feat: borrow pi-sharp patterns — model capabilities, execution mode, extensions#81
Merged
feat: borrow pi-sharp patterns — model capabilities, execution mode, extensions#81
Conversation
…roxy
Introduces a top-level `storage` subcommand that browses and mutates blobs in
the caller's aevatar chrono-storage bucket via the backend's explorer proxy
(`/api/explorer/*`). Reuses the aevatar chat config store for base-URL
resolution (so a single `/aevatar config set-url` also points `/storage` at
the same backend) and the existing NyxID bearer-token flow — no new auth
surface.
Subcommands:
aexon storage ls [prefix] rounded table of the scope manifest
aexon storage cat <key> GET, print as UTF-8 text
aexon storage get <key> [local-path] GET, save to file (or write bytes
to stdout when no path given —
binary fidelity preserved)
aexon storage put <key> <local-path> multipart upload (≤50 MB), infers
content-type from extension
aexon storage put-text <key> read stdin, PUT as text — infers
media type from key extension
aexon storage rm <key> DELETE
`AevatarStorageClient` escapes each path segment so keys with embedded slashes
(`chat-media/abc.png`) survive routing, and wraps failures with the response
body (capped at 500 chars) in `AevatarStorageException` so `403 /
file-not-found / blocked path` surface clearly.
Smoke-tested end-to-end against aevatar mainnet:
- `ls` rendered 17 real files (chat-histories, workflows, chat-media, config.json)
- `cat config.json` round-tripped the actual JSON
- `put-text` → `cat` → `rm` cycle intact
- `get <image.jpeg>` saved 33 KB binary, `file` recognized valid JFIF JPEG
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
… capabilities Plan covers three tasks borrowed from pi-sharp: - Task A: enrich ModelCapability flags and add ModelPricing - Task B: add ToolBatchExecutionMode override (Auto/Sequential/Parallel) - Task C: introduce IAexonExtension + minimal IAexonSessionBuilder Full session API boundary (pi-sharp #2) is deferred to #79.
The LoadMemoryAsync tests that assert "no files → null" were picking up the developer's actual ~/.claude/CLAUDE.md because UserClaudeDirectory / SystemClaudeDirectory defaulted to real paths. Point them at per-test temp directories so the assertion is deterministic. Caught while running the Task A test gate. Co-Authored-By: Codex <noreply@openai.com>
- Extend ModelCapability with ImageInput/Streaming/ToolCalling/Reasoning/PromptCaching - Add ModelPricing record with per-1M rates and EstimateCostUsd helper - Attach Capabilities + Pricing to every ClaudeModelDescriptor - Router now sources model capabilities from descriptor, layers WebFetch/WebSearch on top per provider-kind rules Co-Authored-By: Codex <noreply@openai.com>
…cutor Expose Auto/Sequential/Parallel modes via QueryEngineConfig. Auto is the default and preserves per-tool IsConcurrencySafe behavior; Sequential/Parallel force the whole batch one way or the other for debugging and explicit fan-out.
Extensions run exactly once during CLI startup, before QueryEngine is built, and can register tools, commands, hook observers, and system-prompt fragments through a builder surface. The builder is discarded after all extensions have configured the session. No dynamic DLL discovery yet (in-process registration only). Scope is deliberately minimal; the full session API façade is tracked separately.
# Conflicts: # src/Aexon.Commands/StorageCommand.cs
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Borrow three patterns from pi-sharp into Aexon. Plan in
docs/superpowers/plans/2026-04-18-session-extensions-and-model-capabilities.md.ModelCapability(addImageInput/Streaming/ToolCalling/Reasoning/PromptCaching) and introduceModelPricingrecord attached to everyClaudeModelDescriptor. Router sources model capabilities from descriptors, layeringWebFetch/WebSearchon top per provider-kind rules. Pricing values are copied from the existing authoritative values inBuiltinCommands.cs.ToolBatchExecutionMode(Auto/Sequential/Parallel) onQueryEngineConfig.Autois the default and preserves per-toolIsConcurrencySafebehavior;Sequential/Parallelforce the whole batch one way or the other for debugging and explicit fan-out.IAexonExtension+ minimalIAexonSessionBuilderinAexon.Core/Extensions/. Extensions run exactly once beforeQueryEngineis built and can register tools, commands, hook observers, and system-prompt fragments. ConcreteSessionBuilderlives inAexon.Cli. No dynamic DLL discovery in this task — in-process registration only.Also: 74f41e6 fixes a pre-existing
ContextProvidertest isolation bug (the memory tests were reading the developer's real~/.claude/CLAUDE.md); caught while running the Task A test gate and split out into its own commit perAGENTS.mdscope discipline.Task #2 from the pi-sharp comparison (full
AexonSessionAPI boundary) is deferred to #79.Test plan
dotnet build Aexon.slnx -c Debug --nologo— clean (one pre-existing CS8425 warning, unrelated)dotnet test tests/Aexon.Core.Tests/Aexon.Core.Tests.csproj— 805 passed, 0 failed (+14 new tests across Tasks A/B/C)aexonREPL still starts with zero registered extensions (behavior byte-identical to baseline)Notes for reviewer
ClaudeModelCataloguses named capability constants (StandardCapabilities,ReasoningCapabilities) and named pricing records (Haiku35Pricing,SonnetPricing, etc.) rather than inline literals — easier to audit and update.SessionBuilderthrowsInvalidOperationExceptionif an extension tries to register a tool/command whose name is already taken (loaded or deferred). This is intentional — extensions cannot silently override built-ins.Co-Authored-Bytrailers on e2e96d2 and 74f41e6). Tasks B and C were implemented directly after Codex's rescue flow silently failed mid-flight on Task B.Deferred follow-ups (explicitly out of scope)
ModelPricinginto/costand/statscommandsBuiltinCommands.cshardcoded pricing now that descriptors carry pricingToolBatchExecutionModevia a CLI flag or/modesub-command~/.aexon/extensions/*.dllAexonSessionAPI façade — tracked in RFC: AexonSession as a first-class API boundary (full #2) #79