Restore native model autonomy and capability awareness

Author: gitcommit90

CHANGELOG.md

@@ -7,6 +7,32 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.0.5] - 2026-07-24
+
+### Changed
+
+- Agent turns now receive a compact factual capability map—identity, channel,
+  Linux computer and `/workspace`, callable tools, memory, skill inventory, and
+  authority scope—instead of large behavioral manifestos.
+- The skill arsenal is lazy-loaded. Models can list available skill metadata
+  and explicitly read one full procedure when useful; 1Helm no longer injects
+  the entire playbook library into every turn.
+- Removed runtime prose classifiers that rejected, rewrote, or forcibly
+  re-entered otherwise valid model answers. Security, ownership, and human-only
+  boundaries remain enforced in tool and server code.
+
+### Fixed
+
+- Human display names are no longer prefixed to user messages sent to models,
+  preventing names such as the Captain's from contaminating generated search
+  queries.
+- A public web/news search with no parsed results now retries once with a
+  concise query, so ordinary conversational wording does not consume the turn
+  or force an immediate refusal.
+- Real-model acceptance covers the exact West Hollywood sinkhole prompt,
+  Skipper's computer awareness in `#main`, lazy skill discovery, and autonomous
+  resident work on its Linux machine.
+
 ## [0.0.4] - 2026-07-24
 
 ### Added
@@ -107,7 +133,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   notarization, stapled tickets, Gatekeeper verification, persistent
   Application Support, and isolated Apple container machines.
 
-[Unreleased]: https://github.com/gitcommit90/1Helm/compare/v0.0.4...HEAD
+[Unreleased]: https://github.com/gitcommit90/1Helm/compare/v0.0.5...HEAD
+[0.0.5]: https://github.com/gitcommit90/1Helm/releases/tag/v0.0.5
 [0.0.4]: https://github.com/gitcommit90/1Helm/releases/tag/v0.0.4
 [0.0.3]: https://github.com/gitcommit90/1Helm/releases/tag/v0.0.3
 [0.0.2]: https://github.com/gitcommit90/1Helm/releases/tag/v0.0.2

README.md

@@ -49,7 +49,7 @@ else that deserves an owner. 1Helm provisions a complete world around it:
 | **One permanent resident** | Threads are sessions; the employee identity survives them. |
 | **One private Linux computer** | Chat tools and Terminal use the same persistent `/workspace`. |
 | **Memory with provenance** | Decisions, corrections, preferences, files, and outcomes become continuity. |
-| **A serious skill arsenal** | Relevant playbooks activate automatically; routine tool use is not an approval ritual. |
+| **A serious skill arsenal** | The agent sees what exists and loads a full procedure only when it chooses one. |
 | **Durable obligations** | Follow-ups, timers, workflows, and services can wake the computer back up. |
 | **Skipper at the boundary** | Host work, credentials, fleet operations, and cross-channel work route themselves. |
 
@@ -98,7 +98,9 @@ discovery, memory, research, email, calendar, contacts, messaging, documents,
 spreadsheets, PDFs, meetings, projects, personal operations, travel, finance,
 support, software delivery, data, media, infrastructure, security, and more.
 
-1Helm selects only what the task needs. It can also:
+The model receives a compact inventory of the arsenal—not all 34 procedures in
+every prompt. It can inspect metadata and load one complete skill when useful.
+It can also:
 
 - search the focused SkillsMD catalog of ready GitHub-backed repositories;
 - install ready skills only after immutable revision pinning, bounds,
@@ -156,9 +158,10 @@ does not run on the laptop or phone viewing the web UI.
 - Durable files, threads, curated memory, Mnemosyne long-term recall,
   corrections, follow-ups, and recurring workflows.
 - Direct resident → Skipper escalation and automatic Skipper → resident return.
-- A bounded outcome gate that keeps operational work moving when a model tries
-  to stop at a tutorial, unevidenced blocker, unresolved tool failure, or
-  “Skipper could help.”
+- Compact, factual model context: identity, channel, Linux computer and
+  `/workspace`, callable tools, memory, and the skill arsenal. Security and
+  ownership boundaries are enforced by the tools and server rather than a
+  prose grader rewriting model answers.
 - Outcome-first Activity with expandable work evidence and a tamper-evident
   SHA-256 chain for new operational events.
 - Local-first collaboration through an optional workspace domain routed to the
@@ -228,7 +231,7 @@ A fresh data directory opens first-run setup. The source runtime defaults to
 | `PORT` | `8123` | HTTP/WebSocket control-plane port. |
 | `CTRL_DATA_DIR` | `./data` | Databases, routing state, uploads, and narrow workspace mirrors. |
 | `HELM_CHANNEL_COMPUTER_BACKEND` | `apple` on macOS, `native` elsewhere | Explicit development/test backend override. |
-| `HELM_CHANNEL_MACHINE_IMAGE` | `local/1helm-channel-machine:0.0.4` | Versioned Apple channel-machine image. |
+| `HELM_CHANNEL_MACHINE_IMAGE` | `local/1helm-channel-machine:0.0.5` | Versioned Apple channel-machine image. |
 
