ask-monarch 0.1.60

Thin CLI for the hosted Ask Monarch source plane

Ask Monarch

Teammate Setup

Install Ask Monarch for a teammate with one command. Get a current setup code from Josh or an Ask Monarch admin (don't reuse a code pasted into docs):

uvx ask-monarch setup --code <YOUR-SETUP-CODE>

The code is stable until Josh or an admin rotates it. This command installs or refreshes the persistent CLI, exchanges the setup code, configures Codex and Claude Code, runs doctor, runs the canonical verification query, and reports the source id.

Ask Monarch is the conversational front door to its memory and capabilities.

A teammate should be able to ask Monarch what it knows, inspect the evidence, challenge the answer, and improve the memory as new artifacts appear.

Under the hood, Ask Monarch does not ask the live bucket or a loose pile of files every time. It follows the source contract:

map the source -> access it -> enumerate it -> parse it into atomic units -> verify receipts -> make it queryable

Once a source passes that gate, Ask Monarch answers from the parsed indexes with provenance. That is what makes answers fast, repeatable, and inspectable.

Read ASKMONARCH-SPEC.md for the deeper product contract and future direction.

Architecture

Ask Monarch is artifact-first: raw company artifacts flow through the monarch-source contract into a searchable store. Bucket refreshes also build a canonical experiment register so questions about what ran do not have to infer experiments from files. The ask-monarch CLI is the supported ask path.

Runtime: the GitHub repo manintheandes/ask-monarch is deployed to the ask-monarch-server VM; the CLI runs on the user or agent machine and calls the hosted API on that VM.

Visual Explainers

Ask Monarch's architecture and assay explainers are repo-owned visual assets. Manim scenes live in visuals/manim. Remotion demo videos live in videos. Both are scoped away from the hosted API and CLI runtime.

cd visuals/manim
uv run manim -pql scenes.py AskMonarchArchitecture

For assay visuals, start with Manim's plotting examples and map source-backed rows onto axes, curves, markers, guide lines, and highlighted regions. See visuals/manim/README.md.

For the current visual evidence loop, run the root verifier:

python3 scripts/verify_visual_evidence.py

That checks the source-action hook, overlay resolver, assay-overlay CLI dry-run, Manim brand palette, reusable Manim assay templates, and a smoke still render.

For Remotion video work, use the repo-owned brand tokens in videos/src/brand.ts and validate with npm run validate:brand from the videos folder.

Current State

Ask Monarch currently has twelve verified source lanes. For the fuller human-readable source list, see docs/source-lanes.md.

Source lane	What it covers	Status
bucket	gs://monarch-videos-new, including experiment outputs, CSV rows, spreadsheets, images, videos, PDFs, model metadata, structured sidecars, and a canonical experiment register spanning mapped DART, contact/no-contact, phytotoxicity, oviposition, post-exposure, spatial-affect, and arm-in-cage families	Parsed and queryable
presentations	46 Monarch update decks from the Storyboards_Presentations Google Drive folder, scoped to January 1, 2026 onward	Parsed and queryable
company_profile	Monarch mission, problem, status quo, approach, headquarters, organisms, and crop-testing context	Parsed and queryable
people_directory	Monarch employees, advisors, and partnership facts	Parsed and queryable
company_context	Company context from Monarch_for_askmonarch.pdf, parsed into PDF pages, text blocks, and embedded-image references	Parsed and queryable
investor_faq	Investor-facing FAQ from Monarch Investor FAQ 2.pdf, parsed into PDF pages, text blocks, and embedded-image references	Parsed and queryable
teams_recaps	Recurring Monarch R&D, Monarch Weekly, and Monarch Team Sync Teams meeting transcripts, refreshed daily after midnight	Parsed and queryable
scientific_refs	Scoped scientific reference library, including chemical ecology, receptor/pathway, modeling, assay-prototype, scalability, EPA registration-support, and Teams-discussed journal-article material	Parsed and queryable
milestones_doc	monarch_milestones Google Doc, scoped to one fetched revision and parsed into paragraph units	Parsed and queryable
compound_inventory	Monarch compound / chemical inventory workbook from the user-provided Updated Chemical Inventory.xlsx attachment	Parsed and queryable
chemistry_analysis	gs://monarch-chemistry-analysis, including chemical activity registry rows, docking result rows, JSON score sidecars, protein-structure metadata, model metadata, images, scripts, and text sidecars	Parsed and queryable
ask_insects_context	Hash-pinned immutable public Ask Insects package with exact-species profiles, deep insect-knowledge domains, assay interpretation boundaries, public evidence, and explicit knowledge gaps; it is not digitally signed	Parsed and queryable

Experiment Register

The bucket index represents each real experiment or recorded trial once in experiments. Tube, replicate, timepoint, treatment, measurement, video, and processed-output evidence is attached beneath that row in:

• experiment_observations
• experiment_treatments
• experiment_results
• experiment_artifacts
• experiment_register_coverage

run_date follows strict source precedence: an explicit date in the family's canonical ledger or metadata workbook, then embedded video acquisition time, then an acquisition timestamp encoded in a raw filename or experiment folder. If none is available, run_date remains null and the register records a source gap. Bucket time_created and updated values never become run dates.

