Skip to content

Commit 66b25d4

Browse files
committed
docs(devlog): close audit blockers on routed tool-discovery landing pass
An independent audit of 094_landing_verification_pass.md returned FAIL over four rounds. This closes every blocker it raised. Scope of the central claim (correction 2). "Under code_mode_only MCP tools stay callable in both modes" was unconditional and unsupported. It now reads "an eligible MCP tool", with a table of the three flag-independent removal paths that an override cannot repair: direct_only_tool_namespaces -> DirectModelOnly (spec_plan.rs:210), excluded_tool_namespaces (spec_plan.rs:444), and MCP/App policy filtering applied before exposure classification (mcp_tool_exposure.rs:20). The flag also drives tool_search construction and the deferred-guidance text, so exec.description is not the only delta. Code Mode vs CodeModeOnly. The isolate is built for both (`!matches!(tool_mode, CodeMode | CodeModeOnly)`, spec_plan.rs:459); only ToolMode::Direct has none. 000, 004, 010, 020 and 094 are aligned on this, and 020's differential expectation no longer claims the delta is confined to exec.description. Cursor fence (unresolved #1596 P2). 012 no longer reproduces the slug check. Both construction paths now specify one shared isCursorRoute(entry, providerId) helper that prefers provider identity and falls back to the slug prefix only when no CatalogModel is available. The third parameter becomes an options object { toolDiscoveryMode, providerId } so the existing public callers (src/codex/catalog.ts re-export, tests/parallel-tool-calls-optin.test.ts) keep their positional arguments. 003 updated to match. Draft applicator. patches/apply-draft.mjs writes that same asymmetry while printing "draft seams: OK", because --check asserts seam presence and not policy. Every instruction to run it is withdrawn: the patches README is now marked ARCHIVAL, the Korean summary's application block is removed, and the script's own docblock opens with ARCHIVAL - KNOWN-DEFECTIVE - DO NOT EXECUTE. The body is kept as the seam map it is. Config contract. 011 now requires a load-path warning on degradation and descriptor-before-read rejection of accessor/prototype-polluted model maps, both as first-PR work with tests in 020. Upstream paths. 004 corrected to codex-rs/code-mode/, with spec_plan_tests.rs acknowledged as existing. Verification: independent reviewer VERDICT PASS after four repair rounds; bun test tests/repo-hygiene.test.ts (11 pass / 0 fail); bun run privacy:scan passed; node --check patches/apply-draft.mjs passed; SHA256SUMS regenerated and fully verified.
1 parent c86702a commit 66b25d4

13 files changed

Lines changed: 285 additions & 80 deletions

‎devlog/_plan/260813_routed_tool_discovery_profiles/000_master_plan.md‎

Lines changed: 6 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -12,9 +12,12 @@ Base change: PR #1596, `fix(codex): restore deferred tool discovery for non-Curs
1212
> bundle was authored without a mounted checkout; every claim below was
1313
> re-verified on 2026-08-13 against a real worktree, the upstream `codex-rs`
1414
> source and live GitHub state. Eight corrections were recorded, the most
15-
> consequential being that under `code_mode_only` MCP tools stay callable in
16-
> **both** discovery modes — so `direct` is a comprehension/compatibility lever
17-
> with a payload cost, not a reachability fix.
15+
> consequential being that an **eligible** MCP tool stays callable in **both**
16+
> discovery modes under code mode — so for those tools `direct` is a
17+
> comprehension/compatibility lever with a payload cost, not a reachability fix.
18+
> "Eligible" excludes `direct_only_tool_namespaces`, `excluded_tool_namespaces`,
19+
> and anything removed by MCP/App policy filtering; see `094` for the exclusion
20+
> table and the differential test that must prove the claim.
1821
1922
## 1. Objective
2023

