Introducing a new way to review your PRs.

Read the announcement

thedotmack / claude-mem

Persistent Context Across Sessions for Every Agent – Captures everything your agent does during sessions, compresses it with AI, and injects relevant context back into future sessions. Works with Claude Code, OpenClaw, Codex, Gemini, Hermes, Copilot, OpenCode + More

View on GitHub

47 documents · Updated Nov 20, 2026

docs / plugin-hooks-and-agent-modes

Plugin hooks and agent modes

Claude-Mem connects supported agent products through packaged manifests, lifecycle hooks, skills, modes, MCP tools, and embedded views. These contracts let each host capture activity and restore useful context while the shared worker remains responsible for memory storage and recall. For the wider packaging model, see Agent integrations and distribution; for installation and repair, see Installation and NPX CLI; for host-specific behavior, see Cursor integration, Grok bot integration, and Cowork integration.

Integration boundary

The plugin boundary separates host-specific events from Claude-Mem features. Each host supplies its own event names and metadata, while the integration maps those events to session setup, context delivery, observation capture, summarization, and memory search.

Packaged manifests

The manifests tell an agent product what Claude-Mem is and which packaged resources it can load. The Claude Code manifest carries product identity, while the Codex manifest also declares skills, the MCP server, hooks, and user-facing interface data. Both packaged manifests identify release 13.25.1 (plugin/…/plugin.json• L1-24, plugin/…/plugin.json• L1-46).

The Codex interface metadata gives the host a product label, a short and long description, supported capability labels, default prompts, and brand data. These fields shape how the integration appears to users without changing the worker's memory contract (plugin/…/plugin.json• L28-45).

The package metadata is the source for shared identity fields. scripts/sync-plugin-manifests.js keeps names, versions, descriptions, links, license data, keywords, and author data aligned across the manifests (scripts/sync-plugin-manifests.js• L10-14, scripts/sync-plugin-manifests.js• L24-63). This prevents one host from advertising a different product version or origin from another.

The packaged Codex manifest points to three important contracts:

Manifest resourceProduct role
skillsMakes user-facing workflows available to the agent.
mcpServersProvides recall and related memory tools through MCP.
hooksConnects host lifecycle events to the memory worker.

References: plugin/…/plugin.json• L25-45 .codex-plugin/plugin.json• L25-45 scripts/sync-plugin-manifests.js• L79-98

Distribution build gate

The distribution build turns the shared services and host adapters into packaged launchers, then checks that the files needed by supported products are present. It checks the canonical hook commands, including shell-free Claude Code runtime entries, the Claude Code launcher self-heal, supported Codex events, and the marketplace path to the bundled plugin root (scripts/build-hooks.js• L101-246, scripts/build-hooks.js• L287-433, scripts/build-hooks.js• L1006-1074).

The build also prepares the Pi memory extension and DeepSeek Harness memory plugin. Pi receives its required attribution files, while the DeepSeek Harness package must carry the current project license and notice before either adapter is shipped (scripts/build-hooks.js• L985-994, scripts/harness-plugin-build-options.js• L18-55). The distribution file check also keeps the Qoder manifests, hook launchers, MCP configuration, and other host entry points in the package (scripts/build-hooks.js• L1016-1041).

The same build emits the OpenClaw and OpenCode host plugins when their source entry points are available. This keeps those adapters in the release flow with the hook, MCP, and UI resources instead of treating them as separate source-only integrations (scripts/build-hooks.js• L1065-1103).

The build checks each bundled service before it ships. A bundle must parse and must not include an absolute path from the build machine. This protects installed hooks and MCP services from build-time path leaks or broken generated code (scripts/build-hooks.js• L47-73, scripts/build-hooks.js• L529-747, scripts/build-hooks.js• L776-887).

References: scripts/build-hooks.js• L47-73 scripts/build-hooks.js• L101-246 scripts/build-hooks.js• L287-433 scripts/build-hooks.js• L985-1103 scripts/build-hooks.js• L1126-1197 scripts/harness-plugin-build-options.js• L18-61

Host event bridge

Hooks are the event bridge between an agent product and the memory lifecycle. A host event starts or resumes a session, submits a prompt, records useful tool activity, supplies file context, or asks for a summary; the hook then forwards the event to the shared worker without making the host depend on the full memory process.