Upload and processing questions use the separate file_activity view. Run-date inventory questions use bucket_experiment_register_run_inventory_sql with answer shape experiment_register_run_inventory. That single query returns one row per experiment with treatment or compound, dilution, solvent, control status, every actual result and its status, media status, and exact experiment and child provenance. A validated response is answer-ready and must not trigger follow-up SQL. Agent fast paths use ask-monarch query 'QUESTION' --limit 100 --format markdown, which validates the same one hosted response and emits only the final compact numbered table. Shared source text appears once while every row retains its exact locator. The raw JSON route remains available for programs and diagnostics but is too large for the user-facing Codex fast path. Recent-compound questions that also request the best result and its video use bucket_experiment_register_compounds_media_sql with answer shape experiment_register_compounds_with_winner_media.

That combined route lists every non-control compound in the requested run-date window, ranks only comparable completed measurements, and links media through the winning experiment_id. If no comparable result exists, Ask Monarch says that no winner can be measured yet and returns no arbitrary video.

Two optional high-value release diagnostics are exact. A valid result for what experiments were run yesterday returns complete deduplicated register rows for the resolved absolute date from one hosted source call and renders visibly in Codex in under 30 seconds. A valid result for `list the compounds that were tested in the last 7 days and pull up the video with the most repellent compound

• all in-line` keeps the compound list, date window, ranking status, and media obligation together and does not select a video when indexed results do not establish a winner. Neither diagnostic is a completion gate.

Run the fresh-Codex diagnostic through the Codex desktop app's bundled runtime. Low reasoning is deliberate here: the hosted response is already validated and Codex only has to return it unchanged.

If you choose to run the first optional diagnostic, create a brand-new one-turn Codex desktop task in this repository with the exact question what experiments were run yesterday. After it finishes, export that task with codex_app read_thread, setting includeOutputs=true, maxOutputCharsPerItem=20000, and turnLimit=1. Save the complete raw JSON result without editing it in a current-user-owned regular file with mode 0600. The visible verifier rejects a transcript copied from the UI, missing command output, truncated output, older task history, and a headless-only run.

git fetch origin
REVISION="$(git rev-parse HEAD)"
test "$REVISION" = "$(git rev-parse refs/remotes/origin/main)"
test -z "$(git status --porcelain=v1 --untracked-files=all)"
RELEASE_RECEIPT="$HOME/.codex/evals/ask-monarch/release-receipts/$REVISION.json"
DESKTOP_EVIDENCE="$HOME/.codex/evals/ask-monarch/desktop-task-$REVISION.json"
test -f "$DESKTOP_EVIDENCE"
chmod 0600 "$DESKTOP_EVIDENCE"
PINNED_PYTHON="$HOME/.local/share/uv/python/cpython-3.12.13-macos-aarch64-none/bin/python3.12"
test "$(/usr/bin/shasum -a 256 "$PINNED_PYTHON" | /usr/bin/awk '{print $1}')" = \
  "e2605291e058fdbe3102e8185d0ac5fe0e063398de617010a6af3a42a78f05e3"

"$PINNED_PYTHON" -I -S -B scripts/verify_visible_yesterday_experiments.py \
  --release-receipt "$RELEASE_RECEIPT" \
  --desktop-evidence "$DESKTOP_EVIDENCE" \
  --report "$HOME/.codex/evals/ask-monarch/visible-$REVISION.json"

Run that command from a completely clean checkout whose HEAD and fetched origin/main both equal the revision named by the recent successful production receipt. Copy the receipt without editing it and keep the local file non-writable by group or other users. The verifier derives the expected runtime bundle and installed CLI package from those two release artifacts. It does not accept caller-provided executable hashes, package hashes, a different Codex binary, a different Ask Monarch executable, a different latency limit, or a stale expected row count.

The verifier always uses the Codex desktop app's bundled binary, the canonical ~/.local/bin/ask-monarch derived from the current account record, and the system /bin/zsh. It times the whole Codex turn, audits actual hosted calls, and parses the final visible table. It also binds its sealed headless route diagnostic to the raw one-turn desktop export. CLI-only or headless-only latency is not accepted as a proxy for what the user saw.

Current presentation receipt:

• 46 decks
• 622 slides
• 1,388 slide text or notes units
• 736 media/chart/link references
• 0 explicit gaps

The bucket receipt currently proves all live bucket objects are parsed and that CSV rows are materialized atomically in SQLite. The large CSV-row count is not an animal count; it is source-unit count from processed frame/result files.

Current Teams recap receipt:

• 24 meeting instances across R&D, Weekly, and Team Sync through May 12, 2026
• 24 matching transcript files
• 12,595 speaker/timestamp/utterance units
• 0 explicit gaps

Current company profile receipt:

• 7 company profile units
• 0 explicit gaps

Current people directory receipt:

• 15 people, advisor, and partnership units
• 0 explicit gaps

Current company context receipt:

• 22 PDF pages
• 89 text blocks
• 59 embedded-image references
• 171 total atomic units
• 0 explicit gaps

Current milestones doc receipt:

• 1 Google Doc revision
• 30 paragraph units
• 0 tables
• 0 comments
• 0 explicit gaps

Current compound / chemical inventory receipt:

• 8 workbook sheets
• 1,806 rows with values
• 9,280 non-empty cell units
• 0 explicit gaps
• Original SharePoint link access still failed for the current connector identity; the parsed source is the attached workbook copy.

