@vijaypjavvadi/bdd2pw 3.10.0 → 3.12.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +155 -0
- package/README.md +13 -10
- package/dist/cli.js +10 -7
- package/dist/cli.js.map +1 -1
- package/dist/llm/geminiClient.d.ts +86 -0
- package/dist/llm/geminiClient.js +472 -0
- package/dist/llm/geminiClient.js.map +1 -0
- package/dist/llm/index.d.ts +2 -0
- package/dist/llm/index.js +13 -2
- package/dist/llm/index.js.map +1 -1
- package/dist/llm/openaiClient.d.ts +77 -0
- package/dist/llm/openaiClient.js +447 -0
- package/dist/llm/openaiClient.js.map +1 -0
- package/dist/llm/telemetry.js +14 -0
- package/dist/llm/telemetry.js.map +1 -1
- package/dist/llm/types.d.ts +9 -2
- package/dist/types.d.ts +2 -1
- package/package.json +3 -1
package/CHANGELOG.md
CHANGED
|
@@ -9,6 +9,161 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
|
|
|
9
9
|
|
|
10
10
|
_Nothing yet._
|
|
11
11
|
|
|
12
|
+
## [3.12.0] — 2026-06-19
|
|
13
|
+
|
|
14
|
+
> Note: v3.11.0 was tagged in git but never published to npm. v3.12.0
|
|
15
|
+
> on npm is the first release that includes BOTH the v3.11.0 OpenAI
|
|
16
|
+
> provider AND the v3.12.0 Gemini provider. Anyone who installs
|
|
17
|
+
> `@vijaypjavvadi/bdd2pw@latest` after this release gets all three
|
|
18
|
+
> providers (Anthropic, OpenAI, Gemini) with full parity in one go.
|
|
19
|
+
|
|
20
|
+
### Added — Gemini provider, three-provider parity is complete
|
|
21
|
+
|
|
22
|
+
`LLMClientOptions.provider` now accepts `"gemini"` alongside
|
|
23
|
+
`"anthropic"` and `"openai"`. `createLLMClient` factory routes
|
|
24
|
+
`"gemini"` to a new `GeminiLLMClient` that matches the surface
|
|
25
|
+
area of the other two:
|
|
26
|
+
|
|
27
|
+
- Single-step `generateBinding(input)`
|
|
28
|
+
- Batched `generateBatchBindings(inputs[])` — same per-step cache
|
|
29
|
+
lookup + one batched prompt as the v3.5.0 Anthropic batching
|
|
30
|
+
- Cache hit accounting (`callsMade`, `callsAttempted`,
|
|
31
|
+
`cacheBackendPersistent`, `cacheBackendFallbackReason`)
|
|
32
|
+
- Budget enforcement via `maxCalls`
|
|
33
|
+
- Governance sanitisation through the same sidecar (fail-closed
|
|
34
|
+
if the sidecar is unreachable — identical to Anthropic/OpenAI)
|
|
35
|
+
- Provider timeout enforcement — Gemini's SDK has no native
|
|
36
|
+
timeout option, so we wrap the call in a `Promise.race` against
|
|
37
|
+
`setTimeout`; surfaces a clear `Gemini call exceeded Xms timeout`
|
|
38
|
+
error in logs and the result `error` field
|
|
39
|
+
- LLMLogEvent stream that feeds the v3.9.0 telemetry sidecar
|
|
40
|
+
|
|
41
|
+
### Default model: `gemini-2.5-flash`
|
|
42
|
+
|
|
43
|
+
For bdd2pw's structured-JSON task, Flash is the price/perf sweet
|
|
44
|
+
spot. Pricing snapshot: $0.10 per million input tokens, $0.40 per
|
|
45
|
+
million output. That is the cheapest model across all three
|
|
46
|
+
providers — cheaper than `gpt-4o-mini` ($0.15) and dramatically
|
|
47
|
+
cheaper than `claude-sonnet-4-6` ($3.00). Override via
|
|
48
|
+
`LLMClientOptions.model` or `--llm-model`.
|
|
49
|
+
|
|
50
|
+
### Pricing rows added to `DEFAULT_PRICING`
|
|
51
|
+
|
|
52
|
+
`telemetry.ts` now includes pricing snapshots for: `gemini-2.5-flash`,
|
|
53
|
+
`gemini-2.5-pro`, `gemini-2.0-flash`, `gemini-1.5-flash`,
|
|
54
|
+
`gemini-1.5-pro`. v3.9.0 `llm-stats.json` cost estimates work
|
|
55
|
+
end-to-end for Gemini runs.
|
|
56
|
+
|
|
57
|
+
### CLI
|
|
58
|
+
|
|
59
|
+
`bdd2pw scaffold ... --llm gemini` now works. `--llm-model` help
|
|
60
|
+
text updated to list all three provider defaults.
|
|
61
|
+
|
|
62
|
+
### Auth env vars
|
|
63
|
+
|
|
64
|
+
Reads `GEMINI_API_KEY` first, then falls back to `GOOGLE_API_KEY`
|
|
65
|
+
(both are widely used in the Google AI Studio + Vertex ecosystem).
|
|
66
|
+
|
|
67
|
+
### `optionalDependencies`
|
|
68
|
+
|
|
69
|
+
`@google/generative-ai ^0.21.0` added alongside `@anthropic-ai/sdk`
|
|
70
|
+
and `openai`. Same defensive `require()` loading pattern — teams
|
|
71
|
+
that only use Anthropic do not need to install the Gemini SDK,
|
|
72
|
+
and vice versa. Missing-package error message points to
|
|
73
|
+
`npm install @google/generative-ai`.
|
|
74
|
+
|
|
75
|
+
### Cache compat
|
|
76
|
+
|
|
77
|
+
Cache keys include the model name (since v2.0), so a scaffold
|
|
78
|
+
cached under `claude-sonnet-4-6` does NOT hit the cache when re-run
|
|
79
|
+
under `gemini-2.5-flash` — you get a fresh provider call. Within
|
|
80
|
+
the same model, cache hits work identically regardless of how many
|
|
81
|
+
times you switch providers across scaffold runs.
