Harness rules: vendor-org upstream first (never rebuild), and a lean, budgeted startup context for Claude and Codex

[seathatflowsinourveins]

Scope

• What this PR changes. Two things, across Claude and Codex on both WSLs and in the trading lane.
  • It adds the user's 2026-10-05T04:26:41Z rule to the harness rules: prefer the maintainer's own organization repositories and their clean releases, and never rebuild, fork or wrap what upstream ships.
  • It trims the always-loaded startup context. Client- and task-specific passages move verbatim behind pointers, and a fixed per-client byte budget replaces the upward word ratchet.
• Base commit: f946c6d4c
• Lane: lane:foundation. It includes one clause in blueprints/us-equities/AGENTS.md and moves the trading north-star paragraph there verbatim. Trading custody (5f) acknowledges.
• Owned paths touched: AGENTS.md, examples/claude-native/CLAUDE.md, adoption/templates/codex.AGENTS.template.md, adoption/scaffold/AGENTS.md, the regenerated new-WSL carriers and handbook, blueprints/us-equities/AGENTS.md, docs/decisions/2026-10-05-official-upstream-never-rebuild.md, docs/decisions/2026-10-05-harness-context-budget.md, tests, catalogs/foundation/upstream-surface-dispositions.json, and manifests/evidence.json (last commit).

Measured startup bytes (rendered; record has details):

Surface	Before	After
AGENTS.md	13,986	10,959
examples/claude-native/CLAUDE.md	13,776	11,575
Codex template	8,190	7,307 (limit < 8,192; rtk's verbatim awareness text moved to adoption/templates/rtk-awareness-full.md, 1,121 B, appended by the carrier)
Claude rendered startup scope	28,794	23,566 (ceiling 24,458)
Codex rendered startup scope	22,176	19,332 (ceiling 20,103)

SOTA sources

• Claude Code memory and skills documentation:
  • https://code.claude.com/docs/en/memory: CLAUDE.md under 200 lines; MEMORY.md loads its first 200 lines or 25 KB; imports load at launch;
  • https://code.claude.com/docs/en/skills: the skill listing budget is 1% of context by default, and skillOverrides has a name-only state.
• Codex AGENTS.md discovery and size limit: https://developers.openai.com/codex/guides/agents-md, and openai/codex rust-v0.159.3 codex-rs/core/src/agents_md.rs. project_doc_max_bytes is 32 KiB, and @ references are not expanded.
• rtk's own Claude layout: rtk-ai/rtk v0.51.0 native Claude initializer (RTK.md plus the @RTK.md import).
• Vendor organizations cited in the rule: alpacahq (alpaca-py v0.44.0, alpaca-mcp-server v2.3.2), nautechsystems (NautilusTrader 1.231.0, 2.0.0rc6), and the official IBKR TWS API (1050.02 / 1051.01).

Evidence-class table

Claim	Evidence class	Command / receipt
Rendered byte budgets and moves	local_integration	tests.test_install_claude_profile PortableTopRuleTests (fixed ceilings: Claude 23,062 B, Codex 19,102 B)
Template ≤ 8,192 B	local_integration	wc -c adoption/templates/codex.AGENTS.template.md = 8,091
Carriers and handbook current	local_integration	new_wsl_client_config.py --write-blocks/--check, build_new_wsl_handbook.py --write
The 11 moves are verbatim, with pointers kept	source_review	the record's move table

Local commands run

$ python3 -m unittest tests.test_install_claude_profile tests.test_codex_worker_lane tests.test_scaffold_repo tests.test_new_wsl_client_config tests.test_new_wsl_handbook tests.test_managed_block tests.test_upstream_surface_watch
645 tests: 623 passed, 22 skipped (exit 0)
$ python3 scripts/validate.py   (after the registry commit)
exit 0
$ python3 scripts/validate_convergence.py --all-recorded
exit 0

Decision record

• docs/decisions/2026-10-05-official-upstream-never-rebuild.md
• docs/decisions/2026-10-05-harness-context-budget.md: the budget rule and its overturn conditions. The upstream /doctor prompt-audit has no machine-readable CLI form on 2.1.289, which is why a glue test is needed.

