Token-efficiency profile accepted with the three code-navigation tools; task-to-model routing record (unit A4)

[seathatflowsinourveins]

Scope

• What this PR changes: accepts the token-efficiency adoption profile as the selected practice (2026-09-30) with its 14 pinned components, keeps jcodemunch-mcp, codebase-memory-mcp and ast-grep as the SubagentStart carrier's on-demand lanes (not profile members until each has a reviewed pin on both platforms plus a host receipt: the bootstraps fail closed on an unpinned component), and adds the task-to-model routing record (one table: task class, client, model, effort, enforcement point today; no automatic router; OmniRoute D04/D06/D11 stay held until unit D3's promptfoo A/B; gpt-6.1-sol named only as pending the D4 qualification). Restates the README profile row (accepted; cells describe the 14 pinned components) and the regenerated grand list.
• Base commit: 11227bfd (origin/main at rebase)
• Lane: lane:foundation
• Owned paths touched: adoption/manifest.json (token-efficiency profile label and recipe_paths), adoption/README.md (profile row), docs/decisions/2026-09-30-task-model-routing.md (new), docs/token-practice.md (pointer), tests/test_adoption_status.py, tests/test_task_model_routing.py (new); generated: catalogs/landscape/new-host-grand-list.json, docs/new-host-grand-list.md; manifests/evidence.json (re-registration only, last commit).
• Frozen surfaces (Gate A): none touched; the profile's component set is unchanged, so tools/token-e2e/freeze_snapshot.py captures no new item.
• Round 2 (after the macOS gate failed on the round-1 head): the profile is back to its 14 pinned components; the new test test_every_profile_component_has_a_pin_on_both_platforms enforces it. Token stack winner: one full stack chosen on recorded evidence, provisional until Gate A (decision record) #508 ("Token stack winner", head b7fcc219) agrees for jCodeMunch (owner only once wired) and codebase-memory-mcp (not a member); ast-grep is out only for want of a pin. The record's overturn condition names the follow-up: a reviewed pin in both pin files plus a host receipt per platform, then they join the profile.
• Acceptance covers the selection and its routing only; no host is accepted (each host still runs the coverage check, the pinned-version check and one native call per tool per client).

SOTA sources

• jgravelle/jcodemunch-mcp v1.108.319 @ 8f7b34abe16fb459e0bf1c04747d584216dfe32e (manifests/stack.json)
• DeusData/codebase-memory-mcp v0.11.0 (manifests/stack.json)
• ast-grep/ast-grep 0.45.3 (manifests/stack.json)
• Repository records: catalogs/landscape/foundation.json (code-navigation current choice), adoption/hooks/claude/token-lanes-block.md, docs/decisions/2026-09-23-claude-user-profile.md (jCodeMunch per-project registration addendum), PR Token stack winner: one full stack chosen on recorded evidence, provisional until Gate A (decision record) #508.
• Routing rows: docs/decisions/2026-09-27-model-currency.md, 2026-09-29-sonnet-5-5-dispatch.md, 2026-09-29-max-default-effort.md; PR Preserve OmniRoute feature research and settle October 3 remeasurement promises #423 (OmniRoute routing verdicts, open); Anthropic Sonnet 5.5 system card (https://www.anthropic.com/claude-sonnet-5-5-system-card); Optimizing for cost and intelligence (https://platform.claude.com/docs/en/about-claude/models/optimizing-for-cost-and-intelligence); Sub-agents (https://code.claude.com/docs/en/sub-agents#run-every-subagent-on-one-model); env-vars (https://code.claude.com/docs/en/env-vars).
• Cross-family review (GPT-6 through the OmniRoute gateway) of the first head: one high finding (the profile itself was not delivered) fixed in the repair round; verdict line posted as a PR comment when the gateway review completes.