|
|
82
|
+
|
|
83
|
+
### Tests
|
|
84
|
+
|
|
85
|
+
`tests/unit/v3120Gemini.test.ts`:
|
|
86
|
+
- Default model + DEFAULT_PRICING lookup.
|
|
87
|
+
- generateBinding succeeds with a well-formed JSON response.
|
|
88
|
+
- generateBatchBindings produces N bindings from ONE provider call
|
|
89
|
+
(cache misses → single batched prompt → JSON array → N bindings).
|
|
90
|
+
- Budget exhaustion returns an error without making a provider call.
|
|
91
|
+
- Missing API key short-circuits before the SDK loads.
|
|
92
|
+
- Malformed JSON response surfaces a parse error.
|
|
93
|
+
- Provider timeout aborts a hung call (Promise.race against
|
|
94
|
+
setTimeout works as designed).
|
|
95
|
+
|
|
96
|
+
## [3.11.0] — 2026-06-08
|
|
97
|
+
|
|
98
|
+
### Added — OpenAI provider parity
|
|
99
|
+
|
|
100
|
+
`LLMClientOptions.provider` now accepts `"openai"` alongside
|
|
101
|
+
`"anthropic"`. The factory wires up `OpenAILLMClient` which matches
|
|
102
|
+
the Anthropic client's surface area completely:
|
|
103
|
+
|
|
104
|
+
- Single-step `generateBinding(input)`
|
|
105
|
+
- Batched `generateBatchBindings(inputs[])` — same per-step cache
|
|
106
|
+
lookup + one batched prompt as v3.5.0 batching
|
|
107
|
+
- Cache hit accounting (`callsMade`, `callsAttempted`,
|
|
108
|
+
`cacheBackendPersistent`, `cacheBackendFallbackReason`)
|
|
109
|
+
- Budget enforcement via `maxCalls`
|
|
110
|
+
- Governance sanitisation through the same sidecar
|
|
111
|
+
- Step deadlines + provider timeouts
|
|
112
|
+
- LLMLogEvent stream that feeds the v3.9.0 telemetry sidecar
|
|
113
|
+
|
|
114
|
+
What differs is purely the SDK call shape — OpenAI's
|
|
115
|
+
`chat.completions.create()` instead of Anthropic's `messages.create()`,
|
|
116
|
+
and the response unpacking (`choices[0].message.content` vs
|
|
117
|
+
`content[0].text`, `usage.prompt_tokens` vs `usage.input_tokens`).
|
|
118
|
+
|
|
119
|
+
### Default model: `gpt-4o-mini`
|
|
120
|
+
|
|
121
|
+
bdd2pw's task is structured JSON output over regex-shaped
|
|
122
|
+
deterministic patterns. Mini handles it reliably and runs ~17x
|
|
123
|
+
cheaper than gpt-4o ($0.15 vs $2.50 per million input tokens).
|
|
124
|
+
Override via `LLMClientOptions.model` or `--llm-model`.
|
|
125
|
+
|
|
126
|
+
### Pricing rows added to `DEFAULT_PRICING`
|
|
127
|
+
|
|
128
|
+
`telemetry.ts` now includes pricing snapshots for: `gpt-4o`,
|
|
129
|
+
`gpt-4o-mini`, `gpt-4-turbo`, `o1`, `o1-mini`, `o1-preview`. v3.9.0
|
|
130
|
+
`llm-stats.json` cost estimates work end-to-end for OpenAI runs.
|
|
131
|
+
Prices are Q4 2025 published rates; override via
|
|
132
|
+
`new LLMTelemetry(scaffoldId, version, pricingOverride)` if you
|
|
133
|
+
have negotiated rates.
|
|
134
|
+
|
|
135
|
+
### CLI
|
|
136
|
+
|
|
137
|
+
`bdd2pw scaffold ... --llm openai` now works. The `--llm-model`
|
|
138
|
+
help text updated to reflect both providers' defaults.
|
|
139
|
+
|
|
140
|
+
### `optionalDependencies`
|
|
141
|
+
|
|
142
|
+
`openai ^4.77.0` added alongside `@anthropic-ai/sdk`. Like the
|
|
143
|
+
Anthropic SDK, it's loaded defensively via `require()` so users who
|
|
144
|
+
only want Anthropic don't need to install the OpenAI SDK and vice
|
|
145
|
+
versa. Missing-package error message points to `npm install openai`.
|
|
146
|
+
|
|
147
|
+
### Cache compat
|
|
148
|
+
|
|
149
|
+
Cache keys include the model name (since v2.0), so a scaffold
|
|
150
|
+
cached under `claude-sonnet-4-6` will NOT hit the cache when run
|
|
151
|
+
under `gpt-4o-mini` — you get a fresh provider call. Within the
|
|
152
|
+
same model, cache hits work identically regardless of how many
|
|
153
|
+
times you switch providers across scaffold runs.
|
|
154
|
+
|
|
155
|
+
### Tests
|
|
156
|
+
|
|
157
|
+
`tests/unit/v3110OpenAI.test.ts`:
|
|
158
|
+
- Default model + DEFAULT_PRICING lookup.
|
|
159
|
+
- generateBinding succeeds with a well-formed JSON response.
|
|
160
|
+
- generateBatchBindings produces N bindings from ONE provider
|
|
161
|
+
call (cache misses → single batched prompt → JSON array → N
|
|
162
|
+
bindings).
|
|
163
|
+
- Budget exhaustion returns an error without making a provider call.
|
|
164
|
+
- Missing API key short-circuits before the SDK loads.
|
|
165
|
+
- Malformed JSON response surfaces a parse error.