Current chemistry analysis receipt:

• 11,302 bucket objects
• 550,576 CSV row units
• 30,412 JSON units
• 99,206 text line units
• 3,637 protein-structure metadata units
• 0 explicit gaps

Current Braintrust receipt:

• Project: Andes / ask-monarch
• Project id: 28216333-66fb-42a6-923b-622bda7a2fcc
• Query responses include trace_id, trace_span_id, trace_project, trace_project_id, and trace_url when BRAINTRUST_API_KEY is configured.
• Query traces include spans for trace.user_question, trace.tool_call.1.ask_monarch_source_query, trace.selected_evidence, trace.source_receipts, and trace.final_answer.

How To Ask

In any capable coding or research agent, use the repo-owned Ask Monarch skill:

what is the most effective compound against mosquitoes [$askmonarch](/Users/josh/.codex/skills/askmonarch/SKILL.md)

The user should not need to know which source, database, table, bucket, or deck contains the answer. $askmonarch means: use the skill's routing rules, call the ask-monarch CLI for evidence, answer from the returned rows, and say clearly when the memory does not yet contain enough evidence.

Personal mailbox access is a live overlay, not shared company memory. A mapped Josh-only Ask Monarch token can access the shared Monarch source plane and the mapped [e-mail] mailbox through Microsoft Graph; the shared/team token cannot access personal mail, and personal mail is not indexed into shared SQLite.

For direct shell use, the remote CLI calls the hosted Ask Monarch service:

ask-monarch health
ask-monarch summary
ask-monarch query "what are the most effective repellents?"
ask-monarch query "list the experiments from last week" --limit 10
ask-monarch query "what did today's presentation cover?" --source presentations --limit 8
ask-monarch presentation --latest --open
ask-monarch presentation --date 2026-01-02 --open
ask-monarch query "what did the latest R&D meeting cover?" --source teams_recaps --limit 8
ask-monarch query "what does Monarch do?" --source company_profile --limit 8
ask-monarch query "who works at Monarch?" --limit 50
ask-monarch query "what does the company context say about residues?" --source company_context --limit 8
ask-monarch search "2pp mosquito" --limit 5
ask-monarch search damage --source presentations --limit 5
ask-monarch search "spatial repellents" --source teams_recaps --limit 5
ask-monarch search "acetone" --source compound_inventory --limit 5
ask-monarch company-profile-search mission --limit 5
ask-monarch people-search advisor --kind advisors --limit 10
ask-monarch company-context-search glossary --limit 5
ask-monarch query "62 chemicals" --source milestones_doc --limit 5
ask-monarch milestones-search "62 chemicals" --limit 5
ask-monarch compound-inventory-search acetone --limit 5
ask-monarch search DEET --source chemistry_analysis --limit 5
ask-monarch sql "SELECT COUNT(*) AS rows FROM csv_rows" --source chemistry_analysis
ask-monarch search "oviposition" --source ask_insects_context --limit 5
ask-monarch interpret-experiment EXPERIMENT_ID
ask-monarch interpret-experiment EXPERIMENT_ID --hide-outcomes
ask-monarch csv-search IR3535 --name-like combined_results.csv --limit 5
ask-monarch sql "SELECT COUNT(*) AS slides FROM slide_units" --source presentations
ask-monarch --open-trace query "what does Monarch do?" --source company_profile --limit 5
ask-monarch trace-url TRACE_ID
ask-monarch personal-mail-status
ask-monarch personal-mail-search "mosquito field trial" --days 30 --limit 10
ask-monarch personal-mail-read GRAPH_MESSAGE_ID
ask-monarch personal-mail-draft --to [e-mail] --subject "Subject" --body "Body"
ask-monarch personal-mail-reply-draft GRAPH_MESSAGE_ID --body "Reply body"
ask-monarch personal-mail-send-draft GRAPH_DRAFT_ID

The CLI returns JSON so Codex, Claude Code, shell scripts, and dashboards can consume the same source evidence reliably.

How Corrections Improve Ask Monarch

When Ask Monarch gives a weak answer, the fix should become part of the system, not just part of one chat. There are three places to make that correction.

For source-question regressions, run the live drill:

python3 scripts/run_askmonarch_regression.py

The drill replays known tricky questions against the hosted source service and checks both the returned source values and the $askmonarch skill rules that should route future agents correctly. When a case fails, reproduce it manually, patch the skill/source/parser/CLI layer that caused the miss, then add or tighten the case before rerunning the drill.

Latency counts too. The drill prints wall-clock time per case, summarizes the slowest case, and fails any case over the default per-case budget of 10 seconds. Use --default-max-seconds to tune the budget for a run, or set max_seconds on a case that deserves a tighter or looser threshold.

