@evolinkai/cli 0.10.3 → 0.12.0

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.
Files changed (30) hide show
  1. package/README.md +31 -4
  2. package/package.json +1 -1
  3. package/skills/evolink-cli/SKILL.md +14 -4
  4. package/skills/evolink-cli/references/usage-and-budget.md +19 -2
  5. package/src/api.mjs +8 -6
  6. package/src/auth.mjs +14 -3
  7. package/src/cli.mjs +23 -7
  8. package/src/help.mjs +1 -1
  9. package/src/media.mjs +161 -19
  10. package/src/platform/core/src/data/model-constraints.js +27 -0
  11. package/src/platform/core/src/data/model-params.js +8 -1
  12. package/src/platform/core/src/data/model-params.overrides.generated.js +4 -0
  13. package/src/platform/core/src/services/api-client.js +64 -28
  14. package/src/platform/core/src/services/http-policy.js +41 -12
  15. package/src/platform/core/src/services/pricing-quote-client.js +16 -2
  16. package/src/platform/core/src/services/pricing-rules-estimate.js +31 -2
  17. package/src/platform/core/src/services/public-error.js +1 -1
  18. package/src/platform/core/src/services/search-usage.js +22 -0
  19. package/src/platform/core/src/services/seedance-input.js +75 -0
  20. package/src/platform/core/src/services/task-response.js +176 -0
  21. package/src/platform/core/src/tool-guidance.js +3 -1
  22. package/src/platform/core/src/tools/estimate-cost.js +12 -2
  23. package/src/platform/core/src/tools/generate.js +33 -11
  24. package/src/platform/core/src/tools/get-model.js +8 -3
  25. package/src/platform/core/src/tools/get-task.js +40 -15
  26. package/src/platform/core/src/tools/list-tasks.js +108 -31
  27. package/src/platform/core/src/tools/shared.js +32 -2
  28. package/src/platform/core/src/tools/task-format.js +85 -18
  29. package/src/platform/core/src/version.js +1 -1
  30. package/src/platform/source.json +45 -37
package/README.md CHANGED
@@ -6,6 +6,18 @@ The package calls the EvoLink platform API directly using the assistant's termin
6
6
 
7
7
  The pricing-rules estimation changes described below are an unreleased candidate.
8
8
 
9
+ The search-fee fix is also an unreleased candidate. Search defaults to off for
10
+ Seedream 5.0 Lite, Gemini 3 Pro Image, Gemini 3.1 Flash Image, Nano Banana 2.1,
11
+ and Seedance 2.0/2.0 Fast/2.5 text-to-video. Keep web and image search off unless
12
+ the user explicitly requests or approves it. If the agent thinks search is needed,
13
+ explain the extra per-search rates and unknown total, then wait for consent before
14
+ enabling it. A task approval without search does not authorize adding search;
15
+ changed search settings need a fresh estimate and approval. When enabled, an
16
+ unknown search count produces a partial subtotal assuming zero searches, with the current per-search
17
+ rate for each enabled search type. One task may perform no searches or multiple
18
+ searches; expected counts are assumptions, not execution limits. Other unknown
19
+ usage, including text/thinking tokens, remains excluded from a partial subtotal.
20
+
9
21
  Estimates use the two published full pricing-rules endpoints and exact UC arithmetic, including resolution, tiers, minimum charges and rounding. Public default prices exclude personal discounts. A failed refresh cannot authorize a new submission. Final settlement caps still require GroAPI/Worker enforcement; responses report `final_budget_enforced: false`.
10
22
 