|
|
166
|
+
|
|
12
167
|
## [3.10.0] — 2026-06-06
|
|
13
168
|
|
|
14
169
|
### Added — self-healing stats sidecar + `bdd2pw heal-stats` CLI
|
package/README.md
CHANGED
|
@@ -42,9 +42,12 @@ Point it at a `.feature` file and a URL. It parses the Gherkin, scans the live p
|
|
|
42
42
|
|
|
43
43
|
> **The pitch.** Run `bdd2pw scaffold` and you get a Playwright TS repo where `npx playwright test` runs against the real site. For specific fixtures, all green. For vague ones, the 60-90% you'd otherwise hand-write is done; you finish the rest — or you let the governed LLM finish them.
|
|
44
44
|
|
|
45
|
-
## What's new in v3.
|
|
45
|
+
## What's new in v3.12.0
|
|
46
46
|
|
|
47
|
-
- **NEW** `
|
|
47
|
+
- **NEW** Gemini provider — three-provider parity is now complete. `--llm gemini` or `LLMClientOptions.provider: "gemini"` defaults to `gemini-2.5-flash` ($0.10/M input — the cheapest model across all three providers). Same cache, governance, batching, and telemetry pipeline as Anthropic and OpenAI. Install only when needed via `npm install @google/generative-ai`.
|
|
48
|
+
- OpenAI provider with full parity (cache, governance, batching, telemetry). `--llm openai` defaults to `gpt-4o-mini` (v3.11.0).
|
|
49
|
+
- `bdd2pw heal-stats <repo>` CLI reads `<repo>/artefacts/heal-events.jsonl` and writes a heal-stats.json sidecar with heal rate, top failing fields, retry latency, candidate-selector promotions (v3.10.0).
|
|
50
|
+
- `llmStats: true` writes `<repo>/artefacts/llm-stats.json` with per-call latency, token counts, cache hit rate, and an estimated cost in USD (v3.9.0).
|
|
48
51
|
- Seven domain rule packs (banking, healthcare, insurance, retail, gov, education, telecom) — ~140 industry-specific patterns total.
|
|
49
52
|
- Per-scenario LLM batching cuts Anthropic spend by ~75% on unmatched-heavy scaffolds.
|
|
50
53
|
- `bdd2pw propose-rules` CLI clusters past LLM bindings into draft regex proposals — turning the offline-review pipeline into one command.
|
|
@@ -348,23 +351,23 @@ output/
|
|
|
348
351
|
| v3.6 | Diagnostics + auto-rules | Rule-trace block in `BDD_REVIEW.md`; new `bdd2pw propose-rules` CLI clustering candidate-rules.jsonl into draft regex |
|
|
349
352
|
| v3.7.1 | Regression fixes | PascalCase className shadow fix; `test.step` wrapping regression-asserted |
|
|
350
353
|
| v3.8.1 | More domain packs | Retail / gov / education / telecom packs added — seven total (~140 industry-specific rules) |
|
|
351
|
-
|
|
|
354
|
+
| v3.9.0 | Telemetry | `artefacts/llm-stats.json` sidecar — per-call latency + tokens + cache hit rate + estimated cost. Makes v3.5 batching ROI measurable per scaffold. |
|
|
355
|
+
| v3.10.0 | Heal stats | `bdd2pw heal-stats <repo>` CLI consumes `artefacts/heal-events.jsonl` and writes `heal-stats.json` with heal rate, top failing fields, top error patterns, retry latency, candidate-selector promotions. Self-healing ROI per test run. |
|
|
356
|
+
| v3.11.0 | OpenAI provider | `LLMClientOptions.provider: "openai"` lands with cache + governance + batching + telemetry parity. Default model `gpt-4o-mini` (~17x cheaper than `gpt-4o` for structured-JSON tasks). |
|
|
357
|
+
| **v3.12.0 (current)** | Gemini provider | `LLMClientOptions.provider: "gemini"` lands with full parity. Default model `gemini-2.5-flash` ($0.10/M input — cheapest across all three providers). Three-provider parity complete. |
|
|
352
358
|
|
|
353
359
|
### Next 6 months (Jun 2026 → Nov 2026, ~2 releases/month)
|
|
354
360
|
|
|
355
361
|
| Target | Version | Theme | Headline |
|
|
356
362
|
|---|---|---|---|
|
|
357
|
-
|
|
|
358
|
-
| Jul 2026 | **v3.11.0** | Gemini provider | `LLMClientOptions.provider: "gemini"` lands; full three-provider parity (Anthropic / OpenAI / Gemini). |
|
|
359
|
-
| Jul 2026 | **v3.12.0** | Auto-rules | `bdd2pw apply-proposals --interactive` — walks the human through each v3.6 proposal, validates the regex compiles + matches the sample texts, optionally appends to `stepMatcher.ts`. Closes the propose-rules loop. |
|
|
363
|
+
| Aug 2026 | **v3.13.0** | Auto-rules | `bdd2pw apply-proposals --interactive` — walks the human through each v3.6 proposal, validates the regex compiles + matches the sample texts, optionally appends to `stepMatcher.ts`. Closes the propose-rules loop. |
|
|
360
364
|
| Aug 2026 | **v4.0.0** | Modern defaults | Semver-major reset: bump Anthropic default model, retire v2.x deprecated APIs, switch test-signature defaults to current best practice. Migration guide ships alongside. |
|
|
361
|
-
|
|
|
365
|
+
| Sep 2026 | **v4.1.0** | sel2pw merge | sel2pw migrated onto `@vijaypjavvadi/pw-emit` — three packages share one emitter. All future v3.x emitter improvements automatically flow to Selenium users. |
|
|
362
366
|
| Sep 2026 | **v4.2.0** | Domain packs | Three more opt-in packs: **fintech** (KYC, AML, trading desks), **real-estate** (MLS, listings, escrow), **hospitality** (PMS, reservations, OTA). ~60 more rules. |
|
|
363
|
-
|
|
|
367
|
+
| Oct 2026 | **v4.3.0** | API v2 | WebSocket assertion rules, multipart file-upload patterns, GraphQL query / mutation / subscription rules. Extends v3.0's API testing surface. |
|
|
364
368
|
| Oct 2026 | **v4.4.0** | Coverage analyzer | `bdd2pw coverage` CLI: which deterministic rules fired on which features across a fleet, per-pack hit rate, dead-rule report, suggestion to retire unused rules. |
|
|
365
|
-
|
|
|
369
|
+
| Nov 2026 | **v4.5.0** | CI/CD templates | Optional scaffolder output: GitHub Actions / GitLab CI / Azure Pipelines / CircleCI starter workflows wired to run the generated specs. Opt-in via `--ci <provider>`. |
|
|
366
370
|
| Nov 2026 | **v4.6.0** | Mobile dialect | Touch / swipe / scroll-into-view rules + responsive viewport assertions. Reuses the same emitter pipeline; opt-in via `domains: ["mobile"]`. |
|
|
367
|
-
| Nov 2026 | **v4.7.0** | i18n Gherkin | Accept Cucumber-compatible non-English keywords (Spanish, German, French, Portuguese, Japanese, …). Lets bdd2pw scaffold from `.feature` files written in any of the ~70 languages Cucumber supports. |
|
|
368
371
|
|
|
369
372
|
Each release is shaped to be self-contained, ship in ~2 weeks, and keep `domains: []` byte-stable for existing users so the upgrade path stays painless.
