feat(plugin): structured system-prompt sections by sanchitrk · Pull Request #48 · codeworksh/codework

sanchitrk · 2026-10-06T20:09:45Z

Structured system-prompt sections. Context files (AGENTS.md / CLAUDE.md) build on this next.

What

The Prompt domain no longer stores one string. A prompt plugin writes a foundation (the untagged head) and named sections; the harness renders them as snake_case XML-style tags, following Pi's structure.

import { Plugin, Section } from "@codeworksh/plugin";

const GithubPrRules = Section.define("github_pr_rules", { format: "list" });

export default Plugin.define({
	id: "acme.prompt.github",
	kind: "prompt",
	setup(ctx) {
		ctx.plugin.prompt.sections.append(Section.Rules, "Never push to main."); // merges into the one <rules>
		ctx.plugin.prompt.sections.append(GithubPrRules, "Link the issue in every PR description.");
	},
});

API: prompt.foundation.set/get and prompt.sections.append/set/remove/get, taking Section values, never strings.
Built-ins: Section.Rules (list), Addendum, ProjectContext, Skills (text). Custom sections come from Section.define(name, { format }): snake_case only; tools, cwd, foundation and built-in names are rejected.
Merging: sections merge by name across plugins. A format mismatch on one name fails the snapshot before any request.
Rendering (harness-owned): foundation, <tools>, <rules>, <addendum>, <project_context>, <skills>, <cwd>, then custom sections in first-write order. Tool promptGuidelines lead <rules>; list sections dedupe into - bullets; empty sections are skipped; bodies are never escaped.
codework.prompt.default is reduced to content: foundation, standing rules, caller append.

Rendered example: packages/harness/test/__artifacts__/prompt.sections.txt.

Behaviour changes

The tool index now lists the final tool catalog, not only the tools registered before the prompt plugin ran.
plugins: [] renders the fallback foundation plus <tools>(none)</tools> and <cwd>. Before, it rendered the bare fallback string.
PromptRegistry.set/get(string) is removed outright; there are no external users yet.

Also in this PR

refactor(harness): drop Location.Ref. Ref.directory let a Location point somewhere other than the mount cwd where tools run. Only a test fixture set it, and Location.layerWith / defaultLayer had no callers. location.directory is now always the mount cwd, which is what <cwd> renders.
fix(harness): regenerate retry artifacts from the tests. The committed runner.retry*.json artifacts had been reformatted outside the test, so they never matched what vitest writes. They are regenerated with -u (content unchanged); the tests are untouched.

Verification

vp check and vp run typecheck pass.
Harness suite: 513 passed. The one failure is the live OpenRouter muse model (runner.llm.live), a provider error unrelated to this change.
CLI suite (packages/codework): all non-live tests pass. The live acp tests against muse fail or time out on the provider side.
New E2E: two plugins merging into <rules> with dedupe, plus a custom list section and a custom text section, snapshotted to __artifacts__/prompt.sections.txt. Another E2E checks that a format mismatch fails before any request.
Peer-reviewed with codex. Its findings (CLI fixtures, unchecked hand-built sections, stale README) are fixed.

🤖 Generated with Claude Code

Replace the single-string PromptRegistry with a foundation (the untagged head) and named sections. Plugins write sections by value -- Section.Rules, Addendum, ProjectContext, Skills, or Section.define(name, { format }) -- so two plugins merge into one <rules> instead of overwriting the prompt. The harness now renders the prompt: foundation, <tools>, the built-in sections, <cwd>, then custom sections in first-write order. Tool promptGuidelines lead <rules>; list sections dedupe into bullets; a format mismatch on one name fails the snapshot. codework.prompt.default is reduced to content, and the tool index now covers the final catalog. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Ref.directory let a Location point somewhere other than the mount's cwd, where tools actually run. Only a test fixture set it, and Location.layerWith and defaultLayer took a Ref with no callers at all. Remove all three; the fixture mounts the remote sandbox at the repo instead, as the runner mounts at the session directory. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

