Financial Model Review

Financial Model Review Skill

Help startup founders understand how investors will evaluate their financial model — validating structure, unit economics, runway, and metrics against stage-appropriate standards. Produce a thorough review with actionable improvements. The tone is founder-first: a rigorous but supportive coaching session.

Skill Metadata

• Author: lool-ventures
• Version: managed in founder-skills/.claude-plugin/plugin.json
• Compatibility: Python 3.10+ and uv for script execution. openpyxl required for Excel parsing.
• Imports (optional):
  • market-sizing:sizing.json — validate revenue-to-SOM consistency
  • deck-review:checklist.json — cross-check model-to-deck number alignment
• Exports:
  • report.json → ic-sim, fundraise-readiness, dd-readiness
  • unit_economics.json → metrics-benchmarker, ic-sim
  • runway.json → fundraise-readiness

Skill Execution Model (READ FIRST)

See founder-skills/references/skill-execution-model.md for the full inline-skill execution model (3 dispatch contexts, Mitigation 1+2, producer contract, Cowork quirks, per-symptom triage).

This skill runs inline in the main thread, not as a sub-agent — see the reference above ("Why Inline (Not Forked Sub-Agent)") for the rationale. Sub-agents are deliberately shell-free, so orchestration (producer scripts, artifact persistence) stays in the main thread.

Two dispatch contexts for the sub-agent:

• Context A — Per-step analytical dispatch (Mitigation 1): The INPUTS_REVIEW and CHECKLIST steps dispatch the financial-model-review agent via the Task tool. The agent does deep analysis, WRITES its output JSON to the OUTPUT_PATH given in its prompt (the handoff/ dir), and returns a small receipt. The main thread gates the file with check_handoff.py, then pipes it through the producer script. The sub-agent never writes canonical artifacts — only its hand-off file. (Unit economics and runway are NOT dispatched — those producers consume inputs.json verbatim, so the main thread pipes the file directly.)
• Context B — Post-compose coaching dispatch: The final step dispatches the sub-agent after compose_report.py writes report.md. The sub-agent Reads the staged coaching_payload.json from the hand-off dir (Mitigation 2) — it does NOT read the full report.md — composes the coaching commentary, WRITES it to the OUTPUT_PATH hand-off file, and returns a small receipt. The main thread gates the file (check_handoff.py) and inserts it via the shared insert_coaching.py script (idempotency matrix, uuid-marker replacement, run_id-parity verification — all deterministic). See the reference above for the full Context B contract.

Tolerant JSON extraction protocol (Context B returns; also the Context A message-channel fallback): capture the sub-agent's final assistant message. It should be raw JSON, but may be wrapped in ```json ... ``` fences or carry a prose preamble. Extract tolerantly:

1. If the message is wrapped in a ```json ... ``` (or plain ``` ... ```) fence, strip the fence first.
2. Try to parse the stripped text directly as JSON.
3. If that fails, walk through the text looking for the first { character and try json.JSONDecoder().raw_decode(text[i:]) — this is brace-aware and handles nested objects correctly (unlike regex, which truncates on the first }).
4. If extraction fails entirely, re-prompt the sub-agent with: "Your previous reply could not be parsed as JSON. Return ONLY the JSON object — no markdown fences, no prose preamble."

Context A receipts don't need this protocol by hand — check_handoff.py --receipt-json - applies the same tolerant extraction internally; pass the final message verbatim.

If a sub-agent wrote CANONICAL artifact files directly anyway (anything outside its handoff/ OUTPUT_PATH): do not trust them — take its gated hand-off file (or extract the JSON from its final message on the fallback path), then re-pipe through the producer script as specified; the producer overwrites the file with the validated, run_id-stamped version. For INPUTS_REVIEW specifically: if inputs.json contains the {"corrected": ..., "corrections": ...} wrapper, the sub-agent wrote its reply to disk — feed that wrapper through apply_corrections.py as usual.

Context-pressure note: This skill has the highest context budget of the 5 skills. The win from Mitigation 1 is excluding sub-agent reasoning and the raw extract_model.py output (which can run to megabytes on real models) — which flows through the INPUTS_REVIEW dispatch: the sub-agent reads it in its own context window, returns only the corrected inputs.json. The artifacts themselves still accumulate in the main thread (~80-130K total), but that is manageable.

