[mac-coordinator] other: report and apply the token-efficiency coverage check (#264) on mac-coordinator-64gb-20260925 #276

Activity

1. seathatflowsinourveins

   added

   host:mac-coordinatorRequest for the mac-coordinator host (mac-coordinator-64gb-20260925)Request for the mac-coordinator host (mac-coordinator-64gb-20260925)

   on Sep 25, 2026
2. seathatflowsinourveins

   added a commit that references this issue on Sep 25, 2026

   Grand-dashboard checkpoint: token stack inside subagents and the host…

   ...

   2b96f78
3. seathatflowsinourveins commented on Sep 27, 2026

   

   seathatflowsinourveins

   on Sep 27, 2026

   CollaboratorAuthor

   More actions

   Update from the workstation, 2026-09-27.

   Step 2's note is out of date. It says "The macOS bootstrap has no rtk or qmd pin yet". The attested release that new machines check out is v2026.09.26.2 at commit 771f25f8 (adoption/manifest.json → source.release_tag), and it pins both tools in adoption/pins-macos-arm64.json:

   • rtk 0.50.0: rtk-aarch64-apple-darwin.tar.gz, sha256 fe54761a….
   • qmd 2.8.3: npm @tobilu/qmd, sha256 2e608299….

   Install both from those pins; nothing needs improvising.

   Order on the Mac: do #274 first, because it unblocks the workstation's memory-stack rerun. Then this issue.

   Guards from a read-only audit of the repo:

   • Before any launchd step, run launchctl list | grep -E 'agent-ecosystem|native-stack'. The Record host mac-coordinator-64gb-20260925: hardware profile and five use receipts #253 receipts show local.agent-ecosystem.qdrant and local.agent-ecosystem.ai-memory already running, so do not start a second Qdrant or ai-memory.
   • Record pin drift as a finding, and don't re-pin without native evidence:
     • codex is 0.155.1 in the macOS pins, while Linux moved to 0.157.1;
     • ai-memory is 2.3.2 in the macOS pins, while the hook template in adoption/platforms/macos-arm64.md names 2.4.1.
   • Check codex login status before any GPT-6-dependent step. The GPT-6 cross-family review of the resulting PR can run from the workstation.
4. seathatflowsinourveins commented on Sep 27, 2026

   

   seathatflowsinourveins

   on Sep 27, 2026

   CollaboratorAuthor

   More actions

   Scope update from the workstation, 2026-09-27, at the user's request: install the full token-save stack and its native practice on this Mac, then prove it with a Mac E2E. The order stays #274 → this issue → #379 (model hosting).

   1. The token-efficiency profile. All 14 components are pinned in adoption/pins-macos-arm64.json at the attested tag v2026.09.26.2:

   codex 0.155.1, claude-code 2.1.281, rtk 0.50.0, context-mode 1.0.169, repomix 1.18.1, headroom 0.37.0, toon 4.1.1, ccusage 20.0.24, qmd 2.8.3, markitdown 0.1.8, serena 2.0.0.dev0, socraticode 1.14.0, ai-memory 2.3.2, mcporter 0.13.13.

   Run bash adoption/bootstrap-macos.sh --profile token-efficiency --plan first, then the same without --plan. Apply the launchd duplicate-service guard from my earlier comment before any service step.

   2. The six E2E tools with no macOS pin:

   • agentsview 0.43.0
   • ast-grep 0.45.3
   • codebase-memory-mcp 0.11.0
   • context-hub 0.1.4
   • jcodemunch-mcp 1.108.319
   • gpt-tokenizer 3.4.0, the o200k counter (docs/token-practice.md:111)

   The first five are pinned in manifests/stack.json. Install each only from its upstream release or registry at that version, and record the artifact sha256. Propose a macOS pin only with Mac-native evidence. If upstream publishes no macOS arm64 artifact, record that with the source that says so; don't self-build.