Evidence-class table

Claim	Evidence class	Command / receipt
The profile resolves with 13 commands and 5 recipes on this host	local_integration	python3 scripts/adoption_status.py --profile token-efficiency --json: exit 0, prerequisites_present
Profile membership (14 pinned), landscape selection, README row, routing record binding	local_integration	python3 -m unittest tests.test_token_stack_cards tests.test_adoption_status tests.test_adoption_docs_consistency tests.test_task_model_routing: 172 tests OK (1 data-dependent skip); tests.test_adoption_bootstrap_macos tests.test_adoption_bootstrap tests.test_adoption_launchd: 311 OK (38 environment skips)
Generated reports current	local_integration	python3 scripts/new_host_grand_list.py --check passed; python3 scripts/build_ecosystem.py --check passed
The record's quotes and the profile-to-record binding	local_integration	tests/test_task_model_routing.py: 8 OK; failing-first against the old manifest
macOS bootstrap refusal, #508 rows, freeze-snapshot behaviour	source_review	adoption/bootstrap-macos.sh:30-31; #508 head lines 85/86/103; tools/token-e2e/freeze_snapshot.py:2240-2300
Profile acceptance	structural validation	artifact consistency, not execution or adoption (docs/acceptance-evidence-policy.md)

No unchanged upstream tests are part of this change. No provider run or model call was made.

Local commands run

$ HOME=<scratch> TMPDIR=/dev/shm/ntci-tmp python3 -m unittest tests.test_token_stack_cards tests.test_adoption_status tests.test_adoption_docs_consistency tests.test_task_model_routing
Ran 172 tests   OK (skipped=1)
$ python3 scripts/new_host_grand_list.py --check
{"status": "passed", "layers": 32, "winners": 66}
$ python3 scripts/build_ecosystem.py --check
{"status": "passed", ...}
$ python3 scripts/validate.py
{"components": 69, "hashed_files": 8515, "profiles": 4, "receipts": 176, "status": "passed"}

Decision record

docs/decisions/2026-09-30-task-model-routing.md

Host evidence

No files under evidence/hosts/ changed.

Checklist

• No GitHub Actions changed.
• No workflows changed.
• No secrets printed, logged or committed.
• No paid hosting, subscription or billing surface.
• Peer-owned untracked files and worktrees preserved.

🤖 Generated with Claude Code

[seathatflowsinourveins]

Round 2 (head replaced under the hot-file protocol; pushed with --force-with-lease):

Reverting the round-1 profile change. CI fact: run 36728291629 (head 8ca7895), job validate-macos, step "Gate on the adoption test modules" failed: bootstrap-macos.sh --profile token-efficiency --plan exits 3 with "No pin in adoption/pins-macos-arm64.json for selected component(s): jcodemunch-mcp codebase-memory-mcp ast-grep" (3 TokenEfficiencyPlanTests; reproduced locally, FAILED (failures=3, skipped=38)).

Both bootstraps fail closed on a selected component with no pin (adoption/bootstrap-linux.sh:144-172, adoption/bootstrap-macos.sh:177-223: no component is exempted from a pin by default any more). Neither pin file lists the three, so the profile is again its 14 pinned components, with the accepted label and the routing record kept. jCodeMunch, codebase-memory-mcp and ast-grep stay the carrier's task-appended lanes and the code-navigation current choice, installed on demand from their recipes; this agrees with #508 lines 86 and 103, and ast-grep (#508 line 85) is out only for want of a pin. The record's overturn condition names the follow-up: a reviewed pin in both pin files plus a host receipt per platform, then they join the profile. New test_every_profile_component_has_a_pin_on_both_platforms catches this in the unit's own test set.

[seathatflowsinourveins]

Cross-family review of the repaired head c8f934c (GPT-6 through the OmniRoute gateway, read-only, diff against the merge-base 11227bf):

high, adoption/manifest.json:383, The profile is marked accepted while still omitting `jcodemunch-mcp`, `codebase-memory-mcp`, and `ast-grep` from both component membership and command checks, contrary to deliverable 1; the new routing test explicitly requires their exclusion. Fix: complete the required membership and client-wiring coverage, coordinate the missing bootstrap pins with their owner, and change the tests to require inclusion before declaring acceptance.
verdict: needs_changes — The routing record is present, but the required profile expansion remains unimplemented.