The committed artifacts had been reformatted outside the test (tab-indented, mixed with spaces), so the retry snapshots never matched what vitest writes and failed on every run. Regenerate them with `-u`; the recorded content is unchanged, only the bytes now come from the test. `__artifacts__/**` is already excluded from `vp fmt`. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…tools Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(harness): retry transient provider failures in the loop (#47) * feat(plugin): structured system-prompt sections (#48) * feat(plugin): structured system-prompt sections Replace the single-string PromptRegistry with a foundation (the untagged head) and named sections. Plugins write sections by value -- Section.Rules, Addendum, ProjectContext, Skills, or Section.define(name, { format }) -- so two plugins merge into one <rules> instead of overwriting the prompt. The harness now renders the prompt: foundation, <tools>, the built-in sections, <cwd>, then custom sections in first-write order. Tool promptGuidelines lead <rules>; list sections dedupe into bullets; a format mismatch on one name fails the snapshot. codework.prompt.default is reduced to content, and the tool index now covers the final catalog. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * refactor(harness): drop Location.Ref; location is always the mount cwd Ref.directory let a Location point somewhere other than the mount's cwd, where tools actually run. Only a test fixture set it, and Location.layerWith and defaultLayer took a Ref with no callers at all. Remove all three; the fixture mounts the remote sandbox at the repo instead, as the runner mounts at the session directory. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(harness): regenerate retry artifacts from the tests The committed artifacts had been reformatted outside the test (tab-indented, mixed with spaces), so the retry snapshots never matched what vitest writes and failed on every run. Regenerate them with `-u`; the recorded content is unchanged, only the bytes now come from the test. `__artifacts__/**` is already excluded from `vp fmt`. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * chore(plugin): trim prompt section comments to decisions Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(plugin): allow hyphens in prompt section names Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(harness): tweak default plugin prompt & rules to the registered tools Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * feat(harness): load AGENTS.md and CLAUDE.md into the system prompt (#49) * feat(plugin): structured system-prompt sections Replace the single-string PromptRegistry with a foundation (the untagged head) and named sections. Plugins write sections by value -- Section.Rules, Addendum, ProjectContext, Skills, or Section.define(name, { format }) -- so two plugins merge into one <rules> instead of overwriting the prompt. The harness now renders the prompt: foundation, <tools>, the built-in sections, <cwd>, then custom sections in first-write order. Tool promptGuidelines lead <rules>; list sections dedupe into bullets; a format mismatch on one name fails the snapshot. codework.prompt.default is reduced to content, and the tool index now covers the final catalog. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * refactor(harness): drop Location.Ref; location is always the mount cwd Ref.directory let a Location point somewhere other than the mount's cwd, where tools actually run. Only a test fixture set it, and Location.layerWith and defaultLayer took a Ref with no callers at all. Remove all three; the fixture mounts the remote sandbox at the repo instead, as the runner mounts at the session directory. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(harness): regenerate retry artifacts from the tests The committed artifacts had been reformatted outside the test (tab-indented, mixed with spaces), so the retry snapshots never matched what vitest writes and failed on every run. Regenerate them with `-u`; the recorded content is unchanged, only the bytes now come from the test. `__artifacts__/**` is already excluded from `vp fmt`. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * chore(plugin): trim prompt section comments to decisions Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(plugin): allow hyphens in prompt section names Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(harness): tweak default plugin prompt & rules to the registered tools Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(harness): load AGENTS.md and CLAUDE.md into the system prompt A built-in prompt plugin, codework.prompt.instruction, walks from the sandbox cwd up to the sandbox root and appends the first readable instruction file per directory to <project_context>, closest last. Overrides (AGENTS.override.md, CLAUDE.override.md) beat AGENTS.md and CLAUDE.md in their directory. Files are read only through the sandbox, so local and remote sandboxes behave alike and nothing on the host is read. Inside a linked worktree nested in its main checkout, the main checkout's same-named file is skipped so it is not applied twice. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * test(harness): instruction discovery across host, Daytona and Vercel sandboxes One scenario runs on the host sandbox (every vp test) and on Daytona and Vercel (test:sandbox:e2e), all against a shared artifact: nested files, overrides, a nested git worktree, and a host-dir link whose settings turn the plugin off while host files are never read. A relink case moves one session Daytona -> Vercel -> host, and a real-repository case clones a pinned fixture branch of codeworksh/69th into all three and requires the same context per directory. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * chore: upgrade effect to 4.0.1 and refresh dependencies Move effect and @effect/* from 4.0.0-rc.115 to stable 4.0.1. The rc pin let @effect/platform-node resolve platform-node-shared@4.0.1 on fresh installs, which crashes against effect rc.115. - effect/unstable/* modules are now effect/* - Encoding.encodeHex is Hex.encode from effect/encoding - Schema.isStartsWith is Schema.isStartingWith - serve the RPC server through RpcServer.layerHttp; the shared websocket protocol layer was built twice under 4.0.1 and RPCs over WebSocket hung Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * release: @codeworksh/aikit@0.10.0 * release: @codeworksh/plugin@0.1.0 * release: @codeworksh/harness@0.1.0 * release: @codeworksh/cli@0.1.0 * docs(aikit): release notes for 0.10.0 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(www): drop the CodeWork theme and default to Tokyo Night Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(www): add YouTube video carousel to the watch section Videos show as thumbnails until played, then swap in a youtube-nocookie embed. The rail is full-bleed with neighbours peeking in; arrows and the scroll bar appear once there is more than one video. The lede gets a beating pixel heart. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * chore(www): allow ngrok hosts on the dev server Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(www): spread hero extension tiles to all four corners Tiles now fill the top corners too, keeping the middle band clear. Small screens, where tiles grow and slots are few, fill every slot, and resting tiles are lifted a tenth from dim toward lit so their labels read. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(www): balance hero extension tiles across corners on small screens Small screens have few tile slots, each with a fixed dither threshold, so whole corners went dark. There, tiles stay visible and each corner keeps up to three slots nearest its own corner. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* fix(harness): retry transient provider failures in the loop (#47) * feat(plugin): structured system-prompt sections (#48) * feat(plugin): structured system-prompt sections Replace the single-string PromptRegistry with a foundation (the untagged head) and named sections. Plugins write sections by value -- Section.Rules, Addendum, ProjectContext, Skills, or Section.define(name, { format }) -- so two plugins merge into one <rules> instead of overwriting the prompt. The harness now renders the prompt: foundation, <tools>, the built-in sections, <cwd>, then custom sections in first-write order. Tool promptGuidelines lead <rules>; list sections dedupe into bullets; a format mismatch on one name fails the snapshot. codework.prompt.default is reduced to content, and the tool index now covers the final catalog. * refactor(harness): drop Location.Ref; location is always the mount cwd Ref.directory let a Location point somewhere other than the mount's cwd, where tools actually run. Only a test fixture set it, and Location.layerWith and defaultLayer took a Ref with no callers at all. Remove all three; the fixture mounts the remote sandbox at the repo instead, as the runner mounts at the session directory. * fix(harness): regenerate retry artifacts from the tests The committed artifacts had been reformatted outside the test (tab-indented, mixed with spaces), so the retry snapshots never matched what vitest writes and failed on every run. Regenerate them with `-u`; the recorded content is unchanged, only the bytes now come from the test. `__artifacts__/**` is already excluded from `vp fmt`. * chore(plugin): trim prompt section comments to decisions * feat(plugin): allow hyphens in prompt section names * feat(harness): tweak default plugin prompt & rules to the registered tools --------- * feat(harness): load AGENTS.md and CLAUDE.md into the system prompt (#49) * feat(plugin): structured system-prompt sections Replace the single-string PromptRegistry with a foundation (the untagged head) and named sections. Plugins write sections by value -- Section.Rules, Addendum, ProjectContext, Skills, or Section.define(name, { format }) -- so two plugins merge into one <rules> instead of overwriting the prompt. The harness now renders the prompt: foundation, <tools>, the built-in sections, <cwd>, then custom sections in first-write order. Tool promptGuidelines lead <rules>; list sections dedupe into bullets; a format mismatch on one name fails the snapshot. codework.prompt.default is reduced to content, and the tool index now covers the final catalog. * refactor(harness): drop Location.Ref; location is always the mount cwd Ref.directory let a Location point somewhere other than the mount's cwd, where tools actually run. Only a test fixture set it, and Location.layerWith and defaultLayer took a Ref with no callers at all. Remove all three; the fixture mounts the remote sandbox at the repo instead, as the runner mounts at the session directory. * fix(harness): regenerate retry artifacts from the tests The committed artifacts had been reformatted outside the test (tab-indented, mixed with spaces), so the retry snapshots never matched what vitest writes and failed on every run. Regenerate them with `-u`; the recorded content is unchanged, only the