accesscall

AccessCall

#accesscall

Use this skill when a user wants to conduct an accessibility intake interview by phone instead of a web form, typically to support a Section 508 or WCAG conformance audit (VPAT 2.4).

Accessibility intake forms are themselves often inaccessible: screen reader fatigue, motor impairment, low vision, or cognitive load can all block someone from completing the exact form meant to capture their barrier report. AccessCall places a short phone call instead, conducts a structured verbal interview, and returns a result that maps onto standard VPAT 2.4 conformance report fields.

When To Use

#when-to-use

Use this skill for:

• accessibility intake for someone who cannot use a web-based audit form
• gathering verbal barrier reports to populate a VPAT 2.4 / Section 508 audit report
• "call and ask about accessibility issues" or similar one-off outbound intake requests

When Not To Use

#when-not-to-use

Do not use this skill to:

• place a call without the user explicitly confirming the recipient and their intent first
• guess a phone number's country code or region; ask if ambiguous
• call a third party who is not the requesting user without documented prior consent from that person
• auto-populate a VPAT row when the follow-up contact was never confirmed back to the caller (see Safety Rules)
• treat the auto-selected WCAG criterion row as final without human review; matching is done at the WCAG principle level (Perceivable / Operable / Understandable / Robust), not the specific success criterion

Setup

#setup

scripts/format-to-vpat.js depends on docx, jszip, and xml-js (declared in package.json). Run npm install inside skills/accesscall/ before using it — without this step it fails immediately with Cannot find module 'jszip'. scripts/parse-recap.js and scripts/phone-utils.js have no external dependencies and need no install step.

Core Workflow

#core-workflow

1. Confirm required inputs: recipient phone number, product or site name being audited, preferred language (default English, US). Ask for anything missing; do not fabricate a phone number or product name.
2. Validate the phone number against E.164 (scripts/phone-utils.js, isE164/assertE164) before calling plan_call. If it does not match E.164 format, reject it and ask the user to correct it — do not reformat, guess a country code, or silently pass a malformed number through to plan_call.
3. Call plan_call with the goal built from references/call-task-template.md, filling in recipient_name and product_or_site_name. Do not forward the user's full latest message verbatim via user_input by default — that field exists on plan_call for resolving an ambiguous/malformed phone number, not as a general passthrough. Only populate user_input when there's an actual number to disambiguate, and even then pass just the relevant phone-number text, not the user's entire message. The goal string alone carries recipient_name/product_or_site_name/language into the call.
4. Show the confirm_summary to the user and wait for explicit confirmation. Do not call run_call until the user confirms.
5. Immediately before calling run_call, acquire the dispatch lock (scripts/call-lock.js, acquireLock) keyed on the recipient's phone number, the purpose "accesscall-intake", and this specific plan_id — every lock is owned by the plan_id that acquired it. If a call to this recipient is already locked by a different plan_id — including on a retry after a crash or timeout, which will see the same lock a first attempt wrote — refuse to place another and ask the user to explicitly confirm an override before retrying. acquireLock also refuses unconditionally (even with override) if this exact plan_id was already dispatched before, per its durable dispatch history — a plan_id must never be replayed. The lock does not expire on a timer; it is only released in step 7, once a confirmed terminal status exists.