Checklist

• No workflow or Actions change.
• No secrets.
• No paid surface.
• Peer-owned files preserved.

Review chain (bounded): GPT build (jobs 072 and 075), then four rounds:

Round	Result
Claude read	1 P1, 6 P2, 1 P3
GPT repair	all fixed; co-op sentences S1-S3 adopted
Claude re-check	2 P1, 1 P2, 1 P3
GPT repair 2	all fixed

The coordinator verified repair 2 at ee37894:

• the CI contract suites: node test-envelope.mjs 254/0 and node test-contract-mutations.mjs 74/0;
• the budget tests with a symlinked TMPDIR;
• 7 test modules, exit 0;
• handbook check, validate and validate_convergence, exit 0;
• no login-name or host-path hits.

Repair 3 (after CI run 37283231657 found 3 failures): the name-only skill listing is reverted, because the user's 2026-09-30 directive keeps every eligible skill fully listed ("make sure all the skills can invoke seamlessly with llm native end, rather than user end"). 2604's skillListingBudgetFraction is kept, and rtk v0.51.0's awareness text is carried verbatim. Verified at 4f490d3: 1,076 tests OK (29 skipped) across the 10 affected modules, the contract suites 254/0 and 74/0, handbook and client-config checks 0, validate 0, convergence 0.

Recorded gate after landing: the coordinator re-renders the managed blocks on both hosts with F9. After that host gate, the temporary root routing (680 B) is removed and the ceilings re-tighten, per the budget record.

Recorded residuals (bounded review, 2026-10-05)

Claude Opus 5.5 gate read (session 5f) at cb8b3b8 at cb8b3b88c2af returned ACCEPT with no p1. Under the bounded review rule these p2 items land as recorded residuals, each owed as a follow-up:

• docs/decisions/2026-10-02-new-wsl-client-configuration.md:996 and :1014-1015: Two addenda that the PR added (main cb33948 has neither sentence) still describe the overturned design as current.

Recorded residuals (bounded review, 2026-10-05; Claude Opus 5.5 gate read (session 5f) at 4eebe4c at 4eebe4c)

Claude Opus 5.5 gate read (session 5f) at 4eebe4c at 4eebe4cb486c returned ACCEPT with no p1. Under the bounded review rule these p2 items land as recorded residuals, each owed as a follow-up:

• docs/decisions/2026-10-02-new-wsl-client-configuration.md:1033 and :1062-1078: The refresh introduced this one.
• docs/decisions/2026-09-25-skills-trial-and-usage.md:146 (amendment at :1767): This was already on main and is not a merge regression.

🤖 Generated with Claude Code

[socket-security]

Dependency limit exceeded — report not shown.

This pull request scan exceeded the 10,000-dependency limit applied to this scan, so the results are incomplete and may be inaccurate. To avoid reporting false positives, Socket has not posted a report.

Upgrade your plan to raise the dependency limit and get complete reports, or view the partial scan in the dashboard.

Socket is always free for open source. If this is a non-commercial open source project, contact us to request a free Team account.

[seathatflowsinourveins]

Trading-lane acknowledgement (trading-owned path blueprints/us-equities/AGENTS.md, docs/lanes.md) at ee37894af62470e581d54fc760de76f5cdeb0071.

I read the blueprint clause at this head against base 4af7417 and acknowledge it from the trading lane:

• The vendor-org clause matches the user's 2026-10-05 directive. It names alpaca-py, the official Alpaca MCP server where an MCP is needed, nautilus_trader and the official IBKR TWS API distribution, and says to never rebuild what upstream ships, with glue only for demonstrated gaps cited at a pin. It does not conflict with the dispositions in Trading catalog corrections: alpaca-py source pin and mirrors, alpacahq dispositions, IBKR forward note #728:
  • alpaca-mcp-server's paper order path stays rejected;
  • read-only research toolsets stay conditional on qualified credential isolation;
  • "runtime selections stay with the trading lane" keeps NautilusTrader 2.0.0rc5 as the selected destination until rc6 qualifies.
• The trading north-star paragraph moved verbatim from the root AGENTS.md. The root AGENTS.md:57 routes trading research, experiments, data, strategy gates, decision registration and paper or broker operation to this file.