 ### Agent-first JSON CLI
 
@@ -276,7 +279,7 @@ npm run benchmark:autonomy
 ```
 
 The autonomy benchmark emits deterministic machine-readable JSON covering the
-shipped playbook arsenal, narrow human-blocker gate, bounded outcome gate,
+shipped skill arsenal, compact capability map, narrow human-blocker boundary,
 resident autonomy tools, wakeable recurring work, and audit-chain invariants
 for its fixture. It is deliberately not presented as a live-model success rate
 or a complete security score.

docs/RELIABILITY.md

@@ -9,28 +9,25 @@ runtimes, including Hermes Agent. 1Helm keeps its own identity and product
 model; the useful lesson is to encode dependable behavior in runtime state,
 ownership, recovery, and tests rather than depend on a prompt alone.
 
-## Partner behavior
+## Model capability context
 
-- Stable identity and persona are separate from operating policy and volatile
-  channel, memory, and task context. A new task must not accidentally rewrite
-  who the resident or Skipper is.
-- Skipper and residents are reliable operating partners: calm, candid,
-  resourceful, and accountable for closure. They do not mimic momentary user
-  frustration, another agent's persona, or untrusted recalled text.
-- Safe, reversible action is the default. The agent inspects authoritative
-  state, acts through the capability already available, verifies the observable
-  result, and reports it.
-- `ask_user` is reserved for evidenced human judgment, credentials, external
-  authority, or irreversible commitment. Difficulty, harmless ambiguity, and
-  work the agent or Skipper can inspect are not blockers.
-- Read-only questions never install skills, packages, or integrations. A
-  capability question is answered from authoritative state.
-- Tool results are evidence. Transient failures permit a bounded retry with a
-  changed strategy; an unchanged failed call is not repeated forever; expected
-  human blockers and permanent failures produce one useful final explanation.
-- Memory is provenance-bearing reference data, never a higher-priority source
-  of instructions. Memory and behavior review happen after the user's answer is
-  safely delivered.
+- Every turn identifies the model as Skipper or the channel's resident and
+  states the current channel and authority scope.
+- A resident is told that it owns an isolated persistent Linux computer, that
+  its durable workspace is `/workspace`, and that `run_command` executes there.
+  Skipper in `#main` is told that the lack of a resident VM does not remove its
+  assigned computers or native tools.
+- The model receives the callable tool schemas as its capability surface. Hard
+  isolation, ownership, credential, and destructive-action boundaries live in
+  those tool and server implementations.
+- The prompt names the size and categories of the skill arsenal. Skill metadata
+  is available through `list_skills`; a complete procedure enters context only
+  when the model calls `read_skill` for that skill.
+- Channel memory and agent recall are provenance-bearing reference data, never
+  higher-priority instructions.
+- `ask_user` is a validated tool for evidenced human judgment, credentials,
+  external authority, or irreversible commitment. Routine ambiguity does not
+  satisfy that boundary.
 
 ## Turn lifecycle
 
@@ -89,9 +86,9 @@ Every release that touches agent behavior must keep automated coverage for:
 | Stop | Stop changes only its selected lane while another turn on the same agent completes. |
 | Writers | A stale generation cannot change finalized body or progress state. |
 | Restart | Queued work survives and drains; running work is never blindly replayed. |
-| Blockers | A legitimate connector/credential blocker preserves the first useful response without an outcome-gate rewrite loop. |
+| Blockers | A legitimate connector/credential blocker remains useful without a prose-grading rewrite loop. |
 | Questions | Selecting and submitting a structured option works after a live message repaint. |
 | Connectors | A final reply is returned once; pending delivery recovers; interrupted delivery becomes visibly uncertain rather than duplicating. |
 | Scheduling | Due workflows and follow-ups remain wakeable obligations across restart and computer sleep. |
-| Persona | Stable partner identity/action policy remain unchanged when volatile channel/task context changes. |
+| Capability map | Identity and machine/tool/skill facts remain compact and correct when volatile channel/task context changes. |