Coordinator adjudication: the finding restates the unit brief's original deliverable 1 (add the three code-navigation tools to the profile). That deliverable is superseded on evidence: both bootstraps fail closed on any selected component without a pin (adoption/bootstrap-macos.sh:177-187: "no selected component is exempted from a pin by default any more"; adoption/bootstrap-linux.sh:144-172), neither pin file lists the three, a macOS pin cannot be verified from this host, and the round-1 head failed validate-macos on exactly that rule (run 36728291629). The accepted state is therefore the 14 pinned components, with the three tools recorded as the carrier's on-demand lanes and a named follow-up (reviewed pin in both pin files + host receipt per platform, then membership), enforced by test_every_profile_component_has_a_pin_on_both_platforms. The Gate A owner concurred with this reading. No further change for this finding.

[seathatflowsinourveins]

Rebased onto origin/main 8fc8611 (after #546) with the hot-file protocol: main's manifests/evidence.json taken and this unit's files re-registered in the last commit; python3 scripts/validate.py passed; new head d004de1e, tree otherwise unchanged.

🤖 Generated with Claude Code

[seathatflowsinourveins]

Stacked merge train after #542 merged (main 1f2cdce): this branch is rebased onto #545 3805a59 with the hot-file protocol applied against that head (its evidence.json taken, this unit's files re-registered in the last commit; python3 scripts/validate.py passed). New head 69f3e9ed; the PR's own delta is the commits above that base, the earlier commits belong to the PR(s) before it in the train and vanish from this diff as they merge. Merge order: #539 → #545 → #540 → #547 → #557 → #553 (after its repair) → #549 → #541; each with --match-head-commit once its eight required checks pass.

🤖 Generated with Claude Code

[seathatflowsinourveins]

Accepted foundation runtime enhancement handoff from PR551, head 0e86cb7e7798b497a43672dc67328a12d2793964.

The Claude dispatcher now defaults to the enhanced scoped native SDK home and gates model execution on readiness. Actual bounded acceptance completed: Sol/Max read the selected skill, Context Mode counted the unchanged test, Serena returned add, one native Astra/Max judge ran rtk npm test exactly once with exit0, and Dagu completed in about186 seconds with SDK cleanup closed. Parent/child requested routes were independently observed through the selected OmniRoute Responses lane.

Original-field receipt, scoped record, and native kit. The23 owned paths and additive evidence registrations are synced into the shared checkout. Its full validation and17/16-observation scoped checks pass; unrelated evidence rows were preserved. Your active shared skill manifest was preserved and the trial's exact selection snapshot archived.

The initial Dagu environment failure and300-second parent timeout remain recorded. Only selected skill/MCP calls and a manual native graph are qualified; hooks, schedules, optional services, backend identity and complete provider usage/savings retain their own gates. Earlier native Claude callsite evidence stays tied to archived historical source; the separate Claude SDK bridge stays unqualified.

Token/checkpoint handoff: successful native cumulative scopes are Sol parent153,813 and Astra child66,044, each counted once. Cached input/reasoning are subsets; observer completion sums corroborate those counters and must not be added again. All nine successful-run unclassified flow_error records are retained. Failed-parent usage remains a partial native snapshot, and whole-task savings remains claim:none, coverage:partial. Please use bounded acceptance metadata for the dashboard/checkpoint and keep it separate from emitter freshness and process liveness; the owned observer is already stopped.

[seathatflowsinourveins]

Round 2: routing record reconciled with D4 (#542), restacked on F1

Head 901ed767, stacked on F1 #557's head f29f6ba8 (the PR diff shows F1's commits until #557 merges); manifests/evidence.json changes only in the last commit.

Summary. The record and its test predated D4 (merged as 1f2cdce5). CI job 110072600341 (validate, run 36769683903) failed five checks: two quoted model = "gpt-6-astra" lines, and three GPT-6.1 bindings flagged by the test that expected Sol to be routed nowhere. One of those was bytecode under tools/adoption/__pycache__/.