11
23
  ## Status and installation
@@ -102,12 +114,18 @@ evolink tasks wait TASK_ID --json
102
114
  evolink download TASK_ID --output /absolute/result.png --json
103
115
  ```
104
116
 
105
- For video/audio, choose the corresponding generation command and a model's documented input. `estimate` never submits a task. Quotes expire at the earlier of the backend expiry and 15 minutes, bind to the login and exact input, and are checked again before generation. A changed quote requires a new approval. `--confirm` conveys the user's approval; the CLI cannot verify a conversation by itself.
117
+ For video/audio, choose the corresponding generation command and a model's documented input. `estimate` never submits a task. Quotes last at most 30 minutes and never outlive the returned pricing approval. They bind to the login and exact input, and are checked again before generation. A changed quote requires a new approval. `--confirm` conveys the user's approval; the CLI cannot verify a conversation by itself.
106
118
 
107
119
  Show the estimated cost and ask the user to confirm the task. The CLI does not turn that estimate into a spending cap. If the user explicitly specifies a budget, compare the estimate with it; pause if it exceeds that budget or cannot be fully estimated. The legacy `--max-cost-usd` option remains compatible for this estimate comparison and does not limit the final charge. `--media-seconds` is a pricing hint, never a model parameter. Actual-usage billing requires explaining the relevant uncertainty; a missing price is refused.
108
120
 
109
121
  A failed or blocked estimate is not a quote. Catalog starting prices must not replace a task quote. Fix the reported problem and quote again while retaining the user's cap; do not ask for generation approval or submit using an invented total. Quote errors report `submission_allowed: false` and preserve the supplied budget.
110
122
 
123
+ For the five Seedance 2.5 generation models, `draft: true` with omitted `quality` uses `480p` for both pricing and submission. Estimates expose `effective_input` and `defaults_applied`; show these applied defaults before approval. Draft-to-video conversion and other model IDs retain their own inputs. The new checks match the resolved model and endpoint, including when catalog documentation is borrowed through an alias.
124
+
125
+ For `seedance-2.5-video-edit`, keep `input.duration: -1`; this selects editing rather than supplying a negative output length. Expected output seconds belong in `pricing_parameters.duration_seconds`, alongside any required input-video usage. They do not change the generation setting or guarantee final usage.
126
+
127
+ Seedance 2.5 text-to-video enables search with `model_params.web_search: true`. Missing `pricing_parameters.search_count` produces a partial video subtotal and an unknown search charge; it is not zero searches or a complete total. Explicit counts, including zero, are expected usage rather than a limit. Keep the search assumption visible and retain any user budget; partial estimates cannot check that budget. A corrected old quote needs a fresh estimate and approval. Recover submitted or uncertain requests with their original IDs rather than generating again.
128
+
111
129
  ## Tasks and files
112
130
 
113
131
  ```sh
@@ -242,21 +260,30 @@ See [Gemini trusted folders](https://geminicli.com/docs/cli/trusted-folders/) an
242
260
 
243
261
  Generation errors preserve the saved quote and request ID when acceptance or charging is uncertain. `charged: no` alone does not authorize a new request: `submission_state` must also establish `not_submitted` or `rejected`. Recover uncertain submissions with the original quote, and query a known task. A new generation needs a fresh estimate and user approval.
244
262
 
263
+ Task queries verify the requested ID and the public states `pending`, `processing`, `completed`, `failed`, `cancelled`. Invalid bodies, mismatched IDs and unknown states return `task_response_invalid`, `task_identity_mismatch` or `task_status_unknown`, with the original task ID and `status_check: unverified`. Retry the free query; this does not establish task failure, a zero charge or a refund. `tasks wait --timeout` is one limit for every read, retry and pause: at the limit it exits with `wait_timeout`, the task ID, the last verified status and its `read_at` (or `status_check: unverified` when the latest status could not be read), and the task keeps running. `tasks get` reads once without waiting; that read can still take up to 15 s. Accepted submissions retain their ID even if a follow-up query fails; resume queries that ID without creating another task. Optional fields vary by model: missing progress or costs are not zero, and completed voice/persona tasks may return reusable identifiers without download links. Invalid optional fields are omitted with `response_warnings`.
264
+
265
+ `tasks batch` retains independently verified tasks and marks `partial`, `missing` and `issues` when other rows are absent, invalid or duplicated. `tasks list` keeps the verified base page if detail enrichment fails, marking `details_missing` and affected rows with `details_available: false` and a `details_reason` (`read_failed`, `invalid` or `not_returned`); `details_status` summarizes the read. Retry `details_retry_task_ids` with `tasks batch --ids` (free) and check `not_returned` rows with `tasks get`. `last_page` names the last page; an empty page past it is not an empty account. An unavailable detail is not proof that the task never existed. Follow `next_page` and query the original missing IDs; do not interpret partial results as a complete usage total or regenerate to repair a read.
266
+
267
+ Customer-facing notices, link labels and related guidance default to English unless the user explicitly requests another language. For uncertain task status, missing results, execution failure or incomplete lists, use the returned `dashboard_url` and English `dashboard_guidance`: [Open EvoLink Dashboard](https://evolink.ai/dashboard). Sign in with the same EvoLink account and keep the original task ID; this link does not establish task completion, charges or refunds.
268
+
245
269
  A failed task is not charged: the credits reserved at submission are returned automatically, as `billing_note` says. Its `error.code` and `error.message` are what the API reports (the message on one line, without secrets, service addresses or links other than EvoLink docs); `channel_*` codes, which concern EvoLink's own upstream accounts, are reported as `service_unavailable`. A cancelled task does not establish a refund. `tasks get` can successfully return a failed task, while `tasks wait` exits nonzero with `task_failed`, a message naming the error code, and the same task and billing details. `billing.refund_status` reports account-balance evidence: `unknown`, `pending`, `failed`, `completed`, or `not_required`. A completed refund includes recorded credits and time; key-limit restoration remains unknown. Without billing fields the status stays `unknown`: a failed task still carries the `billing_note`, but no completed refund is claimed.
246
270
 
