Skip to content

Latest commit

 

History

History
714 lines (620 loc) · 51.7 KB

mac-installation-plan.md

File metadata and controls

714 lines (620 loc) · 51.7 KB

First Mac client release

Current disposition — NO GO for protected client rollout

NO GO: keep the optional launch-context guard as a draft, parked pending native acceptance. Accepted historical core evidence and the existing internal pilot are unchanged; live pilot health was not rechecked. Original programme start of 5 November at 13:20 is retained; the previously recorded product repair count was 16, before this bounded whole-candidate repair slice.

2.1.319 observer candidate, 9 November. The explicit local observer contracts are claude-code/2.1.318 and claude-code/2.1.319. This adds a candidate, not native acceptance. The protected rollout remains NO GO and the version-matched native guarded-lifecycle rehearsal gate is unchanged. See the qualification and limits.

See whole-candidate repair and remaining findings.

Observation hardening, 8 November (offline only). The three confirmed observation defects below are now closed in memory_kit/context_authority.py by one shared correction: presence and identity come from the planned metadata that _remember_metadata/_check_plan already maintain, never from a fresh unplanned stat. Managed-preference paths and the excluded credentials entry are recorded in the plan; tree, empty-source, project .claude and .codex collectors use remembered presence; a planned directory that disappears during listing refuses; the export's project device/inode come from the plan, and the plan is rechecked after the last body read. Five fictional regressions failed on the previous source and pass now (three witnesses, plus two from independent review: credentials presence, and identity staying paired after the final check). Context tests 34 PASS, including the controlled-child approval/connect/reconnect/drift-denial/revocation workflow; the original witness script now refuses; repository checks exit 0, ALL PASSED, including 178 installer tests, in native Darwin user-temp. Independent Claude review: ACCEPT WITH LOW FINDINGS; two accepted findings fixed. The third (a vanished tree root raises a bare FileNotFoundError, still fail-closed) is retained. Other fresh-stat refusal checks (.git/worktree, signed remote cache metadata, ancestor binding) were not changed. Implementation: Codex gpt-6.1-sol; Claude did not author product code. No native client, host, two-UID or Keychain step ran.

Before the hardening above, fresh native Darwin user-temp fixtures resolved the earlier environmental test negatives: baseline 143 PASS; candidate repository checks exit 0, ALL PASSED, including 173 installer tests. Exact original archive-run failure IDs, exceptions and call chains now match between baseline and candidate: 28 socket-path length, 4 ownership-disabled-volume and 4 unsafe-ancestor cases. Earlier matching counts alone were insufficient; the new row-level comparison supplies the evidence. These historical environmental negatives remain recorded, not current product failures. Passing offline checks do not establish guard readiness.

Three independent discovery contracts completed with CHANGES_REQUIRED. Their responses each repeat the same three finding identities with shortened wording; they represent nine claims, not eighteen. Source adjudication confirms three defects, including the qualified ABA witness below. The other six claims remain unconfirmed or conservative refusals and are not promoted to demonstrated defects. The earlier Gemini planned-source-io/v1 PASS_FOR_SCOPE remains conditional on the stated, unshown discovery contract. Its exact native ID was explicitly selected, but returned output lacks a server-model field. Accepted revoke and registry bytes are unchanged. Neither scoped pass approves the whole guard. These findings concern consistency and source coverage within the ongoing observation; the approved guarantee still excludes later edits and native consumption, without claiming an immutable filesystem snapshot or universal confinement.

The three cases, as confirmed before the hardening (each now refuses offline):

  • Adversarial FAIL — preference omission: an unsupported preference created after planning is still advertised absent in the returned export. A fresh observation denies. This confirms that preference path, not every signature or git-path variant whose parent metadata may already be watched.
  • Adversarial FAIL — project identity misbinding: replacing the project after the fourth/final plan check returns old instruction content stamped with the replacement directory's inode. This is an observed export/binding defect, not a demonstrated native-client or host exploit.
  • Qualified adversarial FAIL — collector ABA: a fictional full-public-export witness moves a remembered skill directory away and restores it plus the parent's mtime before later checks. Export returns an absent directory slot and omits its existing source. Ordinary moves may be caught; this result depends on the stated controlled ABA/metadata-restoration interleaving.

Fresh Node 24 v24.22.2 version and packaged integrity checks pass. Earlier missing Homebrew-library failures are historical; recovery occurred outside this task, with no host action here. Earlier Node 22 v22.16.0 integrity PASS does not meet the installer's required major 24. Default Claude Code 2.1.319 now has an explicit observer candidate: native guarded-lifecycle acceptance remains unqualified.

Local test passes and witnesses use fictional inputs; native identity/UID/custody/TTY mocks remain labelled. Source-only pre-import loading, broader coverage, token/rollback, real native capture meaning/loading, two-UID/Keychain custody, protected bootstrap, host acceptance and product release remain open. Historical backend/global checks were not rerun. Dated and narrower outcomes below remain historical evidence; this section is the authoritative current go/no-go disposition.

