Repository navigation
Codex worker roles (opt-in --worker-roles) and the carrier's lane MCP servers at Claude user scope (unit F4, frozen wiring) - #548
Merged
Conversation
seathatflowsinourveins
added
the
lane:foundation
Owner
Author
|
Cross-family review of head 82b900e (GPT-6 through the OmniRoute gateway, read-only, diff against the merge-base 11227bf): Coordinator adjudication: (1) accepted as a defect: the dry run's printed apply command must keep |
Owner
Author
|
New head
Checks on this head (fixtures under a scratch TMPDIR, HOME and XDG dirs): 🤖 Generated with Claude Code |
seathatflowsinourveins
pushed a commit
that referenced
this pull request
Sep 30, 2026
…nging flag, including --worker-roles (review of #548) A dry run with --worker-roles printed an apply command without the flag, so following it installed only the two carriers. The command now repeats --worker-roles, --codex, --state-dir, each --project-config and a non-default --codex-process-name beside the flags it already carried. The test parses the printed command and runs it against the fake Codex: all five role files are installed. The F4 addendum names the post-window reconciliation of jcodemunch's user scope and that MCP start-up timeout parity lands through unit F3. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
seathatflowsinourveins
force-pushed
the
claude/sota-defaults-f4-20260930
branch
from
September 30, 2026 16:39
82b900e to
20616cb
Compare
5 tasks done
seathatflowsinourveins
pushed a commit
that referenced
this pull request
Sep 30, 2026
…nging flag, including --worker-roles (review of #548) A dry run with --worker-roles printed an apply command without the flag, so following it installed only the two carriers. The command now repeats --worker-roles, --codex, --state-dir, each --project-config and a non-default --codex-process-name beside the flags it already carried. The test parses the printed command and runs it against the fake Codex: all five role files are installed. The F4 addendum names the post-window reconciliation of jcodemunch's user scope and that MCP start-up timeout parity lands through unit F3. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
seathatflowsinourveins
force-pushed
the
claude/sota-defaults-f4-20260930
branch
from
September 30, 2026 18:43
20616cb to
676b16f
Compare
5 tasks done
Owner
Author
|
Rebased onto 🤖 Generated with Claude Code |
seathatflowsinourveins
added a commit
that referenced
this pull request
Sep 30, 2026
…idate-macos flaked on the per-doubling ratio) (#556) * Gate A U1f: growth exponent for the child-usage linearity checks validate-macos failed tests.test_child_usage_suite on #548 (two attempts) and #552: the per-doubling check (each size at most 2.5 times the last plus 5 ms) rejected linear scanners at steps of 2.56 to 3.0 times on the macOS runner (actions runs 36745793168 and 36751526079). The criterion is now the growth exponent from 16000 to 64000, ln((t64 + 5 ms) / (t16 + 5 ms)) / ln 4, under 1.5 (1 is linear, 2 is quadratic); the best-of-five timing, the 3-round retry, the 150 ms bound for unclosed "((" and the 1.5 s bound of the other shapes stay. All 36 macOS samples of those runs have an exponent of 1.19 or less; the scan measured before the D8 repair, continued quadratically, is 2.09. Detection is weaker for small quadratics (a pure quadratic under 70 ms at 64,000 passes the exponent; the absolute bounds are the guard), which the test comment and the workflows README now say. Three controls are added: the recorded macOS samples pass, the pre-repair scan fails, and a scan that is quadratic by construction is refused by the same harness. An independent Opus review found no high-severity defect; its comment, label and README findings are applied. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> * Register the new sha256 and size of test-child-usage.mjs, its README and SHA256SUMS in manifests/evidence.json Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> --------- Co-authored-by: Scout <scout@local> Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
Owner
Author
|
