OpenAlgo Charts roadmap

Agent brief: paste this into your agent

Task CH-N1 in marketcalls/openalgo-charts: a dry run of OpenAlgo's move from openalgo-charts 2.5.1 to 2.6.
Read CLAUDE.md (the release process, its OpenAlgo consumer dry run, and the rule that a dependency upgrade cannot substitute for a required consumer migration), COMPATIBILITY.md, website/pages/docs/upgrading.mdx, website/pages/docs/openalgo-compatibility.mdx, CHANGELOG.md from 2.5.2 to the current 2.6.x, and the header of scripts/check-openalgo-compat.mjs.
Goal: a classified dry-run report and a linked list of moves. Out of scope: fixing anything you find, and any change to OpenAlgo's code.
Steps:
1. Clone marketcalls/openalgo, then create a scratch linked worktree with git worktree add. The harness refuses a plain clone. Install the frontend's dependencies inside that worktree; do not symlink node_modules.
2. Run npm run build and npm pack here, install the tarball into that frontend, and install the openalgo-script version I name on the issue (the one OpenAlgo will ship).
3. Run node scripts/check-openalgo-compat.mjs --frontend <that frontend> with each flag that applies to OpenAlgo's frontend, once each with --browser chromium, firefox and webkit. --objects applies only if the frontend includes the shared Objects integration. Say which flags you left out and why.
4. Class every failure: library regression, required consumer migration (for example the 2.5.2 drawTools catalogue or the 2.5.3 grouped ObjectsPanel), or flag not applicable.
5. For each library regression, write a failing test and open an issue with it. Do not fix it here.
6. Add a section to openalgo-compatibility.mdx listing the moves from 2.5.1 to 2.6. Link each behaviour change to its upgrading.mdx or CHANGELOG.md entry instead of restating it, and name the setting that restores the old behaviour where one exists.
Rules: no new dependency; no change under src/; name no other product; no emoji, no em or en dashes; never tag, publish or dispatch a workflow.
Done when the three engines' summary lines are pasted with every failure classed, the section links every change, and npm run verify passes.
If a failure needs a decision this brief does not make, stop and ask me.

Agent brief: paste this into your agent

Task CH-N2 in marketcalls/openalgo-charts: prove that documents written by 2.5.1, the version OpenAlgo pins, load and round-trip on the current build, and record what a rollback to 2.5.1 does.
Read CLAUDE.md (Quality gates: Compatibility), COMPATIBILITY.md, website/pages/docs/upgrading.mdx (the 2.5.9 note on version 3 drawings documents), website/pages/docs/state.mdx and workspaces.mdx, and the existing tests for drawings migration and workspace revisions.
Goal:
1. scripts/golden-state.mjs: npm pack openalgo-charts@2.5.1 into a temp folder. Then, in a Playwright page, write through its public API a drawings document, a saved chart state, a layout or workspace document, an alert list and the widget's persisted keys, using bars from a seeded random walk. Commit the output under tests/fixtures/state-2.5.1/ with the script.
2. Tests that load each document on the current build and save it again. Name in the test each field that may change and why.
3. The reverse: write the same kinds from the current build and load them in 2.5.1. Write what loads, what is refused and what is dropped as a rollback section in upgrading.mdx.
Before writing the round-trip tests, delete one required field from one golden document and show me the loader failing on it.
Rules: no change under src/ (a load failure becomes an issue with the failing test, not a fix here); no new dependency; no account data; name no other product; no emoji, no em or en dashes; never publish.
Done when two generator runs match, every golden document round-trips, the rollback section covers every document kind, and npm run verify passes.

Agent brief: paste this into your agent

Task CH-N3 in marketcalls/openalgo-charts: give every open OpenAlgo chart report an outcome on the current 2.6.x.
Read CLAUDE.md (Testing traps) and .github/skills/openalgo-charts/references/pitfalls.md. Then read the reports listed by: gh issue list --repo marketcalls/openalgo --state open --search "chart in:title".
Goal: one outcome per report. Out of scope: fixes.
For each report:
1. Restate it in one sentence and decide which kind it is: a possible library defect, a host issue, a feature request, or something else.
2. For a possible library defect, find the public API path it goes through and write the smallest test that fails for that reason: a unit test, or a Playwright spec for anything that draws. Run it. If it fails, open an issue here with the test as a code block, the command, its output and a link to the report; do not commit the test. If it passes, draft a comment with the 2.6.x steps and the result.
3. For a host issue, draft a comment naming the OpenAlgo file or behaviour where it belongs (for #2136, the symbol search's type badge).
4. For a feature request, link the roadmap task that covers it, or list it for my decision.
Rules: commit no failing or skipped test; use synthetic bars from a seeded random walk, never account data; name no other product; no emoji, no em or en dashes.
Done when the tracking issue for CH-N3 has a table covering every open report with its outcome and links.
Ask me before posting anything on the OpenAlgo repository.

Agent brief: paste this into your agent

Task CH-N4 in marketcalls/openalgo-charts: add a roadmap task issue form, a bug form, a pull request template and a short AGENTS.md.
Read CLAUDE.md, CONTRIBUTING.md, pull request #37 (the shape to copy: problem, change, what is unchanged) and the Contribute section of openalgo.in/charts/roadmap.
Build:
- .github/ISSUE_TEMPLATE/roadmap-task.yml asking for: the task id, a plan in 3 to 5 lines, whether an agent is assisting (yes or no), and an expected finish.
- .github/ISSUE_TEMPLATE/bug.yml asking for: version, browser and device, steps, expected and actual result, and a minimal reproduction. Include a note never to paste account data.
- .github/pull_request_template.md with these sections:
  - task id and issue link;
  - problem, change, and what is unchanged;
  - commands run, with their summary lines;
  - the regression test and 'reverted the fix and watched it fail: yes';
  - Brotli bytes per tier and the chart-only import;
  - lines added and removed;
  - public .d.ts: none, additive (approved in the issue) or deprecation;
  - screenshots for anything that draws;
  - skills updated;
  - original work, or the source and licence of a ported algorithm;
  - agent assisted, and 'I have read every changed line and can explain it'.
- AGENTS.md of at most 30 lines that links CLAUDE.md, CONTRIBUTING.md and .github/skills instead of restating them.
- A 'Working with an agent' section in CONTRIBUTING.md.
Rules: restate no rule that already lives in CLAUDE.md, link it instead; name no other product; no emoji, no em or en dashes; change nothing under src/.
Done when the forms render in the new-issue chooser on your fork (attach a screenshot), a search for U+2013 and U+2014 in the new files finds nothing, and npm run verify passes.