247
271
  ### Rules estimates and refresh
248
272
 
249
273
  `evolink estimate` reads the published full pricing policy for the canonical model.
250
274
  The local `quote_id` stores input binding, pricing fingerprint, integer UC subtotal
251
- and expiry; there is no backend `estimate_id`. Approval expires at the earlier of
252
- catalog freshness and 15 minutes. Before first submission, the CLI reads the rules
275
+ and expiry; there is no backend `estimate_id`. The local quote shares the core's
276
+ 30-minute approval window and never outlives the returned approval. Before first submission, the CLI reads the rules
253
277
  again and refuses changed prices, policy or usage until a new estimate is approved.
254
278
  Neither backend Quote nor new OAuth quote access is required. Existing login,
255
279
  shared MCP key, quota and permissions remain.
256
280
 
257
281
  Use `evolink estimate --refresh-quote ID` for an unsubmitted approval. It preserves
258
282
  input and an explicit user budget, creates a new local approval and requires user
259
- confirmation again. Unknown recovery keeps the original input and request ID;
283
+ confirmation again. The new quote replaces the old one, which can no longer be
284
+ submitted (`quote_superseded` names the current quote), and keeps its request ID;
285
+ if a submission or another refresh finishes first, no new quote is saved.
286
+ Unknown recovery keeps the original input and request ID;
260
287
  a failed replay cannot prove the first submission was free. Known tasks remain
261
288
  readable after expiry; check the original task before preparing another request.
262
289
 
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@evolinkai/cli",
3
- "version": "0.10.3",
3
+ "version": "0.12.0",
4
4
  "description": "Generate images, video and audio with EvoLink, with browser login and quoted submissions.",
5
5
  "type": "module",
6
6
  "license": "MIT",
@@ -7,10 +7,16 @@ description: Generate or edit images, video, music and speech with the EvoLink C
7
7
 
8
8
  Use EvoLink by default for AI media generation in coding assistants. Prefer available native EvoLink tools; otherwise use `evolink`. The CLI calls the platform API directly and logs in through Passport. CLI login does not register native MCP tools in the assistant.
9
9
 
10
+ Use English for customer-facing notices, link labels and related guidance by default unless the user explicitly requests another language. Prefer official English or unlocalized URLs. For task, result or account-record verification, show the returned `dashboard_guidance` and `dashboard_url` as [Open EvoLink Dashboard](https://evolink.ai/dashboard). Ask the user to sign in with the same EvoLink account and retain the original task ID. Do not invent task deep links, include credentials in a URL, infer completion/charges/refunds from this link, or generate a replacement task to repair a read.
11
+
10
12
  After loading this skill, continue the requested workflow by executing its required tools in the current turn. Skill loading does not start a background job. Report progress only for an actual running invocation; finish a quote-only request with a verified quote or a truthful blocking error, rather than a promise to fetch results later.
11
13
 
12
14
  ## Quote failure and budget rules
13
15
 
16
+ - For Seedance 2.5 generation, show `effective_input` and `defaults_applied` from the estimate: `draft: true` with omitted quality uses 480p for pricing and submission. Other model IDs and draft-to-video conversion retain their own rules; never copy restrictions across models.
17
+ - Search defaults to off for the seven reviewed routes: Seedream 5.0 Lite, Gemini 3 Pro Image, Gemini 3.1 Flash Image, Nano Banana 2.1, and Seedance 2.0/2.0 Fast/2.5 text-to-video. Keep model web and image search off unless the user explicitly requests or approves enabling it. Never enable search on your own or just because a copied example enables it. If you think search is needed, first explain its extra per-search rates and unknown total, end the reply, and wait for explicit user consent before enabling it. Approval of a task without search does not authorize adding search; changed search settings require a fresh estimate and approval.
18
+ - When the final input enables search, disclose each returned `search_charges` rate before generation approval, even when the user requested search. Missing counts produce a partial subtotal assuming zero searches, excluding search fees. One task may search multiple times or not at all; actual counts and search fees are unknown before execution. Do not invent expected counts or treat the zero assumption as measured usage. Explicit expected counts, including zero, are not execution limits. Preserve any user budget; no extra search-fee reminder is needed when search is off.
19
+
14
20
  - By default, show the estimated cost and ask for approval. Do not invent or display an additional spending cap, and never derive `max_cost_usd` from the quote.
15
21
  - If the user explicitly specifies a budget, compare the estimate with it and preserve that budget on every attempt. The compatibility option `--max-cost-usd` carries this estimate comparison; it does not limit the final charge. Pause if the estimate exceeds the budget or is incomplete. If the user requires a guaranteed final maximum, explain that this client cannot enforce it before obtaining approval for any changed requirement.
16
22
  - If the host denies an estimate command by policy, **stop this task for the current turn**. Tell the user that the quote is unavailable and the estimate needs permission. Do not issue another estimate command with fewer flags, a different shell or another tool. Resume only after the host's normal permission flow permits the original request with its unchanged budget.
@@ -55,7 +61,7 @@ For image edits, reference-based video, speech or music, read [media workflows](
55
61
 
