AGENTS.md loaded about 33,000 tokens into every agent session, subagent and reviewer, mostly per-feature contracts, and the agent documentation around it had accumulated duplicates, dated logs and history of removed designs. This PR makes AGENTS.md a short set of rules that apply everywhere plus an index of area documents, and leaves AGENTS.md, dev-docs and the repository skills describing only the current state.
Three commits, reviewable separately:
1. **Move (lossless).** Every per-area rule moved verbatim from AGENTS.md into its area document (existing ones, or six new ones: UI_CONVENTIONS, UI_SHELL, COMPOSER, TRANSCRIPT, SETTINGS, PROJECTS). A script split the old file into 208 units, required one destination for each and checked all 199 content units verbatim at their destination. AGENTS.md went from 134,000 to 16,000 characters.
2. **Trim.** A per-area ponytail review removed rules stated two or three times (keeping the more precise copy), dated validation logs, per-version qualification sections below the current minimum, abandoned designs and self-references; seven documents whose surviving rules live elsewhere were deleted. It also fixed statements the code contradicts (locale registration, the Mission play/recovery condition, plugin package updates, cleanup progress UI, timeline index persistence, automation restart/targets, stale pins and minimums).
3. **Current state only.** The new AGENTS.md rule: no history (removed features and their names, former behavior, provenance, dated logs) and no denials; correct a contradicted statement at its source. DEVELOPER_MODE.md is merged into BROWSER_AUTOMATION.md and every Developer Mode mention is gone (including the README and the browser help page in ten locales). Names that only made sense against older designs are renamed with every reference updated, MISSIONS_ENVIRONMENT_ADMISSION is folded into SESSION_ENVIRONMENT, and INDEX lists every document. The server npm README now states the real OpenCode minimum (2.1.2).
AGENTS.md is now about 14,000 characters (roughly 3,500 tokens); AGENTS.md, dev-docs and the vendor README went from about 11,200 lines to under 4,700.
Validation: every relative Markdown link and every cited dev-docs path resolves; no reference to a renamed or deleted document remains; `npm run test:help` passes for all ten locales.
Review: three independent GPT-6.1 Sol (high) reviews by area found binding rules lost in the trim and current-state passes; 25 were restored as present-tense facts or prohibitions and four contradicting statements corrected, each checked against code. A Fable 5.1 gatekeeper review of the whole PR passed; its follow-up routes session recovery from the index, points to dev-docs/INDEX.md and documents the family identity read.
I put Extensions below a separator in the customization popup, collapsed
by default. I combine installed addons and the catalogue in one compact
searchable list and remove the separate modes and floating window.
I make the existing addon checkbox the only profile-wide activation
control. Disabled installed addons stay listed; reset never grants
consent. Explicit trust, digest-bound replacement/removal, offline ZIP
installation and sandbox protections remain intact.
I pass the UI typecheck, 9 server tests and 14 real-component
extension/Status browser tests, including narrow translated/RTL layouts
and real ZIP gestures. I inspect the interactive preview validated by
the user. Windows Tauri release qualification and the final independent
gatekeeper review are in progress; I do not replace the installed app or
stop the shared daemon.
I restore right-panel customization to one compact Extensions action and move catalogue, ZIP installation, activation and trust/removal confirmations into a separate non-modal utility window. Lists stay single-line, metadata is available on hover and keyboard focus, toolbar actions remain reachable at narrow translated widths, and disposal fences late results.\n\nI remove speculative per-folder activation from the UI, client contract and backend. Each installed addon has one general activation control per CodeNomad profile. Existing packages and general grants survive; legacy folder-only consent is never broadened, and obsolete grants are removed only during an explicit successful mutation. Trust, digest-bound replacement/removal and owned-instance executable-panel reads remain enforced.\n\nI publish independent gatekeeper iterations until zero findings on the exact final head. Both typechecks, nine server tests and thirteen real-component extension/Status tests pass, including migration and French/Hebrew/Japanese/Spanish narrow layouts. Windows Tauri/WebView2 command denial passes. I use the requested admin override with the documented CI caveat: four unrelated browser failures also occurred in the final pre-correction run. The installed desktop and shared OpenCode daemon are unchanged.
