[P0][agentProlog-worker] Rebuild Prolog-RLM plugin UX + canonical context meter + bottom-row mode controls #18

New issue

Open

opened 2026-09-11 04:54:15 +00:00 by nsaspy · 0 comments

nsaspy commented 2026-09-11 04:54:15 +00:00

Owner

Copy link

Operator directive

This is the first issue for the AgentProlog worker when a Machine Spirit worker slot is rotated into lost-rob0t/agentProlog.

The current Prolog-RLM/AgentProlog integration UI is not acceptable. It is visually poor, exposes the wrong information hierarchy, and does not make canonical Prolog-RLM context/mode state obvious or trustworthy.

Reuse existing architecture and mode work, especially #9, #7/#11, and the canonical upstream prolog_agent_ui_v1 / conversation/context contracts. Do not create frontend-owned mode or context truth.

Goal

Make the DeepSeek Harness AgentProlog/Prolog-RLM integration feel like a first-class native coding-agent control surface:

• clean and unobtrusive;
• Prolog-RLM actually owns/manages context;
• important RLM state lives in the bottom row/status bar, not as a giant plugin panel;
• one-click enable/disable;
• one-click reasoning-mode selection/cycling;
• live context meter;
• slash commands for all reasoning modes;
• full Playwright, end-to-end, and live LLM verification through llm.starintel.actor.

A — Fix context ownership first

The UI must not fake a context meter over DeepSeek Harness-local history if Prolog-RLM is supposed to be the context authority.

Trace the complete live path:

DSH user turn
-> AgentProlog AgentFactory/plugin
-> persistent bridge/sidecar
-> canonical Prolog-RLM conversation/runtime
-> prompt/context compiler
-> provider request

Prove which layer currently owns:

• conversation/history;
• retained/summarized context;
• prompt/context selection;
• context-window budget;
• current used/remaining context estimate;
• compaction/truncation decisions;
• requested/effective reasoning mode.

Required correction:

• Prolog-RLM is authoritative for the managed conversation/context used for inference.
• DSH session/events remain a presentation projection and must not silently become a second independent history/context authority.
• normal turns, reconnect/resume, mode changes, cancellation and tool activity preserve the same canonical conversation identity where supported.
• if the generic runtime lacks a required context-inspection seam, open/fix the smallest upstream prolog-rlm issue rather than implementing a fake AgentProlog-only context engine.
• no duplicated compaction/summarization policy in the frontend plugin.

Context inspection contract

Expose a bounded typed state sufficient for UI, ideally via existing public protocol/facade semantics or the smallest upstream extension:

context window limit
estimated/actual context used
remaining context
usage percentage
last context-management action
conversation/session identity
requested mode
effective mode

Do not dump entire hidden prompts/context into the UI just to calculate a meter.

B — Replace the ugly plugin presentation

The normal AgentProlog experience should not be a large intrusive Prolog-RLM panel.

Move normal RLM controls/status to a compact bottom row/status bar integrated with the official Harness UI.

Target bottom-row items:

1. RLM enable/disable control

Compact state such as:

RLM: ON
RLM: OFF

Requirements:

• clearly shows canonical state;
• keyboard reachable;
• toggling is explicit and deterministic;
• OFF means the documented non-RLM path, not half-enabled ambiguous behavior;
• if runtime architecture cannot safely support hot enable/disable for an active session, define the smallest truthful session-boundary behavior instead of pretending.

2. Mode control

Compact control showing requested/effective mode, e.g.:

MODE: DIRECT
MODE: SYMBOLIC
MODE: SYMBOLIC-RECURSIVE
MODE: AUTO→SYMBOLIC

Clicking should open/cycle/select among:

• direct
• symbolic
• symbolic-recursive
• auto

Use the canonical mode state from #9/upstream. Do not add a UI-local mode variable.

3. Context meter

Compact bottom-row meter, e.g.:

CTX 42%  132k/315k

or another readable equivalent.

Requirements:

• value comes from Prolog-RLM context/runtime inspection;
• updates after turns/context-management actions;
• hover/click can expose bounded detail such as remaining tokens and last compaction action;
• color/state thresholds may indicate healthy/warning/critical without becoming visual noise;
• never fabricate precision not supplied by the runtime; clearly distinguish estimates if exact provider tokenization is unavailable.