The Claude Code and Codex hook sets use parallel lifecycle concepts but follow each product's accepted event contract. Claude Code includes setup, session start, prompt submit, pre- and post-tool events, failed tool events, stop, and session end; its runtime hooks use direct executable entries, while startup can repair a missing launcher. Codex uses its supported session, prompt, pre-tool, post-tool, stop, and compaction events, with command forms for Windows (plugin/…/hooks.json• L3-89, plugin/…/codex-hooks.json• L1-68). The build verifies these host files against the canonical command generator and the Codex event allowlist (scripts/build-hooks.js• L183-246, scripts/build-hooks.js• L287-433, scripts/build-hooks.js• L1006-1057).

This flow shows the shared purpose of the host adapters.

flowchart LR
    A[Agent product event] --> B[Host hook or plugin adapter]
    B --> C[Session and prompt contract]
    B --> D[Observation contract]
    B --> E[Summary contract]
    C --> F[Claude-Mem worker]
    D --> F
    E --> F
    F --> G[Stored memory]
    F --> H[Context and search]

The important rule is that hooks capture or request work at lifecycle boundaries. They do not become a second agent. The worker remains responsible for turning captured activity into durable memory and for returning context to later sessions (plugin/…/hooks.json• L3-89, plugin/…/codex-hooks.json• L1-68).

References: scripts/build-hooks.js• L101-246 scripts/build-hooks.js• L287-433 scripts/build-hooks.js• L1006-1057 plugin/…/hooks.json• L3-89 plugin/…/codex-hooks.json• L1-68

Shared launch and recall contracts

The packaged hooks and MCP server must find the same installed plugin resources even when a host does not provide a plugin-root variable. Their launchers resolve an available packaged copy and then start the worker-facing command or MCP server. Claude Code can also repair its on-PATH launcher during startup, while the MCP launcher keeps Codex cache fallbacks for hosts that do not inject PLUGIN_ROOT (plugin/…/hooks.json• L9-24, plugin/…/codex-hooks.json• L3-12, plugin/.mcp.json• L2-11, scripts/build-hooks.js• L199-209, scripts/build-hooks.js• L1064-1070).

The MCP side exposes memory search as a host-neutral service. Its recall contract uses a staged path: search returns a compact index, timeline adds nearby context, and observation fetch returns full detail only for selected results. The same service also supports queryable knowledge corpora and code-structure lookup, so an MCP-capable product can use Claude-Mem without knowing how the worker or database is arranged (plugin/…/mcp-server.cjs• L214-238).

References: plugin/.mcp.json• L2-11 plugin/…/mcp-server.cjs• L214-238 scripts/build-hooks.js• L126-176 scripts/build-hooks.js• L199-209 scripts/build-hooks.js• L1064-1070 src/…/index.ts• L276-340

Agent-facing capabilities

The package adds capabilities around the agent rather than replacing the agent. Skills guide explicit workflows, modes shape what the memory system records, MCP provides recall, and embedded views make the resulting memory visible.

Skills

Skills are packaged instruction bundles that give an agent named workflows. They let users ask for a memory search, a codebase learning pass, a plan, a mode change, a report, or an operational action without learning the worker's internal interfaces.

The most important skill groups are:

GroupPurpose
Memory recallSearch prior observations or query a focused knowledge corpus.
Work guidanceExplain the system, learn a codebase, make a plan, or carry out a plan.
Memory customizationCreate a domain mode with its own note types and search tags.
Operations and reportingConfigure sync, prepare digests, produce project reports, or align a worker-watch cache.
Release operationsCoordinate version metadata, build validation, release records, and the final human npm handoff.

The skill contract is descriptive. A skill tells the agent when to use a workflow and what safe sequence to follow; the underlying worker, CLI, or helper performs the action. The memory-search skill uses progressive disclosure so the agent filters compact search results before it fetches full observations or raw tool output (plugin/…/SKILL.md• L6-20, plugin/…/SKILL.md• L51-120). The CCS Align skill is an operational workflow that keeps a seat-owned context cache aligned with the local worker and reports rule conflicts; it does not replace the worker or change the source diary (plugin/…/SKILL.md• L6-18). The timeline-report skill turns a project's complete worker timeline into a narrative report and requires a healthy worker plus a resolved project name (plugin/…/SKILL.md• L6-8, plugin/…/SKILL.md• L21-31). The release skill is a special case because it coordinates package-wide version state and build output, then stops for the maintainer's publish step (plugin/…/SKILL.md• L6-12, plugin/…/SKILL.md• L28-59).