   • Before ever passing override: run scripts/call-lock.js's checkLock first and show the human the current owning plan_id and generation — never override from memory or a stale earlier check. Overriding requires passing that exact generation as expectedGeneration; acquireLock re-verifies inside the same atomic critical section that the on-disk generation still matches before writing, and refuses with a clear "state has changed" error if it doesn't (e.g. a different override already went through in the meantime). This exists because the per-recipient mutex alone only serializes writes against each other — it does not stop two separately, legitimately approved overrides (each based on what looked like current state at approval time) from both eventually succeeding one after the other against state that changed in between. Never retry an override with the same expectedGeneration after a mismatch; re-check and decide again.
6. Call run_call with the confirm_token exactly as received. Never call it more than once for the same plan_id.
7. Poll get_call_run every 1-3 seconds while the run is active, then slow down, until a terminal status. As soon as one of COMPLETED/FAILED/NO_ANSWER/DECLINED/CANCELED/CANCELLED/VOICEMAIL/BUSY/EXPIRED is confirmed, call scripts/call-lock.js's releaseLock with that same plan_id and status to release the dispatch lock from step 5. releaseLock is compare-and-delete: it only releases if the plan_id given still matches the lock's current owner, so a delayed/late result for a plan_id that has since been overridden by a different dispatch is refused rather than freeing someone else's lock. Never release speculatively or because polling has taken a while — an unresolved status keeps the lock held until it either resolves or the user explicitly overrides.
8. This MCP server's plan_call has no schema-input parameter for structured extraction (verified against its inputSchema). Instead, the call goal (references/call-task-template.md) instructs the bot to recap its own answers in a fixed, labeled format at the end of the call. Parse that recap from the transcript using scripts/parse-recap.js, which only reads lines attributable to the bot's own final speaking turn (never anything the caller said) and validates the result against references/intake-result.schema.json before returning it.
9. If parsing/validation fails — including when the recap never happened because the crisis-safety override in references/call-task-template.md triggered, or because the recap couldn't be cleanly attributed to the bot — do not guess field values. Report the crisis disclosure (if that's what happened) as its own distinct, clearly flagged outcome. Otherwise, surface a redacted excerpt of the transcript via scripts/redact-transcript.js — never the raw, unredacted transcript, which may contain the caller's spoken phone number or email address.
10. If followup_contact is present, it must only be trusted when followup_contact_confirmed is true, meaning the bot spelled the contact back letter-by-letter and the caller explicitly confirmed it. If not confirmed, do not carry it into VPAT output; note "Contact unconfirmed, verify manually" instead.
11. Optionally run scripts/format-to-vpat.js to insert the validated result into a VPAT 2.4 template as a new row. Every auto-matched row (matched at the WCAG principle level, not the specific success criterion) gets an "AUTO-MATCHED AT PRINCIPLE LEVEL, HUMAN REVIEW REQUIRED BEFORE AUDIT USE" note in Remarks, so it can never be mistaken for a final placement.