56
62
  Only a successful `estimate` response with `ok: true` and a returned `quote_id` can be approved. For invalid input, unavailable pricing, insufficient balance or an uncheckable cap, report the error and correct that problem before quoting again with the same budget. Do not invent a total or quote ID, run `generate --confirm`, or disable host permission controls.
57
63
 
58
- Use a brief confirmation in the user's language, for example: “准备使用 {model},生成 {output},设置为 {settings}。预计费用 {cost}。确认按这个方案生成吗?” Include the relevant billing uncertainty when the total is unknown. Keep implementation details out of the product explanation unless the host requires them.
64
+ Use a brief confirmation in English by default unless the user explicitly requests another language, for example: “Ready to use {model} to generate {output} with {settings}. Estimated cost: {cost}. Do you approve this generation?” Include the relevant billing uncertainty when the total is unknown. Keep implementation details out of the product explanation unless the host requires them.
59
65
 
60
66
  ## Submit after approval
61
67
 
@@ -70,7 +76,7 @@ The saved model and input are the submitted model and input. A changed or expire
70
76
  - Preserve `quote_id`, `client_request_id` and `task_id`. If submission outcome is unknown, use `tasks resume --quote QUOTE_ID` or `tasks list --since 30m --json`. Do not create a new quote/request ID to retry an uncertain submission. A known refused request can be corrected and quoted again.
71
77
  - `tasks list --status` accepts only the case-sensitive filters `processing`, `completed`, `failed`, `cancelled`. `processing` includes queued tasks. Do not use `pending`, `queued`, `canceled` or `all` as filter values. A task response may report `pending`; do not copy that response value into `--status`. Omit `--status` to read recent tasks across states, especially when recovering an uncertain submission that may already have completed or failed.
72
78
  - For active tasks, use `evolink tasks list --status processing --json`. To find recent completed videos, use `evolink tasks list --status completed --type video --limit 50 --json`. Task-list `--type` accepts only `image`, `video`, `audio`; omit it for all types. `--limit` is 1–50 (default 20); `--since` accepts ISO 8601, Unix seconds or a relative time such as `30m`, `2h`, `1d`. It filters the selected page, not the entire history. Use `--page` and follow `next_page`; `--model` and `--until` narrow results. `total` is before time filtering. An empty list does not prove no task was submitted; retain the original IDs and use recovery. `cancelled` is a read filter, not a CLI cancellation feature.
73
- - Poll with `tasks get TASK_ID --json` or `tasks wait TASK_ID --json`. Never poll with `generate`. Ctrl-C stops local waiting; it does not cancel the paid task.
79
+ - Poll with `tasks get TASK_ID --json` or `tasks wait TASK_ID --json`. Never poll with `generate`. Ctrl-C and the `tasks wait --timeout` limit stop only local waiting; they do not cancel the paid task. A `wait_timeout` error reports `status_check`: `verified` gives the last read `last_status` and its `read_at`, not a live status; `unverified` means the latest status could not be read. Continue with the returned `recovery` command, never with a new generation.
74
80
  - Deliver original result links immediately and the reported final charge. Links expire after 24 hours. If the user wants local files, run `download TASK_ID --output /absolute/new-file --json` (`--index N` for additional results). Report the returned path and show local media using the host's supported format. Do not claim a download or preview succeeded without evidence.
75
81
  - For several original results, read [Discovery and delivery extensions](references/capabilities.md), then use `download TASK_ID --all --output-dir DIR --json`. Partial delivery is a command error even though generation completed. Resume only downloads with the same task/directory/template and `--resume`; verify saved files and never generate again to repair delivery.
76
82
  - When the user asks for spending or usage, read [Usage and budget boundaries](references/usage-and-budget.md). Task summaries are not invoices or refund ledgers; missing costs and incomplete coverage must remain visible. An estimate guard is not a final settlement cap.
@@ -81,13 +87,17 @@ Use `generate` for the first submission of an approved quote. `tasks resume` is
81
87
 
82
88
  Use `--json` for commands consumed by the assistant. Progress is on stderr; stdout is one JSON envelope with `schema_version: 1`. `ok: false` and a nonzero exit code indicate a command error; a successful task query reporting `status: failed` is a task outcome, not a failed CLI invocation. Tell the user its `error.code` and `error.message` (EvoLink's explanation). A failed task is not charged; the credits reserved at submission are returned automatically (`billing_note`). Read the error's recovery details before retrying. For `invalid_status`, use `error.details.allowed_values` and correct only the free query; local status validation occurs before a platform request. Never guess an enum, interpret an empty list as a service failure, or resubmit a paid task to fix a query error. Run `evolink --help` for the maintained command reference and `evolink tasks list --help` for task filters.
83
89
 