Input Formats

Accept any format: Excel (.xlsx), CSV, Google Sheets exports, financial documents, or conversational input. For Excel files, use extract_model.py to parse. For other formats, extract data manually into the inputs.json schema. If multiple copies of the same file exist (e.g., Financials.xlsx and Financials (1).xlsx), use the most recently modified version and note the duplication to the founder. If timestamps are identical, ask the founder which file to use. If the founder cannot be queried, prefer the file without parenthetical suffixes (e.g., (1), (2)) — these typically indicate browser re-download duplicates.

Available Scripts

All scripts are at ${CLAUDE_PLUGIN_ROOT}/skills/financial-model-review/scripts/:

• extract_model.py — Extracts structured data from Excel (.xlsx) and CSV files
• validate_extraction.py — Anti-hallucination gate: cross-references model_data.json against inputs.json to catch mismatches (company name, salary, revenue, cash traceability); run after extraction, before review
• validate_inputs.py — Four-layer validation of inputs.json (structural, consistency, sanity, completeness); supports --fix to auto-correct sign errors
• checklist.py — Scores 46 criteria across 7 categories with profile-based auto-gating
• unit_economics.py — Computes and benchmarks 11 unit economics metrics
• runway.py — Multi-scenario runway stress-test with decision points
• compose_report.py — Assembles report with cross-artifact validation; --strict exits 1 on high-severity warnings (corrupt/missing artifacts)
• apply_corrections.py — Processes founder's downloaded corrections file: coerces types, normalizes ILS→USD, merges overrides, writes corrected_inputs.json and extraction_corrections.json
• verify_review.py — Review completeness gate: checks artifact existence, content quality, and cross-artifact consistency; --gate 1 for after-compose, --gate 2 (default) for final; exit 0 = publishable, exit 1 = gaps remain
• visualize.py — Generates self-contained HTML with SVG charts (not JSON)
• explore.py — Generates self-contained interactive HTML explorer from review artifacts; outputs HTML (not JSON)
• review_inputs.py — Dual-mode review viewer: HTTP server with live validation (Claude Code) or self-contained static HTML with JS sanity metrics (Cowork); outputs HTML

Also available from ${CLAUDE_PLUGIN_ROOT}/scripts/ (shared):

• find_artifact.py — Resolves artifact paths by skill name and filename (used for cross-skill lookups)

Run with: python3 ${CLAUDE_PLUGIN_ROOT}/skills/financial-model-review/scripts/<script>.py --pretty [args]

Available References

Read as needed from ${CLAUDE_PLUGIN_ROOT}/skills/financial-model-review/references/:

• checklist-criteria.md — All 46 checklist criteria with gate definitions
• schema-inputs.md — JSON schema for inputs.json (the artifact the agent writes)
• artifact-schemas.md — JSON schemas for script-produced output artifacts
• data-sufficiency.md — Data sufficiency gate and qualitative path
• extraction-pitfalls.md — 8 common extraction errors (scale denomination, payroll aggregation, collections vs revenue, etc.)

From ${CLAUDE_PLUGIN_ROOT}/references/ (shared): stage-expectations.md, benchmarks.md, israel-guidance.md, revenue-model-types.md, common-mistakes.md

Artifact Pipeline

Every review deposits structured JSON artifacts into a working directory. The final step assembles all artifacts into a report and validates consistency. This is not optional.

Rules:

• Deposit each artifact before proceeding to the next step
• For agent-written artifacts (inputs.json), consult references/schema-inputs.md for the JSON schema
• If a step is not applicable, deposit a stub: {"skipped": true, "reason": "..."}
• Do NOT use isolation: "worktree" for sub-agents — files written in a worktree won't appear in the main $REVIEW_DIR