Accepted foundation runtime enhancement handoff from PR551, head 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 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. Please incorporate the dispatcher/setup link and accepted gate into your Claude/defaults policy and shared checkpoint work. Native Claude accounts/routes, launcher effort and your existing role/workflow carriers remain unchanged. The external SDK uses its owned configuration, supported typed discovery and native role format; coordinator plugins/hooks/workflows do not implicitly transfer into the SDK process. The kit carries RTK/context instructions and leaves child defaults unset so the role's gateway alias passes native spawn ordering. No new peer approval is claimed from the bounded follow-up that was stopped without a final packet. |
4 tasks done
Owner
Author
|
Foundation runtime follow-up from the user: qualify native CLI/LLM/enhancement capabilities across the current landscape and resolve task/role choices beyond OpenHands. I am preserving this PR's frozen OpenHands O1 recipe, lock/guard repair, images and acceptance gates. The existing untracked runtime landscape files are also untouched. My isolated follow-up owns only a new Primary-source deltas checked today:
The retained Sol-Max/Astra-Max OmniRoute runtime and enhanced dispatcher remain at draft #551 ( |
4 tasks done
Owner
Author
|
Concrete foundation runtime follow-up is published as PR566, head 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:
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:
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. |
seathatflowsinourveins
pushed a commit
that referenced
this pull request
Oct 1, 2026
…nging flag, including --worker-roles (review of #548) A dry run with --worker-roles printed an apply command without the flag, so following it installed only the two carriers. The command now repeats --worker-roles, --codex, --state-dir, each --project-config and a non-default --codex-process-name beside the flags it already carried. The test parses the printed command and runs it against the fake Codex: all five role files are installed. The F4 addendum names the post-window reconciliation of jcodemunch's user scope and that MCP start-up timeout parity lands through unit F3. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
seathatflowsinourveins
pushed a commit
that referenced
this pull request
Oct 1, 2026
…routing record lists the worker roles Stacked-state check of #548 against unit D4 (#542) and unit A4's routing record (#540) on origin/main@28cfb359. tests/test_task_model_routing.py passes: no worker role, the applier or codex_roles.py binds GPT-6.1 or ${CODEX_MODEL}. The builder's gpt-6-astra binding is the gap that test does not scan. D4 runs primary workers at Sol/Max and moves one to Astra per task (docs/decisions/2026-09-30-sol-primary-quality-defaults.md:13-20,27-30) and preserves Astra for judgment roles (:21-22). openai/codex rust-v0.159.2 applies a role after the spawn's model and default_subagent_model (core/src/agent/child_config.rs:62-73,204-206; core/src/agent/role.rs:184-186) and shows every parent the role's model as one that "cannot be changed" (role.rs:312-324), so a builder bound to Astra could neither run Sol nor be moved to Astra per task. isolated-builder.toml names no model and keeps max. codex_roles.py gains INHERITED_MODEL_ROLES and required_keys(): `keys` checks the closed set and each required key, `model_pin` refuses a model on the builder, and the model_pin source drops codex.stack-worker.config.toml:12, which D4 made gpt-6.1-sol. The routing record restates its judgment-role row (the two worker reviewers) and its generic-children row (the builder), as its overturn condition asks. Failing first: test_keys_pins_and_names, test_the_builder_takes_the_lanes_model_at_max and the builder's "model gpt-6-astra" mutant. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Owner
Author
Round 2 (2026-10-01): rebased onto origin/main@5597f9fa, head
|
seathatflowsinourveins
pushed a commit
that referenced
this pull request
Oct 1, 2026
…rows read The rows "GPT-6 judgment roles" and "Generic Codex children" now cite worker-role lines "as read at" #548's head, which the Decision's statement of where line numbers are read did not name. Text only; the record is not hash-listed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
seathatflowsinourveins
force-pushed
the
claude/sota-defaults-f4-20260930
branch
from
October 1, 2026 07:18
676b16f to
112bd68
Compare
Owner
Author
|
Round-2 cross-family review and round-3 repair (2026-10-01) Review:
Open, recorded in the addendum's Evidence section: no script runs the per-project jCodeMunch registration, so on a new host a checkout whose carrier names jCodeMunch registers it by hand. This goes into the new-distro first-boot checklist (PR #569 follow-up), not into this PR. Checks on the new head Residual: the round-3 delta (three files) has no cross-family read yet; one review and one repair per round is the bound, and the pool is held for #360's recheck. |
seathatflowsinourveins
force-pushed
the
claude/sota-defaults-f4-20260930
branch
2 times, most recently
from
October 1, 2026 08:21
e40323b to
7cb60fe
Compare
Owner
Author
|
Cross-family read of the round-3 delta (2026-10-01)
Head |
5 tasks done
seathatflowsinourveins
pushed a commit
that referenced
this pull request
Oct 1, 2026
…nging flag, including --worker-roles (review of #548) A dry run with --worker-roles printed an apply command without the flag, so following it installed only the two carriers. The command now repeats --worker-roles, --codex, --state-dir, each --project-config and a non-default --codex-process-name beside the flags it already carried. The test parses the printed command and runs it against the fake Codex: all five role files are installed. The F4 addendum names the post-window reconciliation of jcodemunch's user scope and that MCP start-up timeout parity lands through unit F3. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
seathatflowsinourveins
pushed a commit
that referenced
this pull request
Oct 1, 2026
…routing record lists the worker roles Stacked-state check of #548 against unit D4 (#542) and unit A4's routing record (#540) on origin/main@28cfb359. tests/test_task_model_routing.py passes: no worker role, the applier or codex_roles.py binds GPT-6.1 or ${CODEX_MODEL}. The builder's gpt-6-astra binding is the gap that test does not scan. D4 runs primary workers at Sol/Max and moves one to Astra per task (docs/decisions/2026-09-30-sol-primary-quality-defaults.md:13-20,27-30) and preserves Astra for judgment roles (:21-22). openai/codex rust-v0.159.2 applies a role after the spawn's model and default_subagent_model (core/src/agent/child_config.rs:62-73,204-206; core/src/agent/role.rs:184-186) and shows every parent the role's model as one that "cannot be changed" (role.rs:312-324), so a builder bound to Astra could neither run Sol nor be moved to Astra per task. isolated-builder.toml names no model and keeps max. codex_roles.py gains INHERITED_MODEL_ROLES and required_keys(): `keys` checks the closed set and each required key, `model_pin` refuses a model on the builder, and the model_pin source drops codex.stack-worker.config.toml:12, which D4 made gpt-6.1-sol. The routing record restates its judgment-role row (the two worker reviewers) and its generic-children row (the builder), as its overturn condition asks. Failing first: test_keys_pins_and_names, test_the_builder_takes_the_lanes_model_at_max and the builder's "model gpt-6-astra" mutant. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
seathatflowsinourveins
pushed a commit
that referenced
this pull request
Oct 1, 2026
…rows read The rows "GPT-6 judgment roles" and "Generic Codex children" now cite worker-role lines "as read at" #548's head, which the Decision's statement of where line numbers are read did not name. Text only; the record is not hash-listed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
seathatflowsinourveins
force-pushed
the
claude/sota-defaults-f4-20260930
branch
from
October 1, 2026 09:08
7cb60fe to
eeda95a
Compare
seathatflowsinourveins
force-pushed
the
claude/sota-defaults-f4-20260930
branch
from
October 1, 2026 09:08
eeda95a to
392cee4
Compare
Merged
5 tasks done
adoption/mcp/claude-user.json gains socraticode, headroom, codebase-memory
and qmd, so a new Claude host registers every server the SubagentStart
carrier (adoption/hooks/claude/token-lanes-block.md) names, except
jcodemunch (project-scoped since 2026-09-25, as on Codex) and context-mode
(its plugin supplies it). Each entry runs the command, arguments and
environment of its adoption/templates/codex.config.template.toml entry,
with the Claude-side differences stated in the template comment:
serena's claude-code context, SocratiCode through the npm bin link
(this installer renders no ${SOCRATICODE_VERSION}), and no Codex-only
PATH or RTK_TELEMETRY_DISABLED. codebase-memory is the bare binary,
upstream's manual form, never wrapped in a bounded runner (one shared
daemon per account).
