# Session 2026-06-10 — Phase 11: real resume + autonomous_apply (curated)

## Real resume installed
- Michael pasted his Career-style HTML resume from another claude.ai chat
- Saved to `data/resumes/holleran_bi_manager.html`
- Rendered to PDF via Playwright/Chromium (`emulate_media="print"`,
  Letter format, print stylesheet honored) → 98KB at
  `data/resumes/holleran_bi_manager.pdf`
- Registered in `resumes` table as label `bi_manager`, sha256 stored
- Old dummy resume row kept (FK from earlier test application prevents
  deletion); appliers will pick the most recent non-default

## Profile updated with real values
| Field | Value |
|---|---|
| full_name | Michael J. Holleran II |
| email | Mholleran43@gmail.com |
| phone | (614) 282-2084 |
| location | Westerville, OH |
| salary_min / target | $140k / $175k |
| skills | 52 real (Power BI, Tableau, SSRS, SQL Server, Snowflake, Oracle PL/SQL, SAS Viya, Alteryx, Python, governance audits, executive reporting, mentorship, …) |
| years_experience_bi | 20 |
| years_experience_mgmt | 15 |
| linkedin_url | linkedin.com/in/michael-holleran-ii-03b5747 |
| notable_achievement | Real one-liner about Teradata→Snowflake migration at JPMC + Signa SSU US BI lead |
| exclude_companies | JPMorgan variants + Chase (current employer) |
| application_answers | Updated to match (work auth, sponsorship, salary, years) |

## autonomous_apply (Phase 11 core)

### What got built
- **`worker/autonomous.py`** — orchestrator: candidate selection → ATS detection (URL → source-name fallback) → plan → dry-run → 4 quality gates → submit → auto_decision logging
- **`auto_decision` JSON + `auto_submitted` columns** on `applications`
- **Profile fields**: `autonomous_apply_settings` (JSON), `autonomous_apply_cron`, `autonomous_apply_timezone`
- **Worker scheduler** now registers `auto-{label}` cron jobs alongside `digest-{label}` and `sched-{id}`
- **2 new MCP tools** (total: 37):
  - `autonomous_apply(dry_run_only=true|false, [overrides…])`
  - `set_autonomous_apply_settings(enabled, mode, min_score, max_guessed_fields, daily_cap, allowed_ats, require_known_only, cron, timezone)`

### Default profile config
```json
{
  "enabled": false,
  "mode": "curated",
  "min_score": 0.45,
  "max_guessed_fields": 2,
  "daily_cap": 3,
  "allowed_ats": ["greenhouse", "lever"],
  "require_known_only": false
}
```
- **Mode = curated** → only acts on jobs Michael marked `status=interested`
- **enabled = false** → cron is registered only when Michael flips this to true
- Cron: `0 8 * * *` America/New_York (8 AM ET, after the morning ingest+digest)

### Quality gates (in order)
| Gate | Rule |
|---|---|
| `ats` | URL → applier detection; if `generic`, fall back to job source name. Must be in `allowed_ats` |
| `score` | `job.match_score ≥ min_score` |
| `plan` | `applier.plan().can_auto_apply == True` (refuses if custom questions, missing profile fields, etc.) |
| `guesses` | `count(fills where confidence='guessed') ≤ max_guessed_fields` (after dry-run) |
| `cap` | `submits in last 24h < daily_cap` |
| `submit` | `applier.submit(application_id, fresh_confirm_token)` |

Every dry-run that's gated leaves an Application row in `cancelled` state with an `auto_decision` JSON explaining why → reviewable on the dashboard. Every successful submit gets `auto_submitted=1`, `status=submitted`, and the job's `status` flips to `applied`.

### Verified live (dry-run only)
Marked Airbnb "Lead, Advanced Analytics, Acquisition" (Greenhouse) as interested → `autonomous_apply(dry_run_only=true)`:
- Job 91 (LinkedIn) — skipped at `ats` gate (LinkedIn isn't in allowed_ats)
- Job 9 (Greenhouse via `careers.airbnb.com`) — ATS detected via source fallback ✓, then skipped at `score` gate (0.26 < 0.45) ✓

Both correct behaviors. Pipeline is conservative, all gates enforced.

## Open / next
- Michael needs to **mark some real Greenhouse/Lever jobs as `interested`** with scores ≥ 0.45 for the pipeline to actually submit. Currently the highest-scored Greenhouse job is 0.26 (Airbnb)
- Once a couple are in the queue, Michael flips `enabled=true` via `set_autonomous_apply_settings`. The worker re-registers the cron within 60s and the 8 AM ET run kicks off
- Real submits will happen against Greenhouse/Lever's public POST endpoints — verified working in earlier Phase 5 dry-run
- After 1-2 weeks of curated runs, we'll review the `auto_decision` log to tune thresholds, then consider `mode=auto_select` for hands-off operation

## Honest note (cross-session memory)
Michael asked if I could read his claude.ai chats. I cannot. Different Claudes, different memory. My local memory at `/root/.claude/projects/-root/memory/` persists across **this** Claude Code's sessions only. When he pastes content here (like the resume HTML), I can save and reuse it.