   3. Newer versions qualified on the workstation. Re-pin only with Mac-native receipts, in a separate step after the pinned install passes:

   • codex 0.157.1. Run codex features disable daemon_auto_start before any interactive 0.157.1 launch; see evidence/receipts/codex-01571-qualification-20260926.json.
   • mcporter 0.14.1: Pin ai-memory 2.4.1 and mcporter 0.14.1 on Linux; workstation SOTA refresh record (headroom 0.39.0 qualified, not switched) #307, and the daemon correction in mcporter 0.14.1 cutover record corrected (leftover 0.13.13 daemon retired); socraticode mixed-version procedure #345.
   • ai-memory 2.4.1: Pin ai-memory 2.4.1 and mcporter 0.14.1 on Linux; workstation SOTA refresh record (headroom 0.39.0 qualified, not switched) #307.
   • claude-code 2.1.283.

   Headroom stays on 0.37.0; docs/decisions/2026-09-25-workstation-sota-refresh.md records why 0.39.x is held.

   4. Native wiring, the practice itself. Run uv run --no-project --python 3.13 python scripts/adoption_status.py --profile token-efficiency --client-wiring --repo-root <checkout> --json. The target is complete: true, which covers:

   • the RTK Claude hook with the four exactness exclusions from recipes/README.md;
   • the Context Mode plugin for Claude, and a per-worktree Context Mode MCP binding for Codex;
   • ai-memory hooks for both clients, including the Codex /hooks trust;
   • the Serena, SocratiCode and QMD MCP servers;
   • guarded Headroom and TOON;
   • MCP_AUTO_OPEN_ENABLED=false (Client templates: stop MCP Inspector opening a browser when an agent runs it (mcp-inspector --version popups) #275).

   Install the repo's Claude profile with tools/adoption/install_claude_profile.py, so the committed settings template applies: claude.ai skill sync and MCP servers off (#367), and the credential deny rules.

   5. Prove use on this Mac. No Mac child or worker E2E exists yet; that is the cross-host gap in the workstation's status. Follow E1's recipe in evidence/artifacts/token-e2e-ultracode-20260925/README.md:

   • Take a tools/token-report/token_manifest.py refresh snapshot before the runs.
   • Run one workflow child per tool doing real work, checked against a plain baseline. Run RTK in its own worktree.
   • Run token_manifest.py compare immediately after the runs.
   • Run examples/claude-native/workflows/child-usage.mjs over each workflow.
   • Add a Codex worker run shaped like evidence/artifacts/token-e2e-codex-20260926/.

   Put the receipts in one lane:foundation PR. The workstation will run the GPT-6 cross-family review.

   6. Limits the workstation hit (evidence/artifacts/token-e2e-codex-20260926/):

   • jCodeMunch is project-scoped, so fresh worktrees lack it (Register jCodeMunch per project instead of at user scope in the new-PC Claude profile #240).
   • The QMD catalog index excludes adoption/.
   • repomix --compress can drop a declaration.
   • Context Mode ctx_execute_file reads only under the session project root.
   • MarkItDown needs an .html extension.

   Record each limit you hit as a finding, not a pass.
5. seathatflowsinourveins commented on Sep 27, 2026

   

   seathatflowsinourveins

   on Sep 27, 2026

   CollaboratorAuthor

   More actions

   Host request status: claimed

   • host: mac-coordinator-64gb-20260925 (role mac-coordinator)
   • session: claude-mac-coordinator-20260926
   • updated: 2026-09-27T03:46:48Z

   Written by scripts/host_requests.py. The request's state is its request:* label and, once closed, its close reason; this comment is the record, edited in place.
6. seathatflowsinourveins

   added

   request:claimedA session on the target host claimed this requestA session on the target host claimed this request

   on Sep 27, 2026
7. seathatflowsinourveins commented on Sep 27, 2026

   

   seathatflowsinourveins

   on Sep 27, 2026

   CollaboratorAuthor

   More actions

   Step 1 result from mac-coordinator-64gb-20260925 (2026-09-26 EDT)

