[P0] Project canonical reasoning modes through headless and DeepSeek Harness #10

New issue

Open

opened 2026-09-02 17:24:18 +00:00 by nsaspy · 0 comments

nsaspy commented

2026-09-02 17:24:18 +00:00

Owner

Worker

agentprolog-rage-integration

Goal

Expose the canonical Prolog-RLM reasoning modes consistently through AgentProlog's headless protocol and official DeepSeek Harness integration without creating frontend-owned execution state.

Depends on the reusable upstream reasoning-mode contract in lost-rob0t/prolog-rlm#335.

Current DSH implementation authority remains AgentProlog #7 / PR #8 or whatever newer merged track supersedes it.

Required commands

/direct
/symbolic
/symbolic-recursive
/auto

Product semantics

The command changes canonical AgentProlog/Prolog-RLM session state. DSH reflects that state; it does not maintain an independent mode variable that can drift from the sidecar/runtime.

Required behavior:

explicit mode remains active until changed;
/auto delegates strategy choice to the upstream canonical selector;
DSH/headless inspection shows requested mode, effective mode, and auto selection reason when applicable;
mode changes preserve the same canonical session/conversation identity;
mode changes do not reset history, capabilities, authority, budgets, cancellation state, verifier requirements or durable-effect lineage;
failures/version mismatches are structured and fail closed rather than silently falling back to stock DSH semantics.

Headless/bridge work

Extend the existing AgentProlog bridge/sidecar protocol through the smallest typed operations needed to:

set reasoning mode;
inspect reasoning mode/effective strategy;
receive mode-change/runtime events where the current protocol supports event projection.

Do not create a second command interpreter if the current bridge already has a normalized command/session-operation path.

DeepSeek Harness work

Project the same mode state and commands through the out-of-tree AgentProlog DSH plugin.

retain exactly one AgentFactory;
keep DSH Session/events as presentation projection;
do not give DSH ownership of provider/tool/planner/history/authority/verifier state;
prove one DSH command results in one canonical mode transition;
prove one normal user turn still maps to one canonical AgentProlog/Prolog-RLM trajectory under each effective mode where supported.

Tests

headless set/get for each mode;
direct -> symbolic -> auto transitions preserve session identity;
auto returns/projections match upstream effective selection;
DSH mode display matches sidecar/runtime truth;
reconnect/reopen does not fabricate a different frontend mode;
invalid mode and unsupported protocol version fail structurally;
cancellation/terminal session prevents invalid state mutation;
explicit mode change cannot widen authority/capability/budgets;
no stock DSH fallback executes when AgentProlog mode handling fails.

Acceptance

all four commands work through the canonical headless path;
DSH exposes the same state without a duplicate mode authority;
mode transitions are inspectable and traceable;
current AgentFactory architecture remains intact;
focused tests and exact-head CI/Nix smoke are green;
docs describe mode commands and their strategy-only security boundary.

## Worker `agentprolog-rage-integration` ## Goal Expose the canonical Prolog-RLM reasoning modes consistently through AgentProlog's headless protocol and official DeepSeek Harness integration without creating frontend-owned execution state. Depends on the reusable upstream reasoning-mode contract in `lost-rob0t/prolog-rlm#335`. Current DSH implementation authority remains AgentProlog #7 / PR #8 or whatever newer merged track supersedes it. ## Required commands - `/direct` - `/symbolic` - `/symbolic-recursive` - `/auto` ## Product semantics The command changes canonical AgentProlog/Prolog-RLM session state. DSH reflects that state; it does not maintain an independent mode variable that can drift from the sidecar/runtime. Required behavior: - explicit mode remains active until changed; - `/auto` delegates strategy choice to the upstream canonical selector; - DSH/headless inspection shows requested mode, effective mode, and auto selection reason when applicable; - mode changes preserve the same canonical session/conversation identity; - mode changes do not reset history, capabilities, authority, budgets, cancellation state, verifier requirements or durable-effect lineage; - failures/version mismatches are structured and fail closed rather than silently falling back to stock DSH semantics. ## Headless/bridge work Extend the existing AgentProlog bridge/sidecar protocol through the smallest typed operations needed to: - set reasoning mode; - inspect reasoning mode/effective strategy; - receive mode-change/runtime events where the current protocol supports event projection. Do not create a second command interpreter if the current bridge already has a normalized command/session-operation path. ## DeepSeek Harness work Project the same mode state and commands through the out-of-tree AgentProlog DSH plugin. - retain exactly one AgentFactory; - keep DSH Session/events as presentation projection; - do not give DSH ownership of provider/tool/planner/history/authority/verifier state; - prove one DSH command results in one canonical mode transition; - prove one normal user turn still maps to one canonical AgentProlog/Prolog-RLM trajectory under each effective mode where supported. ## Tests - headless set/get for each mode; - direct -> symbolic -> auto transitions preserve session identity; - auto returns/projections match upstream effective selection; - DSH mode display matches sidecar/runtime truth; - reconnect/reopen does not fabricate a different frontend mode; - invalid mode and unsupported protocol version fail structurally; - cancellation/terminal session prevents invalid state mutation; - explicit mode change cannot widen authority/capability/budgets; - no stock DSH fallback executes when AgentProlog mode handling fails. ## Acceptance - all four commands work through the canonical headless path; - DSH exposes the same state without a duplicate mode authority; - mode transitions are inspectable and traceable; - current AgentFactory architecture remains intact; - focused tests and exact-head CI/Nix smoke are green; - docs describe mode commands and their strategy-only security boundary.

nsaspy referenced this issue

2026-09-02 17:24:22 +00:00

[P0] Implement plugin-only DeepSeek Harness AgentFactory #11