 | Product controls | Skipper uses native inventory/lifecycle tools; terminal paths, file Open/Download, trusted skill search, provider routing, and channel deletion are end-to-end tested. |

docs/USER_GUIDE.md

@@ -193,8 +193,10 @@ stand in for a news photograph.
 
 ## Skills and Learn a new skill
 
-Every resident permanently owns the safe built-in operational arsenal. The
-runtime activates only relevant playbooks for the current task.
+Every resident permanently owns the built-in operational arsenal. The model
+sees a compact inventory, can list the available skills, and loads the full
+procedure for one skill only when it chooses to use it. The procedures are not
+injected wholesale into every turn.
 
 ![Built-in skills and the focused SkillsMD catalog](assets/guide/skills.png)
 
@@ -206,8 +208,9 @@ repository-specific procedure is unavailable, choose **Learn a new skill** so
 Skipper can inspect supplied local sources, web URLs, and notes in a visible
 thread and author a workspace-specific procedure.
 
-A question such as “do we have a calendar skill?” is read-only: it must not
-install a package, install a skill, or open an interview.
+A question such as “do we have a calendar skill?” can be answered directly
+from the agent's actual arsenal inventory without installing anything or
+opening an interview.
 
 ## Gmail
 

docs/VISION.md

@@ -20,12 +20,12 @@ resident, which verifies the original end-to-end outcome. The Captain is
 involved only for consequential human judgment, credentials that only they can
 provide, external authority, or an irreversible commitment.
 
-Prompting is not the only enforcement layer. For operational requests, a
-bounded deterministic outcome gate objects when a model tries to stop at a
-tutorial, an unevidenced blocker, an unresolved tool failure, or a suggestion
-that the Captain call Skipper. It keeps the turn open at most three times,
-leaves the objection in the work log, and yields normally for explanatory
-questions or a genuine structured human boundary.
+The runtime gives the model a compact factual capability map: its identity,
+channel, Linux computer and `/workspace`, callable tools, memory, skill arsenal,
+and authority scope. It does not inject a behavioral manifesto or grade and
+rewrite the model's prose. Tool and server implementations enforce isolation,
+ownership, credentials, destructive-action confirmation, and human-only
+boundaries.
 
 If work depends on time passing, it becomes a persisted obligation. 1Helm does
 not equate a model ending a turn with background execution. Follow-ups and
@@ -34,9 +34,10 @@ and observable completion or failure.
 
 ## Durable growth
 
-Residents start with a substantive safe operational arsenal selected by task,
-not a handful of generic prompt snippets. Skills define activation, execution,
-authority boundaries, recovery, retained state, and verification. Trusted
+Residents start with a substantive operational arsenal, not a handful of
+generic prompt snippets. The model sees concise skill metadata and chooses when
+to load one full procedure. Skills define activation, execution, authority
+boundaries, recovery, retained state, and verification. Trusted
 external procedures may be discovered and installed automatically only through
 bounded, provenance-pinned, scanned, hashed ingestion beneath runtime policy.
 Community popularity is not trust.

package-lock.json

@@ -1,7 +1,7 @@
 {
   "name": "1helm",
   "productName": "1Helm",
-  "version": "0.0.4",
+  "version": "0.0.5",
   "private": true,
   "type": "module",
   "description": "1Helm is the self-hosted home for durable AI employees: one resident, one private computer, compounding memory and skills, and Skipper for every boundary.",

package.json

@@ -34,6 +34,6 @@
 </head>
 <body class="h-screen w-screen overflow-hidden antialiased">
   <div id="app" class="h-full w-full"></div>
-  <script type="module" src="/bundle.js?v=44190c5b5255"></script>
+  <script type="module" src="/bundle.js?v=b053ad3e9c4b"></script>
 </body>
 </html>

public/index.html

@@ -17,7 +17,7 @@ try {
   db = database.db;
   const { q1, run, now, seed } = database;
   const { BUILTIN_SKILLS } = await import("../src/server/builtin-skills.ts");
-  const { outcomeGateObjection, validateAskUserInput, runtimeToolNamesForChannel } = await import("../src/server/bots.ts");
+  const { runtimePromptTiersForChannel, validateAskUserInput, runtimeToolNamesForChannel } = await import("../src/server/bots.ts");