|
|
370
373
|
|
package/dist/cli.js
CHANGED
|
@@ -61,9 +61,9 @@ program
|
|
|
61
61
|
.option("--headed", "Show browser during MCP scan", false)
|
|
62
62
|
.option("--snapshot-file <path>", "Read elements from a JSON file instead of launching a browser")
|
|
63
63
|
.option("--no-discovery", "Skip page discovery entirely (rule-only probing). Field-referencing rules will fall to TODO.")
|
|
64
|
-
.option("--llm <provider>", "Enable LLM fallback for unmatched steps.
|
|
64
|
+
.option("--llm <provider>", "Enable LLM fallback for unmatched steps. v3.12.0: 'anthropic', 'openai', or 'gemini'.")
|
|
65
65
|
.option("--governance-url <url>", "ai-governance sidecar URL — every prompt is sanitised here before leaving the perimeter (fail-closed).", "http://localhost:4900")
|
|
66
|
-
.option("--llm-model <model>", "Override the LLM model
|
|
66
|
+
.option("--llm-model <model>", "Override the LLM model. Defaults: anthropic→claude-sonnet-4-6, openai→gpt-4o-mini, gemini→gemini-2.5-flash.")
|
|
67
67
|
.option("--llm-max-calls <n>", "Max LLM provider calls per scaffold. Default 50. Cache hits don't count.", "50")
|
|
68
68
|
.option("--llm-cache <path>", "SQLite cache path. Default <repo>/.bdd2pw/llm-cache.sqlite. Use ':memory:' for one-shot.")
|
|
69
69
|
.option("--llm-skip-governance", "DO NOT USE in production — bypass the sidecar sanitisation step. Test-only escape hatch.", false)
|
|
@@ -81,10 +81,13 @@ program
|
|
|
81
81
|
try {
|
|
82
82
|
// v2.0 — wire the actual LLM config when --llm is passed. The legacy
|
|
83
83
|
// top-level `llm` field stays for backwards compat; `llmConfig` is the
|
|
84
|
-
// real one used by scaffold().
|
|
85
|
-
const
|
|
86
|
-
?
|
|
87
|
-
|
|
84
|
+
// real one used by scaffold(). v3.11.0 — accept openai too.
|
|
85
|
+
const llmProvider = opts.llm === "anthropic" || opts.llm === "openai" || opts.llm === "gemini"
|
|
86
|
+
? opts.llm
|
|
87
|
+
: null;
|
|
88
|
+
const llmConfig = llmProvider
|
|
89
|
+
? ({
|
|
90
|
+
provider: llmProvider,
|
|
88
91
|
model: opts.llmModel,
|
|
89
92
|
governanceUrl: opts.governanceUrl,
|
|
90
93
|
maxCalls: opts.llmMaxCalls
|
|
@@ -102,7 +105,7 @@ program
|
|
|
102
105
|
governanceTimeoutMs: opts.llmGovernanceTimeoutMs
|
|
103
106
|
? Number(opts.llmGovernanceTimeoutMs)
|
|
104
107
|
: undefined,
|
|
105
|
-
}
|
|
108
|
+
})
|
|
106
109
|
: undefined;
|
|
107
110
|
const result = await (0, index_1.scaffold)({
|
|
108
111
|
feature: path.resolve(feature),
|
package/dist/cli.js.map
CHANGED