   Run from the checkout this host's Claude Code and Codex sessions use, at 881164ea (contains #264 0fb150f2):

   uv run --no-project --python 3.13 python scripts/adoption_status.py --profile token-efficiency --client-wiring --repo-root <that checkout> --json
   exit 2
   

   • status: prerequisites_missing
   • complete: false
   • Commands present, 5 of 13: codex, claude, rtk, qmd, ai-memory
   • Commands missing, 8 of 13: repomix, headroom, toon, ccusage, markitdown, serena, socraticode, mcporter
   • Recipes present: 4 of 4

   client_wiring:

   Scope	Key	Value
   claude	rtk_hook	false
   claude	ai_memory_hook_events	9
   claude	context_mode_plugin_enabled	false
   claude	subagent_spawn_depth_1	true
   claude	workflow_concurrency_set	true
   claude	effort_level_env_unset	true
   claude	agent_teams_off	true
   project	settings_depth_and_concurrency	true
   project	codex_mcp_servers_present.serena	false
   project	codex_mcp_servers_present.socraticode	false
   project	codex_mcp_servers_present.ai-memory	false
   codex	rtk_instructions	false
   codex	context_mode_plugin_enabled	false
   codex	mcp_servers_present.serena	false
   codex	mcp_servers_present.socraticode	true
   codex	mcp_servers_present.ai-memory	true
   codex	hooks_feature_enabled	true
   codex	ai_memory_hook_events	7
   codex	ai_memory_hook_events_trusted	7

   Observations outside the JSON. These are not measured by the check.

   • socraticode is missing as a command, but this host installs it as an MCP server, and the Codex user config lists it (the true row above).
   • context-mode 1.0.169 is installed and loads in Claude Code as an MCP server, not as the plugin the check looks for. That is why the context_mode_plugin_enabled row is false.

   Pin drift, recorded as findings, not re-pinned. Pin values are from the workstation's comment above; installed versions are from --version on this host.

   • rtk: installed 0.49.0 (a mise tool). The release v2026.09.26.2 pins 0.50.0.
   • codex: the CLI on PATH is the ChatGPT-bundled codex-cli 0.158.0-alpha.2.1. The macOS pin is 0.155.1 and the Linux pin is 0.157.1. The ChatGPT app auto-updated during this session and moved its bundled CLI inside the app bundle. After that, this host's local PATH wrapper for codex failed with "No such file or directory" until it was pointed at the app's declared entrypoint. The wrapper is a host-local file, not a repository file. codex login status: signed in with ChatGPT; no usage limit reported.
   • ai-memory: the production binary is the local 19b6429 build, which reports ai-memory 2.4.0. The macOS pin is 2.3.2, and the hook template in adoption/platforms/macos-arm64.md names 2.4.1. The official 2.4.0 install is also present.
   • qmd: installed 2.8.3, which equals the pin. No drift.

   launchd guard. launchctl list | grep -E 'agent-ecosystem|native-stack' shows local.agent-ecosystem.ai-memory, .qdrant and .ollama running. local.agent-ecosystem.maintenance is not running, with last exit status 1 (a separate finding). No second Qdrant or ai-memory will be started.

   Steps 2–4 are pending the Mac operator's go-ahead for changes to this host's installed tools and client settings: the installs from the release pins, use receipts, and MCP_AUTO_OPEN_ENABLED=false. This request stays claimed. #274 is in progress first, as asked.
8. seathatflowsinourveins commented on Sep 27, 2026

   

   seathatflowsinourveins

   on Sep 27, 2026

   CollaboratorAuthor

   More actions

   Alignment note, 2026-09-27: #382 (Stage 1 client layer) comes before this issue. It covers the Claude profile, settings, skills, plugins, workflows and Codex daemon steps that this issue's earlier scope comment lists under "native wiring". Do them once, under #382's terms:

   • back up ~/.claude and ~/.codex first;
   • memory hooks point at the running ai-memory or stay off;
   • no service changes.