Keep the founder informed with brief, plain-language updates at each step. Narrate the founder-visible OUTCOME, never the internal step. That is the test to apply, and it catches more than a word list can: the forbidden thing is not a syntax, it is talking about the machinery. Bad — "Gating and piping the extraction through the producer, then staging the coaching hand-off"; good — "I've checked your numbers and I'm writing up what stood out." Bad — "schema-drift warning on coaching_payload"; good — nothing, because the founder has no stake in it. Never name an internal artifact, field, or token (a payload key, a marker name, an artifact filename, a hand-off dir) even in plain prose with no backticks — a detector keyed on syntax cannot see "gated", "hand-off" or "canonical artifacts", but the founder still reads them and they still mean nothing to them. The between-step progress lines are the primary leak vector, not the final summary. They feel internal — you are narrating what you are about to do — but the founder reads every one of them, and this is where the leaks actually appear: "Now gating the hand-off before piping through the checklist producer", "Gate 1 passes", "Running the final verification gate". Rewrite each pipeline transition as the founder-visible outcome: "Checking your numbers against the 46-point review", "Your inputs look consistent — moving on to unit economics", "Finishing up and putting the report together". If a progress line would mean nothing to someone who has never seen this skill's internals, it does not belong in the channel. Also excluded, as before: file/script names, paths, *.py, --flags, $vars, exit codes ("Exit N", "not found"), W_/E_ codes, JSON, and step/route labels ("Lane N", "Context A/B", "Phase N", "structure detection", "the grid", any ALL_CAPS_TOKEN). After each analytical step (3–6), share a one-sentence finding before moving on. Track progress with at most one batched task tracker (a single TaskCreate), updating it only at phase boundaries — extraction, review gate, scoring, report — never per sub-step: the step narration above is the founder's progress channel, so per-substep TaskCreate/update churn only adds runtime. The task tracker is founder-visible too — the same rule governs its labels. "Gate the inputs review handoff", "Validate inputs.json", "resolve agent namespace paths", "Initialize founder context" are leaks even though each names a real step, and even when the prose around them is clean. Label each task by the founder-visible outcome — "Check your inputs", "Score against the review", "Write up what I found" — never by a file, directory, script, or pipeline stage.

Workflow

Step 0: Path Setup

Every Bash tool call runs in a fresh shell — variables do not persist. Run the block below exactly once: it resolves $PLUGIN_ROOT deterministically, and every later block must substitute the printed value as a literal rather than re-running the resolution — repeating the self-heal search can land on a different mount than Step 0 picked when more than one is present (see why in the block's comments).

Optional, best-effort, and via the Read tool (not a shell command): before the block below, Read ${CLAUDE_PLUGIN_ROOT}/.claude-plugin/plugin.json and note its version field as EXPECT_VERSION. Passing it to select_plugin_root.py below lets an exact version match win over an arbitrary first hit. If the Read fails, skip it and omit --expect-version — selection is still deterministic without it.

SCRIPTS="${CLAUDE_PLUGIN_ROOT}/skills/financial-model-review/scripts"
if [ ! -d "$SCRIPTS" ]; then
  # In Cowork, CLAUDE_PLUGIN_ROOT substitutes to a host-side path absent inside
  # the session VM — self-heal by collecting EVERY candidate mount (a session can
  # have more than one at once: a stale host-side cache, a test marketplace, even
  # a symlink into a different session's tree) and handing them to
  # select_plugin_root.py, which picks ONE deterministically and names the
  # rejects — never trust `find`'s arbitrary first hit, which can silently mix
  # scripts across plugin versions mid-pipeline.
  CANDIDATES="$(find /sessions -type d -path '*/skills/financial-model-review/scripts' 2>/dev/null)"
  [ -n "$CANDIDATES" ] || CANDIDATES="$(find / -type d -path '*/skills/financial-model-review/scripts' 2>/dev/null)"
  PROVISIONAL_ROOT="$(printf '%s\n' "$CANDIDATES" | head -1)"
  PROVISIONAL_ROOT="${PROVISIONAL_ROOT%/skills/*}"
  # Bootstrap order: $SHARED_SCRIPTS isn't known until a root is chosen, so use the
  # provisional root's OWN copy of the selector; an older plugin copy without one
  # falls back to the provisional root unchanged.
  SELECTOR="$PROVISIONAL_ROOT/scripts/select_plugin_root.py"
  if [ -f "$SELECTOR" ]; then
    if [ -n "$EXPECT_VERSION" ]; then
      PLUGIN_ROOT="$(printf '%s\n' "$CANDIDATES" | python3 "$SELECTOR" --expect-version "$EXPECT_VERSION")"
    else
      PLUGIN_ROOT="$(printf '%s\n' "$CANDIDATES" | python3 "$SELECTOR")"
    fi
  else
    PLUGIN_ROOT="$PROVISIONAL_ROOT"
  fi
  SCRIPTS="$PLUGIN_ROOT/skills/financial-model-review/scripts"
fi
PLUGIN_ROOT="${SCRIPTS%/skills/*}"
echo "PLUGIN_ROOT=$PLUGIN_ROOT"   # resolved ONCE, here — paste this literal into every later block; never re-run this resolution
REFS="$PLUGIN_ROOT/skills/financial-model-review/references"
SHARED_SCRIPTS="$PLUGIN_ROOT/scripts"
SHARED_REFS="$PLUGIN_ROOT/references"
# Resolve the canonical artifacts root via a SCRIPT, not inline bash (the agent paraphrases inline
# path computations → outputs/ vs outputs/artifacts/ drift across runs). Deterministic + creates it.
python3 "$SHARED_SCRIPTS/resolve_artifacts_root.py"   # prints ARTIFACTS_ROOT — use the printed path verbatim as ARTIFACTS_ROOT in every later block (a captured var dies in the next fresh shell)