Agent brief: paste this into your agent

Task CH-N5 in marketcalls/openalgo-charts: make the counts and versions in the agent skills come from the build, and fix the stale ones.
Read CLAUDE.md, .github/skills/README.md, .github/skills/openalgo-charts/SKILL.md, .github/skills/openalgo-chart-indicator/SKILL.md, scripts/check-skills-coverage.mjs and the skills step in .github/workflows/ci.yml (it is a separate CI step, not part of npm run verify).
Goal: scripts/check-skills-facts.mjs, run after npm run build. It reads:
- the tier count from package.json exports;
- the reference file count from .github/skills/openalgo-charts/references;
- the indicator and chart type counts from the built dist after importing the indicators tier (registeredIndicators, registeredChartTypes), and the drawing tool count after importing the draw tier (registeredDrawingTools lives there);
- the version from package.json.
It matches count phrases only ('N built-ins', 'N-tier', 'nine tiers', 'N reference files', 'references target X'), in digits or words. It must not flag bare numbers: the skills contain sample bars such as high: 102. It fails when a phrase states a different number, naming the file and line. Add npm run skills:facts and a CI step beside skills coverage. Then fix the text it flags.
Before fixing the text, run the script on master and show me the failures it finds.
Rules: no new dependency; change nothing under src/; name no other product; no emoji, no em or en dashes; never publish.
Done when the script exits 1 on a fixture saying '103 built-ins', ignores the sample bar data, exits 0 after your text fixes, and npm run skills:coverage and npm run verify pass. Paste their summary lines.

Agent brief: paste this into your agent

Task CH-N6 in marketcalls/openalgo-charts: make CI reliable enough that I can require it before a merge.
Read CLAUDE.md (Quality gates: Tests, and the rule never to retry a flaky test into green), playwright.config.ts (projects, retries, workers), every file in .github/workflows, CONTRIBUTING.md, and the Playwright notes in the last three CHANGELOG.md entries.
Goal:
1. scripts/e2e-flake.mjs: run the three-engine e2e suite N times at a given worker count and print a failure rate per spec. Run it five times at 6 workers and five times at 2, on an idle machine, and attach both reports.
2. For each spec that failed at least once, find the cause (for example a wait on a load event that a busy machine misses) and fix it in the test. If you cannot find it, open an issue with the spec, its failure rate and the logs.
3. Make CI print Playwright's flaky list (tests that passed only on retry) in the job summary. Do not change the retry count.
4. In CONTRIBUTING.md, document a worker count that stays green on a four-core laptop and the focused runs for a typical change.
5. Move every node-version in .github/workflows from 20 to 22. Leave the engines field in package.json alone.
I then enable required CI on the default branch; you do not change repository settings.
Rules: never add a retry, raise a timeout or skip a test to make it pass; no change under src/ (a library defect becomes an issue first); no new dependency; name no other product; no emoji, no em or en dashes; never publish.
Done when the before and after flake reports are attached, every flaky spec is fixed or has an issue with its rate, CI passes on Node 22 with the flaky list in its summary, and npm run verify passes.

Agent brief: paste this into your agent

Task CH-N7 in marketcalls/openalgo-charts: script the three gates CLAUDE.md marks as checked at release. One pull request per script; claim one at a time.
Read CLAUDE.md (Quality gates for every release; Keep the code lean), tsconfig.json (the paths map), scripts/line-caps.json and tests/line-caps.test.ts (the shrink-only pattern), and package.json (verify and lint). First check that 2.6.0 has not already added one of these checks; if it has, stop and tell me.
Goal:
1. scripts/check-cycles.mjs: walk runtime imports and re-exports under src/. Resolve relative paths, and resolve package names (openalgo-charts, openalgo-charts/draw and the rest) through the tsconfig paths map. Ignore import type and export type. Print each cycle as a chain of files.
2. scripts/check-unused-exports.mjs: list exports under src/ that no other module imports and no tier entry point re-exports, directly or through another index.
3. scripts/check-skips.mjs: fail when test.skip, describe.skip, it.skip or test.fixme has no reason string. Every skip has one today; this keeps it that way.
Findings that exist today go in a shrink-only allow file per script (a test fails if an entry is added). Wire each script into npm run verify, and make lint run with --max-warnings 0.
Before writing each checker, add a fixture it must reject and a test that expects the rejection. Show me the failure.
Rules: no new dependency (use the TypeScript compiler already in devDependencies if you need a parser); change nothing under src/; never raise a cap or budget; name no other product; no emoji, no em or en dashes; never publish.
Done when each fixture fails its checker, master passes with the allow files, each script runs in under five seconds, and npm run verify passes. Paste the summary lines and the allow lists.

Agent brief: paste this into your agent

Task CH-N8 in marketcalls/openalgo-charts: a committed report of each tier's public API that CI keeps current.
Read CLAUDE.md (Quality gates for every release), COMPATIBILITY.md, scripts/check-dts.mjs, package.json (exports and files), and the dist/*.d.ts files after npm run build. First check that 2.6.0 has not already added such a report; if it has, stop and tell me.
Goal: scripts/api-report.mjs with three modes.
- write: for each of the nine tier entry points, one entry per exported name, sorted by name, with its declaration text. Remove JSDoc comments and private members, and normalise the order of declaration blocks (dist/index.d.ts orders some blocks differently on Linux and Windows). Write api/<tier>.api.txt.
- --check: exits 1 when the built dist differs from the committed reports, naming each added, removed or changed export. Wire it into npm run verify.
- --against <version>: fetches that published version into a temp folder (npm pack openalgo-charts@<version>) and prints additions, removals and narrowed signatures. The release step uses it, and so does OpenAlgo's pin (2.5.1).
Before writing the checker, write a test in which a fixture removes one export and --check is expected to name it. Show me the failure.
Rules: no new dependency (use the TypeScript compiler already in devDependencies to parse); no change under src/; name no other product; no emoji, no em or en dashes; never publish.
Done when two runs of write give byte-identical files, a comment-only change leaves the report unchanged, the fixture fails --check, --against 2.5.1 prints a readable difference, and npm run verify passes. Paste the summary lines and the first 30 lines of the 2.5.1 difference.