‎devlog/_plan/260813_routed_tool_discovery_profiles/003_current_code_map.md‎

Lines changed: 6 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -81,16 +81,20 @@ Change:
8181
normalizeRoutedCatalogEntry(entry, parallelToolCalls)
8282
```
8383

84-
into:
84+
into an options object — the fence needs the provider identity as well as the
85+
mode, and the existing positional arguments must stay put for the public callers
86+
(`src/codex/catalog.ts` re-export, `tests/parallel-tool-calls-optin.test.ts`):
8587

8688
```ts
8789
normalizeRoutedCatalogEntry(
8890
entry,
8991
parallelToolCalls,
90-
toolDiscoveryMode,
92+
{ toolDiscoveryMode, providerId },
9193
)
9294
```
9395

96+
See `012` for the resolved snippet and the shared `isCursorRoute()` helper.
97+
9498
The function still hard-fences Cursor to direct.
9599

96100
### `src/codex/catalog/sync.ts`

‎devlog/_plan/260813_routed_tool_discovery_profiles/004_upstream_codex_code_mode.md‎

Lines changed: 37 additions & 12 deletions
Original file line numberDiff line numberDiff line change
@@ -4,9 +4,12 @@
44
> (`/Users/jun/Developer/codex/120_codex-cli`, `main` @ `4462b9dee`, 2026-07-23).
55
> See `094_landing_verification_pass.md` Correction 2. The mechanism described
66
> below is confirmed, but its consequence was understated: under
7-
> `tool_mode = code_mode_only`, MCP tools are installed on the V8 `tools` /
8-
> `ALL_TOOLS` globals in **both** deferred and direct mode. The flag changes
9-
> where the schemas live, not whether the tools are callable.
7+
> `tool_mode = code_mode_only`, an **eligible** MCP tool is installed on the V8
8+
> `tools` / `ALL_TOOLS` globals in **both** deferred and direct mode. For such a
9+
> tool the flag changes where the schemas live, not whether it is callable.
10+
> "Eligible" excludes `direct_only_tool_namespaces`, `excluded_tool_namespaces`,
11+
> and anything removed by MCP/App visibility or policy filtering — see the
12+
> exclusion table in `094`.
1013
1114
## Verified mechanism
1215

@@ -20,11 +23,17 @@ The public `exec` description states that nested tools are available through the
2023

2124
Relevant upstream files:
2225

23-
- `codex-rs/code-mode-runtime/src/runtime/globals.rs`
26+
- `codex-rs/code-mode/src/runtime/globals.rs`
2427
- `codex-rs/code-mode-protocol/src/description.rs`
2528
- `codex-rs/core/tests/suite/code_mode.rs`
29+
- `codex-rs/core/src/tools/spec_plan.rs`
2630
- `codex-rs/core/src/tools/spec_plan_tests.rs`
2731

32+
Path note (2026-08-13): the runtime crate is `codex-rs/code-mode/`, not
33+
`code-mode-runtime/`. The load-bearing exposure and executor logic lives in
34+
`spec_plan.rs`; `spec_plan_tests.rs` also exists and was correctly named in the
35+
original list. Verified in the checkout at `4462b9dee`.
36+
2837
Repository: https://github.com/openai/codex
2938

3039
## The exposure switch, verbatim
@@ -49,14 +58,30 @@ Two consequences:
4958
namespace tools resolves to `Direct` whatever the catalog says.
5059
2. `Direct` under code mode means every MCP declaration is embedded in
5160
`exec.description` (they enter `enabled_tools` instead of `deferred_tools`).
52-
That is the measured 96,699 → 258,929 character regression, and it is the
53-
*only* material difference — the tools are equally callable either way.
54-
55-
So the `direct` override is a model-comprehension lever (full schemas inline
56-
rather than via `ALL_TOOLS`) with a payload cost. It is **not** a reachability fix
57-
under code mode. E1/E2 in `009` remain open: this reading comes from a 2026-07-23
58-
clone, while #1522 reported against CLI `0.147.0-alpha.6.5` on 2026-08-12, and
59-
app-layer gating outside this clone is unverified.
61+
That is the measured 96,699 → 258,929 character regression. For an eligible
62+
tool the schema placement is the material difference — it stays equally
63+
callable either way.
64+
65+
The flag additionally drives `tool_search` construction and the deferred-guidance
66+
text, and it does **not** govern three independent removal paths