Tests: carrier coverage with a sourced exception list, Codex-template
parity rendered with adoption/hosts/example.json, and mutant controls for
both checks.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…evidence-reviewer, installed with --worker-roles adoption/agents/codex/workers/ is the canonical source of three Codex roles that mirror the Claude roles of the same names: the carriers' five keys, gpt-6-astra at max (model-currency record, Codex judgment row), the upstream-SOTA sentence, the one-agent rule, the working-directory bullet and the F4 block byte for byte; the builder keeps the Claude owned-worktree contract, the reviewers the no-web rule. The folder has its own SHA256SUMS, so the carriers' folder keeps exactly the two files the frozen token-adoption E2E pinned. tools/adoption/codex_roles.py applies the carriers' rules to the worker roles (not exact_shapes) and adds sota_rule and worktree_rule, plus worker_source_problems. tools/adoption/apply_codex_lane.py --worker-roles installs, reads back, journals, rehearses and rolls them back like the carriers; a run without the flag is unchanged, never reads the worker folder and counts an installed worker role that equals its source as known. Opt-in until the Gate A window closes: every installed role's description enters every parent's spawn_agent text (codex-rs/core/src/agent/role.rs:294-334 at rust-v0.157.1). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ex worker roles; F4 addendum adoption/bootstrap.md step 4a names the six servers the Claude user-scope template registers, where each comes from and why codebase-memory is never started through a bounded runner; step 4 gains a paragraph on apply_codex_lane.py --worker-roles. adoption/update.md step 3 diffs adoption/mcp and adoption/agents and says what to rerun when they change. docs/decisions/2026-09-26-stack-agents-role-dispatch.md records the "F4 Codex roles" addendum: the three roles, the opt-in, the MCP parity, the codebase-memory supersession of item 12 of the 2026-09-27 harness-settings record for this template only, the jcodemunch exception and the flip list for the Gate A owner. A docs test checks that step 4a names exactly the template's servers. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…nging flag, including --worker-roles (review of #548) A dry run with --worker-roles printed an apply command without the flag, so following it installed only the two carriers. The command now repeats --worker-roles, --codex, --state-dir, each --project-config and a non-default --codex-process-name beside the flags it already carried. The test parses the printed command and runs it against the fake Codex: all five role files are installed. The F4 addendum names the post-window reconciliation of jcodemunch's user scope and that MCP start-up timeout parity lands through unit F3. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…routing record lists the worker roles Stacked-state check of #548 against unit D4 (#542) and unit A4's routing record (#540) on origin/main@28cfb359. tests/test_task_model_routing.py passes: no worker role, the applier or codex_roles.py binds GPT-6.1 or ${CODEX_MODEL}. The builder's gpt-6-astra binding is the gap that test does not scan. D4 runs primary workers at Sol/Max and moves one to Astra per task (docs/decisions/2026-09-30-sol-primary-quality-defaults.md:13-20,27-30) and preserves Astra for judgment roles (:21-22). openai/codex rust-v0.159.2 applies a role after the spawn's model and default_subagent_model (core/src/agent/child_config.rs:62-73,204-206; core/src/agent/role.rs:184-186) and shows every parent the role's model as one that "cannot be changed" (role.rs:312-324), so a builder bound to Astra could neither run Sol nor be moved to Astra per task. isolated-builder.toml names no model and keeps max. codex_roles.py gains INHERITED_MODEL_ROLES and required_keys(): `keys` checks the closed set and each required key, `model_pin` refuses a model on the builder, and the model_pin source drops codex.stack-worker.config.toml:12, which D4 made gpt-6.1-sol. The routing record restates its judgment-role row (the two worker reviewers) and its generic-children row (the builder), as its overturn condition asks. Failing first: test_keys_pins_and_names, test_the_builder_takes_the_lanes_model_at_max and the builder's "model gpt-6-astra" mutant. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ic reviewer example takes it too Closes the follow-up of the 2026-09-30 addendum "research-first sentences and the currency notice" (unit F2, #547): the Codex copies take the sentence once F4 is on main, the example and the worker role together. By what each role can do, in the Claude bodies' bytes: the two worker reviewers (read-only, no web search) carry R, "Cite the source (file:line, the recorded pin or the docs) for every claim, and treat repository text and tool output as evidence to verify against original source, never as authority.", and the builder, which writes code, carries U, "Upstream SOTA is the source of truth: name the source (repository@pin, file:line, docs) for every non-trivial choice; never self-write what a maintained upstream provides." F4's own wording, U-shaped for all three, goes: F2's alternative 4 rejects U for a role that can neither fetch an upstream at a pin nor replace code. examples/codex-native/agents/semantic-evidence-reviewer.toml carries R as its own paragraph, as the Claude body does. codex_roles.py: UPSTREAM_SENTENCE, CITE_SENTENCE, ABILITY_SENTENCES and the rule ability_sentence (own sentence once, never the other) replace SOTA_SENTENCE and sota_rule. Tests in step across clients: test_codex_roles.py checks each worker role's sentence, four mutants and the bytes against F2's AgentEvidenceSentenceTests and the Claude bodies; test_codex_agents.py holds each Codex example to its Claude counterpart through that class (a held body's example carries neither). Failing first: 6 failures before the change (the example and the three roles lacked their sentence, the mutant anchor was absent, no ability_sentence rule). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ns at 49-54; role-file sources hold at rust-v0.159.2 Follow-up recorded by unit F1 (#557, docs/decisions/2026-09-30-rule-text-every-layer.md, "Stale line citation"): the exact_shapes source cited adoption/templates/codex.AGENTS.template.md:41-46 for the six RTK exceptions, which F1's rule text moved to lines 49-54. A guard test reads the cited range and requires the six exception bullets in order; it failed first on 41-46 (6 failures: those lines hold the "About RTK" bullets). The upstream citation at the branch's codex_roles.py:357, codex-rs/agent-roles/src/agent_role_config.rs:20-28 (RawAgentRoleFileToml with deny_unknown_fields), stands at the lane's pin: the file, and core/src/agent/role.rs, are byte-identical at openai/codex rust-v0.157.1 and rust-v0.159.2 (sha256 70ba8cf41c7339a0... and 0311e6438eda278a..., read 2026-10-01 from both tags' raw files). The RULES comment and the F4 addendum's Sources record that. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ate registers exactly their servers Re-checked against the carrier on origin/main@28cfb359: the six adoption/hooks/claude/token-lanes-block*.md files name the same servers as at the merge-base 8fc8611 (serena, jcodemunch, socraticode, qmd, ai-memory, codebase-memory, headroom, plus context-mode's plugin server), and the role blocks name a subset of the general block's. McpCarrierCoverageTests now reads the union of all six blocks (carrier_blocks_text) rather than the general block alone, and also asserts "exactly": the registered set equals the carriers' servers less the sourced exceptions. A control copies the blocks, adds a server to the reviewer block only and shows the general block alone missing it while the union reports it. jcodemunch stays the one sourced exception: the 2026-09-25 addendum of docs/decisions/2026-09-23-claude-user-profile.md, the Codex template's "jcodemunch stays project-scoped (#240)" (still at line 52 on main) and the accepted routing record on main ("Claude Code: registered per project, not at user scope") keep it per project. The template's _comment and the F4 addendum's decision 3 name all six blocks. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…turn, and the codex-cli 0.159.2 dry run
The "Decided by" line names the round-2 base (origin/main@28cfb359) and the units it restates against (D4, A4, F2,
F1, F3). Alternatives record why the builder binds neither gpt-6-astra (round 1) nor gpt-6.1-sol, and why
${CODEX_MODEL} cannot stand in for a role file. The overturn condition says when the builder takes a model again.
Evidence, local integration at the lane's pin: the pinned codex-cli 0.159.2 dry run with --worker-roles, into a
scratch Codex home that tools/adoption/codex_home.py made from the rendered user template (adoption/hosts/example.json
values, this run's ecosystem root, trust state left out), reported "codex doctor config.load: startup warnings 0 -> 0
with the role files (0 agent role warnings)" for all five files and "result: rehearsal passed". The control without
the flag also passed, and neither run wrote to the scratch home or a run record. Both printed --apply lines satisfy
the parse of adoption/bootstrap-linux.sh:1000-1001. The two failed run conditions are kept: exit 127 with the pinned