When intentionally failure-mining, use plausible teammate questions, not impossible trivia. Count a failure when the hosted source path returns no usable evidence, the route is ambiguous, units cannot be verified, or the answer would require inventing a linkage. Each useful failure should leave behind a durable rule, a regression case, or a named source/tooling gap.

For answer-quality regressions, run the live answer contract gate:

python3 scripts/verify_askmonarch_answer_quality.py

For a clean pass/fail check that does not rewrite the committed ledger outputs:

python3 scripts/verify_askmonarch_answer_quality.py --check-only

That gate checks that composed answers include the requested factual claims, that evidence rows have citations, that broad questions cite multiple source lanes when needed, and that negative/source-gap answers name the checked sources instead of making absolute claims over unqueryable artifacts. It writes:

evals/askmonarch_answer_quality_results.json
evals/askmonarch_answer_quality_findings.md
evals/askmonarch_answer_quality_ledger.html

Each ledger case must explain what failed, why it failed, how the correction was made durable, and the exact verification command that proves the correction. See evals/askmonarch_quality_probe_summary_2026-05-03.md and evals/askmonarch_quality_probe_summary_2026-05-03.html for the May 3 hard probe refresh summary.

1. Fix the agent instructions

Update the $askmonarch skill when Codex or Claude misunderstood the question or chose the wrong source.

Example: if someone asks "how many mosquitoes have we used?" the skill should say to use the assay ledger's n_mosquitoes field, not CSV frame rows.

Example: if someone asks "what compounds does Monarch have in inventory?", the skill should route to compound_inventory and count or list Master!D compound names, not bucket search snippets. If they ask "where is X?", return the matched Master row fields such as Location/Bin, Vendor, CAS, Quantity, SDS, and cell provenance.

Example: if someone asks "what experiments were run today?", the skill should resolve today to an absolute date and query experiments.run_date across every registered assay family. It should not include bucket uploads or processed files unless the user asks about file activity.

2. Fix the shared rulebook

Update the README or spec when the rule should be visible to everyone working on Ask Monarch.

Example: if presentations are queryable only from January 1, 2026 onward, the spec should say that clearly so future agents do not treat every old deck as source-grade.

3. Fix the tool

Add or improve an API/CLI command when the same question will come up again and should not depend on a fresh ad hoc SQL query each time.

Example: if animal-use questions become common, add a command like ask-monarch animal-use mosquitoes that always uses the right ledger field and returns the right evidence.

For most Monarch teammates, the workflow stays simple: ask through Codex, Claude, or another agent and inspect the evidence. Behind the scenes, the agent uses the skill and ask-monarch CLI. Power users can also use the remote CLI directly. Everyone benefits when recurring corrections are captured in these shared layers.

Hosted Service

Ask Monarch is deployed on a Google Compute Engine VM.

Piece	Value
VM	ask-monarch-server
Google Cloud project	gen-lang-client-0407939408
Zone	us-central1-a
Main release	/opt/ask-monarch/main/releases/$REVISION
Main Python environment	/opt/ask-monarch/main/venvs/$REVISION
Experiment-review release	/opt/ask-monarch/experiment-review/releases/$REVISION
Legacy aggregate data root	/home/josh/ask-monarch/artifacts (data only, never served as code)
Public HTTPS endpoint	https://ask-monarch.34.121.138.236.sslip.io
Main local listener	http://127.0.0.1:8787
Review local listener	/run/ask-monarch-experiment-review/review.sock
Systemd services	ask-monarch-http.service, ask-monarch-experiment-review.service

The services are read-only. Public traffic enters through HTTPS and bearer-token auth. Nginx proxies normal routes to the private main listener and only exact POST /experiment-interpretation to the isolated Unix socket. Production code is installed as a root-owned immutable release. It is never served from a mutable checkout.

The canonical deployment and proof sequence is docs/production-release.md.

Health check:

curl -H "Authorization: Bearer $ASK_MONARCH_TOKEN" \
  https://ask-monarch.34.121.138.236.sslip.io/health

Unauthenticated requests should return 401.

Team Onboarding

Do not send broad team invites until the clean-teammate simulation passes and the Codex app, Codex terminal, Claude app, and Claude Code terminal surfaces are verified or explicitly recorded as manually confirmed.

Admin preflight:

python3 scripts/simulate_teammate_onboarding.py --mode temp-home

The simulator creates a fresh one-time setup code, installs into a fake clean home, verifies Codex, shared .agents, and Claude setup, runs ask-monarch doctor, runs source-plane parity checks, and asks canonical Monarch questions for answer parity. It writes reports under output/onboarding-simulations/.

For each teammate, create one personal setup code:

ask-monarch onboarding-code create --email [e-mail] --label "Person setup" --ttl-hours 168

Send that teammate only their own install command:

ASK_MONARCH_SETUP_CODE='MNR-XXXX-YYYY' bash -c 'curl -fsSL https://ask-monarch.34.121.138.236.sslip.io/install.sh | bash'
ask-monarch doctor

After installation, they should open a fresh Codex or Claude thread before asking Monarch questions.