bytes now come from the test. `__artifacts__/**` is already excluded from `vp fmt`. * chore(plugin): trim prompt section comments to decisions * feat(plugin): allow hyphens in prompt section names * feat(harness): tweak default plugin prompt & rules to the registered tools * feat(harness): load AGENTS.md and CLAUDE.md into the system prompt A built-in prompt plugin, codework.prompt.instruction, walks from the sandbox cwd up to the sandbox root and appends the first readable instruction file per directory to <project_context>, closest last. Overrides (AGENTS.override.md, CLAUDE.override.md) beat AGENTS.md and CLAUDE.md in their directory. Files are read only through the sandbox, so local and remote sandboxes behave alike and nothing on the host is read. Inside a linked worktree nested in its main checkout, the main checkout's same-named file is skipped so it is not applied twice. * test(harness): instruction discovery across host, Daytona and Vercel sandboxes One scenario runs on the host sandbox (every vp test) and on Daytona and Vercel (test:sandbox:e2e), all against a shared artifact: nested files, overrides, a nested git worktree, and a host-dir link whose settings turn the plugin off while host files are never read. A relink case moves one session Daytona -> Vercel -> host, and a real-repository case clones a pinned fixture branch of codeworksh/69th into all three and requires the same context per directory. --------- * chore: upgrade effect to 4.0.1 and refresh dependencies Move effect and @effect/* from 4.0.0-rc.115 to stable 4.0.1. The rc pin let @effect/platform-node resolve platform-node-shared@4.0.1 on fresh installs, which crashes against effect rc.115. - effect/unstable/* modules are now effect/* - Encoding.encodeHex is Hex.encode from effect/encoding - Schema.isStartsWith is Schema.isStartingWith - serve the RPC server through RpcServer.layerHttp; the shared websocket protocol layer was built twice under 4.0.1 and RPCs over WebSocket hung * release: @codeworksh/aikit@0.10.0 * release: @codeworksh/plugin@0.1.0 * release: @codeworksh/harness@0.1.0 * release: @codeworksh/cli@0.1.0 * docs(aikit): release notes for 0.10.0 * fix(www): drop the CodeWork theme and default to Tokyo Night * feat(www): add YouTube video carousel to the watch section Videos show as thumbnails until played, then swap in a youtube-nocookie embed. The rail is full-bleed with neighbours peeking in; arrows and the scroll bar appear once there is more than one video. The lede gets a beating pixel heart. * chore(www): allow ngrok hosts on the dev server * feat(www): spread hero extension tiles to all four corners Tiles now fill the top corners too, keeping the middle band clear. Small screens, where tiles grow and slots are few, fill every slot, and resting tiles are lifted a tenth from dim toward lit so their labels read. * fix(www): balance hero extension tiles across corners on small screens Small screens have few tile slots, each with a fixed dither threshold, so whole corners went dark. There, tiles stay visible and each corner keeps up to three slots nearest its own corner. --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* fix(harness): retry transient provider failures in the loop (#47) * feat(plugin): structured system-prompt sections (#48) * feat(plugin): structured system-prompt sections Replace the single-string PromptRegistry with a foundation (the untagged head) and named sections. Plugins write sections by value -- Section.Rules, Addendum, ProjectContext, Skills, or Section.define(name, { format }) -- so two plugins merge into one <rules> instead of overwriting the prompt. The harness now renders the prompt: foundation, <tools>, the built-in sections, <cwd>, then custom sections in first-write order. Tool promptGuidelines lead <rules>; list sections dedupe into bullets; a format mismatch on one name fails the snapshot. codework.prompt.default is reduced to content, and the tool index now covers the final catalog. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * refactor(harness): drop Location.Ref; location is always the mount cwd Ref.directory let a Location point somewhere other than the mount's cwd, where tools actually run. Only a test fixture set it, and Location.layerWith and defaultLayer took a Ref with no callers at all. Remove all three; the fixture mounts the remote sandbox at the repo instead, as the runner mounts at the session directory. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(harness): regenerate retry artifacts from the tests The committed artifacts had been reformatted outside the test (tab-indented, mixed with spaces), so the retry snapshots never matched what vitest writes and failed on every run. Regenerate them with `-u`; the recorded content is unchanged, only the bytes now come from the test. `__artifacts__/**` is already excluded from `vp fmt`. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * chore(plugin): trim prompt section comments to decisions Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(plugin): allow hyphens in prompt section names Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(harness): tweak default plugin prompt & rules to the registered tools Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * feat(harness): load AGENTS.md and CLAUDE.md into the system prompt (#49) * feat(plugin): structured system-prompt sections Replace the single-string PromptRegistry with a foundation (the untagged head) and named sections. Plugins write sections by value -- Section.Rules, Addendum, ProjectContext, Skills, or Section.define(name, { format }) -- so two plugins merge into one <rules> instead of overwriting the prompt. The harness now renders the prompt: foundation, <tools>, the built-in sections, <cwd>, then custom sections in first-write order. Tool promptGuidelines lead <rules>; list sections dedupe into bullets; a format mismatch on one name fails the snapshot. codework.prompt.default is reduced to content, and the tool index now covers the final catalog. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * refactor(harness): drop Location.Ref; location is always the mount cwd Ref.directory let a Location point somewhere other than the mount's cwd, where tools actually run. Only a test fixture set it, and Location.layerWith and defaultLayer took a Ref with no callers at all. Remove all three; the fixture mounts the remote sandbox at the repo instead, as the runner mounts at the session directory. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(harness): regenerate retry artifacts from the tests The committed artifacts had been reformatted outside the test (tab-indented, mixed with spaces), so the retry snapshots never matched what vitest writes and failed on every run. Regenerate them with `-u`; the recorded content is unchanged, only the bytes now come from the test. `__artifacts__/**` is already excluded from `vp fmt`. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * chore(plugin): trim prompt section comments to decisions Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(plugin): allow hyphens in prompt section names Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(harness): tweak default plugin prompt & rules to the registered tools Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(harness): load AGENTS.md and CLAUDE.md into the system prompt A built-in prompt plugin, codework.prompt.instruction, walks from the sandbox cwd up to the sandbox root and appends the first readable instruction file per directory to <project_context>, closest last. Overrides (AGENTS.override.md, CLAUDE.override.md) beat AGENTS.md and CLAUDE.md in their directory. Files are read only through the sandbox, so local and remote sandboxes behave alike and nothing on the host is read. Inside a linked worktree nested in its main checkout, the main checkout's same-named file is skipped so it is not applied twice. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * test(harness): instruction discovery across host, Daytona and Vercel sandboxes One scenario runs on the host sandbox (every vp test) and on Daytona and Vercel (test:sandbox:e2e), all against a shared artifact: nested files, overrides, a nested git worktree, and a host-dir link whose settings turn the plugin off while host files are never read. A relink case moves one session Daytona -> Vercel -> host, and a real-repository case clones a pinned fixture branch of codeworksh/69th into all three and requires the same context per directory. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> * chore: upgrade effect to 4.0.1 and refresh dependencies Move effect and @effect/* from 4.0.0-rc.115 to stable 4.0.1. The rc pin let @effect/platform-node resolve platform-node-shared@4.0.1 on fresh installs, which crashes against effect rc.115. - effect/unstable/* modules are now effect/* - Encoding.encodeHex is Hex.encode from effect/encoding - Schema.isStartsWith is Schema.isStartingWith - serve the RPC server through RpcServer.layerHttp; the shared websocket protocol layer was built twice under 4.0.1 and RPCs over WebSocket hung Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * release: @codeworksh/aikit@0.10.0 * release: @codeworksh/plugin@0.1.0 * release: @codeworksh/harness@0.1.0 * release: @codeworksh/cli@0.1.0 * docs(aikit): release notes for 0.10.0 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(www): drop the CodeWork theme and default to Tokyo Night Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(www): add YouTube video carousel to the watch section Videos show as thumbnails until played, then swap in a youtube-nocookie embed. The rail is full-bleed with neighbours peeking in; arrows and the scroll bar appear once there is more than one video. The lede gets a beating pixel heart. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * chore(www): allow ngrok hosts on the dev server Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(www): spread hero extension tiles to all four corners Tiles now fill the top corners too, keeping the middle band clear. Small screens, where tiles grow and slots are few, fill every slot, and resting tiles are lifted a tenth from dim toward lit so their labels read. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(www): balance hero extension tiles across corners on small screens Small screens have few tile slots, each with a fixed dither threshold, so whole corners went dark. There, tiles stay visible and each corner keeps up to three slots nearest its own corner. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(www): add theme-aware C and cursor logo * feat(www): add brand SVG and browser icon fallbacks * fix(site): use logo wordmark * fixes * feat(www): add solid pixel wordmark with outline depth * fix(www): restore corner tile sizing and animation * fix(www): simplify hero C and animate mobile extension tiles * feat(www): add responsive logo chase for extension tiles * feat(www): add subtle grid behind hero pixel effects * feat(www): simplify hero background to grid and extension tiles --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