References: plugin/…/SKILL.md• L6-20 plugin/…/SKILL.md• L75-120 plugin/…/SKILL.md• L6-18 plugin/…/SKILL.md• L6-8 plugin/…/SKILL.md• L21-64 plugin/…/SKILL.md• L16-32

Agent modes

An agent mode is a memory profile. It defines the kinds of observations worth keeping, the reusable concepts used to find them later, and the language and role used when Claude-Mem creates observations and summaries (docs/…/modes.mdx• L8-16). The bundled profiles cover software work, email fraud investigation, legal study, and Solana meme-token activity. The code profile also includes security alerts, security notes, and sensitive information as distinct record types (plugin/…/code.json• L1-68, plugin/…/email-investigation.json• L1-47, plugin/…/law-study.json• L1-47, plugin/…/meme-tokens.json• L1-47).

Modes use inheritance so a localized or focused mode can change a part of the profile without copying the full base definition. An ID such as code--es means that the base code profile is loaded first and the language-specific profile overrides selected prompts. The same pattern supports custom user modes such as code--architecture-practice (docs/…/modes.mdx• L40-50, docs/…/modes.mdx• L67-100, docs/…/modes.mdx• L115-122).

Mode authoring keeps the contract small and predictable. Observation types describe what a note is; concepts describe why an engineer may search for it later. Custom modes use a safe inherited ID, replace the parent taxonomies, merge prompt overrides, and must provide strict JSON with useful skip rules (plugin/…/mode-authoring.md• L5-19, plugin/…/mode-authoring.md• L78-97).

The mode creator can also configure optional Telegram alerts for selected observation types or concepts. Alerts are off until the user explicitly enables them, and the secure setup flow keeps the bot token out of chat, command arguments, logs, and normal output (plugin/…/SKILL.md• L64-77, plugin/…/SKILL.md• L112-126, plugin/…/telegram.md• L7-18).

Mode selection affects future generated memory. Existing observations keep their original types, and custom modes are stored in the user's Claude-Mem data directory so a plugin upgrade does not remove them (plugin/…/SKILL.md• L79-110, plugin/…/SKILL.md• L128-148).

References: docs/…/modes.mdx• L8-16 docs/…/modes.mdx• L40-50 docs/…/modes.mdx• L115-122 plugin/…/code.json• L1-68 plugin/…/email-investigation.json• L1-47 plugin/…/law-study.json• L1-47 plugin/…/meme-tokens.json• L1-47 plugin/…/code--es.json• L1-23 plugin/…/SKILL.md• L64-77 plugin/…/SKILL.md• L128-148 plugin/…/telegram.md• L7-18

Embedded viewer

The embedded viewer is the visual face of the packaged memory system. It shows live observations, summaries, and prompts from the worker, lets the user filter by project, and provides context settings, logs, help, and theme controls (src/…/App.tsx• L14-31, src/…/App.tsx• L94-147).

The viewer is shipped as a self-contained UI asset set. The package includes the HTML shell and the generated browser bundle with its supporting icons, logos, and fonts; the build creates the viewer before validating the distribution files (plugin/…/viewer.html• L1-16, plugin/…/viewer-bundle.js• L1-16, scripts/build-hooks.js• L516-527, scripts/build-hooks.js• L1016-1041, package.json• L35-51). The UI uses a live event stream and reconnects when the stream fails, so the display can follow newly created memory without a page refresh (src/…/useSSE.ts• L20-45, src/…/useSSE.ts• L47-89).

The generated bundle is the browser payload that users receive. It must remain aligned with the viewer shell and its supporting package assets so the visual surface stays usable after distribution (plugin/…/viewer-bundle.js• L1-16, plugin/…/viewer-bundle.js• L17-65).

Observation TV

Observation TV is a second embedded display for a passive, live view of memory activity. It replays a small recent backlog when the page opens, then gives new observations priority as they arrive. Project and platform filters keep the display focused, while full-screen and picture-in-picture views support a separate monitor or small always-visible window (plugin/…/tv.html• L145-175, plugin/…/tv.html• L292-337, plugin/…/tv.html• L463-501).