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"cli.js","sourceRoot":"","sources":["../src/cli.ts"],"names":[],"mappings":";;AACA;;;GAGG;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAEH,yCAAoC;AACpC,2CAA6B;AAC7B,mCAAuD;AACvD,+BAAqC;AACrC,mDAAuD;AACvD,2CAAwC;AAExC,MAAM,OAAO,GAAG,IAAI,mBAAO,EAAE,CAAC;AAE9B,OAAO;KACJ,IAAI,CAAC,QAAQ,CAAC;KACd,WAAW,CACV,4EAA4E,CAC7E;KACA,OAAO,CAAC,OAAO,CAAC,iBAAiB,CAAC,CAAC,OAAO,CAAC,CAAC;AAE/C,OAAO;KACJ,OAAO,CAAC,UAAU,CAAC;KACnB,WAAW,CACV,iFAAiF,CAClF;KACA,QAAQ,CAAC,WAAW,EAAE,iCAAiC,CAAC;KACxD,cAAc,CAAC,aAAa,EAAE,4BAA4B,CAAC;KAC3D,cAAc,CAAC,eAAe,EAAE,qCAAqC,CAAC;KACtE,cAAc,CAAC,cAAc,EAAE,gDAAgD,CAAC;KAChF,MAAM,CAAC,gBAAgB,EAAE,sDAAsD,CAAC;KAChF,MAAM,CAAC,wBAAwB,EAAE,sCAAsC,CAAC;KACxE,MAAM,CAAC,UAAU,EAAE,8BAA8B,EAAE,KAAK,CAAC;KACzD,MAAM,CAAC,wBAAwB,EAAE,+DAA+D,CAAC;KACjG,MAAM,CACL,gBAAgB,EAChB,8FAA8F,CAC/F;KACA,MAAM,CAAC,kBAAkB,EAAE,
|
|
1
|
+
{"version":3,"file":"cli.js","sourceRoot":"","sources":["../src/cli.ts"],"names":[],"mappings":";;AACA;;;GAGG;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAEH,yCAAoC;AACpC,2CAA6B;AAC7B,mCAAuD;AACvD,+BAAqC;AACrC,mDAAuD;AACvD,2CAAwC;AAExC,MAAM,OAAO,GAAG,IAAI,mBAAO,EAAE,CAAC;AAE9B,OAAO;KACJ,IAAI,CAAC,QAAQ,CAAC;KACd,WAAW,CACV,4EAA4E,CAC7E;KACA,OAAO,CAAC,OAAO,CAAC,iBAAiB,CAAC,CAAC,OAAO,CAAC,CAAC;AAE/C,OAAO;KACJ,OAAO,CAAC,UAAU,CAAC;KACnB,WAAW,CACV,iFAAiF,CAClF;KACA,QAAQ,CAAC,WAAW,EAAE,iCAAiC,CAAC;KACxD,cAAc,CAAC,aAAa,EAAE,4BAA4B,CAAC;KAC3D,cAAc,CAAC,eAAe,EAAE,qCAAqC,CAAC;KACtE,cAAc,CAAC,cAAc,EAAE,gDAAgD,CAAC;KAChF,MAAM,CAAC,gBAAgB,EAAE,sDAAsD,CAAC;KAChF,MAAM,CAAC,wBAAwB,EAAE,sCAAsC,CAAC;KACxE,MAAM,CAAC,UAAU,EAAE,8BAA8B,EAAE,KAAK,CAAC;KACzD,MAAM,CAAC,wBAAwB,EAAE,+DAA+D,CAAC;KACjG,MAAM,CACL,gBAAgB,EAChB,8FAA8F,CAC/F;KACA,MAAM,CAAC,kBAAkB,EAAE,uFAAuF,CAAC;KACnH,MAAM,CACL,wBAAwB,EACxB,wGAAwG,EACxG,uBAAuB,CACxB;KACA,MAAM,CAAC,qBAAqB,EAAE,6GAA6G,CAAC;KAC5I,MAAM,CAAC,qBAAqB,EAAE,0EAA0E,EAAE,IAAI,CAAC;KAC/G,MAAM,CAAC,oBAAoB,EAAE,0FAA0F,CAAC;KACxH,MAAM,CAAC,uBAAuB,EAAE,0FAA0F,EAAE,KAAK,CAAC;KAClI,MAAM,CAAC,2BAA2B,EAAE,6GAA6G,EAAE,OAAO,CAAC;KAC3J,MAAM,CAAC,+BAA+B,EAAE,8DAA8D,EAAE,OAAO,CAAC;KAChH,MAAM,CAAC,iCAAiC,EAAE,4DAA4D,EAAE,OAAO,CAAC;KAChH,MAAM,CAAC,mBAAmB,EAAE,6CAA6C,CAAC;KAC1E,MAAM,CAAC,WAAW,EAAE,2BAA2B,EAAE,KAAK,CAAC;KACvD,MAAM,CAAC,eAAe,EAAE,wBAAwB,CAAC;KACjD,MAAM,CAAC,aAAa,EAAE,uCAAuC,EAAE,KAAK,CAAC;KACrE,MAAM,CAAC,SAAS,EAAE,+BAA+B,EAAE,KAAK,CAAC;KACzD,MAAM,CACL,gBAAgB,EAChB,kNAAkN,EAClN,KAAK,CACN;KACA,MAAM,CACL,aAAa,EACb,yHAAyH,EACzH,KAAK,CACN;KACA,MAAM,CAAC,KAAK,EAAE,OAAe,EAAE,IAAI,EAAE,EAAE;IACtC,IAAI,CAAC;QACH,qEAAqE;QACrE,uEAAuE;QACvE,4DAA4D;QAC5D,MAAM,WAAW,GACf,IAAI,CAAC,GAAG,KAAK,WAAW,IAAI,IAAI,CAAC,GAAG,KAAK,QAAQ,IAAI,IAAI,CAAC,GAAG,KAAK,QAAQ;YACxE,CAAC,CAAC,IAAI,CAAC,GAAG;YACV,CAAC,CAAC,IAAI,CAAC;QACX,MAAM,SAAS,GAAG,WAAW;YAC3B,CAAC,CAAC,CAAC;gBACG,QAAQ,EAAE,WAAgD;gBAC1D,KAAK,EAAE,IAAI,CAAC,QAAQ;gBACpB,aAAa,EAAE,IAAI,CAAC,aAAa;gBACjC,QAAQ,EAAE,IAAI,CAAC,WAAW;oBACxB,CAAC,CAAC,MAAM,CAAC,IAAI,CAAC,WAAW,CAAC;oBAC1B,CAAC,CAAC,SAAS;gBACb,SAAS,EAAE,IAAI,CAAC,QAAQ;gBACxB,cAAc,EAAE,IAAI,CAAC,iBAAiB;gBACtC,mBAAmB;gBACnB,aAAa,EAAE,IAAI,CAAC,gBAAgB;oBAClC,CAAC,CAAC,MAAM,CAAC,IAAI,CAAC,gBAAgB,CAAC;oBAC/B,CAAC,CAAC,SAAS;gBACb,iBAAiB,EAAE,IAAI,CAAC,oBAAoB;oBAC1C,CAAC,CAAC,MAAM,CAAC,IAAI,CAAC,oBAAoB,CAAC;oBACnC,CAAC,CAAC,SAAS;gBACb,mBAAmB,EAAE,IAAI,CAAC,sBAAsB;oBAC9C,CAAC,CAAC,MAAM,CAAC,IAAI,CAAC,sBAAsB,CAAC;oBACrC,CAAC,CAAC,SAAS;aACd,CAAC;YACN,CAAC,CAAC,SAAS,CAAC;QACd,MAAM,MAAM,GAAG,MAAM,IAAA,gBAAQ,EAAC;YAC5B,OAAO,EAAE,IAAI,CAAC,OAAO,CAAC,OAAO,CAAC;YAC9B,GAAG,EAAE,IAAI,CAAC,GAAG;YACb,IAAI,EAAE,IAAI,CAAC,IAAI;YACf,IAAI,EAAE,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC;YAC7B,KAAK,EAAE,IAAI,CAAC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,CAAS,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC;YAC1D,YAAY,EAAE,IAAI,CAAC,YAAY;YAC/B,MAAM,EAAE,IAAI,CAAC,MAAM;YACnB,GAAG,EAAE,IAAI,CAAC,GAAG;YACb,SAAS;YACT,aAAa,EAAE,IAAI,CAAC,aAAa;YACjC,SAAS,EAAE,IAAI,CAAC,SAAS;YACzB,MAAM,EAAE,IAAI,CAAC,MAAM;YACnB,UAAU,EAAE,IAAI,CAAC,UAAU,KAAK,KAAK;YACrC,SAAS,EAAE,IAAI,CAAC,SAAS;YACzB,KAAK,EAAE,IAAI,CAAC,KAAK;YACjB,YAAY,EAAE,IAAI,CAAC,YAAY;YAC/B,6DAA6D;YAC7D,WAAW,EAAE,IAAI,CAAC,SAAS,KAAK,KAAK;YACrC,WAAW,EAAE,IAAI,CAAC,WAAW;YAC7B,6BAA6B;YAC7B,QAAQ,EAAE,IAAI,CAAC,QAAQ,KAAK,IAAI;SACjC,CAAC,CAAC;QACH,eAAM,CAAC,IAAI,CAAC,EAAE,MAAM,EAAE,EAAE,mBAAmB,CAAC,CAAC;IAC/C,CAAC;IAAC,OAAO,GAAG,EAAE,CAAC;QACb,eAAM,CAAC,KAAK,CAAC,EAAE,GAAG,EAAE,EAAE,iBAAiB,CAAC,CAAC;QACzC,OAAO,CAAC,IAAI,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC,CAAC;IACjC,CAAC;AACH,CAAC,CAAC,CAAC;AAEL,OAAO;KACJ,OAAO,CAAC,SAAS,CAAC;KAClB,WAAW,CAAC,qEAAqE,CAAC;KAClF,QAAQ,CAAC,WAAW,EAAE,iCAAiC,CAAC;KACxD,cAAc,CAAC,aAAa,EAAE,aAAa,CAAC;KAC5C,MAAM,CAAC,wBAAwB,EAAE,sCAAsC,CAAC;KACxE,MAAM,CAAC,UAAU,EAAE,8BAA8B,EAAE,KAAK,CAAC;KACzD,MAAM,CAAC,KAAK,EAAE,OAAe,EAAE,IAAI,EAAE,EAAE;IACtC,IAAI,CAAC;QACH,MAAM,MAAM,GAAG,MAAM,IAAA,eAAO,EAAC;YAC3B,OAAO,EAAE,IAAI,CAAC,OAAO,CAAC,OAAO,CAAC;YAC9B,GAAG,EAAE,IAAI,CAAC,GAAG;YACb,YAAY,EAAE,IAAI,CAAC,YAAY;YAC/B,MAAM,EAAE,IAAI,CAAC,MAAM;SACpB,CAAC,CAAC;QACH,OAAO,CAAC,GAAG,CAAC,IAAI,CAAC,SAAS,CAAC,MAAM,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC;IAC/C,CAAC;IAAC,OAAO,GAAG,EAAE,CAAC;QACb,eAAM,CAAC,KAAK,CAAC,EAAE,GAAG,EAAE,EAAE,gBAAgB,CAAC,CAAC;QACxC,OAAO,CAAC,IAAI,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC,CAAC;IACjC,CAAC;AACH,CAAC,CAAC,CAAC;AAEL,OAAO;KACJ,OAAO,CAAC,YAAY,CAAC;KACrB,WAAW,CAAC,oEAAoE,CAAC;KACjF,cAAc,CAAC,eAAe,EAAE,wBAAwB,CAAC;KACzD,cAAc,CAAC,aAAa,EAAE,gBAAgB,CAAC;KAC/C,cAAc,CAAC,cAAc,EAAE,6BAA6B,CAAC;KAC7D,MAAM,CAAC,wBAAwB,EAAE,sCAAsC,CAAC;KACxE,MAAM,CAAC,UAAU,EAAE,8BAA8B,EAAE,KAAK,CAAC;KACzD,MAAM,CAAC,mBAAmB,EAAE,6CAA6C,CAAC;KAC1E,MAAM,CAAC,KAAK,EAAE,IAAI,EAAE,EAAE;IACrB,IAAI,CAAC;QACH,MAAM,MAAM,GAAG,MAAM,IAAA,iBAAS,EAAC;YAC7B,IAAI,EAAE,IAAI,CAAC,IAAI;YACf,GAAG,EAAE,IAAI,CAAC,GAAG;YACb,IAAI,EAAE,IAAI,CAAC,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC;YAC7B,YAAY,EAAE,IAAI,CAAC,YAAY;YAC/B