4. Optional compact runtime indicator

If useful and already available canonically, show one compact status item for busy/idle/error/connecting. Do not turn the bottom row into a telemetry dashboard.

C — Slash commands

Implement/finish the canonical mode commands in the Harness input path:

/direct
/symbolic
/symbolic-recursive
/auto

Requirements:

• same canonical transition path as the bottom-row control;
• no second slash-command-owned mode state;
• command result gives concise confirmation including requested/effective mode;
• mode survives subsequent normal turns until changed;
• invalid or unsupported transition fails visibly and structurally;
• slash mode changes preserve conversation/context identity and do not silently reset context.

Also provide discoverable slash completion/help if the official Harness command surface supports it cleanly.

D — Visual/interaction direction

Keep the official Harness UI as the main shell. AgentProlog should feel integrated rather than bolted on.

Requirements:

• remove/reduce the oversized/ugly plugin surface used for ordinary operation;
• bottom-row controls align with surrounding Harness spacing/typography;
• compact icons/text with useful tooltips;
• clear focus/hover/disabled/error states;
• keyboard navigation works;
• narrow viewport remains usable;
• no giant cards, debug dumps, neon clutter, or duplicate sidebars just to expose RLM state;
• deeper Prolog/runtime inspection can remain in an optional detailed view, but normal operation belongs in the bottom row.

E — Playwright + real E2E gate

Every user-visible behavior in this issue requires Playwright coverage and end-to-end coverage. Unit/component tests are necessary but not sufficient.

Playwright must exercise the actual built official Harness + out-of-tree AgentProlog plugin, not a fake standalone page.

Minimum Playwright scenarios:

1. load the real AgentProlog Harness profile;
2. confirm bottom-row RLM state is visible;
3. toggle RLM OFF/ON and verify canonical state reflected after reload/reconnect as designed;
4. select every reasoning mode through UI controls;
5. select every reasoning mode through slash commands;
6. verify UI mode display matches canonical requested/effective state;
7. submit turns and verify context meter changes from runtime-reported state;
8. verify mode changes do not reset the canonical conversation/context identity;
9. exercise reconnect/reload and verify state reconstruction;
10. exercise context warning/high-usage presentation with a deterministic fixture/runtime seam;
11. verify keyboard navigation/focus of bottom-row controls;
12. verify narrow viewport does not hide or corrupt controls;
13. capture screenshots/video/trace on failure.

End-to-end integration must include the real chain:

Harness UI
-> AgentProlog Cordis plugin / AgentFactory
-> persistent sidecar
-> Prolog-RLM managed conversation/context
-> provider call
-> canonical response/event projection
-> Harness UI

No test may declare success merely because the DOM changed while the sidecar/runtime failed.

F — Live LLM gate via llm.starintel.actor

In addition to hermetic/mock E2E, run an opt-in but mandatory-when-requested live integration suite against the operator's llm.starintel.actor endpoint.

Use the existing provider/config/auth path; do not commit credentials, tokens, internal auth headers, or secrets.

Live test must prove at minimum:

1. Harness starts with AgentProlog plugin and real Prolog-RLM sidecar;
2. one direct turn reaches llm.starintel.actor and returns a real model result;
3. one symbolic turn reaches the same live provider through canonical Prolog-RLM routing;
4. one symbolic-recursive turn executes when supported/configured and reports real canonical state;
5. /auto selects an effective mode and completes a live turn;
6. context usage/state returned after live turns updates the bottom-row meter;
7. mode changes do not create a second independent conversation/history;
8. reconnect/resume continues the same canonical managed conversation where supported;
9. provider failure/timeout is surfaced as failure, never rendered as successful model output;
10. safe evidence records endpoint identity (llm.starintel.actor), model/provider route, session/run IDs, mode, context metrics, status and timestamps without logging credentials or raw secrets.

Live-test configuration law

• Live tests are opt-in for ordinary contributor runs so offline development remains possible.
• CI/manual validation that explicitly requests the live lane must hard-fail if required configuration is missing; never silently downgrade to mocks.
• The worker may not mark this issue accepted without a recent successful live llm.starintel.actor validation receipt from the candidate head.

G — Testing standard for AgentProlog going forward

This issue establishes a product rule for subsequent AgentProlog UI/integration work:

unit/component tests
+ protocol/integration tests
+ Playwright for user-facing UI
+ real end-to-end bridge/runtime tests
+ live llm.starintel.actor validation when the change touches inference/context/modes/provider flow

No UI-only fake test is sufficient for behavior whose truth lives in Prolog-RLM.

