CHANGELOG.mdindex b75d93c0,318fa324..3276b64f--- a/CHANGELOG.md+++ b/CHANGELOG.md@@@ -523,15 -27,16 +523,25 @@@ and [0.x](docs/changelog/0.x.md)that is being deleted, so the queue-file guard failed with `inspectrelay.sqlite-journal: Access is denied`. The guard now retries that answerfor up to about 100 ms; a file that stays unreadable is still refused.+- Fixed SessionEnd consolidation repeating the same expensive prompt on the+ durable queue when the structured LLM response fails deterministically: a+ truncated or empty response, malformed JSON, or an unexpected response+ shape now ends the queue on the first attempt, keeping the heuristic page+ the hook already wrote; pre-send connection and capacity (`503`) failures+ retain the bounded backoff. (#1131)+- Fixed structured LLM responses stopped at the output budget+ (`finish_reason = "length"`) or returned without usable content: they now+ fail with redacted terminal errors, without copying the response. (#1130)+ - Fixed `ai-memory hook` running inside a helper container on Docker-wrapper+ installs (Linux/macOS). A hook config or plugin that called the wrapper paid+ about 280 ms per event, over the 200 ms hook budget, and its default+ `127.0.0.1` server URL pointed at the container instead of the host server.+ The wrapper now runs `hook` and `hook-drain` through the same+ checksum-verified native host client as `mcp-bridge` and `install-hooks`.+ A hook only reuses the installed client, with no update check or download+ on the hook path. When no client is installed yet, the wrapper says to run+ `ai-memory install-hooks` once, skips the event and exits 0, so the agent+ is never blocked.## [2.6.6] - 2026-10-13diff --cc bin/ai-memoryindex a6e68473,03568e85..e1c3c198--- a/bin/ai-memory+++ b/bin/ai-memory@@@ -648,9 -616,24 +668,25 @@@ case "${1:-}" i# has no host networking), so the bridge must run on the host through# the same checksum-verified native client install-hooks uses.NATIVE_HOST_BIN=$(native_host_binary)+ use_volume_tokenexec "${NATIVE_HOST_BIN}" "$@";;+ hook | hook-drain)+ # Lifecycle hooks are client-side for the same reason as mcp-bridge: the+ # default hook URL is the host's loopback server, and the cwd, marker+ # files, capture exclusions and hook spool all live on the host. A helper+ # container would also cost ~280 ms per event, past the hook time bound.+ # Reuse the native client mcp-bridge installs (cached only: no update check+ # or download on the hook path) and never fall back to the container. If+ # that client is unavailable, native_host_binary has already said why on+ # stderr; drop this one event and exit 0 so the agent is never blocked or+ # shown a failing hook.+ if NATIVE_HOST_BIN=$(native_host_binary cached) && [ -x "${NATIVE_HOST_BIN}" ]; then+ exec "${NATIVE_HOST_BIN}" "$@"+ fi+ echo "ai-memory: native host client unavailable; lifecycle hook event not captured" >&2+ exit 0+ ;;esac# ---- normal pass-through to the binary inside a container ----------------diff --cc docs/security-boundaries.mdindex 6c757d88,ad2fbb81..b24f8ca9--- a/docs/security-boundaries.md+++ b/docs/security-boundaries.md@@@ -78,15 -73,11 +78,15 @@@ means the boundary is not built yet| 11e | Native self-upgrade archive confinement (#1025) | `ai-memory-cli/src/commands/upgrade.rs` normalizes every tar/zip path, rejects links and special entries, extracts only the binary and `hooks/`, accepts only the exact non-runtime support layout emitted by `release.yml`, and caps both entry count and declared expanded bytes before writing | `upgrade.rs` `release_workflow_support_assets_are_exact_and_ignored` (shipped support paths admitted, near-miss docs/packaging/template paths refused), `extracts_archive_built_like_release_yml` and `extracts_zip_happy_path_with_hooks` (real workflow-shaped archives succeed but support assets are not written), `validate_rejects_path_traversal`, `zip_extract_rejects_path_traversal`, `tar_extract_rejects_declared_expansion_past_limit`, with binary/hooks extraction as the legitimate control | STRONG || 11f | Native hook executable-path integrity after self-replacement (#1027) | `ai-memory-cli/src/commands/render_shared.rs` `hook_embedded_exe_path` is the single executable resolver for every persisted native hook form (POSIX/Windows command, Windows exec, Zero, ZCode, OpenCode 2 assistant capture): it prefers an existing literal path, strips Linux's trailing ` (deleted)` procfs marker only when the stripped installed binary exists, and otherwise falls back to the platform command name. Upgrade keeps the already-loaded config and refreshes in-process | `render_shared.rs` `embedded_exe_strips_the_marker_when_the_stripped_path_exists`, `embedded_exe_preserves_a_real_file_whose_name_has_the_marker_suffix`, `embedded_exe_falls_back_when_the_resolved_path_is_missing`, `hook_embedded_exe_path_names_an_existing_file_without_marker`; `upgrade_e2e.rs` `native_upgrade_refreshes_hooks_without_a_deleted_path_or_losing_cli_data_dir` reproduces the replacement window through the shipped binary, plants the corrupt suffix, and proves refresh removes it while preserving a CLI-only data dir | STRONG on Linux reproduction plus platform-neutral resolver/render coverage; Windows path rendering remains covered by its existing command/exec tests and full Windows CI || 11g | Cross-project tool attribution cannot bypass source or destination capture, or destination server routing (#932, #1199) | `ai-memory-hooks/src/capture_policy.rs` `absolute_tool_location_paths` accepts only bounded, recognized, absolute locations: file-tool targets, a search/list tool's explicit `path`, or a shell command's structured `workdir` (command text is never parsed for a location); `ai-memory-cli/src/commands/hook.rs` `tool_location_routing_cwd` requires every path to prove the same repository/marker boundary, refuses the reroute when the payload cwd's own capture policy would not keep the event (source `ignore_paths` may name paths outside the source checkout, and a command may name source files while its `workdir` points elsewhere), and resolves that destination before capture policy, allowlist marker, server profile, scope hints, or spool credentials. Relative, mixed-project, unsupported, and arbitrary non-project paths keep the payload cwd | `hook.rs` `absolute_file_target_uses_destination_scope_and_capture_policy` (hostile content under the destination exclusion is never spooled; a public control is routed there), `absolute_file_target_cannot_escape_a_destination_server_profile` (unregistered destination profile drops instead of falling back to source/install default), `file_target_routing_refuses_relative_external_and_mixed_project_paths`, `workdir_reroute_applies_the_destination_capture_policy` (a `workdir`-rerouted command reading a destination-excluded path is dropped, public control kept), **`a_reroute_cannot_escape_the_source_capture_exclusions`** (adversarial: a file read under a source pattern that names another project's directory, and a shell command naming an excluded source file with `workdir` in another project, are both refused; public file and shell controls still reroute; fails with the source-policy check removed), `hermes_tool_location_finds_the_marker_the_process_cwd_lacks` (Hermes `terminal` `workdir`, `read_file` `path` and `search_files` `path` reach the repository marker under allowlist; no-workdir, relative-workdir and relative-search controls stay with the unmarked process cwd); `capture_policy.rs` `absolute_tool_location_paths_cover_search_dirs_and_shell_workdirs` (relative workdir, command-text paths, workdir without a command, and Hermes `process` yield no location) | STRONG for native `ai-memory hook`; generated and compatibility hooks retain cwd-only routing |- | 11h | Docker-wrapper capture-policy capability (#1002) | `bin/ai-memory` routes ordinary Linux/macOS `install-hooks` and upgrade refreshes through the stable checksum-verified native host client, so the hook process executes the capture-policy boundary in 11b before spool/network. Only an explicit `AI_MEMORY_HOOK_PLATFORM=posix|windows` retains the documented script compatibility path and its warning; `AI_MEMORY_NATIVE_BIN` may not point back to the wrapper. For a server initialized on the named volume, `bin/ai-memory` `refresh_volume_token` (run only by `install-hooks`) copies `/data/auth-tokens/default` (only when `/data/config.toml` names it) to an owner-only `native-runner/server-token` through the helper's stdout, and `use_volume_token` points the native client at that copy (the managed host commands and `mcp-bridge` only read it and never enter the container engine) with `AI_MEMORY_AUTH__BEARER_TOKEN_FILE`, so it is a file-sourced token (never rendered into a harness config) and never an argv; a bind-mounted `AI_MEMORY_DATA_DIR` is read directly and an explicit `AI_MEMORY_AUTH_TOKEN` wins | `scripts/check-native-packaging.sh` `Checking host-launch wrapper routing`: a default allowlist install reaches the fake native client while a Docker executable that always fails proves no container fallback occurred; the explicit `posix` control reaches fake Docker and adds no native invocation; the self-referential native override is refused; `Checking the wrapper hands host commands the volume's root bearer`: for `mcp-bridge`, `install-hooks` and `run` the fake native client receives the token file (mode 600, right content) and the token is absent from its argv; controls: a bind-mounted data dir and an explicit `AI_MEMORY_AUTH_TOKEN` make no copy and ask no container (bite: dropping the export fails the check); `auth_default_e2e.rs` `a_host_client_given_the_token_file_by_env_authenticates` (the native side of the handoff, with a no-variable control) | STRONG on the shipped Linux/macOS wrapper. Windows and remote/manual script installers remain explicitly outside capture-policy v1 |-| 11h | Docker-wrapper capture-policy capability (#1002) | `bin/ai-memory` routes ordinary Linux/macOS `install-hooks`, upgrade refreshes, and every `hook`/`hook-drain` invocation through the stable checksum-verified native host client, so the hook process executes the capture-policy boundary in 11b before spool/network; a hook config that calls the wrapper never runs in the helper container, and an unavailable client drops the event with exit 0 instead of falling back to it. Only an explicit `AI_MEMORY_HOOK_PLATFORM=posix|windows` retains the documented script compatibility path and its warning; `AI_MEMORY_NATIVE_BIN` may not point back to the wrapper | `scripts/check-native-packaging.sh` `Checking host-launch wrapper routing`: a default allowlist install reaches the fake native client while a Docker executable that always fails proves no container fallback occurred; the explicit `posix` control reaches fake Docker and adds no native invocation; the self-referential native override is refused. `tests/wrapper_upgrade.sh` `lifecycle hook native host routing`: `hook --event session-start` and `hook-drain` reach the fake native client with the stdin payload while `docker`/`podman` stubs on PATH and in `AI_MEMORY_DOCKER` stay unused, and a non-executable `AI_MEMORY_NATIVE_BIN` exits 0 without reaching the engine | STRONG on the shipped Linux/macOS wrapper. Windows and remote/manual script installers remain explicitly outside capture-policy v1 |++| 11h | Docker-wrapper capture-policy capability (#1002) | `bin/ai-memory` routes ordinary Linux/macOS `install-hooks`, upgrade refreshes, and every `hook`/`hook-drain` invocation through the stable checksum-verified native host client, so the hook process executes the capture-policy boundary in 11b before spool/network; a hook config that calls the wrapper never runs in the helper container, and an unavailable client drops the event with exit 0 instead of falling back to it. Only an explicit `AI_MEMORY_HOOK_PLATFORM=posix|windows` retains the documented script compatibility path and its warning; `AI_MEMORY_NATIVE_BIN` may not point back to the wrapper. For a server initialized on the named volume, `bin/ai-memory` `refresh_volume_token` (run only by `install-hooks`) copies `/data/auth-tokens/default` (only when `/data/config.toml` names it) to an owner-only `native-runner/server-token` through the helper's stdout, and `use_volume_token` points the native client at that copy (the managed host commands and `mcp-bridge` only read it and never enter the container engine) with `AI_MEMORY_AUTH__BEARER_TOKEN_FILE`, so it is a file-sourced token (never rendered into a harness config) and never an argv; a bind-mounted `AI_MEMORY_DATA_DIR` is read directly and an explicit `AI_MEMORY_AUTH_TOKEN` wins | `scripts/check-native-packaging.sh` `Checking host-launch wrapper routing`: a default allowlist install reaches the fake native client while a Docker executable that always fails proves no container fallback occurred; the explicit `posix` control reaches fake Docker and adds no native invocation; the self-referential native override is refused. `tests/wrapper_upgrade.sh` `lifecycle hook native host routing`: `hook --event session-start` and `hook-drain` reach the fake native client with the stdin payload while `docker`/`podman` stubs on PATH and in `AI_MEMORY_DOCKER` stay unused, and a non-executable `AI_MEMORY_NATIVE_BIN` exits 0 without reaching the engine; `Checking the wrapper hands host commands the volume's root bearer`: for `mcp-bridge`, `install-hooks` and `run` the fake native client receives the token file (mode 600, right content) and the token is absent from its argv; controls: a bind-mounted data dir and an explicit `AI_MEMORY_AUTH_TOKEN` make no copy and ask no container (bite: dropping the export fails the check); `auth_default_e2e.rs` `a_host_client_given_the_token_file_by_env_authenticates` (the native