The TV view uses the same worker stream as the memory surface but presents only a compact observation title and its basic context. A shared access token can protect non-loopback requests, and a reconnect loop lets the display recover when the stream breaks (plugin/…/tv.html• L145-154, plugin/…/tv.html• L310-337).

References: src/…/App.tsx• L14-31 src/…/App.tsx• L94-147 src/…/useSSE.ts• L20-45 src/…/useSSE.ts• L47-112 plugin/…/viewer.html• L1-16 plugin/…/viewer-bundle.js• L1-16 plugin/…/viewer-bundle.js• L17-65 plugin/…/tv.html• L145-175 plugin/…/tv.html• L292-337 scripts/build-hooks.js• L516-527 scripts/build-hooks.js• L1006-1041

Product adapters

Each supported product preserves the same memory goals but adapts to its own event model and user interface. The adapter owns host details such as event names, session identifiers, context insertion, and message delivery; the worker-facing paths preserve the common Claude-Mem meaning.

Pi and DeepSeek Harness

Pi and DeepSeek Harness receive dedicated bundled memory plugins during the distribution build. Their outputs are separate from the hook and MCP files, but they remain part of the release contract because the build must package their adapters and preserve the required license and attribution material (scripts/build-hooks.js• L985-994, scripts/harness-plugin-build-options.js• L10-16, scripts/harness-plugin-build-options.js• L46-61).

References: scripts/build-hooks.js• L985-994 scripts/harness-plugin-build-options.js• L10-61

Claude Code and Codex

Claude Code receives prior context at session boundaries and sends prompt and tool activity into the memory lifecycle. Codex follows the same broad pattern through its own hook file and adds Windows command forms in the manifest, which keeps the packaged plugin usable across its supported shells (plugin/…/hooks.json• L17-89, plugin/…/codex-hooks.json• L1-68).

The two products share the same packaged skills, modes, MCP recall service, worker, and memory store. Their difference is the host-side envelope: each product must receive the event names, matcher rules, command shape, and output behavior that its own plugin system accepts (plugin/…/plugin.json• L25-27, .codex-plugin/plugin.json• L25-27, scripts/build-hooks.js• L156-246).

References: plugin/…/hooks.json• L17-89 plugin/…/codex-hooks.json• L1-68 plugin/…/plugin.json• L25-27 .codex-plugin/plugin.json• L25-27 scripts/build-hooks.js• L156-246

OpenClaw

OpenClaw uses a native plugin rather than the Claude Code hook file. It maps OpenClaw session, prompt, tool-result, compaction, and agent-end events to the same worker concepts. It can inject context into an agent system prompt, record tool results, and request summaries while tracking separate agent projects (openclaw/…/index.ts• L620-795, openclaw/…/index.ts• L797-875).

OpenClaw also adds product-specific surfaces. Users can search memory, view recent context, inspect a timeline, check worker status, and optionally stream new observations to a messaging channel. The manifest defines the related settings, including context exclusions, worker location, project scope, and observation-feed targets (openclaw/openclaw.plugin.json• L10-97, openclaw/…/index.ts• L882-1135).

The observation feed uses the worker's live stream and maps supported channel names to the host's channel send functions. This keeps message delivery in OpenClaw while Claude-Mem remains the source of the observation event (openclaw/…/index.ts• L451-524, openclaw/…/index.ts• L526-618).

References: openclaw/openclaw.plugin.json• L1-97 openclaw/…/index.ts• L620-795 openclaw/…/index.ts• L797-875 openclaw/…/index.ts• L882-1135

OpenCode

OpenCode supports two host plugin contracts through one packaged entry point. The older supported contract reads a server callback, while the newer contract reads an identified setup callback; both use the same Claude-Mem behavior instead of maintaining separate memory rules (src/…/index.ts• L1-37).

The shared OpenCode core maps host activity to the worker-facing memory contract. It records tool use, user prompts, and assistant messages, then requests a summary when the session becomes idle or is compacted. It also supplies cached session context and clears that context when compaction requires a fresh view (src/…/core.ts• L278-315, src/…/core.ts• L347-419).

OpenCode also exposes a claude_mem_search tool and related staged recall operations. OpenCode supplies a query or selected observation IDs, while Claude-Mem returns readable search, timeline, or full-observation results. The core marks requests as OpenCode activity and respects the shared setting that can include memory from all supported agent sources (src/…/core.ts• L47-85, src/…/core.ts• L317-345, src/…/core.ts• L421-450).

