Tighten the /permissions report into the same operator-console style used by
other slash commands, and make permission mode changes read like a structured
CLI confirmation instead of a raw field swap.
Constraint: Must keep the real permission surface limited to read-only, workspace-write, and danger-full-access
Rejected: Add synthetic shortcuts or approval-state variants | would misrepresent actual supported modes
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep /permissions output aligned with other structured slash command reports as new mode metadata is added
Tested: cargo fmt --manifest-path ./rust/Cargo.toml --all; cargo clippy --manifest-path ./rust/Cargo.toml --workspace --all-targets -- -D warnings; cargo test --manifest-path ./rust/Cargo.toml --workspace; manual REPL smoke test for /permissions and /permissions read-only
Not-tested: Interactive approval prompting flows beyond mode report formatting
The runtime already knew how to spawn stdio MCP processes, but it still
needed transport primitives for framed JSON-RPC exchange. This change adds
minimal request/response types, line and frame helpers on the stdio wrapper,
and an initialize roundtrip helper so later MCP client slices can build on a
real transport foundation instead of raw byte plumbing.
Constraint: Keep the slice small and limited to stdio transport foundations
Constraint: Must verify framed request write and typed response parsing with a fake MCP process
Rejected: Introduce a broader MCP session layer now | would expand the slice beyond transport framing
Rejected: Leave JSON-RPC as untyped serde_json::Value only | weakens initialize roundtrip guarantees
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Preserve the camelCase MCP initialize field mapping when layering richer protocol support on top
Tested: cargo fmt --all --manifest-path rust/Cargo.toml
Tested: cargo clippy -p runtime --all-targets --manifest-path rust/Cargo.toml -- -D warnings
Tested: cargo test -p runtime --manifest-path rust/Cargo.toml
Not-tested: Integration against a real external MCP server process
The dirty stdio slice had two real regressions in its new JSON-RPC test coverage: the embedded Python helper was written with broken string literals, and direct execution of the freshly written helper could fail with ETXTBSY on Linux. The repair keeps scope inside mcp_stdio.rs by fixing the helper strings and invoking the JSON-RPC helper through python3 while leaving the existing stdio process behavior unchanged.
Constraint: Keep the repair limited to rust/crates/runtime/src/mcp_stdio.rs
Constraint: Must satisfy fmt, clippy -D warnings, and runtime tests before shipping
Rejected: Revert the entire JSON-RPC stdio coverage addition | unnecessary once the helper/test defects were isolated
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep ephemeral stdio test helpers portable and avoid directly execing freshly written scripts when an interpreter invocation is sufficient
Tested: cargo fmt --all; cargo clippy -p runtime --all-targets -- -D warnings; cargo test -p runtime
Not-tested: Cross-platform behavior outside the current Linux runtime
Reformat /compact output for both live and resumed sessions so compaction results are reported in the same structured console style as the rest of the CLI surface. This keeps the behavior unchanged while making skipped and successful compaction runs easier to read.
Constraint: Compact output must stay faithful to the real compaction result and not imply summarization details beyond removed/kept message counts
Rejected: Expose the generated summary body directly in /compact output | too noisy for a lightweight command-response surface
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep lifecycle and maintenance command output stylistically consistent as more slash commands reach parity
Tested: cargo fmt --manifest-path ./rust/Cargo.toml --all; cargo clippy --manifest-path ./rust/Cargo.toml --workspace --all-targets -- -D warnings; cargo test --manifest-path ./rust/Cargo.toml --workspace
Not-tested: Manual terminal UX review of compact output on very large sessions
Reformat /init results into the same structured operator-console style used by the other polished commands so create and skip outcomes are easier to scan. This keeps the command behavior unchanged while making repo bootstrapping feedback feel more intentional.
Constraint: /init must stay non-destructive and continue refusing to overwrite an existing CLAUDE.md
Rejected: Expand /init to write more files in the same slice | broader scaffolding would be riskier than a focused UX polish commit
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep /init output explicit about whether the file was created or skipped so users can trust the command in existing repos
Tested: cargo fmt --manifest-path ./rust/Cargo.toml --all; cargo clippy --manifest-path ./rust/Cargo.toml --workspace --all-targets -- -D warnings; cargo test --manifest-path ./rust/Cargo.toml --workspace
Not-tested: Manual /init run in a repo that already has a heavily customized CLAUDE.md
Extend /config so operators can inspect specific merged sections like env, hooks, and model while keeping the command read-only and grounded in the actual loaded config. This improves Claude Code-style inspectability without inventing an unsafe config editing surface.