• The record's six Codex rows now follow D4's routing record: Sol/Ultra coordinator (the template's ${CODEX_MODEL}, rendered gpt-6.1-sol for Linux 0.159.2 and gpt-6-astra for macOS 0.155.1), Sol/Max primary workers and generic children, Astra/Ultra complex-workflow coordination, Astra/Max escalation, and the Astra role carriers.
• The Claude rows are unchanged. No automatic router exists, and OmniRoute D04/D06/D11 stay off until the D3 promptfoo A/B.
• The test now requires whole-value quotes, checks that Sol is routed exactly where D4 routes it and nowhere else, renders the Codex template per pinned platform, and skips __pycache__/node_modules.

SOTA sources

• docs/decisions/2026-09-30-sol-primary-quality-defaults.md and the model-currency addendum of 2026-09-30 (docs/decisions/2026-09-27-model-currency.md:375-512), both at 1f2cdce5.
• openai/codex rust-v0.159.1 release notes ("Added GPT-6.1 Sol as the default model in the bundled catalog"), and at rust-v0.159.2: models-manager/models.json, protocol/src/openai_models/reasoning_effort.rs, core/src/agent/child_config.rs, config/src/config_layer_source.rs, as D4 cites them.
• Test conventions reused: tests/test_render_config.py CodexModelTests (in-process render with tomllib) and tests/test_adoption_docs_consistency.py:227 (__pycache__ exclusion).

Evidence

All checks are our integration checks; no upstream test and no model run is involved.

• Failing-first: old test on the old record, 4 subtest failures; new test on the old record, 5 failures and 1 error; new test on the new record, 9 OK.
• Mutants: a literal template model line, the macOS pin at 0.159.2, a planted Sol binding, a new *_MODEL_* constant, and a fourth Sol row are each caught. The CI bytecode file is correctly ignored.
• At the head: routing, token-stack-card, docs-consistency and render-config tests 91 OK (1 skip, also present at the base); build_ecosystem.py --check passed; validate.py passed (hashed_files 8609); the three registry tests OK; privacy scan 0 of 530 added lines.

Known residuals: the record cites the CI run by the superseded head 69f3e9ed (the run is durable); the coordinator row couples to the macOS Codex pin by design (a pin move restates the row); the Claude rows' line numbers are dated to e45328d3.

🤖 Generated with Claude Code

[seathatflowsinourveins]

Cross-family review of the round-2 delta (83f2b3bd..901ed767), GPT-6.1 Sol at max through the OmniRoute gateway, read-only codex exec, 2026-09-30 ~23:15Z: verdict needs_changes, one medium finding: tests/test_task_model_routing.py:166, the "nowhere else" check allows every Sol binding in an already-cited file; an in-memory addition of [profiles.security-review] with model = "gpt-6.1-sol" to the user template leaves all nine routing tests passing despite introducing an unrecorded route. Requested fix: validate individual bindings and TOML sections against D4's allowed routes, and add a regression covering an extra profile in an already-cited file. Round 3 is dispatched to the unit's builder; the Gate A owner's go for the head as it stands (static check) is unaffected by the test-strength repair, which touches only the test module and the record if a row is restated.

[seathatflowsinourveins]

Round 3: GPT-6.1 routing checked binding by binding (GPT-6 review of 901ed76: needs_changes, medium)

Head e3acc20b, still stacked on F1 #557's head f29f6ba8; manifests/evidence.json changes only in the last commit.

Summary. The "nowhere else" check compared files only, so a [profiles.security-review] table with model = "gpt-6.1-sol" added to the Codex user template (a file the coordinator row already cites) left all nine routing tests passing.

• Every GPT-6.1 binding in the routing places is now checked individually: a TOML key by its table (parsed with tomllib, on the ID or on ${CODEX_MODEL}), a key or constant elsewhere, and a command's -m named with its -p profile.
• Each binding must be one a Sol row cites by file and name, and one D4 makes: the template's model and [agents] default_subagent_model, CODEX_MODEL_CURRENT, the stack-worker profile's model, and the stack-worker command's -m.
• A regression test plants the profile in memory: with the literal ID, with the placeholder, and cited by its table. Each variant is reported.
• The stricter check found two stack-worker -m bindings that the primary-worker row did not cite: the profile header (adoption/templates/codex.stack-worker.config.toml:3) and the live proof's help (tools/adoption/prove_codex_lane.py:36). That row now cites both. No other record text changed.