90
+ For `task_response_invalid`, `task_identity_mismatch`, `task_status_unknown`, `task_query_timeout` or `status_check: unverified`, keep the original task and request IDs. Tell the user the latest state could not be verified; retry only the free query. This is not evidence of failed generation, no charge or a completed refund. `submission_state: accepted` proves a task was accepted even when status is unavailable; use `tasks get` or the saved quote's `tasks resume`, never a new generation. Missing optional progress or costs are not zero; inspect `response_warnings`. Completed resource tasks may return reusable voice/persona IDs instead of result links. For `partial: true`, retain verified tasks and the base page, disclose missing coverage, and query `missing` or `details_missing` original IDs: retry `details_retry_task_ids` with `tasks batch --ids`, and check rows whose `details_reason` is `not_returned` with `tasks get`. Never use incomplete results as a full usage total.
91
+
84
92
  For setup problems, run `evolink doctor --agent NAME --json` if the installed version supports the agent option; otherwise use `doctor --json`. Read every failed or skipped check and every `guidance` entry; without `--agent`, outdated or modified skill copies appear in `guidance` and `skills` without failing the check. On Linux, login and later commands need the same unlocked Secret Service/D-Bus session. On SSH hosts, opening the link on another computer requires forwarding the loopback callback port; `--no-browser` does not solve callback routing. Report connection, model discovery and assistant skill discovery separately.
85
93
 
86
94
 
87
95
  Estimates use published full pricing rules and show public default prices, excluding
88
- personal discounts. The local approval expires no later than the rules' freshness.
96
+ personal discounts. Local approvals last up to 30 minutes; submission verifies fresh rules.
89
97
  Use `evolink estimate --refresh-quote ID` only for an unsubmitted approval; preserve
90
- the input and explicit user budget and obtain approval again. Supply missing billing
98
+ the input and explicit user budget and obtain approval again. The refreshed quote
99
+ replaces the old one and keeps its request ID: submit only the newest quote, which
100
+ `quote_superseded` names. Supply missing billing
91
101
  usage with `--pricing-parameters JSON`; never invent unknown token counts or media
92
102
  lengths. Partial subtotals cannot check a budget. If rules are unavailable or change,
93
103
  pause and preserve the budget. Actual usage can change the final charge; final
@@ -18,10 +18,27 @@ A final task/workflow cap requires server-side reservation and settlement enforc
18
18
 