Constraint: Config handling must remain read-only and reflect only the merged runtime config that already exists
Rejected: Add /config set mutation commands | persistence semantics and edit safety are not mature enough for a small honest slice
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep config subviews aligned with real merged keys and avoid advertising writable behavior until persistence is designed
Tested: cargo fmt --manifest-path ./rust/Cargo.toml --all; cargo clippy --manifest-path ./rust/Cargo.toml --workspace --all-targets -- -D warnings; cargo test --manifest-path ./rust/Cargo.toml --workspace
Not-tested: Manual inspection of richer hooks/env config payloads in a customized user setup
Reformat /memory into the same structured console style as the other polished commands and enumerate discovered instruction files in ancestry order with line counts and previews. This makes repo instruction memory easier to inspect without changing the underlying discovery behavior.
Constraint: Memory reporting must reflect only the instruction files discovered from current directory ancestry
Rejected: Add memory editing commands in the same slice | presentation polish was a cleaner, lower-risk improvement to ship first
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep instruction-file ordering stable so ancestry-based memory debugging stays predictable
Tested: cargo fmt --manifest-path ./rust/Cargo.toml --all; cargo clippy --manifest-path ./rust/Cargo.toml --workspace --all-targets -- -D warnings; cargo test --manifest-path ./rust/Cargo.toml --workspace
Not-tested: Manual inspection of repos with many nested CLAUDE files
Extend /status with project root and git branch details derived from the local repository so the report feels closer to a real Claude Code session dashboard. This adds high-value workspace context without inventing any persisted metadata the runtime does not actually have.
Constraint: Status metadata must be computed from the current working tree at runtime and tolerate non-git directories
Rejected: Persist branch/root into session files first | a local runtime derivation is smaller and immediately useful without changing session format
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep status context opportunistic and degrade cleanly to unknown when git metadata is unavailable
Tested: cargo fmt --manifest-path ./rust/Cargo.toml --all; cargo clippy --manifest-path ./rust/Cargo.toml --workspace --all-targets -- -D warnings; cargo test --manifest-path ./rust/Cargo.toml --workspace
Not-tested: Manual non-git-directory /status run
Add a minimal runtime stdio MCP launcher that spawns configured server processes with piped stdin/stdout, applies transport env, and exposes async write/read/terminate/wait helpers for future JSON-RPC integration.
The wrapper stays intentionally small: it does not yet implement protocol framing or connection lifecycle management, but it is real process orchestration rather than placeholder scaffolding. Tests use a temporary executable script to prove env propagation and bidirectional stdio round-tripping.
Constraint: Keep the slice minimal and testable while using the real tokio process surface
Constraint: Runtime verification must pass cleanly under fmt, clippy, and tests
Rejected: Add full JSON-RPC framing and session orchestration in the same commit | too much scope for a clean launcher slice
Rejected: Fake the process wrapper behind mocks only | would not validate spawning, env injection, or stdio wiring
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Layer future MCP protocol framing on top of McpStdioProcess rather than bypassing it with ad hoc process management
Tested: cargo fmt --all; cargo clippy -p runtime --all-targets -- -D warnings; cargo test -p runtime
Not-tested: live third-party MCP servers; long-running process supervision; stderr capture policy
Update in-REPL /resume success output to the same structured console style used elsewhere so session lifecycle commands feel consistent with status, model, permissions, config, and cost. This preserves the same behavior while improving operator readability.
Constraint: Resume output must stay grounded in real restored session metadata already available after load
Rejected: Add more restored-session details like cwd snapshot | that data is not yet persisted in session files
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep lifecycle command outputs stylistically aligned as the CLI surface grows
Tested: cargo fmt --manifest-path ./rust/Cargo.toml --all; cargo clippy --manifest-path ./rust/Cargo.toml --workspace --all-targets -- -D warnings; cargo test --manifest-path ./rust/Cargo.toml --workspace
Not-tested: Manual interactive comparison of /resume output before and after multiple restores
Refresh shared slash help and REPL help wording so the command surface reads more like an integrated console, and make successful /clear output match the newer structured reporting style. This keeps discoverability consistent now that status, model, permissions, config, and cost all use richer operator-oriented copy.
Constraint: Help text must stay synchronized with the actual implemented command surface and resume behavior
Rejected: Larger README/doc pass in the same commit | keeping the slice limited to runtime help/output makes it easier to review and revert
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Prefer shared help-copy changes in commands crate first, then layer REPL-specific additions in the CLI binary
Tested: cargo fmt --manifest-path ./rust/Cargo.toml --all; cargo clippy --manifest-path ./rust/Cargo.toml --workspace --all-targets -- -D warnings; cargo test --manifest-path ./rust/Cargo.toml --workspace
Not-tested: Manual comparison of help wording against upstream Claude Code terminal screenshots
Reformat /cost for both live and resumed sessions so token accounting is presented in the same sectioned operator-console style as status, model, permissions, and config. This improves consistency across the command surface while preserving the same underlying usage metrics.