,MAAM,EAAE,IAAI,CAAC,MAAM;YACnB,SAAS,EAAE,IAAI,CAAC,SAAS;SAC1B,CAAC,CAAC;QACH,eAAM,CAAC,IAAI,CAAC,EAAE,MAAM,EAAE,EAAE,qBAAqB,CAAC,CAAC;IACjD,CAAC;IAAC,OAAO,GAAG,EAAE,CAAC;QACb,eAAM,CAAC,KAAK,CAAC,EAAE,GAAG,EAAE,EAAE,mBAAmB,CAAC,CAAC;QAC3C,OAAO,CAAC,IAAI,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC,CAAC;IACjC,CAAC;AACH,CAAC,CAAC,CAAC;AAEL,OAAO;KACJ,OAAO,CAAC,OAAO,CAAC;KAChB,WAAW,CAAC,gCAAgC,CAAC;KAC7C,MAAM,CAAC,eAAe,EAAE,mBAAmB,EAAE,MAAM,CAAC;KACpD,MAAM,CAAC,KAAK,EAAE,IAAI,EAAE,EAAE;IACrB,MAAM,EAAE,WAAW,EAAE,GAAG,wDAAa,UAAU,GAAC,CAAC;IACjD,MAAM,WAAW,CAAC,EAAE,IAAI,EAAE,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;AACjD,CAAC,CAAC,CAAC;AAEL;;GAEG;AACH,SAAS,WAAW,CAAC,GAAY;IAC/B,MAAM,IAAI,GAAI,GAAyB,EAAE,IAAI,IAAI,EAAE,CAAC;IACpD,QAAQ,IAAI,EAAE,CAAC;QACb,KAAK,mBAAmB;YACtB,OAAO,CAAC,CAAC;QACX,KAAK,iBAAiB;YACpB,OAAO,CAAC,CAAC;QACX,KAAK,UAAU,CAAC;QAChB,KAAK,iBAAiB;YACpB,OAAO,CAAC,CAAC;QACX,KAAK,yBAAyB;YAC5B,OAAO,CAAC,CAAC;QACX;YACE,OAAO,CAAC,CAAC;IACb,CAAC;AACH,CAAC;AAED;;;;;;;GAOG;AACH,OAAO;KACJ,OAAO,CAAC,eAAe,CAAC;KACxB,WAAW,CACV,yGAAyG,CAC1G;KACA,QAAQ,CACP,SAAS,EACT,kEAAkE,CACnE;KACA,MAAM,CACL,cAAc,EACd,iEAAiE,CAClE;KACA,MAAM,CACL,wBAAwB,EACxB,qDAAqD,EACrD,GAAG,CACJ;KACA,MAAM,CAAC,KAAK,EAAE,KAAa,EAAE,GAA8C,EAAE,EAAE;IAC9E,MAAM,MAAM,GAAG,MAAM,IAAA,kBAAY,EAAC;QAChC,SAAS,EAAE,KAAK;QAChB,UAAU,EAAE,GAAG,CAAC,GAAG;QACnB,cAAc,EAAE,GAAG,CAAC,cAAc;YAChC,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC,cAAc,CAAC;YAC5B,CAAC,CAAC,SAAS;KACd,CAAC,CAAC;IACH,eAAM,CAAC,IAAI,CACT;QACE,UAAU,EAAE,MAAM,CAAC,UAAU;QAC7B,gBAAgB,EAAE,MAAM,CAAC,gBAAgB;QACzC,eAAe,EAAE,MAAM,CAAC,eAAe;KACxC,EACD,yBAAyB,CAC1B,CAAC;IACF,OAAO,CAAC,MAAM,CAAC,KAAK,CAClB,KAAK,MAAM,CAAC,gBAAgB,2BAA2B,MAAM,CAAC,UAAU,GAAG;QACzE,SAAS,MAAM,CAAC,eAAe,wBAAwB,CAC1D,CAAC;AACJ,CAAC,CAAC,CAAC;AAEL;;;;;;;GAOG;AACH,OAAO;KACJ,OAAO,CAAC,YAAY,CAAC;KACrB,WAAW,CACV,8EAA8E,CAC/E;KACA,QAAQ,CACP,SAAS,EACT,mFAAmF,CACpF;KACA,MAAM,CACL,cAAc,EACd,sEAAsE,CACvE;KACA,MAAM,CACL,WAAW,EACX,2GAA2G,EAC3G,IAAI,CACL;KACA,MAAM,CAAC,KAAK,EAAE,KAAa,EAAE,GAAmC,EAAE,EAAE;IACnE,MAAM,MAAM,GAAG,MAAM,IAAA,4BAAgB,EAAC;QACpC,SAAS,EAAE,KAAK;QAChB,UAAU,EAAE,GAAG,CAAC,GAAG;QACnB,IAAI,EAAE,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,SAAS;KAC5C,CAAC,CAAC;IACH,eAAM,CAAC,IAAI,CACT;QACE,UAAU,EAAE,MAAM,CAAC,UAAU;QAC7B,WAAW,EAAE,MAAM,CAAC,WAAW;KAChC,EACD,sBAAsB,CACvB,CAAC;IACF,OAAO,CAAC,MAAM,CAAC,KAAK,CAClB,uBAAuB,MAAM,CAAC,UAAU,YAAY,MAAM,CAAC,WAAW,eAAe,CACtF,CAAC;AACJ,CAAC,CAAC,CAAC;AAEL,OAAO,CAAC,UAAU,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,KAAK,CAAC,CAAC,GAAG,EAAE,EAAE;IAC7C,eAAM,CAAC,KAAK,CAAC,EAAE,GAAG,EAAE,EAAE,OAAO,CAAC,CAAC;IAC/B,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;AAClB,CAAC,CAAC,CAAC"}
|
|
@@ -0,0 +1,86 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Gemini LLM client — v3.12.0.