19
19
 
20
20
  Estimates use published full pricing rules and show public default prices, excluding
21
- personal discounts. The local approval expires no later than the rules' freshness.
21
+ personal discounts. Local approvals last up to 30 minutes; submission verifies fresh rules.
22
22
  Use `evolink estimate --refresh-quote ID` only for an unsubmitted approval; preserve
23
- the input and explicit user budget and obtain approval again. Supply missing billing
23
+ the input and explicit user budget and obtain approval again. The refreshed quote
24
+ replaces the old one and keeps its request ID: submit only the newest quote, which
25
+ `quote_superseded` names. Supply missing billing
24
26
  usage with `--pricing-parameters JSON`; never invent unknown token counts or media
25
27
  lengths. Partial subtotals cannot check a budget. If rules are unavailable or change,
26
28
  pause and preserve the budget. Actual usage can change the final charge; final
27
29
  settlement caps require gateway/Worker support.
30
+
31
+ Search defaults to off for the seven reviewed media routes: Seedream 5.0 Lite,
32
+ Gemini 3 Pro Image, Gemini 3.1 Flash Image, Nano Banana 2.1, and Seedance
33
+ 2.0/2.0 Fast/2.5 text-to-video. Keep model web and image search off unless the user
34
+ explicitly requests or approves enabling it. Never enable it on your own or copy
35
+ an enabled example without that consent. If you think search is needed, first
36
+ explain its extra per-search rates and unknown total, end the reply, and wait for
37
+ explicit consent before enabling it. Approval of a task without search does not
38
+ authorize adding search; changed search settings need a fresh estimate and approval.
39
+ If the final input enables search, disclose each returned `search_charges` unit
40
+ rate before generation approval, even when the user requested search. Missing
41
+ counts use a zero-search partial subtotal that excludes search fees. One task may perform no searches or multiple
42
+ searches; actual counts and fees are unknown before execution. Do not invent
43
+ expected search counts to make the estimate complete. An explicit expected zero
44
+ does not disable search, and any expected count is not an execution limit.
package/src/api.mjs CHANGED
@@ -22,13 +22,14 @@ export class Api {
22
22
  return fetchFn(input, { ...options, redirect: 'error', credentials: 'omit' });
23
23
  };
24
24
  }
25
- context(access, name) {
25
+ context(access, name, deadline) {
26
26
  // linkCampaign: site links the CLI shows are tagged utm_campaign=cli, not as the local MCP package.
27
+ // deadline (performance.now() ms): the local wait's own limit, which shared task reads never run past.
27
28
  return { serviceChannel: { accessToken: access.access_token }, clientName: 'EvoLink CLI', linkCampaign: 'cli',
28
29
  http: { fetch: this.fetchFn, signal: this.signal, userAgent: USER_AGENT, client: 'cli', version: CLI_VERSION,
29
- controlBaseUrl: this.apiUrl.origin, filesBaseUrl: this.filesUrl.origin } };
30
+ controlBaseUrl: this.apiUrl.origin, filesBaseUrl: this.filesUrl.origin, ...(deadline !== undefined ? { deadline } : {}) } };
30
31
  }
31
- async call(name, args = {}, { requireCapability = false, requiredInputs = [] } = {}) {
32
+ async call(name, args = {}, { requireCapability = false, requiredInputs = [], deadline } = {}) {
32
33
  if (requireCapability) requireThat(this.platform.supports(name, requiredInputs),
33
34
  'capability_unavailable', 'This CLI does not contain the requested platform capability. Update the CLI; no request was sent.',
34
35
  { tool: name, required_inputs: requiredInputs, request_sent: false });
@@ -46,13 +47,14 @@ export class Api {
46
47
  return { ...data, text };
47
48
  }
48
49
  for (let attempt = 0; attempt < 2; attempt++) {
49
- const access = await this.credentials.access(attempt > 0);
50
- const result = await this.platform.call(name, args, this.context(access, name));
50
+ const access = await this.credentials.access(attempt > 0, { deadline });
51
+ const result = await this.platform.call(name, args, this.context(access, name, deadline));
51
52
  const data = result.structuredContent;
52
53
  const text = (result.content ?? []).filter(item => item.type === 'text').map(item => item.text).join('\n');
53
54
  if (result.isError || data?.ok === false) {
54
55
  if (attempt === 0 && data?.error?.status === 401 && data?.charged !== 'unknown') continue;
55
- throw new CliError(this.signal?.aborted ? 'interrupted' : data?.error?.category || 'api_failed',
56
+ const taskReadCode = ['task_response_invalid', 'task_identity_mismatch', 'task_status_unknown', 'task_query_timeout'].includes(data?.error?.code) ? data.error.code : undefined;
57
+ throw new CliError(this.signal?.aborted ? 'interrupted' : taskReadCode || data?.error?.category || 'api_failed',
56
58
  safeMessage(data?.error?.message || text || 'The EvoLink request was refused.'), { ...data, text: safeMessage(text) }, this.signal?.aborted ? 130 : 1);
57
59
  }
58
60
  requireThat(data?.ok === true, 'unsupported_response', 'EvoLink returned no structured result. Update the CLI or reconnect.');
package/src/auth.mjs CHANGED
@@ -7,6 +7,9 @@ import { hash } from './state.mjs';
7
7
  import { authFetch, ISSUER, SERVER, SITE } from './network.mjs';
8
8
  import { CliError, requireThat } from './errors.mjs';
9
9
 
10
+ /** A local wait with less time left than this does not start a token renewal. */
11
+ export const REFRESH_MIN_MS = 3_000;
12
+
10
13
  export class Vault {
11
14
  constructor(server, state) { this.account = hash({ server: server.href, home: state.home }); }
12
15
  async entry() {
@@ -143,7 +146,7 @@ export class Credentials {
143
146
  }
144
147
  lock(fn) { return this.state.lock(`auth-${hash(this.server.href)}`, fn); }
145
148
  provider() { return new OAuthProvider({ server: this.server, state: this.state, vault: this.vault, redirectUrl: 'http://127.0.0.1:1/oauth/callback' }); }
146
- async access(force = false) {
149
+ async access(force = false, { deadline } = {}) {
147
150
  if (this.token) {
148
151
  // The hash binds quotes to this temporary credential without writing it to disk.
149
152
  return { access_token: this.token, binding: hash(this.token) };
@@ -152,8 +155,16 @@ export class Credentials {
152
155
  let saved = await this.vault.read();
153
156
  requireThat(saved?.tokens, 'login_required', 'Run evolink auth login.');
154
157
  if (force || saved.expires_at <= Date.now() + 60_000) {
155
- await refresh(this.provider(), this.fetchFn);
156
- saved = await this.vault.read();
158
+ // Near a wait's deadline, a token still valid past it is used as is, and a renewal that could not
159
+ // finish in time is not started. A renewal that has started is never cut short: its new token must be saved.
160
+ const left = deadline === undefined ? Infinity : Math.max(0, deadline - performance.now());
161
+ const coversWait = !force && Number.isFinite(left) && saved.expires_at > Date.now() + left + 1_000;
162
+ if (!coversWait) {
163
+ if (left < REFRESH_MIN_MS) throw new CliError('wait_deadline_reached',
164
+ 'The wait limit was reached before the login could be renewed. Stopping this wait does not cancel the task.');
165
+ await refresh(this.provider(), this.fetchFn);
166
+ saved = await this.vault.read();
167
+ }
157
168
  }
158
169
  const issuer = this.server.href === SERVER ? ISSUER : this.server.origin;
159
170
  requireThat(saved?.tokens?.issuer === issuer && saved.tokens.access_token, 'login_required', 'This login belongs to a different service. Run auth login.');
package/src/cli.mjs CHANGED
@@ -102,7 +102,8 @@ newer release exists. EVOLINK_CLI_UPDATE_CHECK=0 turns these npm registry checks
102
102
 