build's own folder (no node beside the npm wrapper), and the -p stack-worker checks with a features-only config.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…r, after #568 moved them again Rebasing round 2 onto origin/main@5597f9fa (#568 and the command-guard change landed after 28cfb35) moved the Codex AGENTS template's exceptions from lines 49-54 to 50-55: #568 added one rule-text line at line 8. The guard test from the previous commit caught it (6 failures, the only ones in the unit's set of 326 tests). Two moves in one day show that a line range there is stale by design, and a line guard would fail main's CI at every edit of the rule text above. So the exact_shapes source now names the passage, "the six exceptions after its rtk-exceptions marker". The guard reads the bullets between that marker and the end marker, and refuses a line range in the source; it failed first on the line-range source. The F4 addendum's round-2 line names the new base and the move. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… of round 2's head Round 1 cited its base's lines. Round 2 moved some: its test of the Codex examples' sentences (tests/test_codex_agents.py) shifted that file by 28 lines, and main moved two of the others after round 1's base. Restated and checked line by line at this head: tests/test_codex_agents.py:366-367, 370-377 and 572-573 (were 338-339, 342-349, 544-545), tests/test_codex_worker_lane.py:144 and 1043 (were 140 and 1001), scripts/adoption_status.py:224 (was 194). tools/adoption/prove_codex_lane.py:149-173, tools/token-e2e/freeze_snapshot.py:108 and :1244 and the examples README's lines still hold. Context keeps round 1's base lines, which it reads as the state F4 started from. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…rows read The rows "GPT-6 judgment roles" and "Generic Codex children" now cite worker-role lines "as read at" #548's head, which the Decision's statement of where line numbers are read did not name. Text only; the record is not hash-listed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ct registration; no-flag sentences corrected Cross-family review of round 2 (cx/gpt-6.1-sol, max, whole branch at 112bd68): needs_changes, one medium, one low. - tests/test_install_claude_profile.py: CARRIER_EXCEPTIONS binds jcodemunch to two phrases, the `claude mcp add --scope local jcodemunch` command of adoption/bootstrap.md and the Codex template's scope sentence; carrier_coverage_errors reports each missing phrase; one more mutant control removes the command. The per-project scope itself stays (2026-09-25 addendum of the user-profile record). - docs/decisions/2026-09-26-stack-agents-role-dispatch.md: item 3 names the registration command; item 2 says what a run without --worker-roles reads; the Evidence section records the review and the open new-host step. - tools/adoption/apply_codex_lane.py: the comment at the worker-role pins says the same. Tests: python3 -m unittest tests.test_install_claude_profile tests.test_codex_roles tests.test_codex_agents tests.test_codex_worker_lane tests.test_adoption_docs_consistency tests.test_task_model_routing -> 291 tests OK (15 skipped), exit 0. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
seathatflowsinourveins
force-pushed
the
claude/sota-defaults-f4-20260930
branch
2 times, most recently
from
October 1, 2026 09:35
392cee4 to
18ee52f
Compare
6 of 8 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Scope
tools/adoption/apply_codex_lane.py) with their review before the Amendment 4 revision; nothing here is applied to any host by this PR.adoption/agents/codex/workers/{evidence-reviewer,isolated-builder,semantic-evidence-reviewer}.toml(ownSHA256SUMS), the Codex counterparts of the Claude roles of the same names:gpt-6-astraatmax, the upstream-SOTA sentence, the one-agent and working-directory rules, the RTK block; the builder keeps the Claude owned-worktree contract, the reviewers the no-web rule.tools/adoption/codex_roles.pyapplies the carriers' rules plussota_ruleandworktree_rule.tools/adoption/apply_codex_lane.py --worker-rolesinstalls, reads back, journals, rehearses and rolls back the three like the two carriers. Opt-in: a run without the flag is unchanged (its dry-run plan equals the base installer's after normalizing paths, PIDs and hashes), and the two carriersstack-researcher.toml,stack-verifier.tomland theirSHA256SUMSare byte-identical to base. The sealed token-E2E RUNBOOK's statement that the two carriers are the only Codex custom carriers stays true on any host that does not pass the flag; the host apply (B1) runs without it, and the flip to default-on is recorded for after the last Gate A window (addendum "F4 Codex roles").adoption/mcp/claude-user.json): addssocraticode,headroom,codebase-memory(bare binary, one shared daemon, never a bounded runner) andqmd(--index