side of the handoff, with a no-variable control) | STRONG on the shipped Linux/macOS wrapper. Windows and remote/manual script installers remain explicitly outside capture-policy v1 |+| 11j | Offline recovery journal and transcript replay | `ai-memory-cli/src/commands/recovery.rs` keeps bounded locators, ordered semantic transcript digests, pre-launch cursors and persistent correlated-delivery/loss evidence in an atomic journal, refuses a symlinked data dir and any symlinked component inside it (ancestors above the data dir are trusted, so macOS `/var` → `/private/var` works), sanitizes/caps failure reasons, and bounds each replay plus the whole pass; `hook_spool` coordinates enqueue and quarantine with short filesystem locks, canonically resolves partial scope from the recorded cwd, bounds per-file reads and current global occupancy across the main spool and recovery directories, sweeps completed tombstones, and holds correlated events outside drain until replay succeeds or restores them on failure; degraded resumes use `ai-memory-workstream`'s adapter-validated delta export, whose blocking parse and baseline scans are capped by record, event-count, source-byte and serialized-byte totals; `backfill::import_exact_session` chunks by item and byte limits and requires per-item durable acknowledgements; `run.rs` journals every session launch, including explicit spool-only evidence for unsupported exporters, and journals a resumed export only after capturing its exact pre-launch cursor; `workstream.rs` `/recover/finish` retains owner/project authorization, rejects superseded runs, and validates a finished retry by durable transcript/native/cursor/exit identity without regenerating its checkpoint; validated recovery ingest keys bind session, agent, event kind and sanitized event identity until session purge | `transcript.rs` `persisted_baseline_exports_only_offline_append_and_rejects_rewrite`, parser/baseline cap and ordered-semantic digest tests; CLI `run_offline.rs` real degraded spool-only and overlap/idempotency tests; `recovery.rs` `spool_only_without_correlated_evidence_never_auto_clears`, `a_dropped_spool_event_stays_lost_on_the_next_empty_pass`, loss/time-budget tests, `symlinked_journal_and_ancestor_are_refused`, `a_symlinked_ancestor_of_the_data_dir_is_trusted`, `a_symlink_inside_a_data_dir_under_a_symlinked_ancestor_is_refused`; `hook_spool.rs` partial-scope late enqueue, completion-race, occupancy churn/active-cap, bounded-read and tombstone-sweep tests, plus interval isolation (`completed_tombstone_keeps_a_later_resume_of_the_same_session` proves a post-interval resume hook is delivered, `completing_one_recovery_interval_leaves_the_other_intervals_entries_intact` proves one interval's completion never consumes another's evidence and ambiguous entries fail closed into the main spool); `backfill.rs` `two_sequential_recovery_intervals_on_one_resumed_session_replay_exactly_once` proves the second interval's content and session-end are stored with end-of-session effects while reruns stay pure replays; `transcript.rs` `an_oversized_first_record_is_skipped_by_session_listing_caps` proves the first-record byte caps bite before allocation; store `insert_observation_ingest_dedups_on_key` rejects cross-session and changed-event preclaims; hooks `expired_run_recovery_is_owner_only_and_refuses_superseded_runs` proves changed transcript refusal and changed-checkpoint no-op | STRONG on Unix integration coverage; Unix lock opens use no-follow and atomic replacement rechecks the destination. Windows owner-only ACL/reparse and lock semantics require the PR `windows` lane || 11i | OpenCode executable probing and generated-config cleanup (#1097) | `ai-memory-cli/src/config.rs` captures the full process environment once in typed `RuntimeEnv`; `run.rs` overlays explicit `run --env` values with platform key semantics and uses that same effective environment for executable resolution, probing, and launch without rereading process environment; `opencode_dialect.rs` invokes only direct argv `--version`, caps output and time, and rejects malformed/future majors. Automatic launch skips only OpenCode on probe failure; explicit OpenCode fails. `install_hooks.rs` quarantines the incompatible sibling before reading its exact ownership markers, restores unowned content with an atomic no-replace link, and never deletes or overwrites a concurrent replacement; `install_mcp.rs` owns the shared dual-schema endpoint/token inference used by both installers, rejects conflicting generated V1/V2 entries before mutation, and migrates an alternate-major entry only when endpoint, generated shape, and absent-or-Authorization-only headers match, preserving custom headers and ambiguous entries; auto-wire migrates MCP before hooks infer from that exact config path and skips hook mutation when transition preflight fails | `run.rs` `automatic_opencode_probe_failure_keeps_another_harness_available`, `opencode_executable_probe_selects_v2_for_launch_and_autowire`, effective-child-environment PATH casing controls, fake V1/V2/env-derived-major/malformed/future/timeout executable tests, `setup-agent` V1/V2 no-probe integration controls; `install_hooks.rs` `opencode_plugin_cleanup_removes_only_exact_owned_incompatible_sibling`, `opencode_plugin_cleanup_preserves_a_concurrent_user_replacement`, `opencode_plugin_restore_moves_quarantine_when_destination_is_free`, `opencode_plugin_restore_preserves_an_existing_replacement`, `opencode_plugin_restore_does_not_replace_a_file_created_at_restore_window`, `opencode_plugin_cleanup_does_not_escape_custom_target_directory`; `install_mcp.rs` `opencode_install_reuses_owned_endpoint_and_token_across_major_transitions`, `opencode_install_refuses_conflicting_owned_transition_entries`, `opencode_install_does_not_infer_from_a_user_owned_transition_entry`, `opencode_mcp_migration_moves_only_owned_alternate_shape`, `opencode_mcp_migration_preserves_ambiguous_user_entry`, `opencode_mcp_migration_preserves_custom_headers`, `opencode_mcp_migration_accepts_only_historical_generated_headers` | STRONG for Unix process probes and platform-neutral ownership gates; exact Windows executable resolution remains covered by `run.rs` resolver tests and requires the PR `windows` label |-| 12 | Network/auth posture | `config.rs` loopback `DEFAULT_BIND`; `serve.rs` `validate_http_exposure`, `require_allowed_host`; `auth.rs` `require_bearer`; `serve.rs` mounts `/identity` inside the host and machine-auth gates even with web disabled; `nix/nixos-module.nix` treats only literal `127.0.0.1`/`::1` as loopback, requires exactly one secret environment source for other binds, keeps secret auth and `llm_headers` values out of the Nix store, and brackets IPv6 in `ExecStart` | `serve.rs` host-guard (missing→400 / forged→403), non-loopback-requires-token; `auth.rs` wrong-token 