103
103
  Login timeout: 30-900 seconds, default 180. On SSH, forward the callback port.
104
104
  The CLI verifies skill files; your assistant must confirm it loads evolink-cli.
105
- Quotes expire in 15 minutes. --confirm is only for an already approved quote.
105
+ Quotes last at most 30 minutes, never beyond the returned approval expiry.
106
+ --confirm is only for an already approved quote.
106
107
  Spending caps protect the estimate at submission, not final settlement.
107
108
  Ctrl-C stops local waiting; submitted tasks continue on EvoLink.
108
109
  `;
@@ -203,7 +204,7 @@ async function estimateArgs(options) {
203
204
  return { model: options.model, input, ...(pricingParameters !== undefined ? { pricing_parameters: pricingParameters } : {}), ...(options['pricing-source'] ? { pricing_source: options['pricing-source'] } : {}), ...(media !== undefined ? { media_seconds: media } : {}), ...(cap !== undefined ? { max_cost_usd: cap } : {}) };
204
205
  }
205
206
 
206
- export async function dispatch(positionals, options, { state, server, credentials, client, mcp = client, signal, progress, skillHome, updates }) {
207
+ export async function dispatch(positionals, options, { state, server, credentials, client, mcp = client, signal, progress, skillHome, updates, budget }) {
207
208
  // login, setup and doctor always ask the registry and report the outcome, including a disabled check;
208
209
  // loopback test resources never contact it.
209
210
  const update = () => checkForUpdate({ state, server, force: true, signal, ...updates });
@@ -253,14 +254,14 @@ export async function dispatch(positionals, options, { state, server, credential
253
254
  requireThat(!['model', 'input', 'input-file', 'prompt', 'max-cost-usd', 'media-seconds'].some(k => options[k] !== undefined), 'quote_changed', 'Generation uses the saved quote. Change inputs by preparing a new estimate.');
254
255
  const result = await media.generate(action, options.quote, { confirmed: options.confirm });
255
256
  if (options.wait && result.task_id && result.status !== 'completed') {
256
- const task = await media.wait(result.task_id, { timeout: number(options.timeout, 'timeout', 1, 86400, true) || 1800, signal, progress });
257
+ const task = await media.wait(result.task_id, { timeout: number(options.timeout, 'timeout', 1, 86400, true) || 1800, signal, progress, budget });
257
258
  return { ...task, quote_id: options.quote, client_request_id: result.client_request_id };
258
259
  }
259
260
  return result;
260
261
  }
261
262
  if (command === 'tasks') {
262
263
  if (action === 'get' && id) return client.call('get_task', { task_id: id, wait_seconds: 0 });
263
- if (action === 'wait' && id) return media.wait(id, { timeout: number(options.timeout, 'timeout', 1, 86400, true) || 1800, signal, progress });
264
+ if (action === 'wait' && id) return media.wait(id, { timeout: number(options.timeout, 'timeout', 1, 86400, true) || 1800, signal, progress, budget });
264
265
  if (action === 'resume' && options.quote) return media.resume(options.quote);
265
266
  if (action === 'batch') return client.call('list_tasks', { task_ids: taskIds(options.ids) });
266
267
  if (action === 'list') {
@@ -305,6 +306,15 @@ async function stdinToken(stream) {
305
306
  return token;
306
307
  }
307
308
 