Component Evidence Status Remaining gate
Historical core Council-reviewed bootstrap, restricted service, pending proposals, terminal review, persistence/recall and quarantined restore; later native core capture/recall rehearsal Historical internal acceptance Does not accept optional guard or client rollout
Fresh offline checks Baseline 143 PASS; candidate repository exit 0 / ALL PASSED, including 178 installer tests after hardening PASS in native-temp fixture environment Native gates remain
Shared planned IO Gemini planned-source-io/v1 conditional scoped PASS; 29 context / 68 mocked-boundary binding passes Conditional/local evidence Assumed discovery contract is not accepted
Discovery Preference omission, project identity and qualified collector ABA witnesses now refuse; 5 regressions RED then GREEN; Claude review ACCEPT WITH LOW FINDINGS Fixed offline (fictional fixtures, native identity mocked) Native guarded lifecycle; unchanged fresh-stat refusal checks remain unqualified
Revoke/registry Prior scoped Claude PASS; accepted bytes unchanged Scoped acceptance retained Not whole-guard acceptance
Node 24 Fresh v24.22.2 and packaged integrity PASS Required runtime available No host/install acceptance inferred
Current client Default 2.1.319; explicit contract 2.1.319, with 2.1.318 compatibility 2.1.319 observer candidate; no native acceptance Version-matched actual guarded lifecycle
Host/two-UID/Keychain No new native acceptance OPEN Protected bootstrap, custody and transfer
Pre-import/coverage/meaning Source-only loader, envelope coverage, token/rollback and native meaning/loading unresolved OPEN Scoped implementation/verification as authorised
Existing pilot No changes; health not rechecked Historical only No live health assertion
Protected release Offline PASS; observation defects fixed offline; native gates open NO GO Keep optional guard draft/parked

Status: installer implemented on the candidate branch, with native component evidence; not an installation approval. Follow installer acceptance for the required evidence and report. Complete two-account client use, secure handoff and protected recovery remain release gates.

The public code23/agent-memory-kit repository stays the distribution and documentation home. Preserve its Git history and previous file-vault implementation while replacing the default onboarding only when the new path is actually installable. The client runs memory on their own Mac; no hosted deployment is implied.

Client Claude Code update policy (James, 9 November 2026)

Agreed for when a real client install exists; nothing below is implemented yet.

  • Security releases: a Claude Code release that fixes a security problem reaches client machines within 3 working days of its release.
  • Routine releases: other releases are taken once a month.
  • Who clears a version: the AI, not James. A qualification script triages the release (changelog filter for context-loading changes plus a comparison of the context source names in the old and new binaries), Codex reviews the result, and if both pass the version is added as its own explicit contract entry. James is told the outcome and steps in only when a check fails.

Before the first client install this needs: the qualification script; version control through the requiredMinimumVersion / requiredMaximumVersion managed settings, which the observer must be taught to accept exactly (it refuses non-empty managed policy today); and a safe suspension path when qualification misses the 3-working-day window. Triage is not proof of identical loading behaviour; see the 2.1.319 observer candidate notes.

Frozen acceptance

  • One Mac walkthrough names every required tool, account, permission, credential context and command. It has no undocumented database/role preparation step.
  • Reuse the existing log this entry point and the existing three-model YOLO workflow. Do not create a second memory store or silently reroute personal capture.
  • Ship the tested runtime with complete licensing/provenance and reproducible source access. The six-file runtime verifier remains strict.
  • Setup detects missing prerequisites before mutation, refuses existing nonmatching configuration, stores no credentials in ordinary files or arguments, and preserves business data outside the checkout.
  • Decide and verify the Mac OS administration boundary. A same-account unsandboxed agent is not isolated by environment variables, separate terminals, Keychain labels or Docker.
  • Rehearse installation, pending capture, protected exact-content approval, restart/fresh-session recall, withdrawal, access revocation and service outage through the actual installed command/client.
  • Rehearse the protected backup/recovery and upgrade path. Keep offline archive activation disabled until current authority is reconciled.
  • Run repository checks and independent review before public default-branch replacement. Publish no secrets, private paths, client data or unverified compatibility claims.

Delivery order

  1. Preserve the current public baseline and create the feature branch.
  2. Package the existing tested runtime and licenses; provide its exact source revision and rebuild path.
  3. Complete the Mac provisioning/profile/binding implementation selected by the administration decision.
  4. Make mac-setup.md the exact novice walkthrough, replacing every unfinished step with a checked command or explicit human action.
  5. Verify in a clean non-production Mac context, complete independent review and repository checks, then push/update the public repository with a clear version boundary and legacy/migration guidance.

The before/after quality benchmark is deferred at the user's request. Essential installation, security and regression checks remain part of release acceptance.

Historical baseline instruction and capture compatibility

Blocked: the candidate cannot adopt an existing global instruction or personal capture context. Preserve those files, the current native configuration selectors and the three-model workflow. Do not delete or move instructions, use an empty native configuration directory, or add an acknowledgement flag to bypass the refusal. Successful service installation is not project-context approval.

The current project profile contains coding-user-owned managed-region checksums. They detect inconsistent edits; they are not an operator's approval. That user can regenerate the profile and matching managed blocks. The operator's existing installation root is private and cannot supply a coding-user-readable approval record. No protected context issuer, mandatory record lookup, revocation or reviewed capture-dispatch contract is implemented. A signed file referenced only by the mutable project profile would not solve omission or old-record replay.

The managed block currently treats logthis as business capture in its supported clean test context. It does not prove that a pre-existing personal entry point keeps its default or that explicit business intent wins unambiguously in a real native client. Native selectors are read from the current environment, and the wrapper launches the requested command unchanged. Its checks are not immutable context attestation or confinement of independently launched clients.