## What changes
I add independently distributed right-panel UI extensions and an
official online catalogue in Customize right panel. Users browse/search
available addons, choose one and confirm installation directly from its
exact GitHub Release, without a manual ZIP download, rebuilding
CodeNomad or restarting OpenCode. Manual ZIP installation remains
available.
I publish the curated index, contribution rules and Session info demo at
https://github.com/NeuralNomadsAI/CodeNomad-Extensions. That repository
can host several official addons; the index can also reference
author-owned repositories. Separate means independent of the main
application repository, not a mandatory repository per addon.
I show the manifest and exact archive SHA-256 before trust confirmation.
Installation starts disabled; users enable it for all projects in the
server profile or for the exact opened folder. Replacement requires the
previous digest and revokes every activation grant. Removal is explicit.
Packages and grants stay outside project/OpenCode storage.
I keep author code out of the main renderer. API 1 exposes only session
ID, locale and appearance through a per-mount MessageChannel. It does
not expose filesystem, transcript, images, credentials, native commands
or the OpenCode proxy. Frames are sandboxed without same-origin access
and receive a restrictive CSP; obsolete responses and ports are fenced
on project/session changes, disconnects, revocation and replacement.
I document the package, permissions, API compatibility, scope,
release/checksum and maintainer rules in `dev-docs/PANEL_EXTENSIONS.md`,
with a standalone starter in `examples/panel-extension/`. GitHub's
source ZIP is not an installable package. There is no arbitrary URL
installer or automatic update.
I load catalogue metadata only during discovery. Inspection and
installation each revalidate a fresh index, exact ZIP digest and full
manifest; unsupported APIs/permissions cannot install. Downloads are
bounded, deadline-limited, credential-free HTTPS requests to GitHub's
exact release hosts, with constrained redirects. A withdrawn or changed
selection cannot become a different trusted package. Offline catalogue
failures leave installed addons and ZIP installation available.
## Scope and limits
I keep #801's assets gallery for a separate follow-up. It needs a
bounded, authorized image/history capability rather than access to
internal stores. This PR only establishes external distribution and a
minimal context API.
Windows Tauri may inject native bridge objects into subframes: object
presence is not a grant. My isolated WebView2 fixture proves parent
command execution while direct/raw child calls do not reach it, and
verifies context delivery plus DOM/network restrictions. A trusted
hostile extension can still stall its renderer or attempt
self-navigation; this is not process isolation or a universal offline
guarantee. The installer warns users to trust the author.
I have not run a native Electron/macOS/Linux qualification or changed
the shared daemon. The runtime implementation is shared between desktop
hosts; native GUI validation here is Tauri only.
## Validation
- UI and server typechecks.
- 14 focused server/registry/manifest tests.
- Real right-panel browser tests for ZIP consent, activation scopes,
replacement, revocation, context changes, stale reads and frame
isolation.
- Five catalogue/server checks and seven right-panel browser scenarios
cover metadata-only browse/search, online consent, withdrawal,
incompatibility, offline ZIP fallback, hashes/manifests, redirect
restrictions and size budgets.
- Live public GitHub index and Release download matched SHA-256
`cf80226a40dd077a6a8cc2a4b65bbb46bf33619d62007982bd7dcea34ae4545f`;
installation and activation used a disposable isolated store.
- `node scripts/test-panel-extension-native.mjs` on Windows, using the
application's locked Tauri dependencies and an isolated browser profile:
parent native call allowed; child native calls, parent DOM and network
APIs blocked; session context delivered.
- UI build, Windows Tauri release build (`--no-bundle`) and
packaged-resource smoke passed. Full browser suite results are recorded
in the follow-up validation comment.
Existing oversized files touched only for integration:
`packages/server/src/api-types.ts` (~735 lines),
`packages/server/src/index.ts` (~713),
`packages/server/src/server/http-server.ts` (~2,438).
