fix(hive): backfill the context gauge for Codex agents from their rollout log

[MEKSI8]

What & why

Codex agent cards can remain without a context (ctx %) reading while a worker
is active. The root cause is a provider-coverage gap, not one shared telemetry
failure:

• Claude can send an exact context reading through a status hook.
• The renderer independently polls Claude transcripts every 15 seconds to
  backfill a missing reading (useHive.ts:565-590).
• Fleet tokens/lastTool data comes from a separate OpenTelemetry input.
• Codex rollout logs provide another provider-specific source through
  token_count events.

The shared Codex hook wiring was traced in the architecture, but hook delivery
was not empirically proven as part of this work. This fix does not depend on
that delivery path.

HiveManager now exposes the isolated worker CODEX_HOME at
<hive>/agents/<id>/.codex. The main-process poll reads that location rather
than the registry's project cwd, selects the newest nested rollout, and feeds
the latest valid last_token_usage.input_tokens / model_context_window
reading through HookServer.reportContext(). The renderer needs no change.

The 8-second poll no longer recursively walks all rollout history on every
tick. The selected path, mtime, size, and reading are cached; ordinary ticks
stat only the selected file, changed files are re-tailed, and discovery is
repeated after 60 seconds or immediately if the selected file disappears.

Fleet cost/activity telemetry remains out of scope and unchanged.

Type of change

• Bug fix
• New feature
• Refactor / cleanup
• Docs
• Build / CI

Evidence

Before

The pre-fix candidate was run against a read-only copy of one real worker
rollout. The registry project cwd returned no reading, while the agent
directory happened to work only because the old parser appended .codex
itself:

{"registryCwdReading":null,"agentDirectoryReading":{"tokens":204802,"limit":258400}}

The regression-first focused run against the pre-fix implementation exited 1:
9/15 tests succeeded and 6 failed, including the registry/HiveManager seam,
the exact tail-window boundary, and cache behavior.

After

The same copied rollout, placed under an agent-home-shaped temporary directory,
was read through the corrected CODEX_HOME contract without reading or modifying
the live session:

{"projectCwdReading":null,"codexHomeReading":{"tokens":204802,"limit":258400}}

The focused command node --test test/codex-context.test.cjs exited 0 with
15/15 tests. Its coverage includes the real HiveManager.registry() plus
HiveManager.codexHome() seam, a tail window beginning exactly after a newline,
cache refresh/discovery behavior, and deletion assertions for every test temp
directory.

The repository-wide test command still exits 1 because the base and candidate
both have the same 19 failing test names:

base bdf524e:      831 total / 804 succeeded / 19 failed / 8 skipped
candidate:         846 total / 819 succeeded / 19 failed / 8 skipped
failing-name diff: empty

Since this evidence was gathered, the office was restarted with this fix
applied. The third screenshot above is that running app: Dwight's and Angela's
(Codex) agent cards now carry the same context gauge Claude agents already had.

How I tested it

• Windows 11, Node v24.15.0.
• npm ci --ignore-scripts — exit 0; 870 packages installed. npm reported the
  existing audit state: 22 vulnerabilities (3 moderate, 18 high, 1 critical).
• node --test test/codex-context.test.cjs — exit 0; 15/15.
• npm run typecheck — exit 0 (node and web TypeScript projects).
• node --test --test-reporter=tap test/*.test.cjs on detached base
  bdf524e — exit 1; 19 baseline failures.
• The same full-suite command on the candidate — exit 1; the same 19 failure
  names, with 15 additional successful Codex-context tests.
• npm run build — exit 0; Vite retained its existing dynamic/static import
  chunk warning.
• git diff --check bdf524e..HEAD — exit 0.
• Real-rollout copy reproduction above — exit 0; the temporary copy and
  detached base worktree were removed afterward.

Credit

Discord:

X:

Checklist

• Before and after terminal evidence is recorded above.
• npm run typecheck completed with exit 0.
• npm run test:focused is not claimed: the repository-wide test command
  remains exit 1 with the same 19 baseline failures. The directly relevant
  Node test file completed 15/15 with exit 0.
• npm run build completed with exit 0.
• This PR is limited to the Codex context gauge and supporting tests.
• No debug output, unrelated formatting churn, or live data mutation is
  included.
• N/A — no new UI.
• N/A — no new art.

[github-actions]

✅ Evidence received. Before and after are both attached. Thanks — this is what makes a PR reviewable in one pass.