401; `serve.rs` `machine_identity_remains_authenticated_with_web_disabled_and_expired_keys` (wrong and expired keys refused, valid machine key accepted, web remains disabled); `flake.nix` `nixos-module-eval` refuses `localhost` and `0.0.0.0` without a secret source, refuses two secret sources, `settings.auth` secrets, and `settings.llm_headers`, with literal IPv4/IPv6 loopback and one-source controls | STRONG |+| 11k | Server-side tool drop list (#1241) | `ai-memory-hooks/src/drop_list.rs` `IngestDropList::matches`, checked in `router.rs` `admit_hook` and the `handle_hook_batch` loop right after the client capture protocol and before the rate limiter, the ingest semaphore, the sanitizer and storage; lists from `[ingest] drop_tool_use_id_prefixes` / `drop_tool_names`, validated once in `Config::load` (at most 64 entries of 128 printable characters each). Only tool events match: a tool-call id by prefix, a tool name exactly, both read through the per-agent mapping `capture_policy::tool_payload_fields` that tool metadata uses (generic Claude-style fields for an agent without one; Factory Droid sends no tool-call id, so only its tool name can match). A drop answers 202 and counts as `dropped_by_policy`; the matched value is never logged | `router.rs` `the_ingest_drop_list_drops_before_any_capacity_is_spent` (with the semaphore exhausted a listed event still answers 202; control: an unlisted one gets 429) and `the_ingest_drop_list_applies_to_hook_batches` (listed items acknowledged as policy drops, the unlisted item ingested); `drop_list.rs` unit tests (prefix only at the start, names only exact, non-tool events never match, bounds refused; `each_agent_is_matched_on_its_own_tool_fields` matches OpenCode `callID`, Hermes `extra.tool_call_id` and Antigravity `toolCall.name`, with same-agent controls kept, and fails when only the generic fields are read). Both router tests fail with `matches` forced to `false` | STRONG. The list applies to every project on the server; it narrows what is stored and never widens access |+| 11l | Marketplace plugin front door: a plugin-launched event gets exactly the native hook's capture policy, scope and server routing, is captured once beside settings-file hooks, and sends nothing without the binary (#1166) | `plugins/*/scripts/ai-memory-hook` (rendered from `crates/ai-memory-cli/templates/plugin/ai-memory-hook.sh` by `render_plugins.rs`) has no capture logic: it only finds the binary and `exec`s `ai-memory hook` with `AI_MEMORY_HOOK_ORIGIN=<agent>-plugin`, `hook` first so the Docker wrapper routes it, and without a binary drains stdin, prints `{}` (guidance on `session-start` only) and exits 0; `hook.rs` `is_dropped_plugin_copy` + `plugin_origin.rs` `drops_plugin_event` drop a plugin copy when the user `settings.json` (honouring `CLAUDE_CONFIG_DIR`) or the checkout's `.claude/settings.local.json` carries an ai-memory hook for the same event (`install_hooks::settings_file_carries_ai_memory_hook_for_event`), and drop a folderless session (no `cwd`, or a Claude desktop `scratch-workspaces` cwd); Codex plugin events are never de-duplicated (an untrusted settings hook is skipped silently, so dropping could lose the event); the drift test keeps every plugin hook on the native path (no `hooks/<agent>/*.sh`, no credential in any plugin file) | CLI `tests/suite/plugins.rs` through the committed launcher and the real binary: `ignore_paths_hold_through_the_plugin` (ignored path never spooled; public-file control kept), `allowlist_mode_holds_through_the_plugin` (unmarked checkout emits nothing, prompts included; marked control captured), `marker_scope_routes_plugin_events`, `server_profiles_route_plugin_events` (profile URL and token, never the plugin's URL), `a_settings_hook_for_the_same_event_drops_the_plugin_copy` (user and project settings; unwired-event and no-settings controls), `folderless_sessions_are_not_captured_through_the_plugin`, `without_a_binary_nothing_is_sent_and_the_agent_is_not_blocked`, `foreign_hosts_are_attributed_or_ignored`, `committed_plugins_match_the_renderer`, `plugin_tree_carries_no_credentials`; `plugin_origin.rs` unit tests (`codex_plugin_events_are_never_deduplicated_against_settings`); `hook.rs` `plugin_copy_is_dropped_only_with_the_plugin_origin`; `tests/hooks/test_plugin_launcher.sh`; `tests/e2e/claude_plugin_smoke.py` (real `claude -p`: settings hook plus plugin give one `user-prompt`) | STRONG. Verified 2026-10-13: removing the `is_dropped_plugin_copy` guard fails the de-duplication and folderless tests (and the smoke reports `user-prompt x2`); disabling the native `Drop` disposition or the allowlist gate fails the `ignore_paths` and allowlist tests through the launcher. Not covered: tool events from a live model turn (no credential in CI), Windows Git Bash, Cowork and the desktop Code tab (see `docs/design-marketplace-plugins.md`) |+| 12 | Network/auth posture, including auth on by default for new installs (an existing install keeps loopback no-auth until `ai-memory auth enable`) | `config.rs` loopback `DEFAULT_BIND`; `serve.rs` `validate_http_exposure`, `require_allowed_host`; `auth.rs` `require_bearer`; `serve.rs` mounts `/identity` inside the host and machine-auth gates even with web disabled; `nix/nixos-module.nix` treats only literal `127.0.0.1`/`::1` as loopback, requires exactly one secret environment source for other binds, keeps secret auth and `llm_headers` values out of the Nix store, and brackets IPv6 in `ExecStart`; new-install default: `commands/init.rs` `root_bearer_for_new_config` mints a root bearer only on a fresh data dir (no config file and no database) and only when no bearer is supplied otherwise, via `local_auth.rs` `provision_root_token` (`auth-tokens/` `0700`, file `0600`, atomic, reused when present) and records the relative `[auth].bearer_token_file`; `config.rs` `apply_bearer_token_file` reads it once (`AI_MEMORY_AUTH_TOKEN` wins, inline + file together refused, a missing or blank file leaves no bearer) and `discover_root_token` lets a packaged install's client commands use `<data_dir>/auth-tokens/default` only with no config file at the probed path, a loopback server URL and no registered `default` profile; `serve.rs` `refuse_a_missing_token_file` stops a server whose configured file holds no token; `AuthSettings::renderable_bearer_token` keeps a file-sourced token out of every renderer: `install_mcp.rs` `route_file_token` (Claude Code gets the session-aware `ai-memory mcp-bridge` entry, Codex and Prime Agent name `AI_MEMORY_AUTH_TOKEN`, every other client a `stdio_bridge_entry` that runs `mcp-bridge --no-session-id` in its documented stdio shape, Pi keeps its extension, ZCode (HTTP-only) is refused) and `mcp_bridge.rs`, which authenticates with the config-resolved bearer and connects upstream before serving stdio, so a refused token ends it with `401 Unauthorized` instead of hanging; `uninstall.rs` `mcp_entry_is_ours` / `toml_mcp_entry_targets` recognise