Existing boundary Current behavior
doctor, plan, install, operator status Validate or operate the protected service; do not approve a coding project's instruction/capture context
connect-check, connect, reconnect, run Refuse the unsupported global/capture paths through shared client checks; managed hashes establish consistency only
Rotation and interrupted binding recovery Preserve the existing workspace, exact managed regions and concurrent edits; neither creates reviewed context authority
Operator uninstall Preserves data, configuration and credentials; it is not project unbinding. No client unbind command is implemented

Before this gate can pass, a separately reviewed contract must establish:

  1. An authorized operator's review of the exact full instruction/capture context, including the supported client's loading rules and selected configuration directories. The current path-existence checks never authorize reading private bodies. Such review needs explicitly authorized sources and handling.
  2. Current protected authority outside coding-user-writable state, bound to the coding UID, project identity, workspace/service, software/client contract, selectors and complete approved context. Missing, changed, stripped, revoked or replayed authority must refuse at every affected lifecycle boundary.
  3. Reuse of the existing capture entry point: unscoped/personal capture retains its existing route; only explicit business selection proposes pending facts. No second command, store, model service or automatic publication is required.
  4. Synthetic failure-path evidence followed by the named two-account, unlocked-Keychain and actual-client acceptance. Wrapper tests cannot certify native loading or a host custody boundary that was not exercised.

No filename whitelist, coding-owned review flag or partial profile-format change is approved as a substitute. Until this contract and its fixtures exist, leave the integration stopped in an existing incompatible context. The current safe refusal and isolated clean-fixture tests are not a client-readiness verdict.

What the local evidence establishes

Historical verification recorded offline and packaged runtime passes. The latest full-kit attempt recorded installer failures/errors under strict archive guards; no full-suite pass is claimed for the local observer candidate. The client regression tests exercise clean preflight/connect, a real Python child, rotation/reconnect and another child, preserving fictional model/routing settings and unrelated configuration. Existing global instructions or personal capture cause initial and bound lifecycle checks to refuse without changing fixture bytes or reaching credential/child effects. No actual native coding client or private instruction body is used.

Separate isolated diagnostic witnesses confirm why those checks are not approval: the coding user can regenerate matching profile/managed-region checksums; clean native selectors can drift between calls; and context can appear after verification but before the mocked child boundary. The next check refuses that context, but no immutable launch-custody guarantee follows. These are blocking limitations, not successful compatibility or server-authorization tests. These baseline results are historical. The local observer below adds the separately approved admission mechanism.

Operator-controlled authority: approved local observation contract

The narrower observed-context approval mechanism is authorised. A conservative local Claude Code 2.1.318 observer now has a passing fictional integrated workflow. Independent final code review and full-suite validation remain pending; this is not native-client or installation acceptance. Independent contract challenge found the mechanism viable at the protected launcher's authorisation points, but blocked the stronger requirement to deny every concurrent context edit before child creation. Checks cannot freeze coding-owned files between their final read and the native client's later reads. A cooperative registry lock does not close that interval.

The registry would be a safety interlock on one protected launcher. It would not confine an unsandboxed coding account, certify independently launched clients or turn prompts into an authorisation boundary. Server membership, restricted read/submit tokens, token revocation and protected publication review remain the actual memory authority. Revoking context approval cannot revoke a token already held by a running process.

Candidate What it could establish Missing decision or access
Approval at the protected launcher's decision points Exact reviewed inventory matches the bytes observed during each current approval lookup; admit/revoke replaces blanket refusal Narrower guarantee accepted; complete native loading contract and workflow still unimplemented
Immutable client adapter Supported client consumes only a supplied immutable context Vendor-documented loading/discovery controls and a named authorised native fixture; substituting an empty configuration directory remains forbidden
Operator custody or host sandbox Coding processes cannot alter the context being consumed Separate host/security authority and an agreed change to the coding workflow

The first option is approved with acceptance frozen below. Edits after the final observation, later native loading and independently launched clients are outside its guarantee. The first invented-client draft did not meet that acceptance and remains stopped. The source-backed local observer is a distinct bounded implementation of it:

  • Fixed custody and discovery. A root-owned discovery directory under /Library/Application Support/Agent Memory Kit/Context Authority delegates a client-readable, operator-owned subtree for protected client software and the current registry. Every ancestor must prevent replacement by the coding account. Its real creation needs separately authorised host setup. The private service root stays private. Profile fields and inherited environment cannot choose or omit this authority. Lifecycle commands refuse an unprotected copy of the software. Descriptor-anchored registry reads check the operator owner, file type, inode, ACL and ownership-enforcing volume; the existing stricter project-file/private-root checks must not be relaxed to read public metadata.
  • Complete observed context. Protected software builds an inventory from a versioned loading contract for each supported client. The operator cannot omit sources by selecting a list. It records absent paths and complete supported instruction, import, skill, command, prompt, agent, plugin and settings trees. Unknown loading mechanisms refuse. A familiar filename is not evidence that competing capture instructions are absent. Only explicitly authorised bodies may be read or displayed; current development fixtures use fictional sources exclusively. Credentials are excluded, not copied or hashed into a context packet. Unclassifiable inputs stop intake.
  • Exact review and launch identity. Extend the existing read-only preflight and operator command structure to stream a canonical context inventory over the established transport. The operator re-hashes the displayed authorised bytes and confirms their digest through the protected terminal, without an approval flag or new capture store. Bind coding UID, canonical project path and device/inode, workspace/service, native selectors, context-affecting environment and arguments, resolved client executable/package identity, protected software digest, loading-contract version and complete inventory. Preserve ordinary model/routing arguments; refuse unreviewed context-affecting values, unknown clients and binary/version drift. Binary hashes are observed identity, not custody of a coding-owned executable.
  • Current lookup and effects. Give each approval a current generation in the protected registry. Missing, malformed, stripped, changed, revoked or old authority refuses at connect-check, initial and already-bound connect, reconnect and run, before credential, project-write or child effects at that decision point. Rotation and interrupted recovery do not mint approval. Record all intended prospective managed bytes during review, so binding cannot silently introduce unreviewed instructions. Recheck changes during an operation; report post-session drift separately, without claiming it was prevented.
  • Revocation and races. Revoke publishes a new revoked generation atomically without waiting indefinitely on a coding-readable lock. Launch takes a shared lock only through the bounded authorisation/spawn interval, never while waiting for the session. Operator synchronisation attempts are bounded and report any unconfirmed in-flight launch. Tests must cover parent and symlink replacement, malformed inputs, stale generations, concurrent editing, atomic replacement, maliciously held locks and parallel revocation. A final read still does not freeze files or establish immutable native consumption.
  • Existing capture dispatch. Preserve personal or unscoped capture on its previous route. Only explicitly selected business capture proposes pending facts through the same entry point; it never approves or publishes. Review the exact generated block with the rest of the context. Deterministic tests assert observable effects and structured contracts. A versioned, opt-in LLM meaning rubric and an actual supported native fixture remain separate gates; keyword matching cannot certify intent or personal-route coexistence.

Synthetic acceptance would exercise operator review, preflight, connect, a real controlled child, rotation/reconnect and revoke or changed-context denial, preserving fixture bytes, personal/default dispatch and model choices. It must exercise each failure path above before credential/write/child effects, while labelling mocks of UID/custody checks as mocks. Passing this would not establish a real two-account boundary, unlocked-Keychain transfer or native loading. Those require named authorised fixtures and remain unrun. No host bootstrap, client installation or mutable-profile bypass is part of the evidence. Final independent code review remains required.

Local observer candidate and limits

The invented-client attempt was stopped and removed. The implemented contracts are exactly claude-code/2.1.318 and claude-code/2.1.319, based on the official memory, settings, managed policy, server policy and CLI documentation plus bounded shipped release metadata inspection. Native binary identity binds resolved path, embedded release metadata and observed SHA-256; it is not custody or authenticity attestation. SHA-256 binds reviewed bytes, not vendor authenticity. No invented executable protocol is exposed. The resolved versions/<version> path name and embedded VERSION marker must agree; unknown versions refuse without fallback. Packets record the actual selected contract. Operator packet validation checks only the recorded executable identity's self-consistency, without resolving or reading the coding user's executable. Authorization at launch freshly observes the executable as the coding user and refuses a changed contract, executable version, resolved path or SHA-256; an executable change requires fresh review.

2.1.319 observer candidate: qualification rests on changelog review and binary string comparison. This is triage, not proof of identical loading behaviour. The review observes specific discovery slots; coverage remains unproven. The guarantee remains reviewed local observations during the authorization lookup, without an immutable-filesystem claim or proof of later native consumption. 2.1.319 contains security fixes for a Bash permission bypass and a sandbox escape; 2.1.318 is retained only for compatibility, not recommended. Version control via requiredMinimumVersion/requiredMaximumVersion managed settings is future work because the observer currently refuses nonempty managed policy. The existing version-matched native rehearsal remains required and unrun for this candidate.

The prototype observes its enumerated system/user instructions, managed settings/drop-ins and preference paths, selected native configuration, governing ancestors, descendant discovery slots and supported recursive trees. Independent review found remaining local-source coverage gaps; this is not a complete native context inventory. Missing supported slots are explicit. Known credential files are presence-only exclusions, never hashed or displayed. The local remote-policy cache is observed when empty; active unclassified cache or signature metadata refuses. Remote fetching itself is outside the boundary.

This conservative subset refuses imports/dynamic Markdown/frontmatter, nonempty plugins/managed policy, context-affecting or unknown mixed account-state keys, managed preferences, unknown settings/context arguments/injection environment, linked worktrees and unclassifiable support files. Auto-memory must already be explicitly disabled in the selected native user's settings.json, with no contrary observed setting; the mechanism never disables it or substitutes a clean home. The fixed auto-memory-disabled/v1 entry records only presence of the known projects root. It never enumerates its project names or reads, copies or hashes memory bodies. Its exact enum, boolean type, fixed path and disabled-policy prerequisite are validated and bound in current/prospective approval. It cannot classify an arbitrary active input as inactive. This local disabled-policy observation does not certify what a native client later loads. Native model and permission arguments remain unchanged. Existing native auth stays in the child environment and is absent from review packets/registry.