   const { createWorkflow, setWorkflowStatus, stopWorkflowLoop } = await import("../src/server/workflows.ts");
   const { verifyAuditChain } = await import("../src/server/audit.ts");
 
@@ -45,12 +45,13 @@ try {
   });
   record("structured_human_boundary", !routine.valid && realBoundary.valid, { routine, real_boundary: realBoundary });
 
-  const handHolding = outcomeGateObjection({ request: "Install and verify the CLI", response: "You can run npm install yourself." });
-  const directBoundary = outcomeGateObjection({ request: "Deploy the site", response: "Skipper completed the host change.", successfulTools: ["call_skipper"] });
-  record("bounded_outcome_gate", /rejected an operational reply/i.test(handHolding) && directBoundary === "", {
-    hand_holding_objection: handHolding,
-    completed_boundary_objection: directBoundary,
-    rejection_budget: 3,
+  const prompt = runtimePromptTiersForChannel(botId, channelId, false, "Install and verify the CLI");
+  const promptLength = prompt.identity.length + prompt.operating.length + prompt.context.length;
+  record("compact_capability_map", promptLength < 2_000 && /isolated persistent Linux computer/i.test(prompt.operating) && /skill-arsenal count=/i.test(prompt.context) && !/active-skill-playbooks|### /.test(prompt.context), {
+    characters: promptLength,
+    has_linux_computer: /isolated persistent Linux computer/i.test(prompt.operating),
+    has_skill_inventory: /skill-arsenal count=/i.test(prompt.context),
+    injects_full_playbooks: /active-skill-playbooks|### /.test(prompt.context),
   });
 
   const toolNames = runtimeToolNamesForChannel(botId, channelId);
@@ -89,7 +90,7 @@ const report = {
     validates: [
       "shipped built-in playbook completeness",
       "structured human-blocker validation",
-      "bounded runtime objection to operational hand-holding",
+      "compact factual capability-map delivery",
       "resident autonomy tool availability",
       "wakeable recurring-work persistence",
       "audit-chain integrity for the executed fixture",

scripts/autonomy-benchmark.mjs

@@ -169,11 +169,11 @@ function skillsPanel(): HTMLElement {
     h("div", { class: "flex flex-col gap-3 rounded-lg border border-accent/30 bg-accent-soft p-4 sm:flex-row sm:items-center sm:justify-between" },
       h("div", {}, h("h3", { class: "font-semibold text-fg" }, "Teach 1Helm from your own material"), h("p", { class: "mt-1 text-sm leading-5 text-muted" }, "Give Skipper a folder or file, a web page, pasted notes, or any combination. It will inspect the sources and author one reusable workspace skill in a visible #main thread.")),
       h("button", { class: "btn-primary shrink-0 text-sm", onclick: learnSkillDialog }, icon("sparkles", 15), "Learn a new skill")),
-    h("p", { class: "text-sm leading-6 text-muted" }, "Every resident permanently owns the safe built-in operational library. 1Helm activates only the playbooks relevant to the current task and can search the focused SkillsMD catalog for ready-to-install GitHub-backed skills."));
+    h("p", { class: "text-sm leading-6 text-muted" }, "Every resident permanently owns the built-in operational library. Agents see a compact inventory and load a full procedure only when they choose that skill. You can also search SkillsMD for GitHub-backed skills."));
   void api<{ skills: Array<Skill & { arsenal_locked?: number; arsenal_reason?: string; assigned_agents?: number }>; catalog: SkillCatalogStatus }>("/api/skills").then(({ skills, catalog }) => {
     wrap.append(skillCatalogBrowser(catalog));
     const shipped = h("section", { class: "space-y-3" },
-      h("div", { class: "flex flex-wrap items-end justify-between gap-2" }, h("div", {}, h("h3", { class: "font-display text-lg text-fg" }, "Installed arsenal"), h("p", { class: "text-sm text-muted" }, `${skills.length} complete playbooks · permanently available · selected automatically per task`))));
+      h("div", { class: "flex flex-wrap items-end justify-between gap-2" }, h("div", {}, h("h3", { class: "font-display text-lg text-fg" }, "Installed arsenal"), h("p", { class: "text-sm text-muted" }, `${skills.length} complete procedures · permanently available · loaded on demand`))));
     for (const skill of skills) shipped.append(h("article", { class: `card p-4 ${skill.arsenal_locked ? "opacity-80" : ""}` },
       h("div", { class: "flex flex-wrap items-center gap-2" },
         h("h3", { class: "font-semibold text-fg" }, skill.name),