67+
(`direct_only_tool_namespaces` → `DirectModelOnly`, `excluded_tool_namespaces`,
68+
and pre-classification MCP/App policy filtering). An override cannot repair any of
69+
those.
70+
71+
The isolate is built for `ToolMode::CodeMode` as well as `CodeModeOnly`
72+
(`spec_plan.rs:459`), so the reachability reasoning covers both; `code_mode_only`
73+
additionally hides ordinary nested tools from the top-level list. Only under
74+
`ToolMode::Direct` is there no isolate, and there the flag does govern direct
75+
declaration versus `tool_search`. Every routed OpenCodex row is stamped
76+
`code_mode_only` today.
77+
78+
So for eligible tools under code mode the `direct` override is a
79+
model-comprehension lever (full schemas inline rather than via `ALL_TOOLS`) with a
80+
payload cost, not a reachability fix. E1/E2 in `009` remain open: this reading
81+
comes from a 2026-07-23 clone, while #1522 reported against CLI
82+
`0.147.0-alpha.6.5` on 2026-08-12, and app-layer gating outside this clone is
83+
unverified. `020`/`021` owe a controlled single-variable test before this claim is
84+
treated as settled.
6085

6186
## Architectural implication
6287

‎devlog/_plan/260813_routed_tool_discovery_profiles/010_phase1_profile_resolver.md‎

Lines changed: 6 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -62,10 +62,12 @@ type ResolvedRoutedToolDiscoveryMode = "deferred" | "direct";
6262
- no client-version database;
6363
- no per-session mode mutation;
6464
- no claim that `direct` is cheaper or preferred.
65-
- no claim that `direct` restores *reachability* under code mode. Upstream
66-
installs MCP tools on the `tools`/`ALL_TOOLS` globals in both modes; `direct`
67-
only moves full schemas into `exec.description` at a measured payload cost
68-
(`004`, `094` Correction 2).
65+
- no claim that `direct` restores *reachability* for an **eligible** tool under
66+
code mode. Upstream installs such tools on the `tools`/`ALL_TOOLS` globals in
67+
both modes; `direct` moves their full schemas into `exec.description` at a
68+
measured payload cost and changes `tool_search` construction. It cannot repair
69+
a tool removed by `direct_only_tool_namespaces`, `excluded_tool_namespaces`, or
70+
MCP/App policy filtering (`004`, `094` Correction 2).
6971

7072
## Acceptance
7173

‎devlog/_plan/260813_routed_tool_discovery_profiles/011_phase1_types_and_config.md‎

Lines changed: 16 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -53,6 +53,22 @@ A warning should state that the invalid policy was ignored.
5353

5454
### Write path
5555

56+
> **Amended 2026-08-13 (`094` Correction 5 + audit blocker 3).** Two gaps between
57+
> this section's promises and the draft in `patches/`:
58+
>
59+
> 1. The draft adds `.catch(undefined)` but **no load-path warning**. Silent
60+
> degradation is exactly the observability failure `014` exists to prevent —
61+
> an operator whose emergency override was dropped by a typo must be told.
62+
> The implementation must emit the warning, and a test must assert it.
63+
> 2. The accessor/prototype-pollution rejection below must be established
64+
> **before** the validator enumerates or reads properties. A validator that
65+
> reaches for `candidate.providers[name].modelRoutedToolDiscovery` and only
66+
> then checks for getters has already run attacker-controlled code. Inspect
67+
> with `Object.getOwnPropertyDescriptor` and reject non-data descriptors and
68+
> non-own/prototype-sourced keys first, then read values.
69+
>
70+
> Both are first-PR requirements, not follow-ups.
71+
5672
`validateConfigCandidate()` must inspect the raw candidate before `.catch()` can erase the invalid value. Reject:
5773

5874
- unknown mode;

‎devlog/_plan/260813_routed_tool_discovery_profiles/012_phase1_catalog_patch.md‎

Lines changed: 47 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -30,27 +30,59 @@ Current:
3030
normalizeRoutedCatalogEntry(e, model?.parallelToolCalls === true);
3131
```
3232