Reaching the self-heal branch is normal in Cowork — ${CLAUDE_PLUGIN_ROOT} resolves to a HOST path that does not exist inside the VM, so the [ ! -d "$SCRIPTS" ] test fails by design rather than by misconfiguration. It is not a sign anything is wrong, and it is not worth narrating to the founder.

Outputs mount is append-only. Everything under the promoted outputs mount (.../mnt/outputs/, not just $REVIEW_DIR) is write-allowed and delete-denied by the platform: never rm, move away, or empty anything under it — including files you created yourself. Never create ad-hoc scratch anywhere under the outputs mount (no _src/ copies, no run-state note files); scratch belongs in $STAGING_DIR (a /tmp dir, defined below). Do not "clean up" the outputs folder before delivering — extra working files there are expected and harmless. The uploaded document is already readable in place from the uploads mount; never copy it under outputs to make it readable.

If ARTIFACTS_ROOT resolves to $(pwd)/artifacts but no artifacts/ directory exists at $(pwd): The workspace may not be mounted yet. Use Glob with pattern **/artifacts/founder_context.json to locate existing artifacts, and derive ARTIFACTS_ROOT from the result. If nothing is found, mkdir -p "$ARTIFACTS_ROOT" and proceed — never a relative ./artifacts, which resolves against the shell's cwd (the session root) and lands outside the outputs mount, undelivered.

After Step 1 (when the slug is known), derive REVIEW_DIR. Two modes — pick exactly one:

• Full review (default — the founder attached a model, asked for a review, a report, or the interactive explorer, OR there is no existing full review for this slug): run Steps 2–11. REVIEW_DIR="$ARTIFACTS_ROOT/financial-model-review-${SLUG}".
• Quick-check mode — a single directional question in conversation, no model attached and no request for a review ("with $400k in the bank and $60k/mo net burn, how long do I have?", "is a 4x burn multiple bad at seed?"). Run Step 5-quick instead of Steps 2–11. REVIEW_DIR="$ARTIFACTS_ROOT/financial-model-review-${SLUG}-quickcheck".

Tie-breaker when both bullets seem to fit. Decide on the verb, not the inputs: a request for the work product ("review my model, analyze our runway, I need this for the board") is a full run even when every number is already in hand, while a request for a read ("roughly, ballpark, how long do I have, is X bad") is a quick check even when materials are attached. Complete inputs make the full run faster, not less wanted. When the verb is genuinely absent, default to the full run and say you did — an unwanted full run costs time, an unwanted quick check costs the founder the analysis they came for.

Never answer from your own arithmetic. Quick-check exists because the alternative a model reaches for — computing runway in its head and offering the real review as an opt-in — produces a number with no scenario stress-test, no benchmark provenance, and no record, under this skill's name. Running fewer producers is fine; running none is not.

Step 5-quick: the quick-check path

Run only the producer(s) the question actually needs, with the inputs the founder gave you:

# Runway question -> runway.py alone. Unit-economics question -> unit_economics.py alone.
printf '%s' "$QUICK_JSON" | python3 "$SCRIPTS/runway.py" --stdin --pretty \
  --run-id "$RUN_ID" -o "$REVIEW_DIR/runway.json"

Producers deliberately NOT run: extract_model.py, validate_extraction.py, validate_inputs.py, checklist.py, the producer the question didn't need, compose_report.py, visualize.py, explore.py, verify_review.py, and the Context-B coaching dispatch. No report.md is written.

Same-numbers guarantee. The figures are identical to what the full review would compute from the same inputs — it is the same script reading the same shape. Only the production weight is dropped. What you do not get is what the skipped producers add: the anti-hallucination extraction gate, the four-layer input validation, the 46-item checklist, multi-scenario stress-testing, and the cross-artifact consistency checks.

Presenting it. Label it a quick check, not a review. Give the figure, name the inputs it came from, and state plainly that nothing was validated or stress-tested. Then close with a statement, never a question: "The full review validates the model, stress-tests runway across scenarios, and scores 46 investor criteria — say the word and I'll run it." A question invites a "no" to something the founder would have wanted.

