feat(runtime): stabilize context, interactions, compaction, and capabilities

[Haz3-jolt]

Purpose

Stabilize the daemon-owned model context and complete runtime capabilities that every client can project consistently:

• expose only the active profile-aware tool roster while loading hierarchical project instructions deterministically
• add typed user questionnaires with canonical interaction lifecycle events for TUI and web clients
• add proactive deterministic compaction, one bounded provider-overflow recovery, and queued manual compaction during active responses
• add daemon-owned BM25 capability discovery, Agent Skills, session-persistent activation, and progressively disclosed daemon command tools

Fixes

• Completes OBS-902
• Advances OBS-896 and Build BM25 capability index and capability_search #16

Approach

• Keep prompt construction, resource discovery, capability indexing, activation, interaction ownership, and compaction decisions in the daemon and kernel.
• Discover AGENTS.md from the nearest Git repository root through the session working directory, prefer same-directory AGENTS.override.md, and reject invalid or escaping resources.
• Record exact loaded-resource snapshots canonically, project them through the SDK, retain them across restart, and rediscover them only on /reload.
• Add strict protocol contracts for questionnaires, compaction, and capability search, activation, denial, and resource reads.
• Build one deterministic BM25 index over compact capability metadata. Load full Skill bodies and referenced resources only after activation.
• Discover global Skills from ~/.axl/skills/ and ~/.agents/skills/, plus hierarchical project .axl/skills/ and .agents/skills/ directories.
• Keep activated Skills and tool schemas available for the rest of the session and restore them from canonical events after restart.
• Derive model-callable compact_context and reload_context tools from the authoritative daemon command registry rather than a parallel runtime list.
• Keep slash commands out of Bash and direct model requests for hidden operations through capability discovery.
• Compact before model requests when the configured token boundary is crossed, including between tool rounds. Retry one classified provider overflow only before output is exposed.
• Preserve complete tool groups, split-turn context, cumulative read and modified file tracking, steering and follow-up ordering, and the authoritative JSONL history.
• Advance the wire protocol from 17 to 21. The persisted event format remains version 1.

How was this tested?

• pnpm check: passed with 947 tests passed and 8 environment-dependent sandbox tests skipped.
• uvx reuse lint: passed with 590 of 590 files carrying copyright and license information.
• pnpm audit --audit-level high: passed with no known vulnerabilities.
• git diff --check: passed.
• Live Anthropic testing passed for hierarchical instruction precedence, override rediscovery through /reload, canonical restart preservation, and post-restart reload.
• Live Azure OpenAI Responses testing passed for BM25 Skill discovery, recursive Skill activation, containment-checked resource reads, session-persistent activation, dynamically exposed daemon command tools, context reload, and restoration after a real daemon restart.
• The exact prompt can you run /reload searched for tool:reload-context, activated it, called reload_context, and produced a logged reload boundary.
• The sandboxed Observal flow confirmed that global Skill resources are readable only through the capability broker. Host-installed CLI execution remains intentionally unavailable inside the current sandbox.
• UI screenshots are not attached to this draft and must be added before marking it ready for review.

Learning

• Used the Agent Skills and AGENTS.md conventions as interoperability references while keeping discovery daemon-owned.
• Reviewed OpenCode's Skill implementation at revision 1364769e516289bcd805dd813c36e68391858ab9 as a read-only behavioral reference. Axl retains bounded search rather than embedding the complete Skill catalog.
• Reviewed Pi 0.85.0 compaction documentation and regression behavior as a read-only behavioral reference. The Axl implementation follows its own protocol and architecture boundaries.
• Capability activation works best as session state: discover once, activate once, and retain the exact canonical content and tool binding until the session ends.
• Model-callable daemon commands must derive from the authoritative command registry rather than a second hardcoded capability inventory.

Checklist

• I reviewed the complete diff.
• I added or updated the smallest relevant test for behavior changes.
• I ran the relevant formatting, lint, type-check, test, boundary, and license checks.
• Every new file has SPDX metadata, directly or through REUSE.toml.
• Every commit has a matching DCO Signed-off-by trailer.
• UI changes include screenshots attached to the pull request, not committed to the repository.

AI assistance

• Generative AI materially assisted this change. Tool and model/version: pi coding agent; the model/version was not exposed to the session.
• I manually reviewed, understood, and tested the generated work.