bridge entries by `install_mcp::bridge_entry_url`; `auth_enable.rs` `rewire_clients` re-registers every client whose entry targets this server, `install_hooks.rs` (persisted to the hook credential files, never inlined, not even on the F5 persist-failure fallback), `setup_agent.rs`, `run_autowire.rs`, `setup.rs` (runs `init`, adopts the minted token in memory through `adopt_token_file`, and calls the installers with no explicit token; the rendered user unit carries none; `user_service.rs` `is_rendered_by_setup` lets `uninstall` stop and delete only a unit carrying `setup`'s identity line whose SHA-256 still matches, never a hand-written or edited one); `local_auth.rs` `seed_hook_credential` gives hooks the token through `auth-token`/`auth-header` and keeps a different stored credential; `uninstall.rs` keeps the root bearer (`server_profiles::clear_tokens` `keep`); `commands/auth_enable.rs` is the idempotent opt-in | `serve.rs` host-guard (missing→400 / forged→403), non-loopback-requires-token; `auth.rs` wrong-token 401; `serve.rs` `machine_identity_remains_authenticated_with_web_disabled_and_expired_keys` (wrong and expired keys refused, valid machine key accepted, web remains disabled); `flake.nix` `nixos-module-eval` refuses `localhost` and `0.0.0.0` without a secret source, refuses two secret sources, `settings.auth` secrets, and `settings.llm_headers`, with literal IPv4/IPv6 loopback and one-source controls; `cli/tests/suite/auth_default_e2e.rs` against the shipped binary: `a_fresh_install_requires_the_token_and_its_own_clients_still_work` (after `init`, unauthenticated `/mcp`, `/hook`, `/api/v1/projects`, `/admin/status` → 401 and `/healthz` → 200; the same requests with the token succeed; the native hook delivers with no token in its environment, `status` authenticates through config, `mcp-bridge` lists tools; the token is in neither process's `/proc/<pid>/cmdline`; after `install-hooks`, `install-mcp --apply` for every client it writes (21, by default config path under the test home), a refused ZCode install and `auth enable` twice, a scan of the whole home finds the token nowhere and a scan of the data dir finds it only in the token file and the two hook credential files), `an_existing_install_is_unchanged_until_auth_enable` (control: a pre-default install serves every route open and `status` names the fix; `auth enable` twice is idempotent, rewires Claude Code to the bridge, and after a restart the routes refuse and the hook still delivers), `a_packaged_install_with_an_external_config_wires_its_clients`, `uninstall_keeps_the_root_bearer_and_a_missing_one_stops_the_server`, `a_host_client_given_the_token_file_by_env_authenticates`; `init.rs` `a_fresh_init_mints_a_root_bearer_the_loader_reads`, `init_on_an_existing_install_does_not_turn_auth_on`, `init_with_a_supplied_bearer_mints_nothing`, `re_running_init_keeps_the_same_root_bearer`; `config.rs` `a_relative_bearer_token_file_resolves_against_the_data_dir`, `a_missing_or_blank_bearer_token_file_is_reported_not_guessed`, `the_env_token_wins_over_the_bearer_token_file`, `an_inline_bearer_and_a_bearer_token_file_together_are_refused`, `a_client_without_a_config_file_discovers_the_root_bearer`, `root_bearer_discovery_is_refused_outside_its_narrow_case` (existing config, remote URL, `default` profile, env token); `auth_default_e2e.rs` `a_session_less_bridge_serves_any_client_without_holding_the_token` (`tools/list` and a scoped `memory_query` through `mcp-bridge --no-session-id` against a fresh auth-on server, the token in neither the bridge's `/proc/<pid>/cmdline` nor its `/proc/<pid>/environ`; control: the session-aware mode still refuses to start without `CLAUDE_CODE_SESSION_ID`) and `a_bridge_without_the_token_is_refused_cleanly_not_hung` (exits promptly naming `401 Unauthorized`); `mcp_bridge.rs` `only_the_session_aware_mode_forwards_and_requires_a_session_id`, `a_session_less_upstream_config_carries_no_session_header`, `stdio_bridge_forwards_tools_and_session_header_in_both_http_modes` (control); `install_mcp.rs` `a_file_sourced_token_wires_every_client_without_entering_its_config` (every written client: no token, no header, an entry that reaches the server; ZCode refused and unwritten), `stdio_bridge_entries_follow_each_clients_documented_shape`, `inline_and_explicit_tokens_still_embed_as_before` (control); `uninstall.rs` `bridge_entries_are_recognised_by_their_upstream_url` (JSON, OpenCode array and TOML shapes; a bridge to another server and a foreign command are kept); `install_hooks.rs` `a_file_sourced_token_is_persisted_but_never_embedded`; `auth_enable.rs` `enable_twice_is_idempotent_and_leaks_the_token_into_no_harness_config` (every client pre-wired over HTTP is re-registered, ZCode reported as manual, the second run byte-identical), `enable_leaves_an_entry_for_another_server_alone`, `enable_refuses_while_a_default_profile_would_share_the_token`; `local_auth.rs` `the_root_bearer_is_owner_only`, `seeding_never_replaces_a_different_hook_credential`; `setup_e2e.rs` `setup_installs_wires_is_idempotent_and_uninstall_removes_its_unit` (one `setup --yes` in a fresh temp home: the token is in no file under the home but the token file and the two hook credential files, not in the user unit and not in `setup`'s output; the self-check's `memory_status` and a hook event authenticate through the file; `uninstall` then disables and deletes the unit once and a second `uninstall` calls nothing), `uninstall_leaves_a_unit_it_did_not_write` (adversarial: the packaged unit copied by hand is kept, no `systemctl` call); `user_service.rs` `a_foreign_or_edited_unit_is_left_alone`, `rendered_units_carry_an_identity_that_survives_only_unedited` | STRONG. Bite checks for the new-install default: with fresh-install minting disabled the fresh e2e routes answer without a token and the `init` tests fail; with `renderable_bearer_token` returning the file token, `install-mcp`, `install-hooks`, `auth enable` and the e2e scans find it in harness configs; without the `serve` refusal a server whose token file was deleted keeps running (the e2e times out on it); without the hook seeding the fresh e2e hook is refused; without `clear_tokens` `keep` uninstall deletes the root bearer; with `route_file_token` no longer choosing the stdio bridge the every-client and `auth enable` tests fail; with `auth enable` skipping the re-registration its test fails; with the session-less mode forwarding an inherited id the bridge unit test fails; without the 401 naming the refusal e2e fails; without bridge-entry recognition the uninstall test fails; with `setup` handing the token to `install-mcp` its scan finds it in `.claude.json`; with `is_rendered_by_setup` accepting any unit, the foreign-unit tests fail; with discovery removed the packaged e2e gets 401, and with each discovery guard removed the matching case of `root_bearer_discovery_is_refused_outside_its_narrow_case` fails. Residual: a hand-edited `servers.toml` that adds a `default` profile after the root bearer exists would read the root bearer as that profile's token (`server add` refuses it, `auth enable` refuses while it exists); ZCode takes only HTTP entries and is refused, so it needs a manual entry; the stdio entries for Zero, Muse Code and DeepSeek Harness follow their documented or source-level format but were not run against those clients || 12b | Managed-workstream provenance privacy, stable replay identity and server event namespace | `ai-memory-core/src/workstream.rs` `scrub_workstream_provenance`; `ai-memory-hooks/src/workstream.rs` `finish_run` reserved client-id check and `sanitize_events` before immutable segments, with `sanitize_events` and `append_boundary_events` scrubbing event/checkpoint/losses content *before* the 64 KiB cap so a cap-straddling secret is never cut into an unmatchable prefix (#1113); `ai-memory-store/src/workstream.rs` `finish_run` stable identity check and `stored_event`, scrubbing content with the explicitly supplied configured sanitizer before its 16 KiB store-boundary bound; no changes to the typed observation boundary | `store/tests/suite/workstream_provenance.rs` `workstream_provenance_is_bounded_scrubbed_and_idempotent_before_storage` (hostile and oversized source ids, unknown metadata, legitimate correlation, retry and distinct event ids), `workstream_provenance_legacy_reads_strip_secrets_and_unknown_metadata` (search and tail, malformed/oversized legacy JSON), `workstream_provenance_reused_event_id_retains_data_and_refuses_identity_change` (first indexed version retained, identity refusal and batch rollback), `workstream_provenance_replay_survives_sanitizer_changes` (new extra patterns and removal of a built-in-pattern allowlist exemption, with identity-refusal control), `workstream_provenance_replay_preserves_legacy_rows_above_bounds`, `workstream_provenance_finished_run_ignores_late_events`, `workstream_provenance_retry_with_marker_matching_pattern_is_idempotent`, `workstream_content_is_scrubbed_before_the_store_cap` (a credential straddling the 16 KiB store cap is redacted, not persisted as an unmatchable prefix; benign control verbatim; #1113); `hooks` `workstream_provenance_finish_scrubs_before_raw_segment_and_search` (custom sanitizer, immutable segment, SQL recovery and finished-run no-op), `workstream_provenance_client_cannot_use_server_event_ids` (reserved ids refused before raw writes, normal client event and server-generated checkpoint/loss controls), `finish_run_scrubs_event_content_before_the_byte_cap` and `finish_run_scrubs_boundary_events_before_the_byte_cap` (a credential straddling the 64 KiB event cap is redacted in the raw segment and never reaches the index; checkpoint/losses and benign controls; #1113); `cli/tests/suite/workstream_provenance.rs` (text/JSON and old server DTOs) | STRONG: privacy, stable-identity and reserved-namespace regressions fail with their guards bypassed; an isolated allowlist bypass fails on the unknown key, and bypassing the raw-id search authorization returns 200 where its test requires 403. No provenance is emitted in startup packets, and old raw segments are not rewritten || 12c | Browser cross-origin writes refused on an unauthenticated server (a foreign page in the operator's browser, DNS rebinding, and a LAN host where browsers send no fetch metadata) | `serve.rs` `refuse_cross_origin_browser_requests` / `is_cross_origin_browser_request`, layered on the machine (`/mcp`, `/hook*`, `/handoff`, `/workstream/*`, `/identity`) and admin (`/admin/*`) routes when auth is off: `Sec-Fetch-Site` other than `same-origin`/`none` → 403, or, without fetch metadata, an `Origin` whose authority is not the `Host` → 403; and `is_non_script_browser_request`: any `Sec-Fetch-Dest` other than `empty` (case-insensitive; non-UTF-8 refused) → 403 whatever `Sec-Fetch-Site` says, so links, images, navigations, iframes and form posts are refused even as `same-origin` or `none`. Browsers send fetch metadata over plain HTTP only to a potentially trustworthy host, so on any other `Host` (a LAN address or name in the allowlist, `0.0.0.0`, `host.docker.internal`; `is_loopback_host` admits only `localhost`, `*.localhost`, `127.0.0.0/8` and `[::1]`) `handoff_fetch_lacks_client_marker` also refuses `GET`/`HEAD /handoff` without a non-empty `X-AI-Memory-Client` header (`ai_memory_hooks::CLIENT_MARKER_HEADER`) → 403 whose body names both remedies (re-stage hooks, or set a bearer token). A page cannot add that header without a preflight these routes never answer. `/handoff` is the one state-changing `GET` on these routes; the other `GET`s are reads a foreign page cannot see the answer to, and every cross-origin `POST` carries `Origin`. Every first-party client sends the marker: the native hook (`hook_capture.rs` `post_hook`/`post_batch`/`get_handoff`), CLI subcommands (`http_client.rs` `ServerEndpoint::authenticate`), every curl in `hooks/_lib.sh`, both header tables in `hooks/lib/ai-memory-hook.ps1`, and the generated TypeScript integrations (`render_shared.rs` `ts_auth_headers_fn` plus the spool drain). Loopback hosts keep accepting a fetch without it, so hooks staged before the header existed still work on a default install, except that a request carrying any proxy forwarding header (`Forwarded`, `X-Forwarded-For`/`-Host`/`-Proto`/`-Server`/`-Port`, `X-Original-Forwarded-For`, `X-Real-IP`, `X-Client-IP`, `CF-Connecting-IP`, `True-Client-IP`, `Via`; `PROXY_HEADERS`) counts as non-loopback, since a plain-HTTP proxy can rewrite `Host` to loopback while the browser behind it sent no fetch metadata. No first-party client sends those headers. Hooks, the CLI and MCP clients send none of these headers; the web UI's same-origin `/admin/pending-writes` calls are `fetch()` (`Sec-Fetch-Dest: empty`) and pass. Behind it, axum `Json` and rmcp's `Content-Type: application/json` check refuse the preflight-free content types (`text/plain`, form-urlencoded, multipart) with 415; `require_allowed_host` (and rmcp `allowed_hosts` on `/mcp`) refuses a rebinding `Host`; machine/admin routes carry no CORS layer, so a preflight never comes back permissive. A stored wiki page viewed in the built-in web UI could otherwise turn the guard's same-origin allowance into a claim: `ai-memory-web/src/markdown.rs` `safe_image_url` / `is_safe_image_target` render as `#` any image target that is root-relative, a network path, carries a `..