Agent brief: paste this into your agent

Task CH-N9: the shared eval harness. The OpenScript roadmap uses the same one; one issue in the eval repository covers both.
Read the eval section of openalgo.in/charts/roadmap, the article it links (Automating eval design and hillclimbing), and the README of the eval repository the maintainer links from the issue.
Build, in the eval repository, as three pull requests (runner, statistics, linter):
- A case format: one folder per case holding:
  - task.md: the request as a person would write it, plus a fixed harness contract;
  - meta.json: source, why a human finds it hard, the target failure bucket, the version under test and the approval;
  - grader/: the reference solution, expected values and hidden tests.
- A runner. The agent's model loop runs outside the container, and only its tool calls (shell, file edits) run inside. The container holds only the task, the packed library tarball at a pinned version, the surface bundle under test and an npm registry mirror. It has no secrets, no network except the mirror, and never the grader folder. Model and effort are pinned per campaign.
- Three runs per case. Each run logs its transcript, input, cached and output tokens, wall time, verdict and failure bucket.
- An infrastructure-error verdict only when the case's reference solution fails the same step in the same environment and run (registry down, browser fails to launch). Such runs are retried once, counted, and excluded from the score. A step that fails only for the agent's solution is a fail.
- Statistics:
  - a 95 percent bootstrap interval over cases;
  - a paired comparison of two configurations on the same cases;
  - the A/A noise floor;
  - the keep or revert rule: keep only if train and test both improve beyond noise and no bucket regresses; stop after three rounds with nothing kept.
- A patch linter for surface diffs. It flags identifiers, numbers and titles that appear in a case but not in the published .d.ts files or docs pages, and any run of eight words shared with a case. Public API names such as createChart and addIndicator are allowed.
Rules: no secrets in the repository or the container; case text names no other product and uses no account data; no emoji, no em or en dashes.
Done when:
- one sample case runs end to end twice with the same verdict;
- a registry outage is reported as an infrastructure error, and an agent's bad dependency as a fail;
- unit tests on synthetic results reproduce known intervals and keep or revert decisions;
- a test proves grader/ and secrets are absent inside the container;
- the linter rejects a diff seeded with a case string and passes one that uses only public names.
Paste the outputs.

Agent brief: paste this into your agent

Task CH-N10: the integration eval grader, in the shared eval repository.
Read the eval section of openalgo.in/charts/roadmap. In marketcalls/openalgo-charts, read .github/skills/openalgo-charts/references/pitfalls.md, website/pages/docs/frameworks.mdx, tests/e2e/render-parity.spec.ts (how it decides a canvas painted), website/pages/docs/state.mdx, and the Series interface in src/model/series.ts (getData).
Goal: a programmatic grader that runs these steps in order. The first failure sets the verdict's bucket.
1. Install and build succeed.
2. tsc --strict passes against the published package types. TS2339 and TS2305 map to the invented-API bucket.
3. A headless Chromium page loads with no page errors.
4. The chart canvas painted: at least 10 distinct colours.
5. After a scripted feed tick and a symbol switch, state matches the case's expected values. The harness contract in task.md tells the app to assign its chart to window.__evalChart. Read series, indicator ids, settings and time zone through __evalChart.getState(), and the last bar through the main series' getData(); getState() carries no bars.
6. Mounting, unmounting and mounting again leaves no leaked listeners or canvases.
Then write ten seed cases across two frameworks (vanilla, plus one of React, Vue or Next.js) and five task types. Each has a human-written reference and one mutant that breaks one thing the case is about.
Rules: no LLM judge; the grader never runs inside the agent's container; case text uses synthetic data and names no other product; no emoji, no em or en dashes.
Done when the grader passes 10 of 10 references, fails 10 of 10 mutants in the intended bucket, and gives identical verdicts on a second run of all 20. Paste the verdict table.

Agent brief: paste this into your agent

Task CH-S1 in marketcalls/openalgo-charts: publish the 2.6 performance baseline and the noise floor of the release bench.
Read docs/browser-endurance.md, docs/performance-notes.md (Budgets), scripts/browser-endurance.mjs (its flags), scripts/render-bench-budgets.mjs, benchmarks/releases.json and .github/workflows/nightly.yml.
Goal:
1. scripts/bench-noise.mjs: for each release in benchmarks/releases.json and each cell, print the spread of the five runs of each build (max minus min, divided by min). Flag cells where the release's claimed change falls inside that spread.
2. On an idle machine, run npm run endurance:browser on the current 2.6.x build with --bars 2000, 10000 and 50000. Replace the 2.5.5 table in docs/browser-endurance.md with the new rows, naming device, OS, browser, backend and version. Attach the raw JSON outputs to the pull request.
3. Re-derive the CI budget table from the last several nightly render-bench artifacts (gh run download), using scripts/render-bench-budgets.mjs where it applies, and state the margin. Put the table in the pull request description as a proposal. I decide whether any budget changes; do not edit a budget file.
Out of scope: any change under src/.
Rules: no new dependency; never hand-edit a measured number; never change a budget; name no other product; no emoji, no em or en dashes; never publish.
Done when bench-noise reports a spread of about 57 percent for the 2.5.8 canvas2d 50k tick cell, the endurance table is in the docs, the proposal is in the description, and node scripts/check-render-bench-docs.mjs and npm run verify pass. Paste the summary lines.

Agent brief: paste this into your agent

Task CH-S2 in marketcalls/openalgo-charts: profile the ten-study live tick at 200,000 bars and publish where the time goes.
Read docs/performance-notes.md, tests/e2e/render-bench.perf.ts (the STUDIES list), .github/skills/openalgo-charts/references/pitfalls.md (the entry on longer histories), src/indicators/tail.ts and ARCHITECTURE.md.
Goal: a repeatable profile, not an optimisation.
- Add a mode to the render bench that records a CPU profile of the ten-study tick at 10,000 and 200,000 bars on canvas2d.
- Group JavaScript self time into calc, calcTail, output column copies, plot point comparison, rendering calls and other, at both sizes. A JavaScript profile does not include raster and compositing, so report the remainder of the measured frame time as its own line.
- Add a table to docs/performance-notes.md, plus a ranked list of the costs that grow with history, each naming its file and function.
- Add a bench cell that ticks the studies listed on the CH-S3 issue at 10,000 and 200,000 bars, so incremental tails have their own measure.
Out of scope: any change to src/ beyond what the profile needs in order to run (ideally nothing).
Rules: no new dependency; do not change budgets; name no other product; no emoji, no em or en dashes; never publish.
Done when two runs give group shares within 5 percentage points of each other, the groups add up to the profiled JavaScript time within 10 percent, every named function exists, the new cell runs, and npm run verify passes. Paste the table.