References: src/…/index.ts• L1-37 src/…/core.ts• L1-10 src/…/core.ts• L278-315 src/…/core.ts• L317-450 scripts/build-hooks.js• L1090-1103

Change checklist

When these files or sections are changed, remember to consider these:

  • Keep the root and bundled Claude Code and Codex manifests aligned with package.json; the sync script updates shared identity fields but preserves host-specific resource declarations.
  • When a hook event or command changes, update both the host manifest and the canonical shell-template map in scripts/build-hooks.js so generated launchers and committed JSON do not diverge; keep Claude Code runtime entries in exec form, the launcher self-heal entry valid, and the Codex event allowlist and Windows command form supported.
  • When an OpenCode hook or host contract changes, keep the shared core behavior aligned with both the V1 server entry and the V2 setup entry, then keep the packaged OpenCode build output in the distribution flow.
  • When the MCP search contract changes, keep the staged search, timeline, and full-fetch flow compatible with MCP clients and with adapters that return rendered results.
  • When adding a mode or changing a mode parent, check inheritance, required prompt fields, localized output, taxonomy size, and the user-mode installation path; bundled files and user overrides have different upgrade behavior.
  • When changing mode alerts, preserve explicit opt-in, lowercase type and concept matching, secure token handling, and existing Telegram triggers.
  • When a worker-facing request or response changes, review every adapter that sends it: Claude Code, Codex, OpenClaw, and OpenCode must continue to identify the same session and project.
  • When changing the viewer build or its assets, keep the HTML shell, generated bundle, fonts, icons, TV page, and package file list together so every distributed UI remains complete.
  • When changing a skill's workflow, keep its manifest exposure and its worker-facing commands consistent; a skill can guide an action without implementing that action.
  • When changing timeline-report, preserve worker-port resolution, worktree-to-parent project mapping, full-timeline retrieval, and the confirmation step for large timelines.
  • When changing the release skill or version carriers, keep all package manifests and generated artifacts in the build-and-sync path before the human npm publish handoff.
  • When changing OpenClaw feed settings, keep the manifest schema, channel mapping, live-stream event shape, and missing-channel behavior consistent.
  • When changing a packaged service or launcher, preserve the build gate's parse check, build-path leak check, required-file check, and host-command verification before distribution.
  • When changing the Pi or DeepSeek Harness adapter, keep the build output, version carrier, and required license, notice, and attribution files in sync with the host package contract.

Important failure modes

TriggerConsequenceGuard
A manifest keeps an old version or product identityA host can advertise stale metadata or load a package that does not match the worker resources.Treat package.json as the shared identity source and run the manifest sync path.
A hook uses an event name or matcher that the host does not supportThe host starts normally but does not capture activity or inject context.Keep event names in the host-specific hook contract and the canonical build manifest.
A committed launcher or hook entry differs from the canonical templateThe package can contain a stale command, timeout, matcher, or Windows launcher even though the source rule changed.Regenerate the host files from scripts/build-hooks.js after an intentional template change and pass the canonical-template check.
A Claude Code runtime hook uses a shell entry or an unapproved command formClaude Code can show a console flash, lose output handling, or fail to run the memory action.Keep runtime hooks as command-plus-args entries and allow only the canonical launcher or startup self-heal command.
A Codex hook file uses an unsupported root key or eventCodex may reject the hook file or silently lose part of the lifecycle bridge.Keep the Codex root and event allowlists in the distribution validation path.
An OpenCode package exposes only one incompatible host contractThe plugin fails to load on either the supported V1 or V2 host.Keep one default entry that supplies the supported server and setup shapes.
A launcher cannot resolve the installed plugin rootHooks or MCP recall fail before they reach the worker.Preserve the packaged-root and cache fallback resolution in the generated launchers.
OpenCode omits its platform source or applies the wrong source scopeContext and stored activity can appear under the wrong agent product or disappear from a session.Preserve the OpenCode source marker and the shared all-sources setting when building worker requests.
An MCP client skips the search-to-fetch contractThe agent spends excess context on unrelated memory or receives incomplete context.Use search to select IDs, timeline to add local context, and full fetch only for selected observations.
A mode override omits the parent-compatible prompt or taxonomy contractNew observations can fall back to another mode or fail validation instead of using the intended profile.Follow the mode authoring shape, validate the merged mode, and verify the active mode in startup context.
Telegram triggers are set without explicit observation-alert enablementMatching observations do not send alerts, or a user may misunderstand the privacy boundary.Require explicit opt-in and set the alert switch only after consent.
A Telegram token is passed through chat, logs, or command argumentsThe bot can be taken over by anyone who sees the token.Use hidden terminal input and report only whether credentials are present.
A timeline report uses the wrong project name or an unavailable workerThe report is empty or analyzes another project's history.Resolve the worker port, detect worktrees, and verify the worker before fetching the timeline.
OpenClaw has no feed channel or targetThe observation feed does not start even though memory capture still works.Set both observationFeed.channel and observationFeed.to before enabling the feed.
The viewer bundle, TV page, or a referenced asset is absent from the packageThe worker can run but a visual memory surface is incomplete or cannot load.Keep all UI assets in the package file list and distribution-file validation.
A generated service bundle contains a syntax error or build-machine pathAn installed hook, worker, or MCP service fails at startup or exposes the builder's filesystem layout.Require bundle parsing and absolute-path checks before the package is shipped.
A remote TV stream lacks its access tokenThe display cannot load backlog or live observations from a protected endpoint.Use the bookmark token for remote access and keep loopback access token-free.
A release updates one version carrier but not the othersPublished metadata and bundled plugin resources can refer to different releases.Use the release skill's complete version list and verify every carrier before publishing.
The Pi attribution files are missing or the DeepSeek Harness notices do not contain the current project noticesA host package fails its release or license contract even though its adapter bundle builds.Run the attribution preflight and copy the required notice files beside the Pi extension.