` segment (also percent-encoded), or, for a UI mounted at the base-path root, starts with a server route segment Every `/web` surface (built-in pages including `/web/pending`, the login page, a custom SPA and `/api/v1`) answers `X-Frame-Options: DENY` and `Content-Security-Policy: frame-ancestors 'none'` (`ai-memory-web/src/mount.rs` `deny_framing`), because a framed page's own Approve/Reject `fetch()` is same-origin and the guard rightly lets it through | `cli/tests/suite/cross_origin_e2e.rs` `browser_cross_origin_requests_cannot_write_to_an_unauthenticated_loopback_server`, against the shipped `serve`: every write route (`/hook`, `/hook/batch`, MCP `memory_write_page`/`memory_delete_page`/`memory_message_send`/`memory_handoff_accept`, `/admin/write-page`/`delete-page`/`purge-project`/`messages/send`, `/workstream/runs`) × each simple content type × modern, Origin-only and fetch-metadata-only browser shapes → 4xx; body-less `GET /handoff` (which claims the single-use baton), `POST /admin/profile/rebuild` and `POST /admin/backup` → 403; a rebinding `Host` → 403; preflights answer no `Access-Control-Allow-*`; `GET /handoff` as a same-origin `document` or `image` and as a `Sec-Fetch-Site: none` `document` → 403 with the baton still open. Then a canary scan of the whole data dir, page/project/handoff survival, and legitimate controls (JSON MCP calls, a JSON hook → 202, the web UI's same-origin `/admin/pending-writes` list and approve reach their handlers, the anonymous active-project pointer is unmoved after the refused `/handoff` fetches while the closing hook-style fetch claims the baton and moves it). `serve.rs` `cross_origin_browser_request_detection` and `non_script_browser_request_detection` (header tables) and `cross_origin_browser_guard_applies_only_without_auth` (same-origin `image` and `none` `document` refused, same-origin `empty` passes, both pass with auth on); `ai-memory-web` `markdown.rs` `neutralises_images_that_reach_server_routes` (`/handoff?…`, `/admin/backup`, `//host/…`, `../` and `%2e%2e/` traversal, root-mounted route names refused) with `preserves_relative_images` as the legitimate control. `cross_origin_e2e.rs` `handoff_fetch_on_a_lan_host_requires_the_client_marker`, against a `serve` that allowlists a LAN name and address: `GET` and `HEAD /handoff` through either without the marker, and with an empty one → 403 carrying the remedy text, with the baton still open and the active-project pointer unmoved; controls: a marked fetch through the LAN name claims the baton, and a marker-less fetch through `127.0.0.1` (a hook staged before the marker) claims the next one; then a `127.0.0.1` fetch carrying each proxy forwarding header and no marker → 403 with the baton still open, and the same fetch with the marker claims it. `serve.rs` `loopback_host_detection`, `handoff_fetch_client_marker_decision` (incl. a loopback `Host` with each proxy header, refused without the marker and admitted with it, and a named row per header so none can drop out of the list) and `handoff_client_marker_guard_on_lan_hosts_and_under_a_base_path` (also under a `--base-path` nest; inactive with auth on). Front-door parity: `hook_capture.rs` `every_native_hook_request_carries_the_client_marker` (handoff GET, hook POST, batch POST), `http_client.rs` `authenticate_no_token_sets_only_the_client_marker`, the client-marker block of `tests/hooks/test_lib.sh` (every curl branch of the GET, the POST and the drain, plus a static check that every PowerShell header table carries it and every `Invoke-WebRequest` passes one), `render_shared.rs` `assert_every_ts_request_carries_client_marker` run for OpenCode, OpenCode 2, OMP, Pi and OpenClaw, `external_capture_ts.rs` (the rendered OpenCode plugin under Node sends no unmarked request) and `external_capture_powershell.rs` `every_request_carries_the_client_marker` (Windows leg); `ai-memory-web/src/mount.rs` `web_surfaces_refuse_to_be_framed` (`/web`, `/web/pending`, `/web/login`, `/api/v1/projects` all carry both headers; bites with `deny_framing` neutralized) | STRONG with listed residuals. Bite checks: with the marker check disabled, the LAN e2e marker-less `GET /handoff` returns 200 and the unit layer test fails; with the proxy-header rule disabled, the proxied loopback e2e fetch claims the baton and the unit decision test fails; dropping the marker from the native handoff GET, the shell token-drain curl, a PowerShell header table or the TypeScript spool drain fails the matching front-door test. With the guard removed, the cross-origin `GET /handoff` returns 200, consumes the baton and moves the active-project pointer (the pre-fix behaviour); with the guard treating `same-origin` as foreign, the web UI control gets 403; with the `Sec-Fetch-Dest` check removed, the same-origin `document` `GET /handoff` returns 200 and the unit guard test lets a same-origin image through; with the host layer removed, a rebound `/hook` returns 202; a permissive `CorsLayer` fails the preflight assertion; reverting `is_safe_image_target` to the pre-fix rule fails the markdown test on `/handoff?…`, dropping the percent-decoding fails it on `%2e%2e/`, and emptying the route list fails it on a root-mounted `handoff`. To reproduce the bite on the 415 layer, remove the guard and make `/hook` read a raw body instead of `Json`: a `text/plain` hook then returns 202 and the test fails. Residual, closed by setting a bearer token: a browser that sends no fetch metadata even to a loopback host (Safari before 16.4) sends no `Sec-Fetch-*` header and no `Origin` on an image `GET` or a link click, so such a request to a loopback host still reaches `/handoff` (loopback keeps accepting marker-less legacy hooks) and the other, read-only `GET` routes. On a non-loopback host the marker requirement closes `/handoff` for every browser. A plain-HTTP proxy that rewrites `Host` to its loopback upstream and adds no forwarding header (nginx's default `proxy_pass`) looks exactly like a local client, so in front of an unauthenticated server it must add `X-Forwarded-For` or keep `proxy_set_header Host $host`, or the server must sit behind TLS or use a bearer token (`security.md`, `https-via-proxy.md`). Browsers that send fetch metadata no longer have the earlier link residuals: a link opened from another app (`Sec-Fetch-Site: none`) and a link clicked in a stored page (`same-origin` navigation) both carry a non-`empty` destination and are refused. With auth on, the bearer token and the session CSRF token are the gate and this layer is a no-op |+| 12d | The web UI and `/api/v1` cannot write, and render stored text inert | `crates/ai-memory-web/clippy.toml` `disallowed-methods`: every `Wiki` write and destructive method (`write_page`, `write_page_if`, `apply_batch`, the delete/evict/purge/restore/move/reindex/remove-dir helpers, `commit_all`, `approve_auto_improve_proposal`) and the `writer()` / `git()` accessors are lint errors in the crate, so the `Wiki` held by `WebState` stays read-only under `cargo clippy -D warnings`; only the integration-test module carries the `allow`, for fixtures. Templates are askama with HTML auto-escaping; the history diff (`templates/history.html`) and the session timeline (`templates/session.html`) print page lines, observation