fix(codex): show the context the session actually routed to

[tanglearncode]

The bug

In Codex, the session auto-routes to a context — use_context reports the switch, get_context serves that context's profile — and then $neatcontext:status and $neatcontext:list, run moments later, both report that nothing is connected.

Reproduced on this machine, against the real state the user's Codex window left behind:

$ CODEX_THREAD_ID=01a02647-… node src/codex/neatcontext-cli.mjs status   # before
No context is connected yet. Use `$neatcontext:use` to pick one.

$ CODEX_THREAD_ID=01a02647-… node src/codex/neatcontext-cli.mjs status   # after
Connected context: NeatContext platform control-plane implementation

Why it happened

The two halves of the plugin were reading different files, and nothing in either path could observe it. What Codex actually gives each process — all verified against Codex 0.147 on this machine, not assumed:

process	how Codex starts it	thread id?
skill CLI	shell tool	yes — CODEX_THREAD_ID is in the shell environment (confirmed in Codex's own spawn log)
SessionStart hook	its own parent process	yes — session_id on stdin
MCP bridge	scrubbed environment	no — probed a plugin-shaped MCP server: it receives PATH, TEMP, the platform basics and nothing else; no CODEX_* variable at all

The host-pointer channel that was supposed to bridge that gap keys on process.ppid, and the hook and the bridge do not share one. From a single Codex run on this machine:

plugin-hosts/pid-219336.json         {"sessionId":"01a02675-…","source":"session-start"}   ← hook
plugin-hosts/pid-224804.bridge.json  {"sessionId":null}                                    ← bridge

There is no other route: Codex's MCP client advertises no roots capability and answers roots/list with {"roots":[]}, and a plugin server's cwd must point inside the plugin, so the bridge cannot even name the workspace. Every routing decision the bridge has recorded on this machine says "sessionId": null — it has been writing the machine-wide selection all along, while $neatcontext:use/status wrote and read the per-thread one.

The fix

The Codex adapter stops scoping on a value only half of its processes can see. One selection, one routing mode, one set of declines, shared by the bridge, the hook, and the skills.

That is coarser than per-thread — two Codex windows on a machine share it — but it is the honest scope for what this host exposes, and it is already what the bridge (and therefore all grounding) was doing. The Copilot adapter refuses the same trade for the same reason, in a comment that describes this bug exactly: "a variable only some of this plugin's processes see is worse than none." NEATCONTEXT_SESSION_ID remains the way back for any host that can inject one id into every process.

The machinery built on the old assumption goes with it: the pointer writes, the bridge's published session, and the $neatcontext:use/status drift warnings that could never fire on Codex. pruneHostPointers stays in the SessionStart hook, to sweep the pointer files older versions left on users' machines. The now-inert env_vars: ["CODEX_THREAD_ID"] passthrough is dropped from .mcp.json.

Mode settings were split the same way ($neatcontext:mode manual was invisible to the bridge that enforces it) and are fixed by the same change.

Behavior change

Docs and skill descriptions that promised per-thread isolation are corrected rather than left to mislead: $neatcontext:use, $neatcontext:disconnect and $neatcontext:mode now say the connection is shared across Codex sessions on the machine. No migration: the shared file is the one the bridge was already writing, so whatever a session had routed to stays connected and becomes visible to status immediately.

Tests

• tests/codex-session-drift.test.mjs encoded the old contract and is replaced by tests/codex-session-scope.test.mjs, which spawns the three process kinds the way Codex spawns them — bridge with no thread id, CLI with one, hook with a third host key — and holds them to one answer: what use_context connects is what status and list report, what a skill connects is what get_context serves, the mode a skill sets is the mode the bridge enforces, and a pointer file left by an older version cannot re-split the scope.
• npm test 573 passing, npm run check, e2e:commands, e2e:extensions, e2e:no-nudge all green; diff coverage reports all 60 changed source lines covered.