Source Inspection

You can also pull up assay videos during Q&A. A useful Ask Monarch answer should not stop at a numeric result when the underlying assay is inspectable. When an answer identifies an exact assay row, source video path, or filename, the CLI can resolve the linked raw video from the verified bucket index, download it into the local media cache, and open it:

ask-monarch assay-media --filename IMG_0049_Contact_3PP.mov --open
ask-monarch assay-media --row 1255 --source-name mosquitoes/Mosquito_DART_assay_Results_GCS.xlsx --sheet Assays --open

This is the explicit raw-media exception path. Ask Monarch still answers from verified source indexes first; raw videos are opened only after the indexed source record has been resolved. If a raw video is not linked yet, the command returns diagnostics instead of pretending the video is available.

Assay metric overlays are the next inspectable layer. When an assay has processed metrics, Ask Monarch should be able to move from:

Peak mean avoidance was 0.683 at 05:30.

to:

Peak mean avoidance was 0.683 at 05:30.
Source: advanced_metrics.csv#row=6.
Artifact: raw assay video with the avoidance curve overlaid.

The assay-overlay command does that for supported processed metric CSVs. DART and related tube assays use advanced_metrics.csv with T*_avoidance columns. Contact/non-contact repellency videos use frame_metrics.csv with contact_pi and inside/on-band/outside counts. The command resolves the matching raw video from the verified bucket index and renders a synced time-lapse overlay. The answer source is still the verified SQLite row; the overlay is an inspection artifact. Today this runs from the repo/VM environment where the verified SQLite index is present. With --open, local assay videos and rendered overlays open in a browser player.

ask-monarch assay-overlay \
  --metric-csv 'mosquitoes/processed/_2026_04_28_14_44_46_BHomocyclo3_mov_results/_2026_04_28_14_44_46_BHomocyclo3_mov_results/advanced_metrics.csv' \
  --output videos/out/mosquito-bhomocyclo-overlay.mp4 \
  --speed 15 \
  --open

The renderer automatically discovers T*_avoidance columns, narrows to the assay tubes when source metadata declares them, resolves the raw video object, downloads it into videos/source/, and prints a proof payload with the peak row, timestamp, source CSV provenance, raw media URI, and output video probe.

Fast verification:

python3 scripts/verify_assay_metric_overlays.py

Two verified examples:

# Fly repellency: 2PP, treatment tubes T5-T8, peak 50:30 / 0.654
ask-monarch assay-overlay \
  --metric-csv 'flies/processed/2026-04-22 12-12-14_results/2026-04-22 12-12-14_results/advanced_metrics.csv' \
  --output videos/out/fly-2pp-overlay.mp4 \
  --speed 30 \
  --open

# Mosquito DART: BHomocyclo, tubes T1-T3, peak 05:30 / 0.683
ask-monarch assay-overlay \
  --metric-csv 'mosquitoes/processed/_2026_04_28_14_44_46_BHomocyclo3_mov_results/_2026_04_28_14_44_46_BHomocyclo3_mov_results/advanced_metrics.csv' \
  --output videos/out/mosquito-bhomocyclo-overlay.mp4 \
  --speed 15 \
  --open

# Fly non-contact repellency: 2PP, frame-level contact_pi
ask-monarch assay-overlay \
  --metric-csv 'contact_repel_videos/flies/processed/2PP_non_contact/frame_metrics.csv' \
  --output videos/out/fly-2pp-non-contact-overlay.mp4 \
  --speed 15 \
  --open

The R&D presentation lane is openable too. Indexed Monarch update decks start with the January 2, 2026 presentation, and ask-monarch presentation --open opens the matching Google Slides locator returned by the verified presentations index:

ask-monarch presentation --latest --open
ask-monarch presentation --date 2026-01-02 --open
ask-monarch presentation --query "3-Step Process" --open

Every query row that has source provenance can also return source_actions. Those actions are the provenance resolver layer: they do not change the answer source, which remains verified SQLite3, but they tell a client how to inspect the underlying artifact when one is available. A returned action can point to a raw video, Google Slides deck or slide, transcript annotation, CSV row, workbook cell, PDF page, image, assay metric overlay, or other artifact. Power users can open any returned locator directly:

ask-monarch resolve-source --locator 'https://docs.google.com/presentation/d/.../edit#slide=2' --open
ask-monarch resolve-source --locator 'gs://monarch-videos-new/path/to/video.mov' --open
ask-monarch assay-overlay --locator 'gs://monarch-videos-new/path/to/advanced_metrics.csv#row=6' --open
ask-monarch assay-overlay --locator 'gs://monarch-videos-new/path/to/frame_metrics.csv#row=10' --open

The installer saves the token locally at ~/.config/ask-monarch/config.json. To reconfigure it later:

ask-monarch configure

See docs/remote-cli.md for the full CLI reference.

