fix(codex-runtime): re-running /codex-runtime codex_app_server when already enabled now triggers migration (salvage of #26532 by @simpolism, closes #26529)

[hashbender]

Summary

/codex-runtime codex_app_server now re-runs the migration when the runtime is already set to codex_app_server, instead of silently no-op'ing with "already set".

Root cause: apply() short-circuits on new_value == current before the migration block. That conflates "config value is correct" with "world state is converged" — fine for auto (disabling has no side effects) but wrong for codex_app_server, where the config value can be correct while the ~/.codex/config.toml managed block, the hermes-tools MCP callback, and plugin discovery are missing. Users hit this by pre-setting openai_runtime: codex_app_server in config.yaml (which hermes profile create --clone does automatically), then running the slash command expecting migration — silent no-op, bot falls back to codex_responses.

Changes

• hermes_cli/codex_runtime_switch.py: on codex_app_server re-apply, fall through to the (idempotent) migration block while skipping set_runtime/persist — no config-value change, so config.yaml is not rewritten and prompt cache is untouched. User-visible message becomes "already set to codex_app_server — re-applying migration". auto re-apply keeps the early-return short-circuit.
• tests/hermes_cli/test_codex_runtime_switch.py: test_reapply_codex_app_server_runs_migration — asserts migrate() ran, persist_callback did NOT fire, migration output surfaces, requires_new_session=True.

Validation

scenario	migration runs	config rewritten
first enable (auto→codex_app_server)	yes	yes
re-apply (codex_app_server→codex_app_server)	yes (was: no)	no
re-apply auto→auto	no	no

• RED→GREEN proven: the new test fails on origin/main ("migration must run on reapply") and passes with this fix.
• 31/31 tests in tests/hermes_cli/test_codex_runtime_switch.py pass.

Salvage of NousResearch#26532 by @simpolism (also reporter of NousResearch#26529). Cherry-picked cleanly onto current main, authorship preserved. Duplicate NousResearch#26535 (@aqilaziz) already closed; this is the more thorough implementation.

Closes NousResearch#26529

---

Review + follow-up (hermes-pr-review Phase 2)

Ran the full Phase-2 review (targeted + broader test sweep, real unmocked apply() smoke, 3-angle + Hermes-specific review): no Critical findings. Prompt-cache invariant verified (reapply does NOT call set_runtime/persist and does not touch system prompt/toolsets), auto-disable early-return preserved, migrate() idempotency (strip-before-write, atomic replace) confirmed.

Follow-up commit 6e761547d (mine) folds in the two non-blocking clarity findings: collapse the reapplying_enable predicate to a chained comparison, and dedupe the msg_lines literals. No behavior change; 31/31 tests green.

Also caught + corrected a stale-base artifact before pushing: an earlier working-tree state showed a phantom deletion of an unrelated scripts/release.py AUTHOR_MAP entry (branch was cut before that chore PR merged). Rebased onto current main so the PR shows only the two intended files.

---

Mirror-of: NousResearch#56563
NousResearch#56563

[tenki-reviewer]

Review Complete
No issues found!

Risk: 🟡 Medium (38/100) — no findings · 97 LOC across 2 files

---

PR adds a /codex-runtime slash command to toggle between Hermes' native runtime and Codex app-server mode, with MCP server migration. One medium-severity bug found: the binary check gate incorrectly blocks re-apply migration when codex is already enabled but the binary is missing.

Files Reviewed (2 files)

hermes_cli/codex_runtime_switch.py
tests/hermes_cli/test_codex_runtime_switch.py