opencode-skills-collection 4.1.18 → 4.1.20

package/bundled-skills/outreachgenie-list-audit/references/capture.md

@@ -0,0 +1,53 @@
+# Snapshot contract, version 1
+
+The helper consumes captured results rather than authentication details. Preserve observed tool response data exactly; do not invent missing fields to make it pass. Save only the selected scope in a private local JSON file. The complete fictional example is `assets/fictional-snapshot.json`.
+
+Snapshots must be valid UTF-8 JSON. Duplicate object keys, non-finite numbers and invalid Unicode are rejected rather than silently selecting or coercing ambiguous evidence.
+
+Envelope:
+
+```json
+{
+  "schema_version": 1,
+  "source_kind": "mcp_snapshot",
+  "captured_at": "2026-10-16T12:00:00Z",
+  "workspace_id": "00000000-0000-4000-8000-000000000001",
+  "flow_id": "00000000-0000-4000-8000-000000000002",
+  "selected_list_ids": ["00000000-0000-4000-8000-000000000011"],
+  "calls": [
+    {"tool": "get_workspace_context", "arguments": {}, "result": {"workspace": {"id": "00000000-0000-4000-8000-000000000001"}}}
+  ]
+}
+```
+
+These UUIDs are fictional placeholders; the shortened example is not a complete runnable audit.
+
+## Five captured read tools
+
+| Tool | Arguments | Unwrapped result needed |
+| --- | --- | --- |
+| `get_workspace_context` | `{}` | `workspace.id` |
+| `get_flow` | `{"flow_id":"…"}` | Matching `id`, `nodes`; preserve the remaining observed result |
+| `list_flow_template_fields` | `{}` | `built_in_fields`, each with `key`; workspace `custom_fields` are contextual |
+| `get_prospect_list` | `{"list_id":"…"}` | Matching `id`, `prospect_count`, `custom_field_metadata`, `active_campaign_count` |
+| `list_prospects_in_list` | `{"list_id":"…","query":{"page":1,"page_size":100}}` | `items` and normalized `pagination` |
+
+For product tools (`get_workspace_context`, `get_prospect_list`, `list_prospects_in_list`), take the payload from `structuredContent.result`. Legacy `get_flow` and `list_flow_template_fields` return their payload directly in `structuredContent`. Some hosts present the same JSON as text; parse the actual tool JSON, never prose or a guessed schema. `result` in this snapshot always means the already-unwrapped payload. Do not include error responses as successful calls.
+
+Each selected list needs its own detail response and all intended pages in request order. Use unfiltered membership reads with a fixed page size from 1–100. Start at page 1 and follow `pagination.next_page` while `has_more` is true. Stop on errors, repeated/noncontiguous pages, contradictory totals, a workspace mismatch, or the package's input bound. Report the collection as partial when applicable.
+
+Normalized pagination contains:
+
+```json
+{"total": 2, "page": 1, "page_size": 100, "has_more": false, "next_page": null, "previous_page": null}
+```
+
+Prospect `id` is required. Preserve captured `first_name`, `last_name`, `company`, `title`, `email`, `phone`, `linkedin_url` and `custom_fields` when supplied. Membership is established by the `list_id` used in that page request. `list_count` does not identify another list; do not invent other memberships.
+
+An absent builtin property or absent/unsupported `custom_fields` container means **not captured or unsupported**, not a known blank. Within a captured custom-field dictionary, an absent referenced key is a known empty runtime value. A known `null`, `false` or zero renders empty under the current product rules. Non-scalar values are reported as unsupported by the helper.
+
+## Collection bounds and interpretation
+
+The helper accepts 5 MiB inputs, 20 selected lists, 1,000 captured calls, 1,000 total selected-list memberships, 200 nodes and 500 supported outbound text fields. It stops after 10,000 findings rather than producing an unbounded report. Partial pages may still be audited, but coverage is false. Duplicate prospects within one list count as a coverage issue; the same prospect appearing once in two selected lists is valid and checked for both memberships.
+
+Do not silently remove duplicate rows, change totals, fill missing values or merge fields from another capture. The capture is not transactional: data may change between calls even when the numbers agree. `all_selected_memberships_captured` only describes internal coverage of the supplied snapshot. It cannot authenticate the file, prove live state or certify campaign readiness.