33-
Proposed:
33+
Proposed. The third argument becomes an options object rather than another
34+
positional boolean, because the fence now needs the provider identity as well as
35+
the mode, and `normalizeRoutedCatalogEntry` has public callers (`src/codex/catalog.ts`
36+
re-exports it, and `tests/parallel-tool-calls-optin.test.ts` calls it with one and
37+
two arguments). Keeping both existing positions intact preserves those callers:
3438

3539
```ts
3640
normalizeRoutedCatalogEntry(
3741
e,
3842
model?.parallelToolCalls === true,
39-
model?.toolDiscoveryMode ?? "deferred",
43+
{
44+
toolDiscoveryMode: model?.toolDiscoveryMode ?? "deferred",
45+
providerId: model?.provider,
46+
},
4047
);
4148
```
4249

4350
Inside normalization:
4451

4552
```ts
46-
const isCursorEntry = typeof entry.slug === "string"
47-
&& entry.slug.startsWith("cursor/");
48-
const effective = isCursorEntry ? "direct" : toolDiscoveryMode;
53+
// Provider identity first; the slug prefix is only a fallback for callers that
54+
// have no CatalogModel (see the fence note below).
55+
const isCursorEntry = isCursorRoute(entry, options?.providerId);
56+
const effective = isCursorEntry ? "direct" : (options?.toolDiscoveryMode ?? "deferred");
4957

5058
applyRoutedCodexToolMode(entry);
5159
entry.supports_search_tool = effective === "deferred";
5260
```
5361

62+
The shared helper, used by both construction paths:
63+
64+
```ts
65+
export function isCursorRoute(entry: RawEntry, providerId?: string): boolean {
66+
if (providerId !== undefined) return providerId === "cursor";
67+
return typeof entry.slug === "string" && entry.slug.startsWith("cursor/");
68+
}
69+
```
70+
71+
With no `CatalogModel` the slug prefix remains the only available signal, which is
72+
why it stays as the fallback rather than being deleted.
73+
74+
> **Amended 2026-08-13 (`094` Correction 8).** The original draft here reproduced
75+
> `entry.slug.startsWith("cursor/")` verbatim. That is exactly the asymmetry the
76+
> unresolved #1596 P2 review flags: the template path fences on the public slug
77+
> while the template-less path fences on `model?.provider === "cursor"`, so a
78+
> `cursor/`-aliased combo whose canonical provider is `combo` is classified
79+
> differently depending on whether a template happened to be available.
80+
>
81+
> Since this unit already unifies both paths behind one resolver, it must close
82+
> the asymmetry rather than inherit it. Both paths call one shared helper that
83+
> resolves from provider identity when a `CatalogModel` is present and falls back
84+
> to the slug prefix only when it is not.
85+
5486
Hosted search stays independent:
5587