REVIEW_DIR="${REVIEW_DIR:-$ARTIFACTS_ROOT/financial-model-review-${SLUG}}"              # full review
# REVIEW_DIR="${REVIEW_DIR:-$ARTIFACTS_ROOT/financial-model-review-${SLUG}-quickcheck}"  # quick check
mkdir -p "$REVIEW_DIR"
RUN_ID="$(date -u +%Y%m%dT%H%M%SZ)"
# Context A hand-off dir — PER RUN: sub-agents WRITE their raw output JSON here (the audit trail —
# raw sub-agent output as returned, before producer validation). Permanent by platform design
# (outputs/ mounts are write-allowed / delete-denied); nothing in it is ever a canonical artifact.
# The $RUN_ID segment is load-bearing: it prevents a stale prior-run file from silently passing
# the hand-off gate when a dispatch fails to write.
HANDOFF_DIR="$REVIEW_DIR/handoff/$RUN_ID"
mkdir -p "$HANDOFF_DIR"
# Sub-agents address the SAME dir by a different path (their file tools are rooted at the outputs
# mount in Cowork). Resolve the FULL agent-namespace paths via the script — never hand-splice the
# printed root with a literal skill-name/slug/run-id string yourself (that string-splicing is
# exactly the non-determinism the resolver script exists to remove):
python3 "$SHARED_SCRIPTS/resolve_artifacts_root.py" --handoff-dir-agent \
  --dir-name "financial-model-review-${SLUG}" --run-id "$RUN_ID"   # prints HANDOFF_AGENT verbatim
HANDOFF_AGENT="<printed value>"   # use verbatim in OUTPUT_PATH lines
# Sub-agent READ paths for under-outputs artifacts use the SAME agent namespace (relative — the
# sub-agent's file-tool cwd IS the outputs mount on host-loop; an absolute /sessions/... read is denied):
python3 "$SHARED_SCRIPTS/resolve_artifacts_root.py" --analysis-dir-agent \
  --dir-name "financial-model-review-${SLUG}"   # prints the dir in the agent namespace
REVIEW_DIR_AGENT="<printed value>"   # e.g. model_data.json, inputs.json reads
# Ad-hoc scratch (NOT sub-agent hand-off) lives OUTSIDE the promoted outputs/ tree, in a temp dir
# that is safe to both create and reclaim. Use the printed path verbatim in later steps.
STAGING_DIR="$(mktemp -d "${TMPDIR:-/tmp}/financial-model-review-${SLUG:-fmr}.staging.XXXXXX")"

Pass RUN_ID to all sub-agents. The four producer artifacts (inputs.json, checklist.json, unit_economics.json, runway.json) must carry "metadata": {"run_id": "$RUN_ID"} at the top level — including skipped stubs, whose stub heredoc carries the same "metadata": {"run_id": "$RUN_ID"} block. The producers propagate it from their stdin payloads; never hand-edit script outputs to add it. (model_data.json and extraction_validation.json have no run_id by design.) compose_report.py checks that all present run IDs match — a mismatch triggers a STALE_ARTIFACT high-severity warning, blocking under --strict. Stub artifacts are exempt from the value comparison but still carry the run_id key so the Context B parity grep finds it.

Overwrite-in-place — do NOT delete prior artifacts under $REVIEW_DIR. It is the promoted outputs/ tree in Cowork, where deleting a user-visible path is unsafe (Cowork can deny it; the parity gate flags it). Each producer writes its artifact fresh via -o every run, and RUN_ID is minted fresh per run — so if a prior run left an artifact a later step doesn't regenerate, compose_report.py's STALE_ARTIFACT check (run_ids must match) catches the mismatch. No bulk rm is needed or wanted.

Step 1: Read or Create Founder Context

python3 "$SHARED_SCRIPTS/founder_context.py" read --artifacts-root "$ARTIFACTS_ROOT" --pretty

Three cases based on exit code:

Exit 0 (found, single context): Use the company slug and pre-filled fields. Before proceeding to extraction, use AskUserQuestion to ask the founder for current cash balance and date if not already stated in the conversation — this is the #1 cause of incomplete runway analysis. If files are attached, also ask about monthly burn rate unless the conversation already contains it. Same runtime-labelled shape as the cash/date/burn questions below (an affirmative carrying any already-stated value, plus a "Not stated" fallback) — these are dollar amounts and dates, not a fixed label set. Batch all questions into a single AskUserQuestion call.