Agent brief: paste this into your agent

Task CH-S3 in marketcalls/openalgo-charts: give one more built-in indicator an incremental calcTail. Claim the indicator on the issue first; the issue lists the order. If this is your first task, claim one marked window rerun.
Read CLAUDE.md, src/indicators/tail.ts and src/indicators/steppers.ts (the two shapes: rerun a bounded window, or resume from a checkpoint), one existing tail of the same shape (RSI or BOLLINGER), tests/indicator-tail.test.ts, tests/e2e/indicator-tail-parity.spec.ts, and the indicator's own descriptor and hand-computed tests.
Goal: wrap the descriptor with withTail. After each tick, the tail spliced onto the previous result must equal a full calc value for value (Object.is, NaN and signed zero included). For a checkpoint resume, that means repeating the full calculation's floating-point order; post your plan on the issue and wait for a go-ahead.
First add the indicator to the property test's list and show me the test failing because the descriptor has no calcTail.
Rules: no new dependency; no public .d.ts change; bytes only in the indicators tier; reuse the existing kernels and steppers and never write a second copy of the formula; do not touch src/core/chart.ts or src/core/pane.ts; name no other product; no emoji, no em or en dashes; never weaken the property test, raise a budget or publish.
Done when:
- tests/indicator-tail.test.ts passes over random histories with gaps, NaN, ties and overflow;
- the tail parity spec passes;
- changing one constant in your tail makes the test fail (show it, then restore);
- npm run verify passes.
Report the study-list bench cell before and after, Brotli bytes per tier before and after (npm run size), and lines added and removed.

Agent brief: paste this into your agent

Task CH-S4 in marketcalls/openalgo-charts: remove one per-tick cost that grows with loaded history. Take it from the ranked list in docs/performance-notes.md (CH-S2). One cost per pull request; name it on the issue first.
Read CLAUDE.md (Keep the code lean; Concurrency), docs/performance-notes.md (the profile, and the noise floor from CH-S1), the file and function the cost names, and tests/e2e/render-bench.perf.ts.
Goal: the live tick does bounded work for this cost, whatever the history length, and produces identical output.
If the change needs src/core/chart.ts or src/core/pane.ts, stop and write it up as a proposal for me instead. I am their only writer.
Before changing code, record the ten-study tick cell five times on the current build.
Rules: no new dependency; no public .d.ts change; no new option; nothing added to the base beyond the measured need; no worker and no data-layout rewrite; never accept a new render baseline; name no other product; no emoji, no em or en dashes; never publish.
Done when:
- the targeted cell improves by more than the larger within-session spread of both builds;
- the held-out workloads do not regress beyond their noise (soak heap slope, npm run endurance:browser at 10,000 bars, the tick and tail parity specs, 500 drawings under hover);
- render parity shows zero differing pixels;
- npm run verify passes.
Paste the before and after tables, npm run size and npm run shake.

Agent brief: paste this into your agent

Task CH-S5: the integration eval train set, in the shared eval repository.
Read the case format (CH-N9) and the grader (CH-N10). Also read charts issues #1, #6, #9, #12, #13, #16 and #25, the open and closed OpenAlgo issues with chart in the title (and the CH-N3 table), and .github/skills/openalgo-charts/references/pitfalls.md.
Goal: grow the train set to 20 cases, spread across frameworks (vanilla, React with Vite, Next.js, Vue) and families (setup and lifecycle, streaming updates, indicators, drawings, widget, persistence).
Sources, in this order:
1. Real reports, reframed as integration tasks.
2. Traps a maintainer judges hard: never cache the forming bar; bar times are in seconds; the IST default versus named time zones; teardown on route change; lazy history paging; drawings per instrument.
3. Synthetic variants of those, across frameworks.
Each case has:
- task.md, written the way a developer would ask, with every parameter fixed so the answer is unambiguous, plus the harness contract (the chart exposed as window.__evalChart);
- meta.json, giving its source and why a human finds it hard;
- a reference solution and a mutant.
Rules: run no model on a case before the maintainer approves it; never choose or change a case because a model failed it; no account data and no real credentials; name no other product; no emoji, no em or en dashes.
Done when the maintainer has approved each case on the issue, every reference passes and every mutant fails the grader, and a rerun gives identical verdicts.

Agent brief: paste this into your agent

Task CH-S6 (maintainer only, except the parts marked public): the private test set and the first baseline of the integration eval.
Read the eval section of openalgo.in/charts/roadmap, the harness README (CH-N9), the grader (CH-N10) and the train set (CH-S5).
Goal:
1. Private: write at least 15 test cases in the private repository, stratified by framework and task family like train. Mark 10 as the anchor subset, which never rotates. Each has a reference and a mutant.
2. Private: run the baseline on train and test, three runs per case, twice for the A/A noise floor, plus one run by an agent from another vendor. Score each graded output twice to check that the grader agrees with itself.
3. Private: read 20 scored transcripts across buckets, and note any verdict you disagree with.
4. Public (a contributor may draft this): the report template and the harness configuration for the baseline.
5. Publish the report: train and test scores with 95 percent intervals, noise floor, grader agreement, cost per task, infrastructure-error counts and failure buckets.
Rules: no test case text, identifier or title in any public file; stay within the eval spend cap; no account data; name no other product; no emoji, no em or en dashes.
Done when every private reference passes and every mutant fails twice, the report is published, and the headroom check is met, or harder cases are added and the baseline rerun.