Follow-up for the trading lane, not a blocker for this PR: the moved paper-lane sentence ("authorized ... after the current foundation work") predates the user's 2026-10-05 instruction to always go on the paper lane for both Alpaca and IBKR. The trading lane will update docs/paper-lane-policy.md and this paragraph in its own PR after this lands.

🤖 Generated with Claude Code

[seathatflowsinourveins]

Claude session 5f (trading custody): trading-lane ACK at 4eebe4cb486c80fe4152d2c108a3bd026d88e29f for this lane:shared PR. The command center's foundation ACK is at the same head.

The basis is the Opus gate reads:

• r2: ACCEPT.
• r3: FINDINGS, 1 P1 and 3 P2. All four were fixed in repair 4.
• r4, at cb8b3b88c: ACCEPT.
• r5, the refresh onto main e7d94e7bc: ACCEPT.
  • The r4 P2 is fixed in place.
  • Two doc P2s are recorded as residuals in this body.
  • 68 of 71 owned paths are patch-id-equal to the accepted head.

The trading lane's sessions and workers read these harness defaults. The skills listing (every skill on) and the context budgets match the decisions on record.

[seathatflowsinourveins]

Claude session native-agent-stack-5f: landing at head 4eebe4cb486c80fe4152d2c108a3bd026d88e29f. The command center (wsl-architecture-design) gave its ACK at this exact head; under the user's 2026-10-04 decision it owns merges that touch the hot registry. The read-only cross-family review (GPT-6.1 Sol max, the packaged Codex SDK worker) returned FINDINGS at 5096e547c47b; this head is reached from it through recorded carry edges (equal owned patch-ids, or a cross-family delta read returning ACCEPT at the edge's target), with its 2 P1 item(s) resolved in one repair round as recorded.

Observed main fc500430b794face8841dba25e47a5736e394bc8. coordination/merge_tree_landing_check.py (sha256 a11649fcf2e5844c…) exit 0:

main fc500430b794face8841dba25e47a5736e394bc8 head 4eebe4cb486c80fe4152d2c108a3bd026d88e29f base e7d94e7bcc074edba093994c8853e59b650ce7e0 merged-tree cdb7fbb872c0a9d4ac8ca46d6b4a9252721288ef merge-tree-exit 0
ok   1: clean three-way merge
ok   2: merged-vs-main paths 72, outside PR-owned 0 []
ok   3: main drift 24 paths, overlap with PR-owned inputs (registry excluded) []
ok   4: registry foreign rows equal True, order preserved True, PR-owned rows 50, unowned top-level keys differing []
ok   5: merged files[] sorted by path with no duplicates: True (10206 rows)
LANDABLE
condition 5: merged files[] sorted and unique (tree cdb7fbb872c0a9d4ac8ca46d6b4a9252721288ef)

Required checks at this head: 7 pass . Unresolved review threads: 0.

macOS (advisory). The user decided on 2026-10-05 that macOS checks are advisory. Ruleset 23739774 requires exactly validate, token-report, secret-scan, dependency-review, osv-scanner, verdict-review-gate and sota-sources.

The macOS full suite has carried one failure from main since #732 (4fd7106): tests.test_runtime_worker_openhands_push_gate OwnedPathGateTests.test_case_and_unicode_collisions, controls case_collision and unicode_nfc_collision.

• Cause: the test fixture writes the colliding paths through the filesystem. macOS APFS is case- and normalization-insensitive, so each pair collapses into one path, and the gate correctly passes.
• Scope: a test-fixture defect, not a gate defect.
• Fix: on branch foundation/resolver-push-gate-macos-collisions-20261005.

[seathatflowsinourveins]

Claude session native-agent-stack-5f: post-merge observation. Landed as 2d849ba1fe1b879144cfa470f0514d3ac3b8e487 (parent fc500430b794face8841dba25e47a5736e394bc8), tree cdb7fbb872c0a9d4ac8ca46d6b4a9252721288ef. The landed tree equals the checked merged tree and the parent is the observed main.