Document Codex code-navigation parity and native F9 fixes

[seathatflowsinourveins]

Codex can omit pending navigation servers from its first catalog even when they initialize and return calls later. This draft proposes Codex's native connection-readiness fields and codebase-memory's own cache/runtime forwarding for F9's retained first-turn owners, and documents SocratiCode's own skill delivery path.

The sanitized snapshot separates five enabled registrations, resumed-turn discovery and actual lane transport counts from fresh-task readiness, functional success and organic owner selection. Existing F9/template owners receive the exact delta; this PR does not edit their live-owned templates or apply host configuration.

The co-op forwards this proposal to the command center, which selects the carrying PR and order after #752/#770. This lane does not contact template owners. Related AGENTS-template routing is overlap-token Q2 / draft #749; these startup rows stay separate. Startup/routing changes are harness configuration evidence, not unchanged U1 native-arm evidence or its required own-native routing fix/rerun.

Scope

• Base commit: ecfa112764c664d35377dd66b8cfcb67e5a94d60.
• Lane: lane:foundation; exactly one lane label.
• Owned paths: docs/decisions/2026-10-06-codex-codenav-parity-proposal.md, evidence/artifacts/codenav-parity-20261006/summary.json; their registry entries are the final commit.
• Remains draft for the command center's cross-family read and F9/template-owner incorporation. No default/adoption/exclusion status changes.

SOTA sources

• openai/codex, rust-v0.160.0 / a956835d020762cb2b570053af06f643a11c0ecc: live versus cached readiness and required-server semantics; catalog implementation; optional-grace alternative.
• DeusData/codebase-memory-mcp, v0.11.0, cli.c:3754–3784: its own Codex registration forwards CBM_CACHE_DIR and CBM_RUNTIME_DIR; no custom server or installer is introduced.
• giancarloerra/SocratiCode, 1.15.0 / f6191f076a42405f0d5508139f3a8b505cfef93a, README:108,222–259: pinned own plugin skills with one custom top-level engine; fresh-task lifecycle and MCP-only limitations.
• oraios/serena, 2.0.0.dev0 / c6fbd1c5932df2494ffa0020af5a9fbe80b82143, own Codex integration; jgravelle/jcodemunch-mcp, 1.108.319 / 8f7b34abe16fb459e0bf1c04747d584216dfe32e, CLIENTS.md:34; MinishLab/semble, 0.6.1 / 24497845460960db1839c8485319df189a889225, own Codex integration.

Evidence-class table

Claim	Evidence class	Command / receipt
Five configured, enabled stdio registrations	native_proven: installed CLI observation	Native codex mcp get for each; sanitized summary
Resumed-turn discoverable surfaces	local_integration: current metadata observation	Counts 24/3/2/17/26; no new target RPC or fresh-session acceptance
Lane completed-call counts	native_proven: historical transport telemetry	12 roots +72 native-linked descendants; 59/0/5/3/0 completed plus one failed Serena; independent recount; task success/organic selection ungraded
Proposed startup/forwarding/skill delivery	source_review	Exact upstream pins and files above; no host apply or fresh native rerun

Local commands run

• rtk proxy codex --version: exit 0, 0.160.0.
• rtk proxy codex mcp get <server> for all five: exit 0; native environment values masked.
• rtk proxy codex plugin list --json: exit 0; nine installed entries; SocratiCode not found in those entries. Standalone skills elsewhere remain unverified.
• Native rollout metadata recount: final exit 0; zero duplicate/wrong-thread target records. Initial analysis failed on a source/subagent string variant; guarded parsing corrected it without changing originals.
• rtk python3 scripts/validate.py: exit 0; 69 components, 10274 hashed files, four profiles, 212 receipts. Integrity/scope checks; no live provider or GPU execution.
• rtk git diff --check: exit 0.
• Each of three independently based proposal hunks: native git apply --check --whitespace=error - exit 0; neither template is changed. Blank context lines were formatted without trailing whitespace after the initial whitespace check flagged them.
• Independent read-only completeness critic accepted the draft with no material corrections; residual first-catalog, skill lifecycle, cache/index identity and out-of-scope-call modalities remain documented. Cross-family approval is pending.

Decision record

CI observation: osv-scanner failed. Its six advisory rows name five pre-existing lockfiles, each unchanged from baseline ecfa112; the frozen macOS lock blob matches baseline exactly. This proposal changes none of those sources or scanner policy. The co-op can forward the diagnostic to the existing dependency/CI owner; no CI pass or landing is claimed.

docs/decisions/2026-10-06-codex-codenav-parity-proposal.md names targeted readiness, shared grace and native skill delivery alternatives. Matched upstream-harness evidence can overturn the startup choice; the organic owner map controls which servers remain adopted. No zero-call slot is excluded from this snapshot.

Host evidence

No files under evidence/hosts/ and no host configuration apply. Retained J-MCP diagnosis is reused, not rerun. Fresh intended-profile exposure, actual retrieval/index/graph acceptance and organic reruns remain with F9 and the existing run owners.

Checklist

• No workflow changes or new paid hosting.
• No credentials, raw conversations, tool arguments/results or private host paths published.
• Peer-owned files preserved; template/F9 handoff is explicit.
• No merge/landing; draft until command-center cross-family read.

[seathatflowsinourveins]

F9/template-owner handoff from overlap-codenav: this draft carries the concrete proposal and sanitized counts, leaving live-owned configuration files untouched under docs/lanes.md.

The proposed delta is in this draft's dated decision and evidence: S1/S2/S3 are exact independent unified hunks against origin/main at ecfa112, with per-row (a)-(d) gap evidence and pinned upstream sources. They propose required/live-connection readiness for retained first-turn owners and codebase-memory's upstream CBM_CACHE_DIR/CBM_RUNTIME_DIR forwarding. Preserve existing values and other forwarded names. The SocratiCode skill/server split is separate, with fresh-task lifecycle acceptance pending.

The co-op forwards the proposal to the command center, which selects the carrying PR and application order after #752 and #770. This lane does not contact the owners or alter either template. Related navigation routing text is overlap-token Q2 / draft #749; it stays separate so the command center combines one consistent change. Startup and routing changes are harness configuration evidence, not unchanged U1 native-arm evidence or its required own-native routing fix/rerun. No zero-call tool is excluded by the snapshot.

PR remains draft for the command center's cross-family read. Local validate.py passed; that is integrity/scope evidence, not native readiness or quality acceptance.

[seathatflowsinourveins]

Closed with a record by the PR triage of 2026-10-07 (the command center's ruling). Not merged; the branch codex/ns2604-overlap-codenav-20261005 stays on origin at 50bf3c5.

Folded into #776 at cda3e50. Of this PR's 3 changed files, 2 are byte-identical there and 1 is merged; none is lost. Registry: 2 entries added or changed by this PR are all present there, 0 of them re-registered to the merged file's hash.

Reopen trigger: none while #776 carries the content. Reopen with gh pr reopen 780.