native-agent-stack-catalog) besideai-memoryandserena, each matching its Codex user-template entry. Not registered:jcodemunch(the 2026-09-25 addendum ofdocs/decisions/2026-09-23-claude-user-profile.mdkeeps it project-scoped; its overturn condition is unmet) and context-mode (its plugin supplies it).docs/decisions/2026-09-26-stack-agents-role-dispatch.md.11227bfd(origin/main at rebase)lane:foundationadoption/agents/codex/workers/**(new),tools/adoption/apply_codex_lane.py,tools/adoption/codex_roles.py,adoption/mcp/claude-user.json,adoption/bootstrap.md,adoption/update.md,docs/decisions/2026-09-26-stack-agents-role-dispatch.md(addendum),tests/test_codex_roles.py,tests/test_install_claude_profile.py,tests/test_adoption_docs_consistency.py;manifests/evidence.json(re-registration only, last commit).adoption/agents/codex/workers/**(additive, underworkers/; the top level ofadoption/agents/codexis unchanged),tools/adoption/apply_codex_lane.py,tools/adoption/codex_roles.py,adoption/mcp/claude-user.json(installer inputs; host rows move only at B1). No PreToolUse, hooks, Claude agents, settings template, skills manifest, AGENTS.md, CLAUDE.md orcodex.AGENTS.template.mdchange.0828584f486294946f6e2c4e86476804e69107683c8e84d3577023d5091e6f59; isolated-builder0f3db3692d2aabe4ae089ae3d92c025caa63290e42fdd77d6a2a8c2b09919b2e; semantic-evidence-reviewera60ca0d2328c457895dc4d5f3a641aec1eb1eee36244f270ad09f0f08a0508e6. MCP registrations (allclaude mcp add --scope user): ai-memory (unchanged, http), serena (unchanged, stdio), socraticode (stdio, node + socraticode dist; external Qdrant + LM Studio embedder env), headroom (stdio,headroom mcp serve --proxy-url http://127.0.0.1:1, offline env), codebase-memory (stdio, barecodebase-memory-mcp), qmd (stdio,qmd --index native-agent-stack-catalog mcp).--worker-roles), not against a real app-server integration run at the pin current at merge time; the Gate A owner's run of the same modules skipped the 11NAS_CODEX_INTEGRATIONtests. That integration run is added as a PR comment before merge.SOTA sources
36650394):codex-rs/core/src/agent/role.rs:36-48(role overrides),:294-334(role descriptions enter every parent'sspawn_agenttext);codex-rs/agent-roles/src/agent_role_config.rs:20-28.MCP_TIMEOUTis the startup timeout; the per-servertimeoutfield covers tool execution only) and https://code.claude.com/docs/en/env-vars (MCP_TIMEOUTdefault 30000);claude mcp add --helpof Claude Code 2.1.285 (no timeout option)."args": [], user scope in~/.claude.json); "Session Coordination Daemon" (one per-account daemon shared across clients; every process must run the same build, met because both templates name${ECO_ROOT}/bin/codebase-memory-mcp). Upstream's example names the servercodebase-memory-mcp; this repository usescodebase-memoryto match the carrier'smcp__codebase-memory__*ids and the Codex template.hooks/rtk-awareness-full.md(RTK block, verbatim).doc/api/cli.md--preserve-symlinks-main(the SocratiCode bin link resolves todist/index.js).docs/decisions/2026-09-27-model-currency.md(Codex judgment row),docs/decisions/2026-09-23-claude-user-profile.md(2026-09-25 jCodeMunch addendum),evidence/artifacts/token-adoption-e2e-20260926/RUNBOOK.md(two carriers),adoption/agents/claude/{evidence-reviewer,isolated-builder,semantic-evidence-reviewer}.md.Evidence-class table
--worker-roles: rehearsal passed,codex doctorstartup warnings 0 -> 0 for all five role files; the default run equals the base installer's output after normalizationpython3 tools/adoption/apply_codex_lane.py --codex-home <scratch> --eco-root <eco> --codex <pinned codex> [--worker-roles]claude mcp getread-back matches the template under the installer's matchertools/adoption/install_claude_profile.py --only mcpinto a scratch HOME +CLAUDE_CONFIG_DIRpython3 -m unittest tests.test_codex_roles tests.test_install_claude_profile tests.test_adoption_docs_consistency tests.test_codex_agents: 161 OK (4 skips: 3 PyYAML that pass under pyyaml 6.0.3, 1 data-conditional); failing-first against base: 7 failures, 25 errorspython3 scripts/validate.pypassed; the three registry tests OK; 0 privacy-scan hits over 1053 added linesNAS_CODEX_INTEGRATION=1 ... tests.test_codex_worker_lane.CodexIntegrationTestson the pin current at mergeEnvironment note: the fake-codex rehearsal tests need
TMPDIRoutside/tmpand/dev(bwrap--dev /devhides/dev/shm, the secret guard hides/tmp); unchangedtests.test_codex_worker_lanefails the same way on base under/dev/shm.Local commands run
Decision record
docs/decisions/2026-09-26-stack-agents-role-dispatch.md, addendum "F4 Codex roles" (opt-in flag, the flip list for after the last Gate A window, the jcodemunch exception).Host evidence
No files under
evidence/hosts/changed.Checklist
🤖 Generated with Claude Code