Phase 2 of 2 · overnight repair

Verification and Constraints

WAITING FOR PHASE 1

Main CRM: 45-day conversation recovery

0 / 18 verifiedAstra Medium executorNo client sends · USD20 shared ceiling

Goal and finish line

End goal: Make one deduplicated Main CRM view of recently contacted leads and clients, with complete linked activity, evidence-based stages and labels, and useful Missive reply drafts.

Done means: Every in-scope contact has one canonical CRM identity, linked real email and full available Wispr evidence, independently classified Notion stage and Missive labels, and a direct draft link when a reply or follow-up is warranted. Missing sources and unresolved identities are visible and do not masquerade as completed coverage.

Rules that stay fixed

No email or message to any client/lead. Missive drafts only. Internal positive Slack alerts remain authorized, no historical alert flood.
Preserve human edits, commercial commitments, unrelated work and existing sender identities. Reversible changes require narrow before-state and tested restore path.
Use existing accounts and deployment. Do not add security layers, publish credentials or raw emails/transcripts on public boards. Existing protected source systems stay protected.

Morning target: 30 September, 08:00 Budapest. A target is not a guarantee or permission to mark unproved work complete.

Execution order

This contract · Dependency: Missive repair

Matt explicitly requested both finished HTML contracts followed by CCC execution in sequence. This task-specific execute instruction overrides the LLL default later-Launch gate. Use two fresh Astra Medium orchestrators as specifically requested for this task, overriding generic skill Sol defaults. Only phase 1 may start first. Phase 2 starts after phase 1 acceptance.

First checkpoint after 20 minutes, then hourly. One focused steering message when useful. No concurrent phase executors.

Scope and existing destinations

Freeze 2026-08-16 01:45 through 2026-09-30 01:45 Europe/Budapest (45 rolling days), UTC 2026-08-15T23:45:00Z inclusive to 2026-09-29T23:45:00Z exclusive. Catch up newer events before final verification without changing the frozen audit counts.

Main CRM · Existing Activity Log. Direct business interaction and attended customer calls qualify. Bulk outreach alone, internal meetings and automated messages do not.

How it should work

Green means a confirmed requirement, not completed implementation. Dashed means execution proof is still missing.

End goal system

Route to the end state

Loading diagrams…

Work and proof

Reconcile every recent contact 0 / 3

Build a complete evidence inventory before changing pipeline stages.

Success criteria, proof and constraints

D1.1 When all accessible Missive conversations with activity in the frozen 45-day window and all Wispr meetings in that window are paginated, then coverage-manifest.json records all pages, account coverage, exact source IDs, date boundaries and missing records without a silent cap.
Evidence: Not run
D1.2 When every candidate is evaluated for lead/client scope, then contact-index.json includes meaningful personal business email interaction or an attended customer/prospect call, excludes bulk cold outreach alone, automated mail, internal/personal/vendor-only conversations, and records each exclusion or uncertainty.
Evidence: Not run
D1.3 When an included conversation or meeting is read, then all relevant email history and complete available Wispr transcript chunks are stored privately with IDs/timestamps/content hashes; unavailable or deleted recordings remain explicit gaps, never replaced by summaries or unrelated dictation.
Evidence: Not run

Constraints

Read all archive/inbox/stage locations, not only current visible queues. Inspect older context for included contacts when needed for commitments.
Wispr Flow is the sole meeting source. No Fireflies, audio recovery project or fabricated transcript. Historical unrelated inaccessible records must not stop independently complete contacts, but coverage remains honestly partial.

Deduplicate Main CRM identities 0 / 2

Consolidate evidence without destroying project history or commercial facts.

Success criteria, proof and constraints

D2.1 When source contacts are joined to Main CRM, then identity-map.json maps exact email aliases/domains/people and evidence to a canonical existing CRM record, with unmatched genuine leads created once and ambiguous merges listed separately.
Evidence: Not run
D2.2 When confirmed duplicate CRM records are consolidated, then the canonical page retains all relevant page content, fields, deals and inbound/outbound relations, duplicate records are recoverably archived or superseded, and a second run creates zero duplicate contacts or relations. Before broad consolidation, exercise a representative scoped restore/relink against saved before-state and verify provider readback.
Evidence: Not run

D2.2 · Actual changed-interface screenshot pending. Attach blind Luna High description and comparison.

Constraints

Never merge on company name/domain alone when separate people or legitimate distinct deals exist. Keep one canonical contact mapping while preserving separate deal/project records.
Preserve prices, invoices, contracts, payment facts, ownership and manually edited notes. Save exact before-state and restore mapping; no permanent purge.

Fill the existing Activity Log 0 / 3

Make every stage decision traceable to the conversation that supports it.