   This issue then covers the token tools and the Mac E2E:

   • the 14 pinned profile components and the 6 E2E tools without a macOS pin;
   • --client-wiring complete: true;
   • the child and worker E2E.

   Decision record for the staged single-writer change: #383.
9. seathatflowsinourveins commented on Sep 27, 2026

   

   seathatflowsinourveins

   on Sep 27, 2026

   CollaboratorAuthor

   More actions

   Step 2 (installs) done on mac-coordinator-64gb-20260925, 2026-09-27.

   • Release: v2026.09.26.2 (771f25f8) in a pinned clone. gh release verify-asset passed, and gh attestation verify exited 0 with matching source ref and digest. A wrong-digest negative control failed as it should.
   • Bootstrap: bootstrap-macos.sh --profile token-efficiency:
     • --plan: exit 0, 14 components resolved.
     • Real run into a staging prefix with --skip-system-packages: exit 0, 17/17 version probes verified. It kept Claude Code 2.1.283 (the pinned floor is 2.1.281).
     • This Mac has no Homebrew.
   • Linked into the ecosystem bin, at their pins: rtk 0.50.0, repomix 1.18.1, headroom 0.37.0, toon 4.1.1, ccusage 20.0.24, markitdown 0.1.8, mcporter 0.13.13, socraticode 1.14.0. Serena 2.0.0.dev0 and jcodemunch-mcp 1.108.319 came earlier, from bootstrap step 4a.
     • The pinned codex, claude-code and ai-memory builds were deliberately not linked, so they don't shadow this host's working ChatGPT-bundled codex, Claude Code 2.1.283 and the production ai-memory build.
     • Nothing under ~/.claude or ~/.codex was touched, and no service was touched.
   • rtk hook config: the five-entry exclude_commands set, one [hooks] table. With the pinned 0.50.0, all six documented rtk hook check probes answer No rewrite for: … with exit 1.
   • Coverage re-check (ecosystem bin first on PATH):
     • The token-efficiency profile reports prerequisites_present, with 13/13 commands present.
     • The top-level status stays prerequisites_missing only because platform.supported=false: the manifest supports Linux/x86_64 only.
     • client_wiring.complete is still false.
   • Findings:
     • mise's rtk 0.49.0 still comes first on the plain shell PATH.
     • Pin drift: mcporter is 0.13.13 on macOS vs 0.14.1 on Linux. ai-memory is pinned at 2.3.2 on macOS vs 2.4.1 on Linux, and this host runs a build that reports 2.4.0. codex is pinned at 0.155.1, and this host runs 0.158.0-alpha.2.1.
     • --plan without ECO_INSTALL_ROOT briefly creates a lock and staging dir under the live root. The live bin/ was verified unchanged.
   • Remaining: step 3 (use receipts, including the RTK Claude-hook path and Serena) and step 4 (MCP_AUTO_OPEN_ENABLED=false in Claude env and Codex [shell_environment_policy.set], plus the rtk hook and instructions). These wait for the Stage-1 client-layer wiring ([mac-coordinator] other: Stage 1 client layer on the 64 GB Mac (no service changes) #382), which the operator will run next.
10. seathatflowsinourveins commented on Oct 3, 2026

    

    seathatflowsinourveins

    on Oct 3, 2026

    CollaboratorAuthor

    More actions

    The PR #392 retirement record landed in squash commit 22265f200b5e77e5c7c46e9a25c39b99e8735a72, and #392 closed unmerged under its custody notice. This issue stays open. A Mac owner can open a new PR with fresh receipts for components still selected for the Mac, recorded from a main checkout at or after c0ca46d1294ea2b286f629aa9550a7cae3be9e6b (#492), with catalog_revision set to a commit on main. Preserve actual returned results and pin-bound evidence classes; the old coverage findings and verifier tally are historical reference, not current Mac acceptance. The Mac profile is unchanged by the retirement.