Exit 1 (not found): Use AskUserQuestion (NOT plain chat) to ask the founder for company details AND key financial context. You MUST use the AskUserQuestion tool — do not just list questions in the chat. Gather everything in a single call (one interaction = one chance for the UI to render correctly):

• Company name, stage, sector, geography (required for context creation)
• Current cash balance and date (critical for runway — the #1 cause of incomplete reports)
• Monthly burn rate if not obvious from the provided files

Stage is the one field of the four with a real fixed label set — use it verbatim, do not improvise. Options: Pre-seed / Seed / Series A / Series B+ → pre-seed | seed | series-a | series-b (four options is the tool's max; the shared context script's VALID_STAGES enum has 7 values including series-c/series-d/later, so on a Series B+ pick, ask a plain-text follow-up for the specific stage — do not default to series-b). Company name, sector and geography cannot take fixed labels (a proper noun, an open sector taxonomy, an open location) — shape them per the next paragraph instead.

IMPORTANT: Always use the AskUserQuestion tool for founder questions — never ask as plain chat text. If AskUserQuestion is genuinely unavailable in the host, do NOT skip the ask and do NOT assume the answer: ask the same question in plain chat, state the options explicitly, and wait for an answer before continuing. The ban above is on asking casually WHILE the tool is available — it is not a reason to stall a host that lacks it. The tool provides a structured UI that renders correctly in Cowork. Always provide at least 2 options (the tool requires a minimum of 2). Construct those two options concretely so every question is answerable — never emit a single-option question or a bare free-text prompt (a free-text answer that matches no option dead-ends the run). For each founder question give: (1) an affirmative option carrying the likely value — for the company name that includes "Use what the model file states" (the Step-1 staging branch above resolves that answer safely, so it is a valid choice, not a trap); and (2) a "Not stated — proceed and flag to confirm" fallback so the founder can always move forward. Cash balance, date and burn rate follow the same two-option shape — an affirmative carrying whatever value was already stated in the conversation or files, plus the "Not stated" fallback; these are runtime-labelled (dollar amounts and dates), not a fixed label set.

When there is NO file, "Use what the model file states" is not an answerable option — there is no file to read it from, so offering it costs a round-trip and then a second question. On a conversational or deck-only run, build the company-name question from what you actually have instead: an affirmative option carrying the name as it appeared in the conversation or on the deck's title slide (say where you got it, so the founder is confirming rather than re-supplying), plus the "Not stated" fallback. One question, one answer. The same principle applies to sector and geography: an option the founder cannot possibly choose is a wasted turn.

Why everything upfront: Extraction sub-agents run in parallel and cannot pause to ask questions. Asking early prevents pipeline stalls.

If the founder provides files (Excel/CSV), still ask about cash balance — extraction may miss or misinterpret values, and having the founder's stated number lets the agent cross-check later.

Company name deferred to the model file — stage the extraction FIRST (avoids the slug-ordering deadlock): If the founder does not give a company name and defers it to the uploaded model (e.g. answers the name question with "use the model file"), the name requires the extraction, which normally targets $REVIEW_DIR, which requires the slug, which requires the name — a deadlock. Do not resolve it by improvising a temp file or a provisional review dir under the outputs mount: that mount is append-only (Step 0) and the later rm/mv of the provisional path is delete-denied by the platform. Instead, stage the extraction to $STAGING_DIR (the /tmp dir from Step 0 — safe to both create and reclaim; its ${SLUG:-fmr} default already tolerates being created before the slug is known), derive the name from the staged output, run init below, then cp the staged file into $REVIEW_DIR:

# 1. Stage the extraction OUTSIDE the outputs mount (pre-slug):
python3 "$SCRIPTS/extract_model.py" --file <path> --pretty -o "$STAGING_DIR/model_data.json"
# 2. Read the company name from $STAGING_DIR/model_data.json (company_name field / model header).
# 3. Run `founder_context.py init` (below) with the derived --company-name; it prints the context
#    JSON including "slug" — read the slug from that printed output. Do not capture python output
#    into a shell variable.
# 4. Create $REVIEW_DIR (Step 0's mkdir, now that the slug is known), then copy the staged file in
#    — a plain write, append-only-safe:
cp "$STAGING_DIR/model_data.json" "$REVIEW_DIR/model_data.json"

Then continue from Step 2's periodicity check as normal (the staged extraction already ran). Never create a provisional review dir or temp file anywhere under the outputs mount, and never rename or move a review dir. Extraction runs ONLY via the documented invocations — Step 2's $REVIEW_DIR target or this Exit-1 $STAGING_DIR staging block — never an ad-hoc extract_model.py call with improvised flags or targets.