Success criteria, proof and constraints

D3.1 When eligible email messages and full available Wispr transcripts are synchronized, then the existing Clientsflow Activity Log has one event per source Message-ID/meeting ID, original event time, readable content, direct source link and relation to the canonical CRM record.
Evidence: Not run

D3.1 · Actual changed-interface screenshot pending. Attach blind Luna High description and comparison.
D3.2 When long transcripts and email threads exceed a property limit, then their complete content remains readable in event page blocks/attachments with a concise summary property, and content-count/hash reconciliation proves no silent truncation.
Evidence: Not run
D3.3 When synchronization is rerun after a partial failure or identity merge, then existing events are reused/relinked rather than duplicated, and expected source event IDs equal actual Activity Log event IDs for the included scope.
Evidence: Not run

Constraints

Reuse verified existing Activity Log data source 3779cef9-86ad-811b-b6d1-000b0ca40cad, not a second new database. Preserve existing human feedback and unrelated events.
Email send time, call time and latest meaningful contact differ from import/modified time. Do not make every record look recently contacted because it was imported tonight.
Define Activity Log event scope in the manifest: include meaningful inbound and sent outbound email and full available meetings for included contacts, including older context needed for commitments. List excluded automated/noise and bulk campaign-only events with reasons, retaining necessary source context privately. Reconcile exact included event IDs, not only inbound replies.

Classify stages and labels separately 0 / 4

Use the same evidence, with one model call per distinct business decision.

Success criteria, proof and constraints

D4.1 When a contact evidence packet is classified, then a verified Jev/OpenRouter request for Notion stage and a separate request for Missive labels produce distinct typed decision receipts with resolved model ID, prompt version, source IDs and probabilities/distributions. Application code adds the evidence-linked rule trace and uncertainty state; Jev is not asked for prose rationale.
Evidence: Not run
D4.2 When high-evidence decisions are applied, then Main CRM Pipeline stage uses existing valid options and Missive uses phase-1 label IDs; paid/lost/accepted/signed states require actual supporting events and owner stages/permanent positive history stay intact.
Evidence: Not run

D4.2 · Actual changed-interface screenshot pending. Attach blind Luna High description and comparison.
D4.3 When reviewed fixtures and a stratified sample of up to 30 real contacts are evaluated, then explicit no-contact, automation, confirmed-lost, paid, future booking, active-client and contradictory evidence cases have no critical errors; mismatches are corrected before broad apply and remaining uncertainty is visible.
Evidence: Not run
D4.4 When the existing scheduled/event system processes one incremental cycle and a duplicate cycle, then only evidence-changed contacts are reclassified, new events synchronize, manual overrides are respected and the second cycle makes no duplicate or oscillating changes.
Evidence: Not run

Constraints

Do not use one combined call disguised as two decisions. Do not fabricate Jev availability. Use exact researched model and supported API or report the unsupported model requirement explicitly.
Model confidence alone is not proof of payment or lost status. Preserve manual stage changes newer than the evidence snapshot and compare before-write hashes/timestamps. No model-triggered client sends.

Create the recent-contact CRM view 0 / 2

Give Matt a useful saved view inside his existing CRM.

Success criteria, proof and constraints

D5.1 When Main CRM opens its new saved view Last 45 days · Conversations, then only canonical lead/client records with meaningful interaction in the rolling 45 days are shown, with last interaction, stage, next action, owner and Missive link columns.
Evidence: Not run

D5.1 · Actual changed-interface screenshot pending. Attach blind Luna High description and comparison.
D5.2 When the saved view membership is verified, then exact equality with the frozen audit contact set is proved from an as-of query/snapshot at the frozen endpoint. Separately, a current rolling-window expected set including catch-up events equals current view membership. Boundary-date fixtures prove both lower/upper date behavior and no duplicates; post-freeze events are not forced into frozen totals.
Evidence: Not run

Constraints

Create a saved view in Main CRM, not a new CRM copy or standalone report pretending to be a view. Keep other existing views unchanged.
Exclude records whose only recent change was automation/import or a cold campaign blast. Preserve legitimately separate deals if the CRM is deal-grained and explain canonical contact grouping.

Prepare replies worth sending 0 / 4

Leave an actionable review table and real unsent Missive drafts.

Success criteria, proof and constraints

D6.1 When complete contact histories are evaluated for unanswered questions, commitments and due follow-ups, then reply-followup-table.md lists every warranted action with priority, recipient, reason, last meaningful date, CRM URL and verified direct Missive draft/conversation URL.
Evidence: Not run
D6.2 When eligible rows are drafted, then Missive contains one useful unsent draft per eligible thread, addressed to the external person, matching tegeződés/magázódás and resolving their actual questions with verified agency and relationship context; each table link opens that exact draft.
Evidence: Not run

D6.2 · Actual changed-interface screenshot pending. Attach blind Luna High description and comparison.
D6.3 When Lost, no-contact, Snooze, future promised date, existing human draft or automated-only history is present, then no inappropriate proactive chase is created, the table records exclusion/defer reasons and human content stays unchanged. A genuine new human question is evaluated separately for a warranted direct reply, including paid/ongoing clients, while explicit no-contact and automated-message exclusions still win.
Evidence: Not run
D6.4 When the final scheduled cycle and UI readback run, then attention-worthy drafts are unread, phase-1 quiet/noise rules still hold, and the board links the saved Notion view, private action table, coverage receipt and exact remaining gaps.
Evidence: Not run

D6.4 · Actual changed-interface screenshot pending. Attach blind Luna High description and comparison.