Acceptance

• Prolog-RLM, not Harness-local UI state, owns inference context/history management.
• Typed context inspection is available without exposing giant hidden prompts.
• Normal Prolog-RLM UI is compact bottom-row controls/status rather than the current ugly panel.
• RLM enable/disable control works and reflects canonical state.
• Mode control supports direct, symbolic, symbolic-recursive, and auto through the canonical #9/upstream path.
• /direct, /symbolic, /symbolic-recursive, /auto work from the normal input surface.
• Context meter reflects real Prolog-RLM state and updates after turns/context-management events.
• Mode changes preserve canonical conversation/context identity.
• Reload/reconnect reconstructs the same canonical state rather than frontend-local guesses.
• Playwright covers all main user interactions on the real built Harness integration.
• Real E2E covers UI -> plugin -> sidecar -> Prolog-RLM -> provider -> UI.
• Live LLM validation through llm.starintel.actor passes on the candidate head.
• Live test fails closed when explicitly requested but not correctly configured.
• Screenshots/traces from Playwright and a redacted live-test receipt are attached/referenced before acceptance.

Worker law

When the Machine Spirit completion gate rotates a worker into AgentProlog, claim this issue first. Reconcile #9 and any active PRs before coding so mode work is reused rather than duplicated.

## Operator directive This is the **first issue for the AgentProlog worker** when a Machine Spirit worker slot is rotated into `lost-rob0t/agentProlog`. The current Prolog-RLM/AgentProlog integration UI is not acceptable. It is visually poor, exposes the wrong information hierarchy, and does not make canonical Prolog-RLM context/mode state obvious or trustworthy. Reuse existing architecture and mode work, especially #9, #7/#11, and the canonical upstream `prolog_agent_ui_v1` / conversation/context contracts. Do **not** create frontend-owned mode or context truth. ## Goal Make the DeepSeek Harness AgentProlog/Prolog-RLM integration feel like a first-class native coding-agent control surface: - clean and unobtrusive; - Prolog-RLM actually owns/manages context; - important RLM state lives in the **bottom row/status bar**, not as a giant plugin panel; - one-click enable/disable; - one-click reasoning-mode selection/cycling; - live context meter; - slash commands for all reasoning modes; - full Playwright, end-to-end, and **live LLM** verification through `llm.starintel.actor`. ## A — Fix context ownership first The UI must not fake a context meter over DeepSeek Harness-local history if Prolog-RLM is supposed to be the context authority. Trace the complete live path: ```text DSH user turn -> AgentProlog AgentFactory/plugin -> persistent bridge/sidecar -> canonical Prolog-RLM conversation/runtime -> prompt/context compiler -> provider request ``` Prove which layer currently owns: - conversation/history; - retained/summarized context; - prompt/context selection; - context-window budget; - current used/remaining context estimate; - compaction/truncation decisions; - requested/effective reasoning mode. Required correction: - Prolog-RLM is authoritative for the managed conversation/context used for inference. - DSH session/events remain a presentation projection and must not silently become a second independent history/context authority. - normal turns, reconnect/resume, mode changes, cancellation and tool activity preserve the same canonical conversation identity where supported. - if the generic runtime lacks a required context-inspection seam, open/fix the smallest upstream `prolog-rlm` issue rather than implementing a fake AgentProlog-only context engine. - no duplicated compaction/summarization policy in the frontend plugin. ### Context inspection contract Expose a bounded typed state sufficient for UI, ideally via existing public protocol/facade semantics or the smallest upstream extension: ```text context window limit estimated/actual context used remaining context usage percentage last context-management action conversation/session identity requested mode effective mode ``` Do not dump entire hidden prompts/context into the UI just to calculate a meter. ## B — Replace the ugly plugin presentation The normal AgentProlog experience should **not** be a large intrusive Prolog-RLM panel. Move normal RLM controls/status to a compact **bottom row/status bar** integrated with the official Harness UI. Target bottom-row items: ### 1. RLM enable/disable control Compact state such as: ```text RLM: ON RLM: OFF ``` Requirements: - clearly shows canonical state; - keyboard reachable; - toggling is explicit and deterministic; - OFF means the documented non-RLM path, not half-enabled ambiguous behavior; - if runtime architecture cannot safely support hot enable/disable for an active session, define the smallest truthful session-boundary behavior instead of pretending. ### 2. Mode control Compact control showing requested/effective mode, e.g.: ```text MODE: DIRECT MODE: SYMBOLIC MODE: SYMBOLIC-RECURSIVE MODE: AUTO→SYMBOLIC ``` Clicking should open/cycle/select among: - `direct` - `symbolic` - `symbolic-recursive` - `auto` Use the canonical mode state from #9/upstream. Do not add a UI-local mode variable. ### 3. Context meter Compact bottom-row meter, e.g.: ```text CTX 42% 132k/315k ``` or another readable equivalent. Requirements: - value comes from Prolog-RLM context/runtime inspection; - updates after turns/context-management actions; - hover/click can expose bounded detail such as remaining tokens and last compaction action; - color/state thresholds may indicate healthy/warning/critical without becoming visual noise; - never fabricate precision not supplied by the runtime; clearly distinguish estimates if exact provider tokenization is unavailable. ### 4. Optional compact runtime indicator If useful and already available canonically, show one compact status item for busy/idle/error/connecting. Do not turn the bottom row into a telemetry dashboard. ## C — Slash commands Implement/finish the canonical mode commands in the Harness input path: ```text /direct /symbolic /symbolic-recursive /auto ``` Requirements: - same canonical transition path as the bottom-row control; - no second slash-command-owned mode