Files
paperclip/packages/adapters/codex-local
Jannes Stubbemann 9e750d3e92 feat(codex-local): env-driven gateway routing via PAPERCLIP_CODEX_PROVIDERS config.toml (#7919)
## Thinking Path

> - Paperclip is the open source app people use to manage AI agents for
work
> - The `codex-local` adapter runs the OpenAI Codex CLI; Paperclip
already maintains a managed `CODEX_HOME` per company and ships it to
remote/sandboxed execution targets
> - Deployments increasingly put an OpenAI-compatible LLM gateway
between the harness and the model for cost, governance, or
data-residency reasons: LiteLLM, OpenRouter, Portkey, Kong, a corporate
proxy, self-hosted models (vLLM/Ollama), or region-pinned/sovereign
endpoints. But Codex has no CLI flag or env var for a custom endpoint:
its only mechanism is `[model_providers.<id>]` tables (with `base_url`,
`env_key`, `wire_api`) in `$CODEX_HOME/config.toml`, selected by a
root-level `model_provider` key
> - Today there is no supported way to get such provider config into the
managed `CODEX_HOME`, so gateway routing requires hand-editing files the
adapter owns and regenerates
> - This pull request adds the codex analogue of #7837's opencode
mechanism: a `PAPERCLIP_CODEX_PROVIDERS` JSON env var whose shape maps
1:1 onto codex's TOML schema, merged into the managed `config.toml` so
the existing asset-shipping + `env.CODEX_HOME` mechanics deliver it to
local and sandboxed runs alike; nothing here is specific to one hosting
setup
> - The benefit is Codex works behind any OpenAI-compatible gateway with
config only; with no env set, behavior is unchanged

## Linked Issues or Issue Description

No existing issue; describing in-PR (feature / adapter enhancement).

- **Gap:** there is no supported way to register a custom/gateway
`[model_providers.*]` endpoint for `codex-local`. Codex's only
custom-endpoint mechanism is `config.toml` (`base_url` + `env_key` +
`wire_api`, selected via the root `model_provider` key), and the adapter
owns/regenerates the managed `CODEX_HOME`, so operators cannot durably
hand-edit it.
- Related: #7837 (the opencode-local analogue of this change, same
env-driven gateway-routing pattern). Searched for duplicate/related PRs:
no existing codex-local gateway/provider-routing PR found.

> Note on ROADMAP: this is adapter-level, opt-in config (defaults
unchanged) that *enables* gateway routing for one harness; it is not the
core "Cloud / Sandbox agents" platform work itself.

## What Changed

- New `prepareCodexRuntimeConfig()`
(`packages/adapters/codex-local/src/server/runtime-config.ts`): reads
`PAPERCLIP_CODEX_PROVIDERS` (run env first, then `process.env`), shaped
as `{"providers": {"<id>": {base_url, env_key, wire_api, ...}},
"model_provider": "<id>"}`, and merges it into the managed
`CODEX_HOME`'s `config.toml`. No-op when unset or empty.
- A malformed value (invalid JSON, not a JSON object, no `providers`
object, no usable provider entries, or individual entries with empty
names or non-object values, which are skipped by name) is never silently
dropped: each case surfaces a distinct, user-visible note (via the
prepare notes, which flow into command notes + `onLog`) and unusable
input leaves `config.toml` untouched.
- Merge is marker-delimited and TOML-correct: existing `config.toml`
content is preserved between two managed blocks. Root keys (e.g.
`model_provider`) are prepended **before the first table header** (TOML
root-region rule), `[model_providers.*]` tables are appended.
Pre-existing same-name provider sections and root `model_provider` keys
are excised so the managed definitions win without duplicate-table parse
errors.
- `{env:VAR}` placeholders are expanded server-side for
literal-credential fields; `env_key` indirection remains the preferred
path.
- Crash-safe restore: prepare writes a pre-run backup
(`config.toml.paperclip-backup`) before the merged file; `cleanup()`
restores the original in the execute `finally` and removes the backup.
If a run never reaches `cleanup()` (a throw during the setup between
prepare and execution, or SIGKILL), the next prepare restores the
original from the backup with full fidelity, including user
`[model_providers.*]` sections the merge excised (review feedback, P2);
plain block-stripping remains the fallback for pre-backup state.
- An explicit adapter-config `env.CODEX_HOME` override is treated as
user-managed: no merge, surfaced as a command note.
- Dependency-free hand-emitted TOML (strings/numbers/booleans, arrays of
scalars, plain objects as inline tables); basic strings escape
U+0000-U+001F and U+007F per TOML 1.0 (review feedback, P2). Merged
output was additionally validated locally with python tomllib during
development; the committed tests assert the structural invariants.
- `execute.ts` wiring: `prepareCodexRuntimeConfig` runs after
`prepareManagedCodexHome` (before the home ships to the remote target),
notes surface via `onLog` + command notes, and the `finally` calls
`cleanup()`.