Current registry lookup is mandatory for preflight, initial and already-bound connect, reconnect and run. Review includes exact prospective managed bytes. Approval/revocation increments generation atomically. The operator lock must be 0600; packet-free revoke consults protected state. After terminal confirmation it attempts the exclusive launch lock for at most 0.2 seconds before publication. When acquired, that lock spans atomic publication. Otherwise publication proceeds and uncertainty stays true even if a shared lock releases during publication. A held readable shared lock cannot indefinitely block revocation publication. The launch lock is released after child creation, before waiting. in_flight_launch_uncertain covers only the authorization/child-creation window; false does not mean no session is running. Revocation never stops an existing session or removes the token already inherited by that process. This first version supports one active project per delegated registry; changing project requires explicit revoke. The fixed discovery/bootstrap remains an unrun host gate.

The integrated fictional workflow passed: canonical terminal review, issue, read-only preflight, initial/bound connect, actual controlled Python child, rotation/reconnect, another child, context-edit denial before effects and packet-free revoke/denial. Native resolver/source roots/UID/custody are explicit mocks. Structured capture directives retain default/personal on the previous route and select only explicit business memory_submit pending proposals. This checks the fixture protocol, not native LLM meaning. Final full-suite validation, independent code review, actual native loading and real two-UID/Keychain gates remain open. Historical failures/repairs and the missed first milestone remain retained in external task evidence.

Targeted privacy and revoke corrections; prototype still held

Focused regressions reproduced both confirmed defects before correction: unrelated project-memory enumeration, and false revoke certainty when a separate process releases its shared lock during publication. The repaired guarded fixture preserves the full approve/preflight/connect/bound-connect/real controlled Python child/rotation/drift-denial/packet-free-revoke workflow. Neither foreign text nor project names appear in its packet or registry. Additional tests reject malformed inactive entries, re-enabled or missing selected-user policy and inactive tags for other paths. A separate process proves the exclusive lock spans uncontended approval and revoke publication; a malicious held lock cannot prevent bounded publication. These are explicitly mocked native identity/root/UID/custody/TTY fixtures, not native or two-account acceptance.

The bytecode investigation remains an open pre-import gate. A harmless fictional module demonstrated that Python -I -B reads a matching existing .pyc while source bytes and the current software_hash() remain unchanged. The launcher imports kit code before authority(); a later tree check cannot undo code already executed. No launcher, host or security implementation was changed in this repair. A trusted pre-import bootstrap/source-only loading path and protected provisioning need a separate bounded implementation and independent validation before this prototype can claim protected software execution.

Independent observer/authority/lifecycle reviews returned changes needed. Local OS preference/git coverage and aggregate bounds, possible token-reference reuse, and inherited rollback/post-effect reporting concerns remain open or unqualified. The restrictive-umask usability gap remains open; the claim that umask002 turns 0644 into 0664 was incorrect and is not a defect being patched. No new exclusions hide local active inputs. The approved observed-launch boundary and server role/token/publication authority are unchanged. This partial correction is not whole-context completion, code approval, installation or release readiness.

Scoped review closure: namespace admission and revoke failure paths

Independent privacy review found the metadata-only memory-root repair insufficient: a selected custom native config below a valid project could still be reached by its generic descendant walk. That approach is stopped. The bounded recovery now refuses config/project overlap before any context body read or source walk, using canonical directory identities and no-follow metadata opens. Config aliases and linked ancestors also refuse. Packet intake applies the same admission. Nothing is pruned or relocated, no native selector is changed, and the inactive tag still cannot hide an active source. The regression uses a valid project and poison CLAUDE.md/AGENTS.md foreign cache files; the review's project-equals-HOME example would already be refused by the public project validator.

The independent revoke-certainty review passed its scope. Additional hardening now ensures revoke publishes its revoked next generation with uncertainty even if launch-lock opening/validation or non-contention flock acquisition fails. Approval still refuses these failures. Publication/flush/fsync errors are outside that catch and still propagate; repairing the lock does not restore old approval. The private operator lock, packet-free revoke, bounded wait and lock ownership across successful publication remain intact. Independent native Claude Opus 5.5 review passed both the certainty correction and the acquisition-error hardening, strictly within their stated scopes. No overall prototype approval follows.

Final disposition: privacy approach stopped

A further isolated fixture exposed a remaining alias route: a native projects cache link can point into the coding project, and the project walk attempts the foreign instruction read before the later cache-link rejection. No real private data was accessed. The privacy approach is stopped and remains unapproved; the passing scoped regressions do not establish the general privacy invariant.

Keep the prototype uninstalled. The revocation corrections have scoped review acceptance; privacy, broader local-source coverage and real host acceptance remain open. The bytecode fixture establishes missing integrity coverage where files are writable, not an exploit of a fresh protected installation: the fresh copier excludes caches. Source-only loading is a reviewed feasible design, not an implemented or certified loading boundary.

Read-order planner redesign (local candidate)

The earlier late-root checks are stopped as insufficient. A recorded fictional witness showed config/projects linked into a project cache reaching poison instructions before rejection. The observer now builds a metadata-only plan: no-follow/type admission of the inactive root and all fixed source, empty-only and traversal roots precedes enumeration. Recursive discovery records slots and absences without _classified or binding reads; only the completed plan is materialized through the existing _read/classification protections. Metadata is rechecked before reads and around binding preparation/verification. Profile and managed binding paths are planned too. Early profile/fallback readers use the same metadata planner before reading; external CLI/API options stay unchanged.