5688
```ts
@@ -60,14 +92,22 @@ else entry.web_search_tool_type = "text_and_image";
6092

6193
## Template-less path
6294

63-
The fallback in `src/codex/catalog/sync.ts` currently duplicates policy. It must consume the same resolved value:
95+
The fallback in `src/codex/catalog/sync.ts` currently duplicates policy. It must
96+
consume the same resolved value **and the same fence helper** — using
97+
`isCursorFallback` here is what created the asymmetry in the first place:
6498

6599
```ts
66-
const mode = isCursorFallback
100+
const isCursorEntry = isCursorRoute(entry, model?.provider);
101+
const mode = isCursorEntry
67102
? "direct"
68103
: model?.toolDiscoveryMode ?? "deferred";
69104
```
70105

106+
`isCursorRoute()` collapses the old `isCursorFallback`
107+
(`model?.provider === "cursor"`) and the old template-path slug check into one
108+
expression, which is the point of Correction 8: both paths must reach the same
109+
verdict for the same row.
110+
71111
Regression tests must cover both paths because #1596 already established this as a dual seam.
72112

73113
## Combo path

‎devlog/_plan/260813_routed_tool_discovery_profiles/020_phase2_unit_tests.md‎

Lines changed: 19 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -2,6 +2,25 @@
22

33
Status: READY TO IMPLEMENT AFTER PR A
44

5+
> **Added 2026-08-13 (audit blocker 1).** One case in this phase is load-bearing
6+
> for the whole unit's rationale and must not be skipped:
7+
>
8+
> **Single-variable code-mode differential.** Under otherwise byte-identical
9+
> `tool_mode = code_mode_only` conditions, toggle **only** `supports_search_tool`
10+
> and assert an eligible MCP tool remains callable in both states. The expected
11+
> delta is `exec.description` content and size **plus** `tool_search`
12+
> construction and the deferred-guidance text — do not assert the difference is
13+
> confined to `exec.description`. Until this runs, the claim in `004`/`094` that
14+
> `direct` is a comprehension lever rather than a reachability fix rests on a
15+
> source reading of a 2026-07-23 upstream clone rather than on an executed
16+
> differential.
17+
>
18+
> Two further cases from the same audit:
19+
>
20+
> - the load-path warning fires when a malformed discovery value is degraded;
21+
> - the write-boundary validator rejects accessor/prototype-polluted model maps
22+
> *without* first reading the attacker-controlled property.
23+
524
## Test pyramid
625

726
### Layer 1 — pure resolver

‎devlog/_plan/260813_routed_tool_discovery_profiles/094_landing_verification_pass.md‎

Lines changed: 64 additions & 10 deletions
Original file line numberDiff line numberDiff line change
@@ -74,10 +74,16 @@ consequences the plan must absorb:
7474
- The diagnostic in `014` must not claim the resolved catalog mode is the
7575
effective runtime mode. It can only report what OpenCodex advertises.
7676

77-
## Correction 2 — under `code_mode_only`, MCP tools stay reachable in BOTH modes
77+
## Correction 2 — under `code_mode_only`, an *eligible* MCP tool stays reachable in BOTH modes
7878

7979
This is the most important correction, and it strengthens the plan's default.
8080

81+
**Scope, stated first.** This correction holds for an **ordinary eligible MCP
82+
tool** — one that already survived MCP/App visibility and policy filtering, is not
83+
in `direct_only_tool_namespaces`, and is not in `excluded_tool_namespaces`. It is
84+
not an unconditional statement about every tool on every surface. The exclusions
85+
are enumerated below and each is independent of `supports_search_tool`.
86+
8187
`004` implies deferred discovery is what keeps tools reachable. In fact, under
8288
`ToolMode::CodeModeOnly` the nested tool specs are handed to the `exec` handler
8389
and installed as V8 globals in either mode:
@@ -100,15 +106,63 @@ let exposure = if search_tool_enabled { ToolExposure::Deferred } else { ToolExpo
100106

101107
With `supports_search_tool = false`, the tools' **full declarations** move into
102108
`exec.description` because they enter `enabled_tools` rather than `deferred_tools`.
103-
That is precisely the 96,699 → 258,929 character regression #1596 measured. So the
104-
`direct` escape hatch buys **no additional callability under code mode** — it buys
105-
schema text inside `exec.description`.
106-
107-
This reframes the override honestly: it is a *model-comprehension* lever (the model
108-
sees full schemas inline instead of having to consult `ALL_TOOLS`), not a
109-
*reachability* lever. `010`'s non-goal list already says "no claim that `direct` is
110-
cheaper or preferred"; it must also say direct is not a reachability fix under code
111-
mode. `044`'s weak-model fallback rationale is the honest use case.
109+
That is precisely the 96,699 → 258,929 character regression #1596 measured. So for
110+
