PHASE 0 EPIC: Debloat tool, skill, and prompt context without RLM #3

New issue

Open

opened 2026-08-22 22:30:16 +00:00 by lost-rob0t · 0 comments

lost-rob0t commented

2026-08-22 22:30:16 +00:00

(Migrated from github.com)

Parent: #2

Goal

Reduce the largest remaining bootstrap/context cost: Agent Zero currently loads the full prompt payload for every allowed tool. Do this without Prolog-RLM so Phase 0 establishes a clean baseline first.

Current problem

The tool system enumerates agent.system.tool.*.md, reads every allowed prompt, concatenates them, and injects the entire result. Tool authorization filtering exists, but an allowed tool still pays its full prompt cost whether or not the current task needs it.

Phase-0 design constraints

No Prolog-RLM dependency.
Existing _tool_access remains execution authority.
Visibility reduction must not grant or revoke execution permission accidentally.
Prefer existing plugin/extensible hooks and a small deterministic Python discovery layer over core loop edits.
Unknown tasks need a bounded discovery fallback; never solve uncertainty by restoring the entire catalog.

Work order

Measure tool-prompt token contribution for default profiles.
Build a compact tool catalog from existing prompt/tool metadata.
Separate always-visible essential tools from task-selectable tools.
Select a bounded candidate set using plain Agent Zero/Python signals for Phase 0.
Load full prompt bodies only for selected candidates.
Add a bounded discovery/retry path when the initial set is insufficient.
Measure task-success and token delta against unmodified Agent Zero.

RAGE loop

Review exact tool prompt builder, tool policy, profile prompt paths and representative tasks -> Analyze smallest behavior-preserving visibility cut -> Generate bounded catalog/selection slice + tests -> Execute token and task regressions -> inspect misses -> repeat.

Loop breakers

If a task fails because a needed tool was hidden, record the miss and improve discovery; do not simply re-enable all tools.
Same selection miss twice -> re-plan the selector or add explicit metadata.
Never let a prompt-selection mechanism override execution authorization.

Acceptance

Default root no longer receives every full allowed tool prompt on every turn.
Essential response/control tools remain reliably available.
Full contracts are loaded only for a bounded candidate set.
Unknown capability discovery is bounded and tested.
Tool authorization behavior is unchanged.
Token telemetry reports catalog vs selected full-contract cost.
Representative coding/research/simple tasks do not materially regress.
Works with zero Prolog-RLM installation.

Exit

Finish this Phase-0 behavior and measure it before replacing the selector with or augmenting it using the Prolog-RLM prompt compiler in Phase 1.

Parent: #2 ## Goal Reduce the largest remaining bootstrap/context cost: Agent Zero currently loads the full prompt payload for every allowed tool. Do this **without Prolog-RLM** so Phase 0 establishes a clean baseline first. ## Current problem The tool system enumerates `agent.system.tool.*.md`, reads every allowed prompt, concatenates them, and injects the entire result. Tool authorization filtering exists, but an allowed tool still pays its full prompt cost whether or not the current task needs it. ## Phase-0 design constraints - No Prolog-RLM dependency. - Existing `_tool_access` remains execution authority. - Visibility reduction must not grant or revoke execution permission accidentally. - Prefer existing plugin/extensible hooks and a small deterministic Python discovery layer over core loop edits. - Unknown tasks need a bounded discovery fallback; never solve uncertainty by restoring the entire catalog. ## Work order 1. Measure tool-prompt token contribution for default profiles. 2. Build a compact tool catalog from existing prompt/tool metadata. 3. Separate always-visible essential tools from task-selectable tools. 4. Select a bounded candidate set using plain Agent Zero/Python signals for Phase 0. 5. Load full prompt bodies only for selected candidates. 6. Add a bounded discovery/retry path when the initial set is insufficient. 7. Measure task-success and token delta against unmodified Agent Zero. ## RAGE loop **Review** exact tool prompt builder, tool policy, profile prompt paths and representative tasks -> **Analyze** smallest behavior-preserving visibility cut -> **Generate** bounded catalog/selection slice + tests -> **Execute** token and task regressions -> inspect misses -> repeat. ### Loop breakers - If a task fails because a needed tool was hidden, record the miss and improve discovery; do not simply re-enable all tools. - Same selection miss twice -> re-plan the selector or add explicit metadata. - Never let a prompt-selection mechanism override execution authorization. ## Acceptance - [ ] Default root no longer receives every full allowed tool prompt on every turn. - [ ] Essential response/control tools remain reliably available. - [ ] Full contracts are loaded only for a bounded candidate set. - [ ] Unknown capability discovery is bounded and tested. - [ ] Tool authorization behavior is unchanged. - [ ] Token telemetry reports catalog vs selected full-contract cost. - [ ] Representative coding/research/simple tasks do not materially regress. - [ ] Works with zero Prolog-RLM installation. ## Exit Finish this Phase-0 behavior and measure it before replacing the selector with or augmenting it using the Prolog-RLM prompt compiler in Phase 1.