This refuses linked roots, ancestor skill aliases and inactive/recursive namespace overlap rather than pruning active sources. The inactive directory is inspected through its parent without opening/listing private children. Exact alias and no-body-before-plan regressions, phase-boundary drift/addition checks and the controlled Python-child lifecycle provide local evidence only. Accepted revoke code is unchanged. Native loading, real custody/host setup and the other coverage and loader findings remain unqualified; independent Claude code review is still required. The design review returned aborted streaming and NO VERDICT.

The read-order candidate preserves accepted revoke and existing read/classifier function bytes. Controlled-child and targeted alias checks passed before the implementation stop; full-kit and independent code review remain lead-owned.

Shared planned IO correction (held for independent review)

Independent review and a fictional witness showed a regular private directory could replace an admitted source between the plan check and body read. Current inode checks alone were insufficient. The new shared ContextVar scope supplements existing checks: every opened directory must match its planned identity before listing/yield, and each file must match its planned parent/name and leaf metadata before body reads. Both observer reads and binding snapshots use this boundary. It resets in finally/nested scopes, and is absent during effectful binding writes. Discovery listings use one descriptor-anchored scandir helper, retaining the open parent while consuming names/stat. No Path.iterdir listing remains in the observer.

Whole-plan checks run at phase boundaries; per-read checks cover their own chain and leaf. A deterministic several-dozen-directory count regression guards against quadratic whole-plan scans. When an ancestor .claude is the selected global config, its already-classified global slots/categories apply without treating history/projects as unknown project-local sources. A real fictional HOME-ancestor fixture exercises the unmocked scope algorithm and privacy guard.

The exact body-move witness first failed, then denied before foreign bytes. Final focused context tests:29PASS, including two actual controlled Python-child lifecycles, directory/listing and binding snapshot moves, public-run early ordering, ContextVar reset, aliases and drift. Legacy binding tests:68PASS with explicit boundary mocks. Accepted operator/revoke/registry helpers and other non-owned files remain byte-identical; existing read safety checks are supplemented, not removed. No whole-context/native/host readiness follows. Independent Claude code review is required. Latest lead full run167tests9FAIL27ERROR lacks failure IDs, so identical-baseline attribution is unproved. Node24 cannot load libsimdjson; lead reported Node22v22.16.0 integrity PASS. Neither runtime nor host repair was run here. Loader/token/rollback/broader coverage and real native/twoUID/Keychain acceptance remain open.

Whole-candidate repair, 9 November 2026 (offline)

The guarantee remains reviewed local observations during the authorization lookup. No immutable filesystem or later native-consumption claim is made. Rollout stays NO GO. The real two-UID boundary is untested because the native rehearsal is not being run (James, 9 November 2026). No installation, account, host-permission or live-pilot change is part of this repair.

Operator validate_packet checks typed recorded identity, canonical absolute path spelling, namespace separation and packet self-consistency only. It never opens or stats the coding user's project/configuration/executable paths. The coding user's decision() independently re-observes sources before registry comparison. The simulation denies coder-path opens, body reads and metadata probes throughout operator validation/publication; it then restores the coder's view and verifies changed instructions refuse. The real authority() has direct unit coverage for delegation key set/version, both UID floors/types, distinct UIDs and exact coding-UID match; root custody and account identities remain mocked.

Mixed Claude Code account-state classification

For ~/.claude.json, $CLAUDE_CONFIG_DIR/.claude.json and the selected <config>.json sibling, non-empty context-affecting values refuse. The exact empty defaults admitted at either scope are allowedTools, mcpContextUris, enabledMcpjsonServers, disabledMcpjsonServers, enabledMcpServers, disabledMcpServers, ignorePatterns and exampleFiles as lists, and mcpServers as an object. Wrong types refuse. Acceptance flags hasTrustDialogAccepted, hasTrustDialogHooksAccepted, hasClaudeMdExternalIncludesApproved, hasClaudeMdExternalIncludesWarningShown and bypassPermissionsModeAccepted must be booleans. They are classified as acceptance bookkeeping, not context; changing a valid boolean does not change approval. Their values stay local.

Unknown keys still refuse at the top level and in every project entry. projects must be an object and every entry is classified, not only the one whose key matches the launch folder: Claude Code keys project state by the git root (or folder) it resolves, so an ancestor or differently-cased key can carry context for a launch (9 Nov 2026 review; an earlier selected-entry-only check failed open and was withdrawn). A machine with an old project that still holds non-empty MCP servers, tools or unknown keys is therefore refused until that entry is cleared. Entries are never used as context or exported. Inert account/UI/metrics values remain opaque: neither their text, values, project names nor changing-value digests enter the packet or registry. The fixed inert-claude-state / keys/v1 marker records successful classification, including the boolean type checks. Full planned metadata checks still catch concurrent state-file changes during observation.

Source: structure-only inspection of a real local Claude Code 2.1.319 ~/.claude.json, 9 November 2026: key names, nesting, types and empty/default forms only. No account values or project path keys were printed, logged or stored. The observed key sets are recorded below, separately from the supported allow-lists. bypassPermissionsModeAccepted was requested for compatibility; it was not present in this observation. The official settings files and precedence, settings reference and MCP scopes and enablement remain loader references, not a complete published internal-state schema.

The fictional state fixture mirrors the observed supported project default/metrics structure, with invented account/metric values and another project's active configuration. A controlled Python-child test admits a second launch after native-shaped defaults and acceptance flags are saved. It does not prove native acceptance. The real local file also has active global configuration and unknown caches: those still refuse. Do not delete state or select a clean profile to bypass refusal.