state; - command result gives concise confirmation including requested/effective mode; - mode survives subsequent normal turns until changed; - invalid or unsupported transition fails visibly and structurally; - slash mode changes preserve conversation/context identity and do not silently reset context. Also provide discoverable slash completion/help if the official Harness command surface supports it cleanly. ## D — Visual/interaction direction Keep the official Harness UI as the main shell. AgentProlog should feel integrated rather than bolted on. Requirements: - remove/reduce the oversized/ugly plugin surface used for ordinary operation; - bottom-row controls align with surrounding Harness spacing/typography; - compact icons/text with useful tooltips; - clear focus/hover/disabled/error states; - keyboard navigation works; - narrow viewport remains usable; - no giant cards, debug dumps, neon clutter, or duplicate sidebars just to expose RLM state; - deeper Prolog/runtime inspection can remain in an optional detailed view, but normal operation belongs in the bottom row. ## E — Playwright + real E2E gate **Every user-visible behavior in this issue requires Playwright coverage and end-to-end coverage.** Unit/component tests are necessary but not sufficient. Playwright must exercise the actual built official Harness + out-of-tree AgentProlog plugin, not a fake standalone page. Minimum Playwright scenarios: 1. load the real AgentProlog Harness profile; 2. confirm bottom-row RLM state is visible; 3. toggle RLM OFF/ON and verify canonical state reflected after reload/reconnect as designed; 4. select every reasoning mode through UI controls; 5. select every reasoning mode through slash commands; 6. verify UI mode display matches canonical requested/effective state; 7. submit turns and verify context meter changes from runtime-reported state; 8. verify mode changes do not reset the canonical conversation/context identity; 9. exercise reconnect/reload and verify state reconstruction; 10. exercise context warning/high-usage presentation with a deterministic fixture/runtime seam; 11. verify keyboard navigation/focus of bottom-row controls; 12. verify narrow viewport does not hide or corrupt controls; 13. capture screenshots/video/trace on failure. End-to-end integration must include the real chain: ```text Harness UI -> AgentProlog Cordis plugin / AgentFactory -> persistent sidecar -> Prolog-RLM managed conversation/context -> provider call -> canonical response/event projection -> Harness UI ``` No test may declare success merely because the DOM changed while the sidecar/runtime failed. ## F — Live LLM gate via llm.starintel.actor In addition to hermetic/mock E2E, run an **opt-in but mandatory-when-requested live integration suite** against the operator's `llm.starintel.actor` endpoint. Use the existing provider/config/auth path; do not commit credentials, tokens, internal auth headers, or secrets. Live test must prove at minimum: 1. Harness starts with AgentProlog plugin and real Prolog-RLM sidecar; 2. one `direct` turn reaches `llm.starintel.actor` and returns a real model result; 3. one `symbolic` turn reaches the same live provider through canonical Prolog-RLM routing; 4. one `symbolic-recursive` turn executes when supported/configured and reports real canonical state; 5. `/auto` selects an effective mode and completes a live turn; 6. context usage/state returned after live turns updates the bottom-row meter; 7. mode changes do not create a second independent conversation/history; 8. reconnect/resume continues the same canonical managed conversation where supported; 9. provider failure/timeout is surfaced as failure, never rendered as successful model output; 10. safe evidence records endpoint identity (`llm.starintel.actor`), model/provider route, session/run IDs, mode, context metrics, status and timestamps without logging credentials or raw secrets. ### Live-test configuration law - Live tests are opt-in for ordinary contributor runs so offline development remains possible. - CI/manual validation that explicitly requests the live lane must **hard-fail** if required configuration is missing; never silently downgrade to mocks. - The worker may not mark this issue accepted without a recent successful live `llm.starintel.actor` validation receipt from the candidate head. ## G — Testing standard for AgentProlog going forward This issue establishes a product rule for subsequent AgentProlog UI/integration work: ```text unit/component tests + protocol/integration tests + Playwright for user-facing UI + real end-to-end bridge/runtime tests + live llm.starintel.actor validation when the change touches inference/context/modes/provider flow ``` No UI-only fake test is sufficient for behavior whose truth lives in Prolog-RLM. ## Acceptance - [ ] Prolog-RLM, not Harness-local UI state, owns inference context/history management. - [ ] Typed context inspection is available without exposing giant hidden prompts. - [ ] Normal Prolog-RLM UI is compact bottom-row controls/status rather than the current ugly panel. - [ ] RLM enable/disable control works and reflects canonical state. - [ ] Mode control supports `direct`, `symbolic`, `symbolic-recursive`, and `auto` through the canonical #9/upstream path. - [ ] `/direct`, `/symbolic`, `/symbolic-recursive`, `/auto` work from the normal input surface. - [ ] Context meter reflects real Prolog-RLM state and updates after turns/context-management events. - [ ] Mode changes preserve canonical conversation/context identity. - [ ] Reload/reconnect reconstructs the same canonical state rather than frontend-local guesses. - [ ] Playwright covers all main user interactions on the real built Harness integration. - [ ] Real E2E covers UI -> plugin -> sidecar -> Prolog-RLM -> provider -> UI. - [ ] Live LLM validation through `llm.starintel.actor` passes on the candidate head. - [ ] Live test fails closed when explicitly requested but not correctly configured. - [ ] Screenshots/traces from Playwright and a redacted live-test receipt are attached/referenced before acceptance. ## Worker law When the Machine Spirit completion gate rotates a worker into AgentProlog, **claim this issue first**. Reconcile #9 and any active PRs before coding so mode work is reused rather than duplicated.