Agent brief: paste this into your agent

Task CH-S7: run the first hillclimbing campaign on the charts skills text, against the integration eval.
Read the eval section of openalgo.in/charts/roadmap, the harness README (CH-N9), the baseline report (CH-S6), .github/skills/README.md and scripts/check-skills-coverage.mjs.
The maintainer runs every scored run (train and test) on the maintainer's account within the campaign cap. You and your agent see train transcripts and verdicts, and only the aggregate test score. You may run train cases on your own account while drafting a patch, but only scored runs decide.
Each round:
1. Your agent reads the train failures only and groups them by bucket.
2. It proposes one patch to one file under .github/skills that fixes one root cause as a general rule. The patch contains no case-specific text, identifiers, numbers or titles.
3. The patch linter must pass; then the maintainer scores train and test against the baseline and the noise floor.
4. Keep the patch only if train and test both improve beyond the noise floor and no bucket regresses. Otherwise revert it and log it.
Stop after two or three rounds with nothing kept, or at the cost cap, and write a root-cause note. If failures point at a missing library API, file an issue for a human decision; never change the library during this campaign.
Rules: never read the private test set or ask for test transcripts; npm run skills:coverage stays at 100 percent and every name in the skills exists in dist; name no other product; no emoji, no em or en dashes.
Done when the kept patches are merged as pull requests, the keep or revert log is in the eval repository, and a report gives the test score with its 95 percent interval before and after, cost per task, flip rate and bucket counts. A flat campaign is a finished campaign.

Agent brief: paste this into your agent

Task CH-S8: a bug-fix eval mined from past fixes in marketcalls/openalgo-charts.
Read CLAUDE.md (Testing traps: write the regression test, then revert the fix and watch it fail), the harness README (CH-N9), and the output of git log --grep '^fix' in the charts repository.
Goal, in the eval repository:
- Scripts that list the fix commits touching tests/. For each one, they check mechanically that the new or changed test fails on the parent commit and passes on the fix, and record the exact test command. Expect about 45 to 50 candidates between 2.5.5 and 2.5.9 and a lower yield after the check; go back before 2.5.5 if you need more.
- A case skeleton for each verified commit: the parent as an archive with no git history, the hidden test in grader/, and a slot for a bug report.
- A registry mirror per case that holds no openalgo-charts version published after the parent commit.
- 30 verified cases, unit-level first. Each has a bug report written in user terms by a person, not the commit message.
- A temporal split: train from the older releases, test from the newest. Hand the test cases to the maintainer for the private repository.
Rules: no case sandbox may contain history, the fix, the hidden test or a later library version; no account data; name no other product; no emoji, no em or en dashes.
Done when:
- each of the 30 cases has a log showing a fail on the parent and a pass on the fix;
- a test proves the sandbox has no .git folder and no grader files, and the mirror no later version;
- one case runs end to end through the harness;
- the maintainer has approved the bug reports.

Agent brief: paste this into your agent

Task CH-S9 in marketcalls/openalgo-charts: find out which built-ins lack hand-computed cases, then add cases for them three at a time.
Read tests/parity-momentum.test.ts (the convention: every number is worked by hand in a comment and never read back from the code), the other tests/parity-*.test.ts and tests/indicator-audit-*.test.ts files, and registeredIndicators in src/model/indicator-registry.ts.
Goal:
1. scripts/indicator-coverage.mjs: from the built dist, list every registered indicator id with the number of hand-computed cases that name it across the test files, and list the ids with fewer than three.
2. Then open one pull request per three uncovered built-ins. Add at least three cases for each, including a warmup case and an absent-value case, worked from the published definition with the arithmetic in comments.
Before adding cases, show me the coverage output and check five ids by hand.
Rules: never derive an expected value by running the indicator; no change under src/ (if a case exposes a defect, stop and open an issue with the failing case); no new dependency; name no other product; no emoji, no em or en dashes.
Done when the script agrees with your manual check, changing one constant in each covered indicator makes its new cases fail (show it, then restore), and npm run verify passes.

Agent brief: paste this into your agent

Task CH-S10 in marketcalls/openalgo-charts: run an accessibility audit with people who use assistive technology every day.
This task needs people; your agent helps you prepare and write up. Read website/pages/docs/interactions.mdx and keyboard-shortcuts.mdx, website/pages/docs/widget.mdx, and the chart's live region in src/core/chart.ts (read only).
1. With your agent, draft a session script. It covers tasks on a bare chart page and on the widget demo: find the latest price, read the last five bars, add an indicator, change the interval, draw a line and remove it, and move between charts in a grid. It also says what to observe, and includes a consent note. Post it on the issue for the maintainer to approve.
2. Run sessions with at least three people: a desktop screen reader user, a phone screen reader user, and a keyboard-only or screen magnification user. Record notes, not identities. If you cannot find three people within eight weeks, run what you can and say so.
3. File each finding as an issue with steps, expected result, actual result and severity.
4. Write website/pages/docs/accessibility.mdx: what was tested, on which version, what works and what does not, and any gap in who was tested, with links to the issues.
Rules: no personal data anywhere; no change under src/; name no other product; no emoji, no em or en dashes.
Done when the sessions are summarised (or the gap is stated), every finding is an issue, and npm --prefix website run build passes with the page in the docs navigation.

Agent brief: paste this into your agent

Task CH-S11 in marketcalls/openalgo-charts: read the chart bar by bar from the keyboard.
Read CLAUDE.md (Keep the code lean; Concurrency), src/input/shortcuts.ts (DEFAULT_KEYMAP and the reserved combinations), _runShortcut in src/core/chart-input.ts, the live region and its summary in src/core/chart.ts (read only), website/pages/docs/keyboard-shortcuts.mdx and interactions.mdx, playwright.config.ts (projects), and any findings from CH-S10.
Goal: a new ShortcutManager command that enters bar-reading mode. It appears in the keymap editor and can be rebound. Propose its default binding on the issue; it must collide with no default or reserved combination. While the mode is on:
- Left and Right move a focus bar and announce its date, open, high, low, close and volume through the existing live region;
- Home and End jump to the first and last loaded bar;
- Escape leaves the mode and gives the arrow keys back to their bound commands (pan by default).
The text comes from a formatter the host can replace, so it can be localised. Put the mode and the announcer in a new module, reached through the existing shortcut path. If you still need a change in src/core/chart.ts, describe it in the pull request instead of editing the file; I am its only writer. Propose the public option on the issue and wait for approval before adding it.
Measure first: if the feature adds more than 0.5 kB Brotli to the base, it belongs in the widget tier instead.
Before writing it, add a Playwright spec that enters the mode, presses Right three times and expects three announcements. Add it to the widget-loading-* testMatch in playwright.config.ts (and to the chromium project's testIgnore, as vue-integration is) so it runs once in each of chromium, firefox and webkit. Show me it failing.
Rules: no new dependency; no public .d.ts change beyond the approved option; reuse the existing bar lookup helpers; name no other product; no emoji, no em or en dashes; never raise a budget or publish.
Done when the spec passes in three engines, the announced values match getData() on the main series, rebinding works, Escape restores panning, keyboard-shortcuts.mdx documents the command, and npm run verify passes. Paste npm run size, npm run shake, and lines added and removed.