Agent Skills

This repo owns the agent skills/instructions that make the experience feel like asking the company rather than running a database query:

skills/askmonarch/SKILL.md
skills/braintrust/SKILL.md
skills/correction/SKILL.md
skills/monarch-source/SKILL.md

The hosted installer writes:

• Codex global instructions: ~/.codex/AGENTS.md
• Codex skills: ~/.codex/skills/
• Claude global instructions: ~/.claude/CLAUDE.md
• Claude skill: ~/.claude/skills/ask-monarch/

Install or refresh repo-owned Codex skills on the VM with:

python3 scripts/install_codex_skills.py

monarch-source defines when a Monarch artifact becomes source-grade. $askmonarch is the conversational front door over source-grade Monarch memory. correction defines what to do when an Ask Monarch answer needs a durable routing or answer fix. braintrust defines the Ask Monarch trace inspection and verification workflow. The canonical text is host-agnostic: Codex, Claude, Gemini, Pi, OpenCode, and other agents should follow the same source-routing contract when they can run the CLI or call the hosted service.

Source Lanes

Bucket

• Source id: monarch_videos_new
• Bucket: gs://monarch-videos-new
• Project: gen-lang-client-0407939408
• Parser: scripts/source_pipeline.py
• Verification: scripts/verify_source_complete.py
• Index on VM: /home/josh/ask-monarch/artifacts/monarch-videos-new/source_index.sqlite (the lane directory resolves to the external large-artifact root)

Useful commands:

python3 scripts/parse_monarch_bucket.py --manifest-only
python3 scripts/source_pipeline.py init
python3 scripts/source_pipeline.py parse --lane structured
python3 scripts/source_pipeline.py parse --lane video
python3 scripts/source_pipeline.py parse --lane model
python3 scripts/source_pipeline.py parse --lane pdf
python3 scripts/source_pipeline.py parse --lane image
python3 scripts/source_pipeline.py build-derived
python3 scripts/source_pipeline.py status
python3 scripts/verify_source_complete.py

build-derived creates the canonical experiment register and its child observation, treatment, result, artifact, and coverage tables across mapped assay families. It also creates the separate file_activity view. A run date comes only from the canonical ledger/metadata date, embedded source-video acquisition time, or a timestamp encoded in the raw filename/folder. Missing dates remain explicit source gaps.

experiment_artifacts is the canonical source-grade bridge from an exact experiment_id to raw media and processed outputs. Legacy assay_media_links remains available for row-based callers. The remote CLI exposes that older path as ask-monarch assay-media ... --open, which can cache and open the video for inspection during an Ask Monarch session.

Presentations

• Source id: monarch_storyboards_presentations_2026
• Folder: Storyboards_Presentations
• Folder id: 1-dL59DdeoPySGbqE5cjGqednyp4CIzch
• Boundary: Monarch update decks from January 1, 2026 onward
• Freshness: daily Drive enumeration and verified hosted promotion after the VM refresh is inactive
• Refresher: scripts/refresh_monarch_drive_sources.py
• Parser: scripts/parse_storyboards_presentations.py
• Query helper: scripts/query_storyboards_presentations.py
• Index on VM: /home/josh/ask-monarch/artifacts/storyboards-presentations-2026/source_index.sqlite
• Receipt: docs/storyboards-presentations-2026-source.md

Useful commands:

ask-monarch presentation --latest --open
ask-monarch presentation --date 2026-01-02 --open
ask-monarch presentation --query "3-Step Process" --open
python3 scripts/parse_storyboards_presentations.py
python3 scripts/query_storyboards_presentations.py 2PP --limit 5
python3 scripts/query_storyboards_presentations.py damage --limit 5

ask-monarch presentation resolves decks through the hosted presentations source index and opens the Google Slides locator stored in deck or slide provenance. This source does not claim the entire historical presentation folder. Older decks, PDFs, notebooks, interviews, Teams messages, emails, academic papers, and dashboards are future source candidates until they pass the same source gate.

Monarch Teams Recaps

• Source id: monarch_recurring_teams_recaps
• Source lane: teams_recaps
• Source system: Microsoft Teams / Microsoft Graph
• Boundary: recurring Monarch R&D meetings from January 2, 2026 onward, Monarch Weekly meetings from May 5, 2026 onward, and Monarch Team Sync meetings from May 6, 2026 onward
• Freshness: VM refresh runs just after midnight America/Los_Angeles and includes meetings through the previous Pacific calendar day
• Refresher: scripts/refresh_teams_recaps_source.py
• Fetcher: scripts/fetch_monarch_r_and_d_teams_recaps.py
• Parser: scripts/parse_monarch_r_and_d_teams_recaps.py
• Query helper: scripts/query_monarch_r_and_d_teams_recaps.py
• Index on VM: /home/josh/ask-monarch/artifacts/monarch-r-and-d-teams-recaps/source_index.sqlite
• Receipt: docs/monarch-r-and-d-teams-recaps-source.md

Useful production-safe commands:

sudo systemctl start ask-monarch-refresh.service
python3 scripts/query_monarch_r_and_d_teams_recaps.py "spatial repellents" --limit 5

The manual service start refreshes every scheduled VM lane through the shared production lock. The Teams child refresher is an internal implementation detail and must not be invoked directly against production.

This lane is meeting-transcript memory, not the whole Teams universe. Teams messages, channels, company-wide email sources, papers, dashboards, and older meeting transcripts remain future source candidates until they pass the same source gate. Personal mailbox access, when mapped for a personal token, is a private live overlay and does not promote those emails into shared company memory.