Top-level inert allow-list (INERT_STATE_KEYS):

announcementImpressions, anonymousId, birthdayHatAnimationCount, btwUseCount changelogLastFetched, claudeCodeFirstTokenDate, diffSidebarOpen, effortCalloutDismissed effortCalloutV2Dismissed, feedbackSurveyState, firstStartTime, hasCompletedOnboarding hasIdeOnboardingBeenShown, hasSeenEffortMediumNudgeByModel, hasSeenStashHint, hasSeenTasksHint hasUsedBackslashReturn, hasUsedStash, ideHintShownCount, installMethod lastClawdEntranceVersion, lastOnboardingVersion, lastReleaseNotesSeen, lastShownEmergencyTip machineID, migrationVersion, numStartups, oauthAccount promptQueueUseCount, seenNotifications, shiftEnterKeyBindingInstalled, showSpinnerTree subscriptionNoticeCount, tipLifetimeShownCounts, tipsHistory, tipsHistoryByCommand userID

Per-project inert allow-list (INERT_PROJECT_STATE_KEYS):

lastAPIDuration, lastAPIDurationWithoutRetries, lastCost, lastDuration lastFpsAverage, lastFpsLow1Pct, lastGracefulShutdown, lastLinesAdded lastLinesRemoved, lastModelUsage, lastSessionId, lastSessionMetrics lastSessionModified, lastStartTime, lastToolDuration, lastTotalCacheCreationInputTokens lastTotalCacheReadInputTokens, lastTotalInputTokens, lastTotalOutputTokens, lastTotalWebSearchRequests lastVersionBase

Context-key inventory (CONTEXT_STATE_KEYS); only the exact empty defaults and boolean acceptance flags described above are exceptions:

additionalDirectories, allowedMcpServers, allowedTools, autoConnectIde autoInstallIdeExtension, autoMemoryDirectory, autoMemoryEnabled, autoUpdates autoUpdatesProtectedForNative, chromeExtension, claudeInChromeDefaultEnabled, defaultToAgentsView deniedMcpServers, disabledMcpServers, disabledMcpjsonServers, enableAllProjectMcpServers enabledMcpServers, enabledMcpjsonServers, enabledPlugins, env exampleFiles, exampleFilesGeneratedAt, externalEditorContext, extraKnownMarketplaces hasClaudeMdExternalIncludesApproved, hasClaudeMdExternalIncludesWarningShown, hasTrustDialogAccepted, hasTrustDialogHooksAccepted hooks, ignorePatterns, managedMcpServers, mcpContextUris mcpServers, model, permissions, remoteControlSpawnMode

Settings outside this mixed-state file retain their existing conservative classification: models/limited permissions/disabled auto-memory are reviewed; other settings, hooks, plugins, environment sources, imports, managed preferences, remote caches and unsupported sources refuse. This repair adds no context-source bypass and never reads native credentials to classify them.

Observed top-level key set (names only; presence does not imply support):

additionalModelCostsCache, additionalModelOptionsAnsweredAt, additionalModelOptionsCache, agentPushNotifEnabled, announcementImpressions anonymousId, apiCreditLinkStatusCache, autoCompactWindowsCache, autoModeEnvSetup, autoPermissionsNotificationCount autoUpdates, autoUpdatesProtectedForNative, birthdayHatAnimationCount, btwUseCount, cachedArtifactRoster cachedChromeExtensionInstalled, cachedDynamicConfigs, cachedExperimentData, cachedExperimentFeatures, cachedExtraUsageDisabledReason cachedGrowthBookFeatures, cachedGrowthBookFeaturesAt, cachedStatsigGates, cachedUsageUtilization, changelogLastFetched chromeExtension, claudeAiMcpEverConnected, claudeCodeFirstTokenDate, claudeInChromeDefaultEnabled, clientDataCacheSlots closedIssuesLastChecked, deepLinkTerminal, diffSidebarOpen, effortCalloutDismissed, effortCalloutV2Dismissed fableOverageConsentV2, fallbackAvailableWarningThreshold, feedbackSurveyState, firstStartTime, fullscreenUpsellSeenCount githubRepoPaths, githubWebConnectionStatusCache, groveChoiceSeenAt, groveConfigCache, hasAvailableSubscription hasCompletedClaudeInChromeOnboarding, hasCompletedOnboarding, hasIdeOnboardingBeenShown, hasOpusPlanDefault, hasRemoteEnvironment hasResetAutoModeOptInForDefaultOffer, hasSeenAutoDefaultNudge, hasSeenAutoModeEntryWarning, hasSeenAutoModeOutsideReadPrompt, hasSeenEffortMediumNudgeByModel hasSeenStashHint, hasSeenTasksHint, hasShownOpus45Notice, hasShownOpus46Notice, hasUsedBackslashReturn hasUsedRemoteControl, hasUsedStash, hasVisitedPasses, ideHintShownCount, installMethod isQualifiedForDataSharing, lastClawdEntranceVersion, lastOnboardingVersion, lastPlanModeUse, lastReleaseNotesSeen lastShownEmergencyTip, lspRecommendationIgnoredCount, machineID, mcpNeedsAuthNoticed, mcpServers migrationVersion, modelAccessCache, numStartups, oauthAccount, officialMarketplaceAutoInstallAttempted officialMarketplaceAutoInstalled, opus45MigrationComplete, opus46FeedSeenCount, opusProMigrationComplete, orgModelDefaultCache overageCreditGrantCache, passesEligibilityCache, passesLastSeenRemaining, passesUpsellSeenCount, penguinModeOrgEnabled pluginUsage, pluginUsageLspGraceAppliedIds, projects, promoStartupSeenCount, promptQueueUseCount rcLongTurnNudgeSeenCount, rcLongTurnNudgeSeenKey, recommendedSubscription, remoteControlMachineId, remoteControlReadyPushCount remoteControlReadyPushKey, remoteControlSurfacesSeen, remoteDialogSeen, replBridgePlaceholders, resumeReturnDismissed s1mAccessCache, seenNotifications, shiftEnterKeyBindingInstalled, showSpinnerTree, skillUsage slackAppInstallCount, sonnet1m45MigrationComplete, sonnet45MigrationComplete, sonnet45MigrationTimestamp, subscriptionNoticeCount thinkingMigrationComplete, tipLifetimeShownCounts, tipsHistory, tipsHistoryByCommand, toolUsage transcriptShareDismissed, unpinFable5LaunchEffort, unpinOpus47LaunchEffort, unpinOpus48LaunchEffort, userID voiceFooterHintSeenCount