Agent brief: paste this into your agent

Task CH-S12 in marketcalls/openalgo-charts: a tested recipe for Indian number formats and localised month names on the series axes.
Read website/pages/docs/constants.mdx (volume compaction), website/pages/docs/series-and-styling.mdx, the priceFormatter and timeFormatter options in src/core/chart-types.ts (read only), examples/vue for the example layout, and playwright.config.ts (projects).
Goal: examples/axis-locale/ with:
- an Intl-based priceFormatter and timeFormatter, with Hindi month names as the sample;
- a volume formatter in lakh and crore for the series axes (for example, 1,50,000 shows as 1.5 L and 2,50,00,000 as 2.5 Cr).
Add a recipe page, website/pages/docs/axis-localization.mdx, that shows the same code and says that indicator legend values, the measure tool and the footprint still use K, M and B. List those three places on the issue for my decision. The library does not change.
Before writing the formatters, write unit tests for them, including the boundaries at 1 L and 1 Cr, and show me they fail.
Rules: no change under src/; no new dependency; example bars come from a seeded random walk, never a sine wave; name no other product; no emoji, no em or en dashes.
Done when:
- the unit tests pass;
- a Playwright spec, added to the widget-loading-* testMatch in playwright.config.ts, loads the example in chromium, firefox and webkit and, through a wrapper around the formatters, confirms the chart received lakh or crore and Hindi month labels;
- npm run size shows no tier change;
- npm --prefix website run build passes.

Agent brief: paste this into your agent

Task CH-S13 in marketcalls/openalgo-charts: put real phones and tablets on record.
Read website/pages/docs/mobile.mdx, COMPATIBILITY.md (the runtime boundary) and website/pages/docs/_meta.ts.
Your agent drafts; you run the checks on real devices.
1. With your agent, write website/pages/docs/device-checks.mdx.
   It has a numbered checklist:
   - open the widget demo;
   - pinch to zoom, and pan;
   - draw a trend line and move it;
   - open and close each mobile sheet;
   - on a tablet, confirm the desktop layout in auto mode;
   - open a grid of four charts;
   - watch one minute of live updates on the live demo and note any stutter.
   It also has a results table with these columns: device, OS, browser, device pixel ratio, backend, library version, date, result.
2. Run the checklist on your own devices. The horizon needs at least three rows: a low-end Android phone, an iPhone and an iPad. One device per pull request is fine.
3. File each failed step as an issue with a screenshot or screen recording, and link it from the table.
Rules: no change under src/; no personal data in screenshots; name no other product; no emoji, no em or en dashes.
Done when npm --prefix website run build passes, the page appears in the docs navigation, and your device row is filled in with links to any issues.

Agent brief: paste this into your agent

Task CH-S14 in marketcalls/openalgo-charts: publish the feed adapter contract as a copy-in kit.
Read docs/adapter-conformance.md, tests/conformance/adapter-contract.ts, controlled-transport.ts and reference-adapters.ts, tests/adapter-conformance.test.ts (its fake timers and the getTimerCount teardown check), COMPATIBILITY.md (host responsibilities), tsconfig.json, the ESLint config and package.json (files).
Goal: kit/feed-conformance/, a folder a host copies into its own tests. It holds the contract rewritten against the public types exported by openalgo-charts (no src/ paths). It takes three things from the host's runner: an assert function, a deepEqual, and a clock adapter that can install a fake clock, set the time, advance timers asynchronously and count pending timers. Write a vitest adapter for this repository's own test, which then runs through the kit so the two cannot drift. Add kit/ to the tsconfig include list and to lint. Document copy-in use in docs/adapter-conformance.md and on a website page.
Out of scope: a separately published package, vendor adapters, and any runtime code.
Before moving code, add a broken adapter fixture that skips duplicate repair, and show me today's contract failing it.
Rules: no new dependency; no change to the library's runtime or public .d.ts; name no other product; no emoji, no em or en dashes; never publish.
Done when:
- npx vitest run tests/adapter-conformance.test.ts passes through the kit;
- a copy of the kit in a scratch project using node:test and its mock timers passes against a reference adapter;
- the broken fixtures fail with named checks (duplicate repair, and a timer left running);
- npm pack --dry-run lists no kit file;
- npm run verify passes.

Agent brief: paste this into your agent

Task CH-S15 in marketcalls/openalgo-charts: a tested React example, following the pattern of the Vue one.
Read website/pages/docs/frameworks.mdx (the React section), examples/vue/, tests/e2e/vue-integration.spec.ts and how playwright.config.ts runs it (the chromium project's testIgnore and the widget-loading-* testMatch), the existing rollup config, and .github/skills/openalgo-charts/references/pitfalls.md (lifecycle).
Goal: examples/react/ with a hook and a component that:
- create the chart on mount;
- handle resize and theme;
- subscribe to a live feed;
- switch symbol;
- remove everything on unmount.
Write them with createElement so no JSX transform is needed, and keep the JSX form in frameworks.mdx, linking the example from there. React 19 on npm is CommonJS only, so bundle the example for the test with a small rollup config using @rollup/plugin-node-resolve and @rollup/plugin-commonjs, built by one npm script before the spec runs. This card approves react, react-dom and those two plugins as dev dependencies only.
Before writing the component, write tests/e2e/react-integration.spec.ts. It mounts twice (strict mode), switches symbol, unmounts, and checks that no listeners or canvases remain. Register it the way vue-integration is registered, so it runs once in each of chromium, firefox and webkit. Show me it failing.
Rules: no other new dependency; no change under src/; bars from a seeded random walk; name no other product; no emoji, no em or en dashes; never publish.
Done when the spec passes in chromium, firefox and webkit, the bundle builds from a clean checkout, npm pack --dry-run shows the package contents unchanged, npm run size shows no tier change, and npm run verify passes.