Constraint: Cost output must continue to reflect cumulative tracked usage only, without claiming real billing or currency totals
Rejected: Add dollar estimates | there is no authoritative pricing source wired into this CLI surface
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep /cost focused on raw token accounting until pricing metadata exists in the runtime layer
Tested: cargo fmt --manifest-path ./rust/Cargo.toml --all; cargo clippy --manifest-path ./rust/Cargo.toml --workspace --all-targets -- -D warnings; cargo test --manifest-path ./rust/Cargo.toml --workspace
Not-tested: Manual terminal UX review for very large cumulative token counts
Rework /permissions output into the same operator-console format used by status, config, and model so the command feels intentional and self-explanatory. Switching modes now reports previous and current state, while inspection shows the available modes and their meaning without adding fake policy logic.
Constraint: Permission output must stay aligned with the real three-mode runtime policy already implemented
Rejected: Add richer permission-policy previews per tool | would require more UI surface and risks overstating current policy fidelity
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep permission-mode docs in the CLI consistent with normalize_permission_mode and permission_policy behavior
Tested: cargo fmt --manifest-path ./rust/Cargo.toml --all; cargo clippy --manifest-path ./rust/Cargo.toml --workspace --all-targets -- -D warnings; cargo test --manifest-path ./rust/Cargo.toml --workspace
Not-tested: Manual operator UX review of /permissions flows in a live REPL
Move the default Agent artifact store out of rust/crates/tools so repeated Agent runs stop generating noisy crate-local files, normalize explicit Agent names through the existing slug path, and ignore any crate-local .clawd-agents residue defensively. Keep the slice limited to the tools crate and preserve the existing manifest-writing behavior.
Constraint: Must not touch unrelated dirty api files in this worktree
Constraint: Keep the change limited to rust/crates/tools
Rejected: Add a broader agent runtime or execution model | outside the final cleanup slice
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep Agent persistence defaults outside package directories so generated artifacts do not pollute crate working trees
Tested: cargo test -p tools
Not-tested: concurrent multi-process Agent writes to the default fallback store
Replace terse /model strings with sectioned model reports that show the active model and preserved session context, and use a structured switch report when the model changes. This keeps the behavior honest while making model management feel more intentional and Claude-like.
Constraint: Model switching must preserve the current session and avoid adding any fake model catalog or validation layer
Rejected: Add a hardcoded model list or aliases | would create drift with actual backend-supported model names
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep /model output informational and backend-agnostic unless the runtime gains authoritative model discovery
Tested: cargo fmt --manifest-path ./rust/Cargo.toml --all; cargo clippy --manifest-path ./rust/Cargo.toml --workspace --all-targets -- -D warnings; cargo test --manifest-path ./rust/Cargo.toml --workspace
Not-tested: Manual interactive switching across multiple real Anthropic model names
Require an explicit /clear --confirm flag before wiping live or resumed session state. This keeps the command genuinely useful while adding the minimal safety check needed for a destructive command in a chatty terminal workflow.
Constraint: /clear must remain a real functional command without introducing interactive prompt machinery that would complicate REPL input handling
Rejected: Add y/n interactive confirmation prompt | extra stateful prompting would be slower to ship and more fragile inside the line editor loop
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep destructive slash commands opt-in via explicit flags unless the CLI gains a dedicated confirmation subsystem
Tested: cargo fmt --manifest-path ./rust/Cargo.toml --all; cargo clippy --manifest-path ./rust/Cargo.toml --workspace --all-targets -- -D warnings; cargo test --manifest-path ./rust/Cargo.toml --workspace
Not-tested: Manual keyboard-driven UX pass for accidental /clear entry in interactive REPL
Add a minimal runtime MCP client bootstrap layer that turns typed MCP configs into concrete transport targets with normalized names, tool prefixes, signatures, and auth requirements.
This is intentionally scaffolding rather than a live connection manager: it creates the real data model the runtime will need to launch stdio, remote, websocket, sdk, and claude.ai proxy clients without prematurely coupling the code to any specific async transport implementation.
