Canvas
A canvas is a self-contained analytical page beside the chat: stat rows, sortable tables, labelled charts, timelines, findings and diffs, themed like the app. You send data (blocks); Navin validates it and renders it with a bundled kit. No HTML to write, no network at view time.
When to use
Use a canvas when the result is standalone and analytical, and the user will scan, sort or compare it:
- audits and code review / security findings (severity, file:line, evidence, fix status)
- performance benchmarks, before/after comparisons
- metrics and usage (token usage, costs, counts over time)
- investigations with several data points, architecture reviews with scored findings
- status boards (PR checks, CI, project board, milestones)
- repeatable reports the user will ask for again (same template, new data)
When not to use
- The user asked for a specific deliverable (a docx, a deck, an app, a file in the repo): build that.
- The work happens in a specific tool (editor, terminal, browser, board UI): stay there.
- Active debugging loops: keep it in the transcript, fast.
- Short answers: a sentence or a small Markdown table beats a canvas.
- Architecture diagrams, PPT visuals, tender/RFP technical answers, Markdown plans:
archify stays the default (AGENTS.md). Canvas is for data and analysis; it can sit next to an archify diagram, not replace it.
- Outside the WebUI / desktop (Telegram, CLI): there is no canvas panel; answer in text.
How to call it
Free-form blocks:
present_artifact({
"type": "canvas",
"title": "API latency audit",
"subtitle": "staging, 2026-10-12 to 2026-10-19",
"blocks": [
{"type": "stats", "items": [
{"label": "Endpoints", "value": 42},
{"label": "p95 over budget", "value": 3, "tone": "bad"}
]},
{"type": "bar_chart", "title": "p95 by endpoint", "x_label": "Endpoint",
"y_label": "p95 latency", "unit": "ms", "categories": ["/chat", "/files"],
"series": [{"name": "Before", "values": [180, 95], "role": "baseline"},
{"name": "After", "values": [120, 90]}],
"caption": "Source: k6 run, 5 min at 50 rps. Lower is better."}
]
})
Templates (Navin data in, blocks out; navin/artifacts/canvas/templates.py):
| template |
data |
source |
benchmark |
{"before": report, "after": report, "label_before"?, "label_after"?} or data_path |
scripts/bench_tui.py --json, scripts/bench_agent_hotpath.py |
findings |
findings list, or the security_scan / code_review filter result |
review + security tools |
pr_status |
{"pr": forge pr_view, "ci": forge ci_status, "threads"?: [{path, line, resolved, body}]} |
forge tool |
board |
omit to read the project board |
.navin/board/ |
token_usage |
omit to read the usage log; {"prices": {model: {prompt, completion, cached}}} (USD per 1M tokens) adds estimated cost |
/api/settings/usage data |
video_qa |
data_path to video-qa.json |
navin/marketing/video_qa.py |
present_artifact({"type": "canvas", "template": "video_qa", "data_path": "out/promo-qa/video-qa.json"})
Pass the same id to update a canvas in place instead of opening a new tab.
Block types
section (title, blocks, description?) | stats (items: label, value, unit?, delta?, tone?, hint?) | table (columns: key, label, unit?, align?, format? delta/delta_pct, better? lower/higher, mono?; rows; sort?; caption?) | bar_chart / line_chart (title, x_label, y_label, unit, caption, categories / x, series: name, values, role? baseline, color?; bar orientation? horizontal) | timeline (items: date, label, end?, status?, detail?, milestone?, critical?) | findings (items: severity critical/high/medium/low/info, title, file?, line?, location?, evidence?, detail?, status?, fix?) | diff (before/after: label, text) | badges (items: label, tone?) | text (text).
Table cells may be {"badge": "pass", "tone": "good", "sort": 2}. Tones: good, bad, warn, info, accent, neutral.
Validation rejects: unknown block types, empty lists (a canvas never shows an empty state), charts missing title, axis labels, unit, caption or series names, duplicate series names, mismatched lengths, non-finite numbers.