Company Profile

• Source id: monarch_company_profile
• Source lane: company_profile
• Raw source: sources/monarch-company-profile/company_profile.json
• Parser: scripts/parse_company_lanes.py
• Index on VM: /home/josh/ask-monarch/artifacts/monarch-company-profile/source_index.sqlite
• Receipt: docs/monarch-company-profile-source.md

Useful commands:

python3 scripts/parse_company_lanes.py
ask-monarch company-profile-search mission --limit 5
ask-monarch search "where is Monarch based" --source company_profile --limit 5

This lane is durable company framing, not an automatically discovered corporate knowledge base. Update the JSON source when the company framing changes, then rerun the parser.

People Directory

• Source id: monarch_people_directory
• Source lane: people_directory
• Raw source: sources/monarch-people-directory/people_directory.json
• Parser: scripts/parse_company_lanes.py
• Index on VM: /home/josh/ask-monarch/artifacts/monarch-people-directory/source_index.sqlite
• Receipt: docs/monarch-people-directory-source.md

Useful commands:

python3 scripts/parse_company_lanes.py
ask-monarch people-search Avinash --limit 5
ask-monarch people-search advisor --kind advisors --limit 20
ask-monarch sql "select name, role, lane, provenance_grain from units where lane = 'employees' order by name" --source people_directory --limit 50
ask-monarch sql "select name, role, lane, provenance_grain from units order by lane, name" --source people_directory --limit 50

This lane covers people facts explicitly entered into the source JSON. It should not be treated as a full HRIS, inbox, or LinkedIn scrape.

Company Context

• Source id: monarch_company_context
• Source lane: company_context
• Raw source: sources/monarch-company-context/Monarch_for_askmonarch.pdf
• Boundary: exactly this one 22-page PDF file
• Parser: scripts/parse_monarch_company_context.py
• Index on VM: /home/josh/ask-monarch/artifacts/monarch-company-context/source_index.sqlite
• Receipt: docs/monarch-company-context-source.md

Useful commands:

python3 scripts/parse_monarch_company_context.py
ask-monarch company-context-search glossary --limit 5
ask-monarch search "USDA Insecticide Residue Report" --source company_context --limit 5
ask-monarch sql "select page_number, unit_index, text, provenance_grain from text_units where text like '%USDA%'" --source company_context --limit 20

This is company context sourced from PDF intake evidence. Current employee/advisor questions should use people_directory first; older names in the PDF do not override current people-directory truth.

Investor FAQ

• Source id: monarch_investor_faq
• Source lane: investor_faq
• Raw source: sources/monarch-investor-faq/monarch-investor-faq-2026-06-16.pdf
• Boundary: exactly this one 5-page PDF file
• Parser: scripts/parse_monarch_investor_faq.py
• Index on VM: /home/josh/ask-monarch/artifacts/monarch-investor-faq/source_index.sqlite
• Receipt: docs/monarch-investor-faq-source.md

Useful commands:

python3 scripts/parse_monarch_investor_faq.py
ask-monarch investor-faq-search burn --limit 5
ask-monarch search "commercialization partner" --source investor_faq --limit 5
ask-monarch sql "select page_number, unit_index, text, provenance_grain from text_units where text like '%break-even%'" --source investor_faq --limit 20

This lane is investor-facing FAQ evidence. It is useful for FAQ claims on raise, burn, break-even target, commercialization partner, distribution, product results, market framing, risks, and partnership discussions. Current employee/advisor questions should still use people_directory first unless the user explicitly asks what the investor FAQ says.

Milestones Doc

• Source id: monarch_milestones_doc
• Source lane: milestones_doc
• Source system: Google Docs / Google Drive connector
• Boundary: exactly the monarch_milestones Google Doc revision fetched on May 1, 2026
• Parser: scripts/parse_monarch_milestones_doc.py
• Query helper: scripts/query_monarch_milestones_doc.py
• Index on VM: /home/josh/ask-monarch/artifacts/monarch-milestones-doc/source_index.sqlite
• Receipt: docs/monarch-milestones-doc-source.md

Useful commands:

python3 scripts/parse_monarch_milestones_doc.py
ask-monarch query "phytotoxicity" --source milestones_doc --limit 5
ask-monarch milestones-search "62 chemicals" --limit 5
ask-monarch sql "select paragraph_index, text, provenance_grain from paragraph_units" --source milestones_doc --limit 5

This lane is a document-revision snapshot. Refetch the Google Doc and rerun the parser before using it for claims about later milestone edits.