SOTA sources

• docs/decisions/2026-09-30-sol-primary-quality-defaults.md and docs/decisions/2026-09-27-model-currency.md (addendum of 2026-09-30, :414-422 and :484-487), both at 1f2cdce5.
• Python tomllib (standard library, PEP 680) for the structural TOML parse.
• Test conventions reused: tests/test_render_config.py CodexModelTests and tests/test_adoption_docs_consistency.py:227.

Evidence

All checks are our integration checks; no upstream test and no model run is involved.

• Failing-first: the planted profile passes the round-2 routing tests (exit 0) and fails the round-3 tests (exit 1, "profiles.security-review.model binds GPT-6.1 and no Sol row cites it"). The mutant table is M0–M7; M7a and M7b flip from pass to fail.
• At the head: routing, token-stack-card, docs-consistency and render-config tests 92 OK (1 pre-existing skip); routing module 10 OK; validate.py passed; the three registry tests OK; privacy scan 0 of 653 added lines above F1's head.

Residuals: the test now carries a second copy of D4's allowed bindings (D4_BINDINGS) beside SOL_ROUTES, to change only when D4 does; the record does not yet describe the header-scoped citation syntax the test accepts; any new -m gpt-6.1-sol mention under the routing places fails until a Sol row cites it (by design); the test module needs Python 3.11+ for tomllib, as eight sibling modules already do.

🤖 Generated with Claude Code

[seathatflowsinourveins]

Concrete foundation runtime follow-up is published as PR566, head 343f5e579ddb9c775efc6dd3e788bccd46d1fe8b, lane:foundation.

The19-role packet resolves the broader landscape from352 public stars/four maintained awesome lists into task-specific choices. Retain the scoped Codex SDK/Dagu/Sol-Max/Astra-Max lane accepted by the earlier Claude dispatcher handoff. New native OpenHands CLI1.16.0, SDK/tools1.50.1 and DeepAgents0.7.21 recipes are separate from your frozen O1/global default work.

Actual results:108 CLI tests;184 final private SDK tests;361 DeepAgents tests plus1 expected failure. One native OpenHands Sol-Max Responses exact-output request passed. DeepAgents selected skill, one returned specialist and fresh-process SQLite continuation passed after a preserved24-step failure and one32-step saved-context repair. Same-oracle negative controls failed as intended. Exact executed sources, native counters and failed usage remain in the receipt. CLI strict cache isolation and production MCP/Conversation/condenser/extension qualification remain held/separate.

Primary review paths:

• docs/native-runtime-role-resolution-20260930.md
• catalogs/foundation/native-runtime-role-resolution-20260930.json
• evidence/receipts/native-runtime-role-resolution-20260930.json
• evidence/artifacts/native-runtime-role-resolution-20260930/{experiment,completeness-critic,execution-source-bindings}.json

Completed Astra architecture and independent OH setup reviews remain scoped. Native Claude Opus/Max read-only review timed out180s with no verdict; later peers reached account usage limits. Root re-executed the unchanged post-provider oracles and independently reproduced27 unique DeepAgents AI-message records. No completed broader peer acknowledgement is claimed.

For the dashboard owner, the concrete checkpoint payload is:

• gate foundation-native-runtime-capabilities: scoped CLI/SDK install and native task interoperability/continuation passed; CLI isolation held; source-reviewed alternatives conditional.
• worker foundation-persistent-research: selected-skill/single-specialist/saved-state continuation passed within frozen fixture; original failure preserved.
• evidence_ref evidence/receipts/native-runtime-role-resolution-20260930.json.
• progress meaning: recorded acceptance metadata, not current worker/process liveness or emitter freshness. Owned observer stopped after28 requests and port25371 closed.

Please reconcile these rows through your owned shared checkpoint and review the concrete role/default boundaries when the native peer is available. Shared O1, skill manifests, gateway, active parents and trading paths were preserved.