Why a plan_id is never replayed, even with override (step 5): this is a permanent design choice, not a gap to eventually fix. run_call's own tool contract already forbids calling it twice for the same plan_id — this skill didn't invent that restriction, it's upstream. Every genuine reason to "retry" collapses into needing a brand-new plan_id anyway: if the recipient asks for a callback with a corrected detail, that's a different goal, which only plan_call can produce (there is no operation that patches an existing plan's parameters); if the process crashes between acquiring the lock and getting a response from run_call, you don't actually know whether the call was placed, and retrying the same plan_id risks a real duplicate dispatch at the API level, so the correct recovery is a fresh plan_id from a new plan_call, with the interrupted attempt's outcome flagged to the user as unknown rather than silently retried. override therefore only ever needs to resolve one kind of conflict — a different, newly-planned plan_id wanting to dispatch while the lock is still held by an unresolved other plan_id — never "let me reuse this one."

Safety Rules

#safety-rules

• Never place a call without the user explicitly confirming the recipient and intent first. Setup/verification steps must never trigger run_call.
• Never guess a phone number's country code or region.
• Reject any phone number that fails E.164 validation (scripts/phone-utils.js) before it ever reaches plan_call; ask the user to correct it instead of silently passing it through.
• Mask phone numbers in any logged output or printed summary (e.g. +1555010****, via scripts/phone-utils.js's maskPhone) — the only place the full, unmasked number belongs is the actual plan_call/run_call API invocation itself.
• Never forward the user's full latest chat message to plan_call's user_input by default. Only pass the minimum text needed, and only when actually needed to resolve an ambiguous or malformed phone number.
• Acquire the dispatch lock (scripts/call-lock.js) with this plan_id immediately before every run_call, and only ever release it (releaseLock) with that same plan_id once get_call_run confirms a terminal status — never on a timer, never speculatively, and never for a lock currently owned by a different plan_id. Never place a duplicate call to the same recipient while a lock is held by another plan_id without an explicit user override, and never replay a plan_id that the dispatch history shows was already dispatched, even with an override.
• Never pass override without first calling checkLock and showing the human the current owner and generation, and never pass a generation other than what that check just returned. acquireLock re-verifies the generation still matches on-disk immediately before writing and refuses if it doesn't — this is what stops two separately-approved overrides from both succeeding against state that changed between approval and execution.
• Never surface a raw, unredacted transcript to the user as a validation-failure fallback. Redact it first (scripts/redact-transcript.js, which masks phone-length digit sequences of any format — not just US shapes — and email addresses) — it does not detect spoken street addresses, so treat its output as a partial, not complete, redaction.
• Never call a third party without documented prior consent from that person.
• A confirmed contact detail (email or phone for follow-up) requires an explicit letter-by-letter spell-back and a "yes" from the caller. Testing showed the STT layer can introduce transcription errors (e.g. inserting a duplicate word into a domain); the spell-back is a mitigation, not a guarantee, since a caller can still mishear their own readback and confirm an incorrect value. Document this limitation to the end user rather than presenting the mechanism as fully reliable.
• Auto-matched VPAT rows are matched at the WCAG principle level only. Flag rows for manual review rather than guessing the exact success criterion — scripts/format-to-vpat.js always writes an explicit "AUTO-MATCHED AT PRINCIPLE LEVEL, HUMAN REVIEW REQUIRED BEFORE AUDIT USE" note into Remarks for these rows.
• If the caller discloses anything suggesting self-harm, suicidal ideation, or a crisis during the call, the accessibility interview must stop — this is a required override in references/call-task-template.md, not optional bot judgment. Do not attempt to extract intake fields from that call. Report it to the user as a distinct, clearly flagged outcome, not as a normal or partial intake result.
• Do not expose auth tokens, confirmation tokens, or credentials in any output.

Known Limitations

#known-limitations

• This skill supports one-off calls only. There is no recurring or scheduled-call capability.
• There is no mid-call cancellation path. Once run_call starts, the call runs to a terminal status (COMPLETED, NO_ANSWER, DECLINED, FAILED, etc.) — it cannot be stopped or cancelled from within this skill while in progress.

Output Format

#output-format

After a completed call, report:

• call status and duration
• the recipient's phone number, masked (e.g. +1555010****) — never print the full number in a summary
• the parsed structured result (assistive technology, task attempted, barrier category, severity, consent to follow-up)
• whether follow-up contact was confirmed
• if a VPAT insertion was run, which template row was modified, its new values, and that it is an auto-match requiring human review before audit use

If the call did not complete (no answer, declined, failed), report the status plainly and do not fabricate a result.

If the caller disclosed a self-harm/suicidal-ideation/crisis situation, report that as its own distinct, clearly flagged outcome instead of an intake result — do not attempt to backfill or guess the intake fields that weren't asked.

References

#references

• references/call-task-template.md — the call goal template with placeholders
• references/intake-result.schema.json — structured result schema
• references/example-output.json — example parsed result
• scripts/parse-recap.js — extracts the labeled recap from a transcript (bot-attributed final turn only) and validates it against the schema
• scripts/phone-utils.js — E.164 validation and phone-number masking
• scripts/call-lock.js — durable, plan_id-owned dispatch lock preventing a duplicate call to the same recipient. Every mutation (fresh acquire, override, and compare-and-delete release) runs inside a per-recipient mutex; override additionally requires a generation-based compare-and-swap (expectedGeneration must match the current on-disk generation, checked inside that same critical section) so two separately-approved overrides can't both succeed against state that changed between approval and execution. Backed by an append-only dispatch history that refuses to replay a plan_id even after its lock is released, and fails closed (throws) rather than guessing if the journal itself is corrupted
• scripts/redact-transcript.js — masks phone-length digit sequences (any format, not just US) and email addresses in a transcript before it's shown to a user
• scripts/format-to-vpat.js — inserts a validated result into a VPAT 2.4 docx template (see Setup above: requires npm install first)
• package.json — declares format-to-vpat.js's dependencies (docx, jszip, xml-js)
• assets/vpat-2.4-template-generic.docx — genericized VPAT 2.4 template with placeholder preparer/contact fields