Querying Locally

The hosted path is preferred for shared use. Local scripts are still useful for debugging and parser development:

Braintrust Tracing

Ask Monarch query requests emit Braintrust traces when the runtime has the Python dependency and an API key:

python3 -m pip install -r requirements.txt
export BRAINTRUST_API_KEY=...
export ASK_MONARCH_BRAINTRUST_PROJECT=ask-monarch
export ASK_MONARCH_BRAINTRUST_PROJECT_ID=28216333-66fb-42a6-923b-622bda7a2fcc

Tracing is on by default for /query when the SDK and API key are present, and fails open: if either is missing, source queries still run and simply omit trace fields. Set ASK_MONARCH_BRAINTRUST_TRACE_ALL=0 only for local debugging that must suppress default traces. When tracing is active, every /query response includes a Braintrust full trace URL. The local Braintrust context is stored in .bt/config.json.

python3 scripts/query_source.py --summary
python3 scripts/query_source.py --q "Preference Index" --kind pdf_page --limit 5
python3 scripts/query_source.py --csv-q IR3535 --name-like combined_results.csv --limit 5
python3 scripts/query_source.py --sql "select name, row_number, row_json, provenance_grain from csv_rows_with_provenance where row_json like '%IR3535%' limit 5"

See docs/querying-monarch-source.md.

Deep Research Max

Normal Ask Monarch questions should query verified source indexes directly. Deep Research Max is only for long-form external research, literature comparison, scientific synthesis, market/context scans, or visual analysis grounded in Monarch evidence.

Deep Research does this:

research question -> verified multi-source evidence bundle -> Gemini Deep Research Max -> cited report

The bundle currently includes verified lanes such as bucket, presentations, teams_recaps, company_profile, people_directory, and company_context when relevant. The milestones_doc lane is queryable through the hosted HTTP API and remote CLI, but should be added to Deep Research bundles before relying on it in that mode. If a required source lane is missing or has unverified receipts, Deep Research stops before launch instead of producing a partial internal-evidence report.

Bundle-only:

python3 scripts/deep_research_monarch.py "Compare our 2PP fly and mosquito assay signal against published 2-propylphenol repellency literature."

Launch Gemini Deep Research Max:

GEMINI_API_KEY=... python3 scripts/deep_research_monarch.py "Compare our 2PP fly and mosquito assay signal against published 2-propylphenol repellency literature." --launch --wait

Outputs live under:

reports/deep-research/

See docs/deep-research-mode.md.

Daily Refresh

The VM should refresh parsed scheduled source indexes daily just after midnight America/Los_Angeles. The VM-scheduled lanes are bucket (gs://monarch-videos-new), chemistry_analysis (gs://monarch-chemistry-analysis), and teams_recaps recurring Teams meeting transcripts. A separate launchd job refreshes presentations and scientific_refs, then waits for the VM refresh to become inactive before promoting those Drive-backed indexes.

The main bucket path is incremental: reuse the current verified SQLite index, enumerate the live bucket, invalidate only added/changed/deleted objects, parse those objects, then verify before replacing the live SQLite files:

live source -> staged incremental parse -> receipts/gaps -> verification -> atomic swap

Repo-owned pieces:

• scripts/refresh_source_index.py
• scripts/refresh_chemistry_analysis_source.py
• scripts/refresh_teams_recaps_source.py
• scripts/refresh_bucket_sources.py
• scripts/refresh_monarch_drive_sources.py
• scripts/parse_scientific_reference_library.py
• deploy/systemd/ask-monarch-refresh.service
• deploy/systemd/ask-monarch-refresh.timer
• deploy/launchd/com.ask-monarch.drive-refresh.plist
• docs/daily-source-refresh.md

Operations

VM service commands:

sudo systemctl status ask-monarch-http.service
sudo systemctl restart ask-monarch-http.service
sudo journalctl -u ask-monarch-http.service -n 100 --no-pager
sudo systemctl status nginx
sudo nginx -t

Do not update production with a pull inside a long-lived VM checkout. Install the exact remote main merge into revision-specific code and virtual environment directories, move the active links atomically, and run the release verifier. A successful deploy writes /var/lib/ask-monarch/release-receipts/$REVISION.json. After that, publish and refresh the installed CLI and skill copies, verify Codex, shared .agents, and Claude parity, then run Reality Eval #4, the 50-question sole completion gate. One unchanged 50/50 run passes it. Merge, deployment, required source refresh, and installed-agent parity are prerequisites for that counted run. Recording, privacy review, and sharing are its proof. The exact visible latency question and fresh 200-case suite are optional supporting diagnostics. They can localize defects, but they are not separate gates and do not block or complete the goal. Follow docs/production-release.md for the release path and diagnostic commands.

Repo Layout

ASKMONARCH-SPEC.md              product/source contract
config/source-map.yaml          source registry and query-plane map
docs/                           operating docs and source receipts
scripts/                        parsers, query tools, HTTP service, CLI, refresh
skills/                         repo-owned Codex skills
deploy/systemd/                 VM service and refresh units
reports/deep-research/          generated Deep Research bundles and reports

Generated source-plane outputs are intentionally not committed to GitHub. Keep the repo as code, specs, scripts, skills, and docs. Keep large parsed artifacts on the VM or in artifact storage.

Source Contract

A source is not queryable just because it exists. Ask Monarch treats a source as queryable only when:

• it is listed in config/source-map.yaml
• it can be accessed through an authenticated route
• its live contents are enumerated
• its artifacts are parsed into atomic units
• each unit has source-grade provenance
• receipt and gap files are written
• verification passes
• query tools can return evidence with exact locators

That contract is the core of the system: do the hard parsing once, verify it, store it in a parsed index, and query the index fast.