Key files

FileWhy to read it
package.jsonShared product metadata and the list of plugin resources included in the published package.
scripts/sync-plugin-manifests.jsShows which manifest fields must stay aligned across package roots and agent products.
scripts/harness-plugin-build-options.jsDefines Pi and DeepSeek Harness build contracts plus license and attribution checks.
plugin/…/plugin.jsonClaude Code package identity.
plugin/…/plugin.jsonCodex package identity and links to skills, MCP, hooks, and interface data.
plugin/…/hooks.jsonClaude Code lifecycle events connected to memory capture and context delivery.
plugin/…/codex-hooks.jsonCodex lifecycle contract and cross-platform command forms.
plugin/.mcp.jsonHost-neutral memory search launcher.
plugin/…/mcp-server.cjsPackaged MCP service for staged memory recall, corpora, and code-aware tools.
scripts/build-hooks.jsBuilds packaged services and UI assets, then checks launcher, hook, bundle, and distribution contracts.
plugin/…/SKILL.mdAgent workflow for progressive memory recall and raw-detail disclosure.
plugin/…/SKILL.mdOperational workflow for worker-watch cache alignment and rule reporting.
plugin/…/SKILL.mdUser workflow for safe custom mode creation, activation, verification, and optional alerts.
plugin/…/mode-authoring.mdRules for mode taxonomies, inheritance, prompt overrides, and validation.
plugin/…/telegram.mdPrivacy, matching, setup, and troubleshooting rules for mode alerts.
plugin/…/SKILL.mdWorkflow for turning a complete project timeline into a report.
plugin/…/SKILL.mdRelease contract for version alignment, build verification, and the human publish handoff.
plugin/…/code.jsonGeneral development memory profile and security-aware record types.
plugin/…/email-investigation.jsonDomain profile for email fraud investigation.
plugin/…/law-study.jsonDomain profile for legal study and exam preparation.
plugin/…/meme-tokens.jsonDomain profile for Solana meme-token activity and trading signals.
docs/…/modes.mdxConceptual model for bundled, inherited, localized, and custom modes.
openclaw/openclaw.plugin.jsonOpenClaw plugin identity and configuration schema.
openclaw/…/index.tsOpenClaw event, context, search, status, and live-feed adapter.
src/…/index.tsOpenCode entry contract shared by supported host plugin versions.
src/…/core.tsOpenCode session, capture, context, summary, and recall behavior.
src/…/App.tsxMain viewer behavior and user-facing memory controls.
plugin/…/viewer.htmlPackaged viewer shell and embedded visual assets.
plugin/…/viewer-bundle.jsGenerated browser payload shipped with the plugin.
plugin/…/tv.htmlStandalone live observation display with replay, filters, and display modes.