sanchitrk force-pushed the feat-prompt-sections branch from a8baa75 to 9b9dbca Compare October 6, 2026 20:19

sanchitrk changed the title ~~feat(plugin): structured system-prompt sections (COD-150)~~ feat(plugin): structured system-prompt sections Oct 6, 2026

sanchitrk and others added 3 commits October 7, 2026 01:56

sanchitrk force-pushed the feat-prompt-sections branch from 9b9dbca to 0f47e95 Compare October 6, 2026 20:27

chore(plugin): trim prompt section comments to decisions

745b1a8

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

sanchitrk mentioned this pull request Oct 6, 2026

feat(harness): load AGENTS.md and CLAUDE.md into the system prompt #49

Merged

sanchitrk and others added 2 commits October 7, 2026 14:20

feat(plugin): allow hyphens in prompt section names

2b55bfc

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

feat(harness): tweak default plugin prompt & rules to the registered …

e03303a

…tools Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

sanchitrk merged commit bddb812 into canary Oct 7, 2026

sanchitrk deleted the feat-prompt-sections branch October 7, 2026 12:13

sanchitrk mentioned this pull request Oct 8, 2026

canary (#50) #51

Merged

Navigation Menu

Repository navigation

feat(plugin): structured system-prompt sections - #48

feat(plugin): structured system-prompt sections#48
sanchitrk merged 6 commits into
canaryfrom
feat-prompt-sections

sanchitrk commented Oct 6, 2026 •

edited

Loading

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

sanchitrk commented Oct 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What

Behaviour changes

Also in this PR

Verification

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

sanchitrk commented Oct 6, 2026 •

edited

Loading