|
|
3
|
+
*
|
|
4
|
+
* Implements the same `LLMClient` interface as `AnthropicLLMClient` and
|
|
5
|
+
* `OpenAILLMClient`: single-step `generateBinding`, batched
|
|
6
|
+
* `generateBatchBindings`, cache hit accounting, budget tracking,
|
|
7
|
+
* governance sanitisation, step deadlines, provider timeouts.
|
|
8
|
+
*
|
|
9
|
+
* SDK-shape differences from Anthropic / OpenAI:
|
|
10
|
+
*
|
|
11
|
+
* Anthropic OpenAI Gemini
|
|
12
|
+
* -------------------------------- ------------------------------ ------------------------------------
|
|
13
|
+
* anthropic.messages.create({ openai.chat.completions.create({ model.generateContent({
|
|
14
|
+
* system: SYSTEM_PROMPT, messages: [ systemInstruction: SYSTEM_PROMPT,
|
|
15
|
+
* messages: [ { role: 'system', ... }, contents: [
|
|
16
|
+
* { role: 'user', content } { role: 'user', ... } { role: 'user',
|
|
17
|
+
* ] ] parts: [{ text: prompt }] }
|
|
18
|
+
* }) }) ]
|
|
19
|
+
* })
|
|
20
|
+
*
|
|
21
|
+
* resp.content[0].text resp.choices[0].message.content resp.response.text()
|
|
22
|
+
* resp.usage.input_tokens resp.usage.prompt_tokens resp.response.usageMetadata.promptTokenCount
|
|
23
|
+
* resp.usage.output_tokens resp.usage.completion_tokens resp.response.usageMetadata.candidatesTokenCount
|
|
24
|
+
*
|
|
25
|
+
* The prompt machinery (`SYSTEM_PROMPT`, `buildUserPrompt`,
|
|
26
|
+
* `buildBatchUserPrompt`, `cacheKey`), the parser (`parseBindingJson`),
|
|
27
|
+
* the cache (`openSqliteCache`), and the governance client are shared
|
|
28
|
+
* with the other two providers. Cache keys include the model name —
|
|
29
|
+
* cross-provider hits only fire when the model strings match.
|
|
30
|
+
*
|
|
31
|
+
* Default model: `gemini-2.5-flash`. Reasoning: bdd2pw's task is
|
|
32
|
+
* structured JSON output over regex-shaped deterministic patterns.
|
|
33
|
+
* Flash handles it reliably and runs even cheaper than gpt-4o-mini.
|
|
34
|
+
* Override via `LLMClientOptions.model`.
|
|
35
|
+
*
|
|
36
|
+
* Note on JSON output: Gemini's structured-output mode would let us
|
|
37
|
+
* request a JSON schema directly. We deliberately don't — the prompt
|
|
38
|
+
* already constrains the response shape, and going through the same
|
|
39
|
+
* text-then-parse pipeline as the other providers keeps the
|
|
40
|
+
* parseBindingJson behavior provider-agnostic.
|
|
41
|
+
*/
|
|
42
|
+
import type { BindingCache } from "./cache";
|
|
43
|
+
import type { GenerateBindingInput, GenerateBindingResult, LLMClient, LLMClientOptions } from "./types";
|
|
44
|
+
export declare class GeminiLLMClient implements LLMClient {
|
|
45
|
+
private model;
|
|
46
|
+
private apiKey;
|
|
47
|
+
private maxCalls;
|
|
48
|
+
private skipGovernance;
|
|
49
|
+
private log;
|
|
50
|
+
private governance;
|
|
51
|
+
private cache;
|
|
52
|
+
private cachePath;
|
|
53
|
+
private callsCounter;
|
|
54
|
+
private attemptsCounter;
|
|
55
|
+
private genai;
|
|
56
|
+
private modelHandle;
|
|
57
|
+
private cachePersistent;
|
|
58
|
+
private cacheFallbackReason;
|
|
59
|
+
private providerTimeoutMs;
|
|
60
|
+
constructor(opts: LLMClientOptions, cachePathDefault: string);
|
|
61
|
+
ensureCache(): Promise<BindingCache>;
|
|
62
|
+
budgetExhausted(): boolean;
|
|
63
|
+
callsMade(): number;
|
|
64
|
+
callsAttempted(): number;
|
|
65
|
+
cacheBackendPersistent(): boolean | null;
|
|
66
|
+
cacheBackendFallbackReason(): string | undefined;
|
|
67
|
+
/**
|
|
68
|
+
* Single-step generation. Cache → governance → Gemini → parse.
|
|
69
|
+
* Mirrors `AnthropicLLMClient.generateBinding` and `OpenAILLMClient.generateBinding`.
|
|
70
|
+
*/
|
|
71
|
+
generateBinding(input: GenerateBindingInput): Promise<GenerateBindingResult>;
|
|
72
|
+
/**
|
|
73
|
+
* Batched generation — same per-step cache lookup + single batched
|
|
74
|
+
* prompt as the Anthropic / OpenAI clients. JSON array response
|
|
75
|
+
* parsed slot by slot through `parseBindingJson`.
|
|
76
|
+
*/
|
|
77
|
+
generateBatchBindings(inputs: GenerateBindingInput[]): Promise<GenerateBindingResult[]>;
|
|
78
|
+
/**
|
|
79
|
+
* Wraps a promise with a deadline. Gemini's generateContent has no
|
|
80
|
+
* native timeout option (unlike Anthropic/OpenAI SDKs), so we race
|
|
81
|
+
* the SDK call against a setTimeout.
|
|
82
|
+
*/
|
|
83
|
+
private withTimeout;
|
|
84
|
+
private lazyModel;
|
|
85
|
+
close(): Promise<void>;
|
|
86
|
+
}
|