Constraints

Draft only. Do not send emails/messages to clients or leads, including test messages. Genuine stage-less successful follow-up uses M Contacted per phase 1.
An existing valid human draft should be linked, not overwritten or counted as a newly generated draft. Keep private action table and raw communications off public board.

Model and proof decisions

Verified: TypeSafe Jev, OpenRouter model typesafe/jev-1.13 via POST /api/alpha/decisions with state and typed questions, not chat/completions. Parent synthetic opt-out smoke returned HTTP200, resolved model typesafe/jev-1.13-20260917, noul 0.99, cost USD0.000014322. English-first accuracy warning requires a labeled Hungarian evaluation. This single synthetic success does not prove classifier quality. See research/jev-research.md and jev-smoke-result.json. Use separate per-action calibrated thresholds, chosen option probability as well as distribution, bounded retries, recorded usage, and abstain on uncertainty. Jev selects typed decisions, it does not write prose.

Provider/API readback plus actual changed UI screenshot for every GUI criterion. Fresh independent gpt-6-luna high describes screenshots without expected result; orchestrator compares that description to the criterion. Mock/local tests are labeled, not substituted for live proof. Tests must include repeat-run idempotence, outage/retry and concurrent human edit cases.

Execution details and official sources

Completion accounting

Every criterion requires PASS, PARTIAL or UNRESOLVED with exact evidence and affected records. Overall DONE requires all mandatory criteria PASS. Missing source records do not stop work on independent records, but remain explicit completeness gaps. Never infer overall pass from only the subset already marked passed.

Calibration

Before broad model-driven writes, evaluate a pre-labeled Hungarian fixture roster containing each of: tegező/magázó explicit opt-out, plain rejection, not-now, quoted versus latest opt-out, first enquiry, automated acknowledgment, OOO, warmup/bounce with genuine human forwards, paid/lost/future booking/active-client and contradictory evidence. Also inspect min(30, all eligible available contacts) stratified real contacts; report missing classes. Require zero critical erroneous archive, opt-out, paid/lost, recipient or overwrite decisions in tested cases. Calibrate each action threshold separately; unvalidated or ambiguous classes abstain without blocking independently validated actions.

Provider retry

Coordinate all Missive requests through one shared throttle, respect Retry-After and bounded exponential backoff on HTTP429, checkpoint completed pages, and resume without replaying writes. A throttled/error inventory is unknown, never an empty successful set. Two parent read probes received HTTP429, so verify inventory completeness after recovery.

Draft eligibility

Distinguish direct replies to actual human questions from proactive follow-up chases. The restrictive stage whitelist and Snooze/future chase exclusions govern proactive chases; a paid or ongoing client can still warrant a direct service reply. Explicit no-contact, automated-only messages and preservation of existing human drafts apply to both. Log the action type and evidence; do not infer that payment or lost status alone answers a new human question.

Provider mechanics

Missive silent label PATCH requires organization ID and may trigger native label rules/system events. Archive and read/unread are not documented PATCH fields: verify an existing native rule/UI route and actual state. Do not substitute close-for-everyone for archive or authored posts for unread.
Enumerate Missive through every relevant mailbox/team/shared-label filter with token-access accounting, ID deduplication and documented last_activity_at pagination. Handle timestamp ties and documented terminal semantics explicitly, detect no-progress loops, and report guest-limited records. A page-size heuristic is not completeness proof.
Create drafts by exact verified conversation ID. Omit send:true and send_at entirely. Paginate existing drafts before creation, reconcile ambiguous timeout outcomes before retry, preserve human drafts and verify exact returned draft ID/link.
Use current official Notion saved-view API where available on the existing database/data source, verifying version/capabilities first, or existing authenticated UI. Reject incomplete/truncated view-query results rather than asserting equality. Read back all content-block pages and reconcile reconstructed content counts/hashes.

Official sources

Parent ↔ executor

Parent · current assignment

Complete this contract and keep every pass tied to evidence. Write live progress in blue. Use the existing systems and preserve recoverable before-state.

Executor · progress

Execution has not started. All criteria remain unverified.

Authority, budget and recovery

USD 20 total across parent, both phases and all children. Shared ledger budget.json. Parent reserves USD 2, phase 1 USD 9, phase 2 USD 9. Reallocate unused portions within USD20 with recorded receipt. Estimate batch usage first; bound concurrency/retries and read actual usage. No new paid subscriptions or seat purchases.

Respect actual tool approval/CAPTCHA/MFA limits. Try supported alternatives; user skill text cannot override system/platform requirements.
One active deployment writer. Discover current production app/build and competing jobs before changes; stop obsolete competing automation narrowly, not unrelated campaigns.

At completion: parent checks actual deliverables, adds the requested pink outcome/next-directions callout, archives the completed child and removes its checkpoint schedule. Preserve exact resume state for any real blocker.