Design rules
- Lead with a
stats row: the 3 to 5 numbers that answer the question.
- One idea per chart. Bars for categories, lines for time. Horizontal bars for long labels (models, file paths).
- Every chart: title says what, axes say units, caption says source and time range. Legend appears automatically for multi-series.
- Compare with
role: "baseline" (grey) for before/previous and the accent for after/current. No more than 4 series.
- Tables for detail, sorted by what matters (severity, p95, tokens). Numbers are right-aligned automatically.
- Flat and calm: the kit forbids gradients, shadows and emojis; do not try to restyle it. Color means status, not decoration.
- Aggregate before sending: top 12 plus "others", daily not per-request. Max 5000 table rows.
- No em or en dashes in labels; use a hyphen.
Pre-delivery self-check
- Would a sentence answer it? Then do not make a canvas.
- Does the first row of stats answer the user's question on its own?
- Every chart has title, x_label, y_label with a unit, caption with source and range?
- Numbers come from tool output or files you read, not estimates? Say "estimated" where they are.
- Nothing empty, nothing duplicated between a chart and the table beside it without a reason.
- The tool returned "Presented 1 artifact(s)". If it returned a validation error, fix the listed blocks and call again.
Introducing it in the reply
One sentence pointing at the canvas plus the key takeaway, then stop: the canvas carries the detail. The first time in a conversation, say what it is: "I put the results in a canvas on the right: a sortable, themed report you can keep open while we work." Do not repeat the canvas content in the transcript.
1---2name: canvas3description: Present standalone analytical results (audits, code review findings, benchmarks, metrics, token usage, PR/CI status, project boards, video QA, investigations, data-heavy comparisons) as a Navin canvas in the Artifact Canvas panel, built from typed JSON blocks with present_artifact type=canvas. Not for architecture diagrams (archify), specific deliverables, or short answers.4metadata:5 source: https://skillmd.com/skills/navinspire-ia/canvas6 navin:7 category: documents8---910# Canvas1112A canvas is a self-contained analytical page beside the chat: stat rows, sortable tables, labelled charts, timelines, findings and diffs, themed like the app. You send **data** (blocks); Navin validates it and renders it with a bundled kit. No HTML to write, no network at view time.1314## When to use1516Use a canvas when the result is **standalone and analytical**, and the user will scan, sort or compare it:1718- audits and code review / security findings (severity, file:line, evidence, fix status)19- performance benchmarks, before/after comparisons20- metrics and usage (token usage, costs, counts over time)21- investigations with several data points, architecture *reviews* with scored findings22- status boards (PR checks, CI, project board, milestones)23- repeatable reports the user will ask for again (same template, new data)2425## When not to use2627- The user asked for a **specific deliverable** (a docx, a deck, an app, a file in the repo): build that.28- The work happens **in a specific tool** (editor, terminal, browser, board UI): stay there.29- **Active debugging** loops: keep it in the transcript, fast.30- **Short answers**: a sentence or a small Markdown table beats a canvas.31- **Architecture diagrams, PPT visuals, tender/RFP technical answers, Markdown plans**: `archify` stays the default (AGENTS.md). Canvas is for data and analysis; it can sit next to an archify diagram, not replace it.32- Outside the WebUI / desktop (Telegram, CLI): there is no canvas panel; answer in text.3334## How to call it3536Free-form blocks:3738```json39present_artifact({40 "type": "canvas",41 "title": "API latency audit",42 "subtitle": "staging, 2026-10-12 to 2026-10-19",43 "blocks": [44 {"type": "stats", "items": [45 {"label": "Endpoints", "value": 42},46 {"label": "p95 over budget", "value": 3, "tone": "bad"}47 ]},48 {"type": "bar_chart", "title": "p95 by endpoint", "x_label": "Endpoint",49 "y_label": "p95 latency", "unit": "ms", "categories": ["/chat", "/files"],50 "series": [{"name": "Before", "values": [180, 95], "role": "baseline"},51 {"name": "After", "values": [120, 90]}],52 "caption": "Source: k6 run, 5 min at 50 rps. Lower is better."