Observed per-project entry key union (names only; project path keys omitted):

allowedTools, disabledMcpServers, disabledMcpjsonServers, enabledMcpServers, enabledMcpjsonServers exampleFiles, exampleFilesGeneratedAt, hasClaudeMdExternalIncludesApproved, hasClaudeMdExternalIncludesWarningShown, hasTrustDialogAccepted hasTrustDialogHooksAccepted, hasUnseenTeamArtifacts, ignorePatterns, lastAPIDuration, lastAPIDurationWithoutRetries lastCost, lastDuration, lastFpsAverage, lastFpsLow1Pct, lastGracefulShutdown lastHintSessionId, lastLinesAdded, lastLinesRemoved, lastModelUsage, lastSessionFirstPrompt lastSessionId, lastSessionMetrics, lastSessionModified, lastStartTime, lastToolDuration lastTotalCacheCreationInputTokens, lastTotalCacheReadInputTokens, lastTotalInputTokens, lastTotalOutputTokens, lastTotalWebSearchRequests lastVersionBase, loggedAuthoredArtifactPaths, mcpContextUris, mcpServers, reactVulnerabilityCache remoteControlSpawnMode

Native executable custody also checks the opened executable and its opened parent with the existing descriptor-based macOS ACL checker. This adds no immutable-filesystem or later native-consumption claim. Host install and daemon validate root custody before entering startup cleanup. Cleanup requires a PID file and the existing owned-cluster status verification, stops with check=False, and preserves startup errors and signal-handler restoration. A stop requested during PostgreSQL startup returns before spawning the server. Missing registry metadata has a bootstrap-specific refusal; unreadable metadata retains the mode repair diagnostic. Unreachable legacy binding review modes have been removed.

Cost and credential exposure

Observation now makes three full plan checks: after discovery, immediately before source bodies, and after the final binding/source read. The two middle checks were removed; the shared per-path/descriptor identity guards remain active. A run still performs an early metadata plan and two decision() observations: seven full checks instead of eleven. Each check still walks every planned path's ancestors; cost grows with path count and depth, and native executable hashing still occurs per observation. No large-repository timing claim is made. Supported projects are ordinary owned directory trees without descendant symlinks; common node_modules/.bin and .venv/bin links can still refuse the whole project.

The token is stored in Keychain and passed in AGENT_MEMORY_TOKEN in the native child's environment. Its subprocesses inherit it and can print it into plaintext Claude transcripts or send it elsewhere. The installer does not prevent this. The bundled MCP server exposes read tools and, only for read_submit with submissions enabled, pending submission. Its API requires the credential's exact single workspace and supplies no approval/publication capability to that bot. These are source observations, not a new live token/authorization test.

Only software files on disk are protected by delegation. A cooperative coding process performs the interlock; refusing PYTHON* inside Python does not undo startup imports or protect the already-running interpreter.

Open findings after this repair

  • Context 4 residual: parent walks and conservative symlink refusal remain; a larger discovery/performance redesign is outside this bounded repair.
  • Context 6 residual: revoke output gains no new field/message; the existing field is documented as authorization-window certainty to avoid changing the CLI response contract.
  • Context 8 residual: the explicit 0644 registry remains readable by all local users and discloses project metadata, PATH and argv; a coding-user-only ACL needs separately authorized permission/bootstrap work.
  • Context 11: the bounded canonical JSON terminal view remains cumbersome; review presentation needs a separate operator-UX change.
  • Host 3: PostgreSQL support files/dylibs and Python site-package trust need a wider prefix/loader audit; host changes here are limited to requested findings 1, 2 and 5.
  • Host 4: lifecycle test custody mocks remain; changing them is outside the explicitly limited host repair, and does not replace real two-UID acceptance.
  • Native loading/lifecycle, discovery completeness, protected bootstrap, two-account Keychain transfer and host recovery remain untested release gates; the rehearsal is not being run.
  • No new model review was requested or run, per this repair's instruction; the earlier whole-candidate reviews remain the source of these finding IDs.