package/bundled-skills/outreachgenie-list-audit/references/checks.md

@@ -0,0 +1,35 @@
+# Checked behavior and report interpretation
+
+The contract is based on OutreachGenie's saved flow/list interface as reviewed on 16 October 2026. These are product limits, not universal LinkedIn limits. Recheck the contract when the service changes; the helper rejects an unfamiliar builtin-field catalog.
+
+| Text-bearing saved action | Field | Limit and unit |
+| --- | --- | --- |
+| `send_connection_request` | `invite_note` (or `inviteNote` fallback) | Optional, 200 UTF-16 units |
+| `send_message` | `message` | Required, 1,000 Python Unicode code points |
+| `message_sequence` | Each `messages[i].message` | Required, 1,000 code points |
+| `comment_post` | `comment_text` (or `commentText` fallback) | Required, 1,000 code points; may appear in older saved flows |
+
+Emoji outside the basic multilingual plane use two UTF-16 units. The helper checks raw invitation notes and raw message/sequence text lengths, then checks per-prospect rendered text. Missing required text is reported; optional empty invitation notes are contextual because the product omits the note. Oversized rendered text is a product execution problem. Do not imply the rest of a campaign would execute correctly.
+
+The helper checks sequence count and message-ID shape, but **does not** validate sequence delays, all action config, graph reachability, branch conditions, reply checks, workspace limits or schedules. Non-text actions are listed under `unchecked_actions`. Every saved supported action node is inspected, including nodes a particular prospect might never reach.
+
+When both snake_case and camelCase names exist for the same field, the helper reports `ambiguous_text_field_aliases` and skips that text field rather than assuming the order used by runtime key normalization.
+
+## Template behavior
+
+Builtins: `first_name`, `last_name`, `company`, `title`, `email`, `phone`, `linkedin_url`. `{{note}}` is forbidden. Custom fields can use `{{custom_fields.topic}}` or `{{topic}}`. The custom prefix form normalizes its key to lowercase; bare aliases must already follow the lowercase identifier rules. Tokens cannot contain internal whitespace. Malformed/unbalanced tokens are reported and per-record rendering conclusions for that text are skipped.
+
+Campaign activation compares referenced custom keys to **each selected list's raw `custom_field_metadata` keys**. The workspace catalog is a union and does not prove that every list has a field. A catalog formatter may normalize a key while activation still compares the stored raw key; the helper follows the latter when auditing list metadata.
+
+Values substitute once, then the result is stripped. A saved value containing `{{first_name}}` stays literal after substitution; the helper flags this as context rather than recursively expanding it. Empty, `null`, `false` and zero values render empty. Missing values can leave a nonempty but poor message, so report them even when length passes.
+
+## Findings and next decisions
+
+- `coverage_issue`: missing, inconsistent or duplicate collection evidence. Finish or recapture the selected scope before making completeness claims.
+- `template_issue`: malformed tokens, a custom field missing from selected-list metadata, empty required text, or raw/rendered product-limit failures. Identify the exact list/prospect/node/field combination for review.
+- `unknown`: uncaptured/unsupported values or missing active-campaign context. Obtain the missing read evidence; do not treat it as a blank or a pass.
+- `review`: a captured template value is empty. Decide whether to complete the source data or revise the wording in a separate task.
+- `context`: active campaigns on a selected list, empty optional notes, literal braces or privacy-sensitive `email`/`phone` references. These are context, not automatic errors or permission to modify active campaigns.
+- `unsupported`: unavailable configuration or no supported outbound text fields. The report cannot stand in for a full flow check.
+
+Reports omit raw saved field values and rendered text. They contain identifiers and field names, and therefore are still private work products. Zero findings means only that these checks found none in the supplied evidence. It says nothing about deliverability, truth, targeting quality, current LinkedIn profile state, permission to contact or provider policy compliance.

package/bundled-skills/outreachgenie-list-audit/references/setup.md

@@ -0,0 +1,25 @@
+# Account, host and privacy prerequisites
+
+The offline demonstration uses synthetic records and does not require signup, a LinkedIn account, credentials or a network connection. The Python helper uses only the standard library and requires Python 3.10 or newer.
+
+A real audit requires:
+
+- An [OutreachGenie account](https://app.outreachgenie.ai) with access to the intended workspace, one saved flow and the selected saved prospect lists.
+- A host that can connect to the hosted MCP endpoint `https://api.outreachgenie.ai/mcp`, read the selected tools, store a private local snapshot and run Python.
+- Authorization to read those records. Prefer an OAuth connection limited to `crm.read` when the host supports it. The read tools used by this skill do not require campaign activation or write scopes. Existing workspace API keys may be broader; a written skill is not a server-side permission boundary.
+
+Use the current [developer documentation](https://www.outreachgenie.ai/developers) and [MCP connection documentation](https://app.outreachgenie.ai/docs/mcp/claude) for connection setup. Let the host perform OAuth login and consent. Do not paste tokens or passwords into chat, source files, fixtures, reports or command arguments. This package does not configure credentials or register an OAuth client.
+
+The helper itself makes no HTTP requests. During connected collection, the MCP server and the chosen AI host/model provider process the returned records; host retention and processing terms still apply. Do not describe the whole workflow as browser-only, offline, or zero third-party processing. Keep snapshots outside git and shared/synced folders unless the user deliberately chose a suitable location. Inputs contain saved personal data even though report outputs omit raw values and message text. IDs, field names and flow structure can still be sensitive.
+
+## Compatibility evidence
+
+The package uses the Agent Skills directory format (`SKILL.md`, `scripts/`, `references/`). This distribution omits the upstream optional Codex UI metadata; it is not needed to run the helper. Local synthetic execution is testable without an MCP host. Format compatibility is not a claim that every host executes the skill or authenticates successfully.
+
+The offline helper and fictional fixture have been validated locally.
+
+On October 16, 2026, Codex CLI 0.160.0 on Linux with Python 3.11.8 completed an explicit `$outreachgenie-list-audit` invocation in an isolated workspace using the included fictional fixture. The host read the installed instructions, ran the local helper with an absolute Python interpreter path and explained the generated report. No OutreachGenie account or MCP connection was used.
+
+This verifies one explicit fixture workflow only. Real account connections, hosted OAuth, real-workspace capture, automatic skill selection, Claude Code and ChatGPT/Claude web execution remain unverified. A web chat host without a local interpreter cannot use the helper as written; it should report the missing capability, not imitate the checks or claim a tested result.
+
+The audit is read-only and does not need the OutreachGenie Chrome companion to be running. It does not discover current LinkedIn profile data, verify current connections, or exercise delivery. If the user has no saved data, explain that the demo is available; do not silently import or enrich prospects to make the audit run.