**Note for reviewers:** current codex removed `wire_api = "chat"`
(openai/codex#10157, Feb 2026), so gateway provider configs must use
`wire_api = "responses"`, i.e. the gateway must speak `/v1/responses`.
The adapter passes the value through verbatim; this is a codex-side
constraint worth knowing when configuring it.

## Verification

- `pnpm --filter @paperclipai/adapter-codex-local build` and
`typecheck`: tsc clean against current `master`
- `pnpm exec vitest run packages/adapters/codex-local`: 45 passing
(incl. 17 `runtime-config` tests: fresh-merge + cleanup restore,
root-region placement, same-name provider override, inline
tables/arrays, DEL escaping, `{env:}` expansion from run env +
`process.env`, per-case malformed-input notes with `config.toml`
untouched, skipped-entry notes alongside a successful merge, silent
no-op when unset/empty, explicit-`CODEX_HOME` skip note, backup restore
of excised user sections after an interrupted run, backup removal on
cleanup, stale-block self-heal, re-run replacement)
- Verified end-to-end: a codex agent in a hardened Kubernetes (gVisor)
sandbox completed a real task routed through an OpenAI-compatible
gateway's `/v1/responses`, with a billed usage row recorded on the
gateway. That deployment supplies the verification evidence; the
mechanism is gateway-agnostic.

## Risks

Low. Entirely env-driven and opt-in; with `PAPERCLIP_CODEX_PROVIDERS`
unset the adapter never touches `config.toml` and behavior is
byte-identical to before. The merge preserves user content, restores the
original file on cleanup, and survives interrupted runs via the pre-run
backup; malformed input surfaces a visible note and is ignored without
touching `config.toml`. No migration/UI impact.

## Model Used

Claude Opus 4.8 (`claude-opus-4-8`, 1M context), extended thinking +
tool use, via Claude Code.

## Checklist

- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md (adapter-level opt-in config enabling
gateway routing; not the core sandbox-platform work, noted above)
- [x] I have searched GitHub for duplicate or related PRs and linked
them above (#7837 is the opencode analogue; no codex-local duplicate
found)
- [x] I have either (a) linked existing issues OR (b) described the
issue in-PR following the relevant issue template
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] If this change affects the UI, I have included before/after
screenshots (n/a, no UI)
- [ ] I have updated relevant documentation to reflect my changes (env
var documented inline; no central doc references the adapter env yet)
- [x] I have considered and documented any risks above
- [x] All Paperclip CI gates are green (green on the previous head;
re-running on the final note-copy polish commit)
- [x] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
(both review P2s are fixed at head: the interrupted-run restore via the
pre-run backup and the U+007F escaping; a re-review is requested for the
note-copy polish)
- [x] I will address all Greptile and reviewer comments before
requesting merge

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-06-10 21:13:31 -07:00
..
2026-03-12 13:09:22 -05:00