nsaspy referenced this issue 2026-09-19 04:27:35 +00:00

[P0][Workbench] Full coding-agent desktop UX #30

nsaspy referenced this issue 2026-09-19 04:27:36 +00:00

[EPIC] AgentProlog Workbench: desktop UI + Android + Termux runtime #33

Sign in to join this conversation.

No Branch/Tag specified

Branches Tags

main

fix/zara-profile-closed-set

fix/zara-profile-runtime-advertisement

feat/zara-profile-handshake-1046

workbench/client-issue-26-20260919

fix/zara-profile-health-coherence-1046

master

security/cve-2026-82533-dsh-0.1.2-rc.1

feat/zara-runtime-agentprolog-profile-1046

feat/zara-runtime-profile

codex/text-streaming

codex/recursive-runtime-options

feat/coding-expert-tools

security/cve-2026-82533-harness

agent/issue-7-dsh-agent-factory

agent/issue-2-deepseek-plugin-design

agent/headless-workflow-conformance

agent/issue-1-flake-baseline

No results found.

Labels

Clear labels   

accessibility

Barrier affecting people with disabilities

  

bug

Something isn't working

  

documentation

Improvements or additions to documentation

  

duplicate

This issue or pull request already exists

  

enhancement

New feature or request

  

good first issue

Good for newcomers

  

help wanted

Extra attention is needed

  

invalid

This doesn't seem right

  

question

Further information is requested

  

security

  

testing

  

wontfix

This will not be worked on

No labels

accessibility

bug

documentation

duplicate

enhancement

good first issue

help wanted

invalid

question

security

testing

wontfix

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

nsaspy/agentProlog#18

No description provided.