Agent brief: paste this into your agent

Task CH-S16 in marketcalls/openalgo-charts: one guide for a broker or platform taking the library to production.
Read COMPATIBILITY.md, SECURITY.md (supported versions and the broker-owned backend), website/pages/docs/performance-and-operations.mdx, instruments.mdx, openalgo-compatibility.mdx, upgrading.mdx (the rollback section), docs/adapter-conformance.md and docs/browser-endurance.md.
Goal: website/pages/docs/adopting-in-production.mdx, a go-live checklist covering:
- what the host supplies (transport, symbol resolution, session metadata);
- the credential boundary;
- testing in sandbox mode or analyzer mode before live orders;
- bounded history per device class, with the published numbers;
- pinning a version, and testing an upgrade with a consumer harness like the OpenAlgo one;
- rollback, and what an older build does with newer saved documents;
- what support means.
Link each fact to its source page instead of restating it.
Rules: no change under src/; for simulated trading, say sandbox mode or analyzer mode and use no other term; name no other product, broker or platform; no emoji, no em or en dashes.
Done when npm --prefix website run build passes with the page in the docs navigation, every claim links its source, and a person who has integrated a broker or platform has reviewed the pull request and had their comments resolved.

Agent brief: paste this into your agent

Task CH-S17 in marketcalls/openalgo-charts: let the host say which hyphenated names are symbols, so the chart expression parser reads them as one instrument (issue #16).
Read CLAUDE.md, CONTRIBUTING.md, issue #16 with its comments, src/transform/expression.ts (SYM_BODY, the tokenizer, parseExpression and isPlainSymbol) and website/pages/docs/transforms.mdx. Hosts call isPlainSymbol before parseExpression, so both need the option.
Goal: an optional options argument, { knownSymbol?: (name: string) => boolean }, on parseExpression and on isPlainSymbol. When a run of symbol characters joined by '-' is a known symbol, it is one instrument. Otherwise '-' stays a subtraction, exactly as today. Out of scope: any list of symbols inside the library.
Before changing code, write unit tests with a made-up symbol such as ABC-DEF that fail for this reason, including isPlainSymbol('ABC-DEF', { knownSymbol }) expected to be true. Run them and show me the failures. Also test that without the option both functions return exactly what they return today.
Rules: no new dependency; the only public .d.ts change allowed is this optional argument on these two functions; bytes only in the transform tier; extend the existing tokenizer, never add a second one; do not touch src/core/chart.ts or src/core/pane.ts; name no other product; no emoji, no em or en dashes; never skip or weaken a test, raise a budget or cap, or publish.
Done when npm run verify and npm run skills:coverage pass, and the docs and the skills reference describe the option. Paste Brotli bytes per tier before and after (npm run size), npm run shake, and lines added and removed.

Agent brief: paste this into your agent

Task CH-S18 in marketcalls/openalgo-charts: check the project's writing rules with one script, and clean up what it finds. Start only after I have added the user-interface wording rule to CLAUDE.md and decided the missing-value placeholder on the issue.
Read CLAUDE.md (writing rules and the new wording rule), .github/workflows/ci.yml (the product-name step, which scans src and tests), COMPATIBILITY.md (Deprecated APIs), src/alerts/types.ts and src/alerts/document.ts (alert states are stored values), and the encoded word list I post on the issue.
Goal: scripts/check-writing.mjs. It scans src, tests, scripts, website/pages, docs, examples and .github for four things:
- product names from the encoded list, stored encoded so the file names nobody in plain text;
- em and en dashes;
- emoji;
- the wording rule in user-interface strings.
It exempts identifiers, stored values, public types and names on the Deprecated APIs list. It replaces the grep step in ci.yml and runs in npm run verify. Then fix every finding, except in src/core/chart.ts and src/core/pane.ts; their lines go in scripts/writing-allow.json, which may only shrink, and I will clear those. The canvas placeholder for a missing value (the Buy and Sell quantity chip, and the footprint) changes only as I decided on the issue.
Before the cleanup, run the checker on master and show me its counts by rule and folder.
Rules: no new dependency; rewrite each sentence with a comma, colon, parentheses or full stop; change no behaviour, stored value or public type; name no other product; never publish.
Done when:
- fixtures for each rule fail the checker, master passes it, and a product name in tests/ is still caught;
- the public API report shows no change;
- render parity shows zero changed pixels, or each changed glyph or label is listed with before and after screenshots;
- npm run size states the bytes per tier that changed;
- npm run verify passes.
Paste the summary lines.

Agent brief: paste this into your agent

Task CH-S19 in marketcalls/openalgo-charts: seeded fuzz tests for every parser of untrusted input.
Read SECURITY.md (scope: bar data, WebSocket frames, broker responses, saved chart state, clipboard payloads, restored drawings) and CLAUDE.md, then the parsers: series setData and update; the OpenAlgo WebSocket frame handling in src/feed/openalgo-ws.ts; mapHistoryResponse in src/feed/openalgo-rest.ts and decodeOrder in src/feed/openalgo-trade.ts; restoreState; drawings fromJSON and migrateDrawings; clipboard payloads; workspace documents; alert lists. Note each parser's documented error or report shape.
Goal: a small seeded generator in the test folder, with no new dependency. It mutates valid documents (drop, duplicate, retype, truncate, deep nesting, huge numbers, NaN, prototype keys) and feeds each parser.
- Three pull requests: saved state and drawings; then bar data, WebSocket frames and broker responses; then the rest.
- CI runs a fixed seed with 10,000 inputs per parser; the nightly workflow runs a longer random seed.
If you find a crash, stop. Do not open a public issue; send it to me as SECURITY.md describes.
Rules: no change under src/ in these pull requests; no new dependency; name no other product; no emoji, no em or en dashes; never publish.
Done when every parser handles its inputs with no throw outside its documented shape, all six kinds are covered, a failure prints the seed and replays exactly, the fuzz tests add under 30 seconds to npm run verify, and npm run verify passes.