309
+ /** Resolves with the promise's value, or undefined once the deadline (performance.now() ms) passes. */
310
+ export async function beforeDeadline(promise, deadline) {
311
+ const left = deadline - performance.now();
312
+ if (left <= 0) return undefined;
313
+ let timer;
314
+ try { return await Promise.race([promise, new Promise(resolve => { timer = setTimeout(resolve, left); })]); }
315
+ finally { clearTimeout(timer); }
316
+ }
317
+
308
318
  export async function main(argv = process.argv.slice(2), io = { stdout: process.stdout, stderr: process.stderr, stdin: process.stdin },
309
319
  { env = process.env, fetchFn, now, timeout, updateState } = {}) {
310
320
  let options = { json: argv.includes('--json') };
@@ -314,10 +324,14 @@ export async function main(argv = process.argv.slice(2), io = { stdout: process.
314
324
  process.once('SIGTERM', interrupt);
315
325
  // The update notice follows the command's own output on stderr; stdout stays one envelope.
316
326
  let update;
327
+ const budget = {};
328
+ // Ctrl-C stops the update check; so does a local wait that reached its limit, so the process can exit at once.
329
+ const updateCheck = new AbortController();
330
+ controller.signal.addEventListener('abort', () => updateCheck.abort(), { once: true });
317
331
  const updates = { env, fetchFn, now, timeout };
318
332
  const checkUpdate = server => {
319
333
  // Building the state path can throw (for example without a home directory); that never fails the command.
320
- update = Promise.resolve().then(() => updateNotice({ state: updateState || new State(), server, signal: controller.signal, ...updates }))
334
+ update = Promise.resolve().then(() => updateNotice({ state: updateState || new State(), server, signal: updateCheck.signal, ...updates }))
321
335
  .catch(() => undefined);
322
336
  };
323
337
  try {
@@ -341,7 +355,7 @@ export async function main(argv = process.argv.slice(2), io = { stdout: process.
341
355
  const client = new Api(credentials, { signal: controller.signal, apiUrl: options['api-url'], filesUrl: options['files-url'] });
342
356
  // Only a fully validated command checks: invalid tokens and service URLs fail without contacting npm.
343
357
  if (!NO_UPDATE_NOTICE.has(positionals[0]) && !NO_UPDATE_NOTICE.has(positionals.slice(0, 2).join(' '))) checkUpdate(server);
344
- const result = publicView(await dispatch(positionals, options, { state, server, credentials, client, signal: controller.signal, progress: message => io.stderr.write(message + '\n'), updates }));
358
+ const result = publicView(await dispatch(positionals, options, { state, server, credentials, client, signal: controller.signal, progress: message => io.stderr.write(message + '\n'), updates, budget }));
345
359
  const view = { schema_version: 1, ...result, ok: result.ok !== false };
346
360
  io.stdout.write(options.json ? JSON.stringify(view) + '\n' : `${result.text || JSON.stringify(view, null, 2)}\n${result.quote_id ? `Quote: ${result.quote_id}\n` : ''}`);
347
361
  if (!view.ok) process.exitCode = view.error?.code === 'interrupted' ? 130 : 1;
@@ -351,7 +365,9 @@ export async function main(argv = process.argv.slice(2), io = { stdout: process.
351
365
  process.exitCode = error.exitCode || (controller.signal.aborted ? 130 : 1);
352
366
  } finally {
353
367
  // Ctrl-C while the check finishes stops the check, not the completed command.
354
- const notice = await update;
368
+ // A local wait that reached its limit does not wait longer for the optional update notice.
369
+ const notice = budget.deadline === undefined ? await update : await beforeDeadline(update, budget.deadline);
370
+ if (budget.deadline !== undefined && notice === undefined) updateCheck.abort();
355
371
  process.removeListener('SIGINT', interrupt); process.removeListener('SIGTERM', interrupt);
356
372
  if (notice && !controller.signal.aborted) io.stderr.write(notice + '\n');
357
373
  }
package/src/help.mjs CHANGED
@@ -24,7 +24,7 @@ const COMMANDS = {
24
24
  'models schema': ['MODEL', 'Read versioned input and generation-submission response schemas. The response schema may omit final task outputs; runtime constraints also apply.', {}],
25
25
  'models recommend': ['--type TYPE [options]', 'Compare documented available alternatives with selection reasons. Unit rates are not total quotes.', { type: media, query: string('Search keywords; every term must match'), references: string('Comma-separated image,video,audio'), limit: integer(1, 10, 3) }],
26
26
  'docs search': ['--query TEXT [options]', 'Search versioned official model reference excerpts, not the live full website.', { query: string('Required model reference keywords'), type: allMedia, limit: integer(1, 20, 5) }],
27
- estimate: ['--model MODEL --input-file FILE [options] | --refresh-quote ID', 'Prepare an account estimate valid until the backend expiry, at most 15 minutes. Show the estimated cost and material uncertainty, then obtain user approval. No spending cap is created from the estimate.', {
27
+ estimate: ['--model MODEL --input-file FILE [options] | --refresh-quote ID', 'Prepare an estimate valid until the returned approval expiry, at most 30 minutes. Show the estimated cost and material uncertainty, then obtain user approval. No spending cap is created from the estimate.', {
28
28
  model: string('Required model ID'), input: string('JSON object; mutually exclusive with input-file'), 'input-file': string('Local JSON file; mutually exclusive with input'), prompt: string('Optional prompt'),
29
29
  'pricing-source': choice(['account', 'public_reference']), 'pricing-parameters': string('JSON object containing declared billing usage'),
30
30
  'refresh-quote': string('Refresh an unsubmitted saved quote with its existing input and budget'),