titles and bodies escaped, with no `|safe`, and no captured value is interpolated into a script | `ai-memory-web` `history_replay.rs` `session_timeline_renders_captured_markup_inert` (a `<script>` title and body plus `"><img onerror>` and `' onmouseover='` payloads, checked to survive ingest first, render escaped) and `history_diff_renders_page_markup_inert` (the same payloads in a page version's diff lines); `mount.rs` `web_surfaces_refuse_to_be_framed` covers the history, timeline and history-API routes (200 with both anti-framing headers) | STRONG. Bite checks: a `Wiki::write_page` and a `Wiki::delete_page` call added to `routes/history.rs` fail `cargo clippy -p ai-memory-web --all-targets -- -D warnings` with `use of a disallowed method`; `|safe` on the observation body or a diff line fails the matching inert-render test. The lint is per crate: a write reached through another crate's helper is outside it, which review and the `/api/v1` "zero writer calls" rule still cover || 13 | Managed-run transcript attribution (concurrent launches in one checkout, invariant #16) | `ai-memory-store/src/workstream.rs` `link_native_session` stamps `native_session_linked_at` on its own run only, both `finish` updates drop the stamp when the session changes, `run_status` reports it; `ai-memory-cli/src/commands/run.rs` `resolve_native_session_after_run` takes a linked session only when `ai-memory-workstream` `native_session_in_checkout` holds it for this checkout (OpenCode by recorded directory), never falls back to a link it set aside, and turns an `AmbiguousNativeSession` into a warning with nothing imported; for Claude it moves the run's own session to a background session only when `ai-memory-workstream` `claude_continued_session` finds a `sessionKind: "bg"` record that names that exact session as its attached client and this checkout as `cwd` (#1050), and to a `/clear`ed session only when that session's own transcript records the `/clear` command and names the run's session as its launch `session_id` from this checkout or below it; `ai-memory-workstream/src/transcript.rs` `discover_crush` claims only the one top-level session created (or, with `--continue`, touched) during the run, and in a data directory outside the project only one that edited a file in it | `multi_session.rs` `a_session_linked_by_one_managed_run_is_not_another_runs`; `run.rs` `a_session_linked_during_the_run_wins_over_discovery` (concurrent newer session, another checkout's link refused, no fallback to it, unlinked control); `transcript.rs` `claude_continued_session_follows_only_this_checkouts_bg_records` (a concurrent launch's background session, another checkout's, and a non-background record refused; own attachment control); `transcript.rs` `claude_continued_session_follows_only_this_launchs_clear` (an uncleared transcript, a quoted `/clear`, another checkout's clear, a concurrent launch's clear, a subagent file, and a pre-run clear refused; own clear and the newest of two as controls); `run.rs` `a_claude_run_that_attached_to_a_background_session_follows_it`, `a_claude_run_that_was_cleared_follows_the_new_session`; `transcript.rs` `native_session_in_checkout_checks_the_opencode_directory`; `store/src/lib.rs` `managed_run_status_reports_a_link_made_during_the_run`; `run.rs` `ambiguous_crush_discovery_keeps_the_run`; `transcript.rs` `crush_discovery_claims_only_the_session_the_run_created`, `crush_discovery_in_a_shared_store_claims_only_an_edit_here` | PARTIAL: a run whose child links nothing still falls back to discovering the newest session in the checkout; Crush, which has no hooks, relies on that discovery alone and claims nothing when it is ambiguous || 13a | Managed Codex stale-daemon run recovery (#987; concurrent launches, project/worktree isolation, operator isolation) | `managed_runs.owner_user` (V71) uses the topology-aware `IdentityKey` owner stamp; `router.rs` resolves and authorizes the SessionStart repository before handling the untrusted run id; store `link_or_adopt_native_session` selects and links in one writer transaction and admits only the sole live, undelivered, never-linked Codex run in the exact workspace/project/checkout-cwd/owner bucket. An active supplied id must itself match every boundary; another harness, an ambiguity, or a second native session is refused | `multi_session.rs` `stale_codex_run_recovery_cannot_cross_project_owner_or_session_boundaries` (foreign owner, foreign checkout, and cross-project candidates refused; project-local controls and racing rebind), `stale_codex_run_recovery_fails_closed_on_ambiguity_and_active_mismatch` (two candidates and active foreign id refused, exact control); `workstream.rs` `prepare_stamps_the_operator_bucket_used_by_session_start_recovery` (HTTP actor stamp with foreign and legitimate controls); `router.rs` `managed_session_start_delivers_ledger_and_pending_handoff` supplies a finished stale id and proves the new run is linked and delivered | STRONG for SessionStart recovery. No candidate is selected from a run id's stale scope, another checkout, or another operator bucket; ambiguity deliberately falls back to no managed packet || 13b | Managed-run lease exclusivity and explicit recovery (one active run per workstream, invariant #16) | `ai-memory-store/src/workstream.rs` `prepare_run` expires lapsed leases and otherwise refuses an `active` run inside the writer transaction (`StoreError::WorkstreamBusy`). The explicit `force_unlock` path may expire and replace that row in the same transaction only when its durable `owner_user` equals the incoming actor's attribution; different operators remain isolated. `heartbeat` renews only `active` rows. `ai-memory-cli/src/commands/run.rs` normally uses `wait_out_held_lease`; `--force-unlock` instead asks once and fails immediately when refused. Neither path treats the informational `lease_owner` label as authority | `multi_session.rs` `force_unlock_replaces_only_the_same_operators_active_run` (foreign operator refused with live control, same operator succeeds, old run loses heartbeat, unattributed single-user control); hooks `workstream.rs` `force_unlock_cannot_evict_another_operator` exercises the HTTP actor boundary; CLI `run.rs` `force_unlock_asks_once_instead_of_waiting_out_the_lease`; existing controls `managed_workstream_batches_are_idempotent_and_release_the_lease`, `a_renewed_lease_is_reported_as_a_live_owner_not_taken_over`, `interactive_launch_waits_out_a_lapsing_lease_then_proceeds`, and `ctrl_c_aborts_the_held_lease_wait_immediately` | STRONG for exclusivity. The `lease_owner` label (`host:pid`) is informational only and not unique inside ai-jail (every jailed launcher reports `ai-sandbox:<ns-pid>`), so it must never become an ownership key. A shared unattributed/root context is intentionally one operator bucket; deployments needing person-level refusal must authenticate DB users |