Agent brief: paste this into your agent

Task CH-L1 in marketcalls/openalgo-charts: finish the 3.0.0 deprecation list and remove it, one row per pull request. Claim one row on the issue.
Read COMPATIBILITY.md (Deprecated APIs), ARCHITECTURE.md (the rule for major releases), docs/migrating-to-2.md (the tone and shape to copy), website/pages/docs/drawing-tools.mdx (the chrome icon section), src/draw/icons.ts and src/draw/icon-svg.ts, and every use of the API in src, tests, examples, website and .github/skills.
The glyph row has two steps:
- In a 2.x release: measure what an import of openalgo-charts/draw alone pays for the glyphs only the widget uses. Mark those ids @deprecated naming 3.0.0 and add a row to the Deprecated APIs table. chromeIconIds and chromeIconSvg keep serving every id until 3.0.0.
- On the 3.0 branch: move those glyphs into the widget tier.
Every other row, on the 3.0 branch I name:
- remove the API and its uses;
- move its row to a Removed in 3.0.0 table;
- add a before and after entry to docs/migrating-to-3.md and to the website migration page.
Remove depth_level only after I confirm OpenAlgo's server no longer sends it.
Rules: remove nothing else; add no new API in this pull request; no new dependency; do not touch src/core/chart.ts or src/core/pane.ts unless I name you their writer for this row; name no other product; no emoji, no em or en dashes; never publish.
Done when:
- the public API report shows exactly this removal (or no removal, for the 2.x deprecation);
- the migration entry's after code compiles in a test against the 3.0 branch;
- the OpenAlgo consumer harness passes, or its failure is listed in the migration entry for OpenAlgo;
- every tier is the same size or smaller, and for the glyph row draw shrinks with zero differing pixels in tests/e2e/icon-raster.spec.ts;
- npm run verify passes.

Agent brief: paste this into your agent

Task CH-L2 in marketcalls/openalgo-charts: measure WebGL2 against Canvas 2D on real GPUs.
Read docs/performance-notes.md (the webgl2 rows time an emulated GPU), tests/e2e/render-bench.perf.ts, scripts/render-bench.mjs, README.md (what the GPU backend draws) and the noise floor from CH-S1.
Goal:
- A bench page that runs the render bench's pan, zoom and tick cells for both backends in the visitor's own browser (desktop or phone) and prints the results as JSON. Reuse the existing bench code; add no dependency.
- A results page, website/pages/docs/gpu-measurements.mdx, with rows from at least three desktops or laptops (discrete and integrated GPUs) and two phones, five runs per cell, with the spread.
- A closing recommendation for me: stay manual, or switch automatically above a stated visible-bar count. Cite only cells whose difference exceeds their spread.
Rules: no change to the default backend or to the rendering code under src/; never hand-edit a measured number; name no other product; no emoji, no em or en dashes.
Done when every row states device, OS, browser, GPU, device pixel ratio, backend and library version, the raw outputs are attached, and npm --prefix website run build passes.

Agent brief: paste this into your agent

Task CH-L3 in marketcalls/openalgo-charts: widget text in other languages as data files, starting with Hindi.
Read docs/widget-localization.md, website/pages/docs/widget.mdx, src/widget/localization.ts (WidgetBuiltinMessage, WidgetMessageKey and its schema keys), the settings schemas, package.json (files), playwright.config.ts (projects) and CLAUDE.md (Keep the code lean).
Goal:
- A build-time script that uses the TypeScript compiler (already in devDependencies) to list every member of the WidgetBuiltinMessage union, plus the schema keys found in the settings schemas, into a JSON file that only tests read. Never add a runtime array of keys to the widget.
- A catalog format: one JSON file per language (dist/locales/<lang>.json) holding every key. It is copied into dist at build and never imported by any tier, so no host pays for a language it does not load.
- A test that fails when a catalog is missing a key, has an extra key, or changes a placeholder.
- A loader recipe in docs/widget-localization.md: fetch or import the JSON, then pass it to the translate callback.
- hi.json, drafted with your agent and reviewed by a native Hindi speaker who is not you.
Before writing the catalog, write the parity test against an empty hi.json and show me it failing.
Rules: no new dependency; no public .d.ts change; no tier bundle may grow; name no other product; no emoji, no em or en dashes; never publish.
Done when:
- the parity test passes, and adding a message to the union makes it fail until the catalog follows;
- npm run size and npm run shake show no tier change;
- npm pack --dry-run lists dist/locales/hi.json;
- a spec added to the widget-loading-* testMatch reads three Hindi labels from the widget DOM in chromium, firefox and webkit;
- the reviewer has approved the pull request.

Agent brief: paste this into your agent

Task CH-L4: a joint eval with OpenAlgo's in-app agent, which draws on the operator's chart.
Read services/agent/chart_contract.py and the agent's chart tools in marketcalls/openalgo; in this repository, read website/pages/docs/drawing-tools.mdx and state.mdx; and read the harness README (CH-N9). OpenAlgo owns the agent's prompts and tool descriptions; this task builds only the grader and the cases.
Goal, in the eval repository: a grader that:
1. loads fixture bars into the chart;
2. applies the draw and clear operations the agent emitted into named groups;
3. recomputes, from the bars, the levels the request asked for (for example, the previous day's high and low);
4. compares them with the drawing document the chart holds, including requests the contract must reject.
Write 20 cases: requests written by maintainers, plus requests operators choose to share after a privacy review that strips account data. OpenAlgo collects no telemetry, and this task adds none.
Rules: no account data, no credentials, no real orders; synthetic bars from a seeded random walk; name no other product; no emoji, no em or en dashes.
Done when the grader passes 20 references and fails a mutant for each, a rerun gives identical verdicts, and the OpenAlgo maintainer has approved the case set.

Agent brief: paste this into your agent

Task CH-L5 in marketcalls/openalgo-charts: make the chart's built-ins agree with OpenScript's library, or say why they do not.
Read docs/integrating/library-vectors.md, docs/integrating/numerical-audit.md and spec/vectors/library/ in marketcalls/openscript. In this repository, read the indicator docs and scripts/indicator-coverage.mjs (CH-S9).
Goal:
- A table of the built-ins that overlap an OpenScript library function, with their parameters mapped (length, source, smoothing).
- Tests here that feed each vector's inputs to the built-in and compare the outputs. Copy the vectors under tests/fixtures and record their source commit.
- Each difference classed as either a documented default difference (written in both projects' docs) or a defect (an issue with the failing case).
Rules: no change under src/ in this pull request; no new dependency; never loosen a comparison to make it pass; name no other product; no emoji, no em or en dashes.
Done when every overlapping pair has a test or a written reason, every defect has an issue, and npm run verify passes.