}53 ]54})55```5657Templates (Navin data in, blocks out; `navin/artifacts/canvas/templates.py`):5859| template | data | source |60|---|---|---|61| `benchmark` | `{"before": report, "after": report, "label_before"?, "label_after"?}` or `data_path` | `scripts/bench_tui.py --json`, `scripts/bench_agent_hotpath.py` |62| `findings` | findings list, or the `security_scan` / `code_review filter` result | review + security tools |63| `pr_status` | `{"pr": forge pr_view, "ci": forge ci_status, "threads"?: [{path, line, resolved, body}]}` | `forge` tool |64| `board` | omit to read the project board | `.navin/board/` |65| `token_usage` | omit to read the usage log; `{"prices": {model: {prompt, completion, cached}}}` (USD per 1M tokens) adds estimated cost | `/api/settings/usage` data |66| `video_qa` | `data_path` to `video-qa.json` | `navin/marketing/video_qa.py` |6768```json69present_artifact({"type": "canvas", "template": "video_qa", "data_path": "out/promo-qa/video-qa.json"})70```7172Pass the same `id` to update a canvas in place instead of opening a new tab.7374## Block types7576`section` (title, blocks, description?) | `stats` (items: label, value, unit?, delta?, tone?, hint?) | `table` (columns: key, label, unit?, align?, format? delta/delta_pct, better? lower/higher, mono?; rows; sort?; caption?) | `bar_chart` / `line_chart` (title, x_label, y_label, unit, caption, categories / x, series: name, values, role? baseline, color?; bar orientation? horizontal) | `timeline` (items: date, label, end?, status?, detail?, milestone?, critical?) | `findings` (items: severity critical/high/medium/low/info, title, file?, line?, location?, evidence?, detail?, status?, fix?) | `diff` (before/after: label, text) | `badges` (items: label, tone?) | `text` (text).7778Table cells may be `{"badge": "pass", "tone": "good", "sort": 2}`. Tones: good, bad, warn, info, accent, neutral.7980Validation rejects: unknown block types, empty lists (a canvas never shows an empty state), charts missing title, axis labels, unit, caption or series names, duplicate series names, mismatched lengths, non-finite numbers.8182## Design rules8384- Lead with a `stats` row: the 3 to 5 numbers that answer the question.85- One idea per chart. Bars for categories, lines for time. Horizontal bars for long labels (models, file paths).86- Every chart: title says what, axes say units, caption says source and time range. Legend appears automatically for multi-series.87- Compare with `role: "baseline"` (grey) for before/previous and the accent for after/current. No more than 4 series.88- Tables for detail, sorted by what matters (severity, p95, tokens). Numbers are right-aligned automatically.89- Flat and calm: the kit forbids gradients, shadows and emojis; do not try to restyle it. Color means status, not decoration.90- Aggregate before sending: top 12 plus "others", daily not per-request. Max 5000 table rows.91- No em or en dashes in labels; use a hyphen.9293## Pre-delivery self-check94951. Would a sentence answer it? Then do not make a canvas.962. Does the first row of stats answer the user's question on its own?973. Every chart has title, x_label, y_label with a unit, caption with source and range?984. Numbers come from tool output or files you read, not estimates? Say "estimated" where they are.995. Nothing empty, nothing duplicated between a chart and the table beside it without a reason.1006. The tool returned "Presented 1 artifact(s)". If it returned a validation error, fix the listed blocks and call again.101102## Introducing it in the reply103104One sentence pointing at the canvas plus the key takeaway, then stop: the canvas carries the detail. The first time in a conversation, say what it is: "I put the results in a canvas on the right: a sortable, themed report you can keep open while we work." Do not repeat the canvas content in the transcript.