Then create:

python3 "$SHARED_SCRIPTS/founder_context.py" init \
  --company-name "Acme Corp" --stage seed --sector "B2B SaaS" \
  --geography "US" --artifacts-root "$ARTIFACTS_ROOT"

If the script prints a sector_type warning but exits 0, that's non-fatal — proceed without retrying. However, a null sector_type may suppress sector-specific checklist gating downstream. If you know the correct type, re-run with --sector-type (valid values: saas, ai-native, marketplace, hardware, hardware-subscription, consumer-subscription, usage-based, transactional-fintech, retail).

Exit 2 (multiple context files): Present the list to the founder and ask which company via AskUserQuestion (labels are the runtime company names found on disk — necessarily runtime-labelled, no fixed set can exist here), then re-read with --slug.

Execution checkpoint — END OF STEP 1, READ BEFORE CONTINUING

You now have enough to run. Invoking this skill is not the same as running it. From here, every number that reaches the founder must come out of a producer script. Concretely:

• Never compute a figure in chat. Not TAM, not runway, not a ratio, not a benchmark comparison — not even one you are confident about. An in-chat number has no provenance, no range, no artifact, and nothing downstream can contradict it. That is worse than a slow answer and worse than no answer.
• Never benchmark against a figure you recalled. Benchmarks live in the reference files and the producers read them. If you find yourself writing "typically around X for this stage", stop: either a producer sourced it or it does not go in front of the founder.
• A what-if, a sensitivity illustration, or "roughly what would X give" is NOT an exemption. This is the exemption a live run invented: having correctly produced the real figure, it then wrote "using the current count would shave TAM to roughly €249M rather than €270M" — a second number, computed in chat, from an input the founder never gave. An illustrative figure is read exactly as confidently as a computed one, and the founder cannot tell which came from the pipeline. Two ways to answer a what-if: re-run the producer with the alternate input and quote its output, or give no number and say which direction it moves. Never arithmetic in prose.
• Never offer the real run as an opt-in after answering. "Here's a rough estimate — I can run the full analysis if you want" is the failure. The founder cannot tell that what they just read was not the analysis, so they will not ask for it.
• Two ways to finish, and only two: run the full pipeline to completion, or run the quick-check path (Step 5-quick), which still runs the producer the question needs. Both end with real artifacts on disk. Anything else is not a finished run.
• If you are blocked, say BLOCKED and say why. A missing input, a failed hand-off, an unreadable document — name it and stop. Do not substitute your own reasoning for the pipeline and present the result as its output.

Artifact existence is the proof of execution: if no canonical artifact was written, the skill did not run, whatever the transcript says.

Step 2: Extract Model Data

When Excel (.xlsx) or CSV files are provided, run extract_model.py directly in the main thread:

python3 "$SCRIPTS/extract_model.py" --file <path> --pretty -o "$REVIEW_DIR/model_data.json"

Check the periodicity_summary and per-sheet periodicity fields. If periodicity is quarterly or annual, all flow metrics (burn, revenue, expenses — anything measured per period) must be divided by 3 or 12 respectively in the next step. Do NOT convert stock metrics (cash balance, headcount, customer count, ARR — point-in-time snapshots). If periodicity is unknown, flag it.

When documents (PDFs, data room dumps, Google Sheets exports) are provided: Extract what you can directly from the documents, consulting ${CLAUDE_PLUGIN_ROOT}/skills/financial-model-review/references/schema-inputs.md for the schema and ${CLAUDE_PLUGIN_ROOT}/skills/financial-model-review/references/data-sufficiency.md for sufficiency assessment. Write a provisional inputs.json.

When conversational input is provided (no files): Gather all needed fields within Step 1 through normal conversation. Consult references/schema-inputs.md for the full schema.

Context A hand-off protocol (file transport + gate)

Every Context A dispatch prompt carries an OUTPUT_PATH: line built from $HANDOFF_AGENT. The sub-agent WRITES its output JSON to that path with its Write tool and returns only a small receipt: {"status": "complete", "output_path": "<echo of OUTPUT_PATH>"}. The payload leaves the model exactly once (into the Write call) — never re-type sub-agent JSON into a heredoc.

