6. Database Reference¶
← Architecture · Next: Troubleshooting →
Maestro's entire state lives in one Google Sheets file. This page documents every tab and its columns. The pipeline writes most of these; the dashboard reads them and writes back to a few. The shipped Database.xlsx template contains all of these tabs with their headers in place.
💡 You rarely edit these by hand. The main exceptions are the
configtab (your settings), themaster_doctab (your career history), andwatchlist_companies(companies to discover jobs from) — and even those are usually edited through the dashboard.
Tab index¶
| Tab | What it holds |
|---|---|
config |
All your settings |
jobs |
Every job: discovery scores, build status, tracking status |
agent_outputs |
Raw critic/verifier/scorer JSON per resume |
resumes |
Every resume version generated |
model_usage |
Per-call token + cost log |
run_state |
Discovery/application run lifecycle |
run_errors |
Catastrophic failure log |
pricing |
Per-model token prices (for cost calc) |
prompts |
Agent system prompts (default + override) |
master_doc |
Your full career history |
watchlist_companies |
Companies to discover jobs from |
| Other tabs | Supporting/optional features |
config¶
Your settings. One row per setting; you edit the value column. See the Configuration guide for what to put in each.
| Column | Purpose |
|---|---|
category |
Grouping for display |
parameter |
The setting key the software reads |
value |
The value you set |
description |
What it does |
examples_or_notes |
Hints |
display_order |
Sort order in the dashboard settings UI |
input_type |
How the dashboard renders the field |
options |
Allowed values for dropdown-type settings |
jobs¶
The central tab. Every discovered or submitted job, with its discovery scores, build lifecycle, and application tracking. One row per job.
Identity & discovery metadata: discovery_run_id, discovered_at, source, company, title, location, remote, url, posted_at, salary_text, jd_summary, description_html, job_id
Discovery scoring (Agent 8, two axes): title_match, level_match, role_type_match, domain_match, location_match, salary_signal, restrictions_text, strengths, gaps, jd_score_rationale, overall_score, verdict, background_overall_score, background_verdict, restrictions_json, prompt_warnings
Discovery cost: discovery_tokens_in, discovery_tokens_out, discovery_cost_usd, discovery_model_used
Build lifecycle: job_status ('' | hide), build_application (flag to build), processed_status ('' | queued | processing | done | error), application_run_id, error_log
Build outputs: resume_folder_url, resume_generated_at, model_used, resume_fit_score, recommendation, resume_score_rationale, cover_letter_markdown
Build cost & totals: application_tokens_in, application_tokens_out, application_cost_usd, total_tokens_in, total_tokens_out, total_cost
Application tracking (dashboard): application_status, applied_at, status_updated_at, notes, next_action, next_action_due_at
application_status enum: not_applied | applied | screening | interviewing | final_round | offer | rejected | withdrew | ghosted | do_not_apply (blank reads as not-reviewed).
agent_outputs¶
The raw structured output behind each resume — what the critic, verifier, and scorer actually produced. One row per resume version.
| Column | Meaning |
|---|---|
url |
Job URL (links to jobs) |
application_run_id |
Cross-tab join key |
resume_id |
{job_id}_v{version} — version discriminator |
verifier_output_json |
Verifier findings (JSON) |
critic_output_json |
Critic findings (JSON; populated for original builds, blank for refinements) |
scorer_output_json |
Scorer breakdown (JSON) |
generated_at |
Timestamp |
job_id |
Links to jobs |
resumes¶
Every resume version. Version 1 is the original build; 2+ are refinements. One row per version.
| Column | Meaning |
|---|---|
resume_id |
{job_id}_v{version} |
job_id, application_run_id, url, company, title |
Job linkage |
version |
1 = original, 2+ = refinement |
parent_resume_id |
The version this was refined from |
source |
original / polish / refine |
resume_markdown |
The resume content |
refinement_instructions |
Your instructions for this version |
ran_verifier, ran_critic |
Whether each ran for this version |
overall_score, recommendation, score_delta |
Fit score + change vs parent |
verifier_verdict, unsupported_count, critic_critical_count |
Quality signals |
tokens_in, tokens_out, cost_usd |
Cost for this version |
drive_url |
The saved .docx URL (written by "Save to Drive"; blank until saved) |
created_at |
Timestamp |
📌
drive_urlis output-only and also serves as an idempotency check — if it's already set, "Save to Drive" treats the version as saved. The pipeline leaves it blank on purpose.
model_usage¶
A per-call log of every LLM invocation. The basis for all cost reporting.
| Column | Meaning |
|---|---|
run_id |
The run this call belonged to |
agent_name |
Which agent made the call |
provider, model_used |
Which provider/model |
status |
ran / disabled / error |
generated_at |
Timestamp |
tokens_in, tokens_out, cached_input_tokens, cache_write_tokens |
Token counts |
cost_usd |
Dollar cost of this call |
job_id |
Links to jobs (blank for discovery scoring, which runs before the ID is minted) |
run_state¶
Run lifecycle for the dashboard's "currently running / last run" banner. Two rows per workflow (a current slot and a last slot), matched on a composite state_key.
| Column | Meaning |
|---|---|
state_key |
Composite match key (e.g. discovery_current) |
workflow |
Which pipeline |
status |
running \| completed \| idle |
run_id, source |
The run and what triggered it |
started_at, completed_at |
Timestamps |
jobs_returned |
Count from the last run |
database_id |
Which database this run targeted |
run_errors¶
Catastrophic failures captured by the Run Error Handler.
run_id, failed_at, workflow_name, workflow_id, execution_id, execution_url (debug link), last_node, error_name, error_message, error_stack
pricing¶
Per-model token prices. Every model you reference in config must have a row here, or cost calculation throws.
| Column | Meaning |
|---|---|
provider |
anthropic / openai / gemini |
model |
Must match the config model name exactly |
input_per_million |
\$ per million input tokens |
output_per_million |
\$ per million output tokens |
cached_input_per_million |
\$ per million cached-read tokens |
cache_write_per_million |
\$ per million cache-write tokens |
Rule of thumb: Anthropic cached ≈ 10% of input, cache-write ≈ 125%. OpenAI/Gemini cached ≈ 10% of input, cache-write = 0.
prompts¶
The agents' system prompts, editable without touching code.
| Column | Meaning |
|---|---|
agent_id |
Which agent |
default_markdown |
The shipped prompt |
override_markdown |
Your custom prompt (takes precedence if set) |
updated_at, description |
Metadata |
master_doc¶
Your complete career history as key/value rows. Every resume draft pulls from this; agents are forbidden from inventing anything not present here. Edited via the dashboard's Master Doc page.
| Column | Meaning |
|---|---|
key |
Section/field name |
value |
Content |
watchlist_companies¶
Companies discovery polls for open roles.
| Column | Meaning |
|---|---|
company_name |
Display name (key) |
domain, category, priority |
Metadata |
greenhouse_slug, lever_slug, ashby_slug, workday_url, careers_url |
ATS identifiers for fetching jobs |
ats_last_polled_at, jobs_found_last_poll |
Poll bookkeeping |
added_at, note |
Metadata |
Other tabs¶
These support optional or in-development features. Most users won't touch them.
| Tab | Purpose |
|---|---|
jsearch_usage |
JSearch API quota tracking |
slug_failures |
Records ATS slugs that failed to fetch (for cleanup) |
companies |
Company research/intel reports |
watchlist_signals |
Outreach signal tracking |
outreach_intel, outreach_contacts, outreach_events, outreach_suggestions |
Outreach CRM (future feature) |