an eligible tool the `direct` escape hatch buys **no additional callability under
111+
code mode** — it buys schema text inside `exec.description`.
112+
113+
Confirmed by the upstream suite from both directions: a direct-exposure MCP tool
114+
reaches the globals (`core/tests/suite/code_mode.rs:3248`), and a deferred one does
115+
too (`core/tests/suite/code_mode.rs:644`).
116+
117+
### What else the flag changes, and what it never controls
118+
119+
Schema placement is the main difference but not the only one. Deferred exposure
120+
also drives `tool_search` construction and the deferred-tool guidance text
121+
(`spec_plan.rs:932`, `description.rs:10`).
122+
123+
And three mechanisms remove a tool from the Code Mode globals **independently of
124+
this flag** — an override cannot fix any of them, and a reporter hitting one will
125+
look like a discovery failure:
126+
127+
| Mechanism | Effect | Citation |
128+
|---|---|---|
129+
| `direct_only_tool_namespaces` | exposure becomes `DirectModelOnly`; the namespace stays top-level and is **excluded from Code Mode** | `spec_plan.rs:210` |
130+
| `excluded_tool_namespaces` | nested tools removed outright | `spec_plan.rs:444` |
131+
| MCP/App visibility and policy filtering | applied *before* exposure classification, so a filtered tool never reaches either path | `mcp_tool_exposure.rs:20` |
132+
133+
The isolate is not exclusive to `code_mode_only`. `build_code_mode_executors()`
134+
returns early only when the mode is neither:
135+
136+
```rust
137+
if !matches!(tool_mode, ToolMode::CodeMode | ToolMode::CodeModeOnly) {
138+
return vec![];
139+
}
140+
```
141+
142+
`spec_plan.rs:459`. So plain `ToolMode::CodeMode` also builds the `tools`/
143+
`ALL_TOOLS` globals, and the reachability reasoning holds there too; what
144+
`code_mode_only` adds is hiding ordinary nested tools from the **top-level** tool
145+
list. Only under `ToolMode::Direct` is there no isolate — and there
146+
`supports_search_tool` genuinely governs whether MCP tools are declared directly
147+
or must be found through `tool_search`. OpenCodex stamps every routed row
148+
`code_mode_only`, so `Direct` is not a routed-row concern today, but the
149+
distinction matters if that stamp is ever relaxed.
150+
151+
### Required before the claim is load-bearing
152+
153+
`020`/`021` must add one controlled test that toggles **only**
154+
`supports_search_tool` under otherwise identical `code_mode_only` conditions and
155+
asserts the tool remains callable in both, so this correction rests on an
156+
executed differential rather than on a source reading. The expected delta is
157+
`exec.description` content/size **plus** `tool_search` construction and the
158+
deferred-guidance text — not `exec.description` alone.
159+
160+
This reframes the override honestly: for an eligible tool under code mode it is a
161+
*model-comprehension* lever (the model sees full schemas inline instead of having
162+
to consult `ALL_TOOLS`), not a *reachability* lever. `010`'s non-goal list already
163+
says "no claim that `direct` is cheaper or preferred"; it must also say direct is
164+
not a reachability fix for eligible tools under code mode. `044`'s weak-model
165+
fallback rationale is the honest use case.
112166

113167
Caveat recorded rather than assumed away: the inspected clone is dated 2026-07-23,
114168
while #1522 reported against CLI `0.147.0-alpha.6.5` on 2026-08-12. Whether the

‎devlog/_plan/260813_routed_tool_discovery_profiles/KOREAN_SUMMARY.md‎

Lines changed: 10 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -33,8 +33,8 @@ Cursor: direct discovery
3333
- 060–069: `search/describe/call` 3개 meta-tool 폴백
3434
- 070–079: 단계적 배포·관측·canary
3535
- 080–089: 롤백·장애 분류·설정 예시
36-
- 090–093: 최종 권고·PR 스택·완료 기준·실행 순서
37-
- `patches/apply-draft.mjs`: 현재 `dev` 코드 조각이 정확히 존재하는지 먼저 검사한 뒤 패치하는 스크립트
36+
- 090–094: 최종 권고·PR 스택·완료 기준·실행 순서·랜딩 검증
37+
- `patches/apply-draft.mjs`: seam 위치를 읽기 위한 보관용 스크립트 (실행 금지, 아래 참고)
3838
- `prototype/`: 의존성 없는 정책 프로토타입과 벤치마크
3939
- `results/`: 실행 결과와 한계
4040

@@ -61,13 +61,14 @@ Cursor: direct discovery
6161

6262
실제 작업트리에서 실행할 명령은 `scripts/run-repo-validation.sh`에 묶었습니다.
6363

64-
## 적용 순서
64+
## 적용 순서 (2026-08-13 철회됨)
6565

66-
```bash
67-
node patches/apply-draft.mjs --check /path/to/opencodex
68-
node patches/apply-draft.mjs /path/to/opencodex
69-
cd /path/to/opencodex
70-
/path/to/bundle/scripts/run-repo-validation.sh
71-
```
66+
`patches/apply-draft.mjs` 실행 절차는 삭제했습니다. 이 초안은 `094` 정정 8이
67+
지적한 Cursor 판정 불일치를 그대로 재현하면서도 `draft seams: OK`를 출력합니다.
68+
seam 위치를 읽는 참고 자료로만 쓰고, 커밋할 저장소에는 적용하지 마십시오.
69+
자세한 내용은 `patches/README.md`에 있습니다.
70+
71+
실제 구현은 provider 신원을 먼저 보는 공용 헬퍼 하나로 템플릿 경로와
72+
템플릿 없는 경로를 모두 처리하므로, 이 초안과 diff가 일치하지 않습니다.
7273

7374
적용 뒤에는 반드시 `git diff`를 리뷰하고, `042_codex_app_deepseek_browser.md`의 A/B 시나리오로 #1522의 정확한 조합을 확인해야 합니다.

0 commit comments

Comments
 (0)