$HANDOFF_AGENT and $HANDOFF_DIR name the SAME directory by two different paths — they are not interchangeable. $HANDOFF_DIR is the absolute VM path your shell uses (python3, check_handoff.py, producer pipes). $HANDOFF_AGENT is the relative path a sub-agent's file tools resolve against the outputs mount, and it is the ONLY one that goes in a dispatch prompt. Putting $HANDOFF_DIR in an OUTPUT_PATH line hands the sub-agent an absolute /sessions/... path the host-loop gate denies; putting $HANDOFF_AGENT in a shell command resolves it against the wrong cwd. Rule of thumb: agent namespace in prompts, shell namespace in bash.

The receipt is the ONE exemption from the never-re-type rule. "Never re-type" governs the payload — the extraction JSON, the coaching commentary, anything the founder's numbers pass through. The receipt is a two-field acknowledgement the sub-agent returns in its final message, and reading output_path out of it to pass to check_handoff.py --agent-path is expected, not a violation. If it were forbidden, the hand-off could not be gated at all.

Path idiom for dispatch prompts (host-loop path gate): OUTPUT_PATH and any under-outputs artifact READ path a sub-agent is given are relative to the sub-agent's file-tool cwd (the outputs mount) — built from the resolve_artifacts_root.py --agent namespace ($HANDOFF_AGENT / $REVIEW_DIR_AGENT). Never hand a sub-agent an absolute /sessions/... path for a file-tool Read/Write — the host-loop path gate denies it (steering shell work to the bash tool instead). Bundled references/*.md are the one exception: pass them as the literal ${CLAUDE_PLUGIN_ROOT}/skills/financial-model-review/references/... token (it is pre-resolved to a host-readable path); do NOT substitute a find /sessions-discovered $REFS (a shell path a file tool can't read).

After EVERY Context A dispatch, gate before piping (<step> = the dispatch's file stem):

printf '%s' '<agent final message verbatim>' | \
  python3 "$SHARED_SCRIPTS/check_handoff.py" "$HANDOFF_DIR/<step>_output.json" \
    --agent-path "$HANDOFF_AGENT/<step>_output.json" --receipt-json -

Branch on the exit code (complete state machine — do not improvise):

• Exit 0 → pipe the file through the producer: cat "$HANDOFF_DIR/<step>_output.json" | python3 "$SCRIPTS/<producer>.py" ...
• Exit 3 (missing/empty file — receipt may be fabricated) → redo-dispatch: fresh Task, same prompt plus one line: "your receipt claimed a file at <path> but none exists; use Write to create exactly that path."
• Exit 4 (file exists, invalid JSON) → repair-dispatch: fresh Task: "Read <OUTPUT_PATH>; it fails JSON parsing with <verbatim detail from the diagnostic>; fix and rewrite it; return the receipt."
• Exit 5 (receipt echoes a different path) → repair-dispatch telling the agent the exact expected OUTPUT_PATH (it wrote somewhere else).
• Exit 6 (receipt unparseable / no output_path key) → redo-dispatch with "return ONLY the receipt JSON — no fences, no prose."
• Producer schema rejection (the pipe fails next) → repair-dispatch with the producer's stderr verbatim.
• Exit 8 (path_namespace_mismatch) → the sub-agent complied; the agent-namespace prefix was wrong. Its relative OUTPUT_PATH resolved against the outputs mount instead of the session root, so the file landed at the doubled path reported in found_at. Do NOT treat this as a fabricated receipt (that is exit 3), and do NOT read the hand-off from found_at — it is diagnostic only. Re-run resolve_artifacts_root.py --agent, rebuild the agent-namespace prefix from the printed value, and re-dispatch. Counts against the same 2-dispatch retry budget.
• Any other exit (script crash etc.) → STOP with the stderr.
• After ANY corrective dispatch, resume from check_handoff.py — never pipe to the producer unchecked.

Retry budget: max 2 corrective dispatches per step, of any kind, in any combination (max 3 total dispatches). After the second corrective dispatch fails any gate: STOP and report the exact diagnostic to the founder. The main thread MUST NOT author or patch analytical content itself — filling in the JSON is the fabrication failure mode this architecture exists to prevent. A status: "blocked" return is not a gate retry, but it is bounded: at most ONE input-fix re-dispatch per step; a second blocked return STOPs with both reasons quoted.

Graceful degrade (fleet heterogeneity): if the FIRST corrective dispatch also exits 3 while the agent's receipt claims complete with the correctly echoed path, treat the host's filesystem topology as hand-off-incompatible: fall back to message-channel transport for the REST of this run (sub-agent returns full JSON in its final message; apply the tolerant JSON extraction protocol; stage to `$ST

…(truncated)