Constraint: Keep the slice real and minimal without adding connection lifecycle complexity yet
Constraint: Runtime verification must stay green under fmt, clippy, and tests
Rejected: Implement live connection/session orchestration in the same commit | too much surface area for a clean foundational slice
Rejected: Leave bootstrap shaping implicit in future transport code | would duplicate transport mapping and weaken testability
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Build future MCP launch/execution code by consuming McpClientBootstrap/McpClientTransport rather than re-parsing config enums ad hoc
Tested: cargo fmt --all; cargo clippy -p runtime --all-targets -- -D warnings; cargo test -p runtime
Not-tested: live MCP server processes; remote stream handshakes; tool/resource enumeration against real servers
Reformat /status and /config into sectioned reports with stable labels so the CLI surfaces read more like a usable operator console and less like dense debug strings. This improves discoverability and parity feel without changing the underlying data model or inventing fake settings behavior.
Constraint: Output polish must preserve the exact locally discoverable facts already exposed by the CLI
Rejected: Add interactive /clear confirmation first | wording/layout polish was cleaner, lower-risk, and touched fewer control-flow paths
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep CLI reports sectioned and label-stable so future tests can assert on intent rather than fragile token ordering
Tested: cargo fmt --manifest-path ./rust/Cargo.toml --all; cargo clippy --manifest-path ./rust/Cargo.toml --workspace --all-targets -- -D warnings; cargo test --manifest-path ./rust/Cargo.toml --workspace
Not-tested: Manual terminal-width UX review for very long paths or merged JSON payloads
Teach Skill path resolution to accept the common $skill invocation form in addition to bare names and /skill prefixes. Keep the behavior narrow and add regression coverage using the existing help skill fixture.
Constraint: Must not touch unrelated dirty api files in this worktree
Constraint: Keep the change limited to rust/crates/tools
Rejected: Canonicalize the returned skill field to the resolved name | would change caller-visible output semantics unnecessarily
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep invocation-prefix normalization aligned with how prompt and skill references are written elsewhere in the CLI
Tested: cargo test -p tools
Not-tested: CODEX_HOME layouts with unusual symlink arrangements
Accept case-insensitive domain filters and URL-style allow/block list entries so WebSearch behaves more forgivingly for caller-provided domain constraints. Keep the change small and limited to host matching logic plus regression coverage.\n\nConstraint: Must not touch unrelated dirty api files in this worktree\nConstraint: Keep the change limited to rust/crates/tools\nRejected: Add full public suffix or hostname normalization logic | too broad for this parity slice\nConfidence: high\nScope-risk: narrow\nReversibility: clean\nDirective: Preserve simple host matching semantics unless upstream parity proves a more exact domain model is required\nTested: cargo test -p tools\nNot-tested: internationalized domain names and punycode edge cases
Make title-focused WebFetch prompts prefer the real HTML <title> value when present instead of always falling back to the first rendered text line. Keep the behavior narrow and preserve the existing summary path for non-title prompts.\n\nConstraint: Must not touch unrelated dirty api files in this worktree\nConstraint: Keep the change limited to rust/crates/tools\nRejected: Broader HTML parsing dependency | not needed for this small parity slice\nConfidence: high\nScope-risk: narrow\nReversibility: clean\nDirective: Preserve lightweight HTML handling unless parity requires a materially more robust parser\nTested: cargo test -p tools\nNot-tested: malformed HTML with mixed-case or nested title edge cases
Tighten the PowerShell tool to surface a clear not-found error when neither pwsh nor powershell exists, and mark explicit background execution as user-requested in the returned metadata. Harden the PowerShell tests against PATH mutation races while keeping the change confined to the tools crate.\n\nConstraint: Must not touch unrelated dirty api files in this worktree\nConstraint: Keep the change limited to rust/crates/tools\nRejected: Broader shell abstraction cleanup | not needed for this parity slice\nConfidence: high\nScope-risk: narrow\nReversibility: clean\nDirective: Keep PowerShell output metadata aligned with bash semantics when adding future shell parity improvements\nTested: cargo test -p tools\nNot-tested: real powershell.exe behavior on Windows hosts
Add runtime MCP helpers for name normalization, tool naming, CCR proxy URL unwrapping, config signatures, and stable scope-independent config hashing.
This is the fastest clean parity-unblocking MCP slice because it creates real reusable behavior needed by future client/transport work without forcing a transport boundary prematurely. The helpers mirror key upstream semantics around normalized tool names and dedup/config-change detection.
Constraint: Must land a real MCP foundation without pulling transport management into the same commit
Constraint: Runtime verification must pass with fmt, clippy, and tests
Rejected: Start with transport/client scaffolding first | would need more design surface and more unverified edges
Rejected: Leave normalization/signature logic implicit in later client code | would duplicate behavior and complicate testing
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Reuse these helpers for future MCP tool naming, dedup, and reconnect/change-detection work instead of re-encoding the rules ad hoc
Tested: cargo fmt --all; cargo clippy -p runtime --all-targets -- -D warnings; cargo test -p runtime
Not-tested: live MCP transport connections; plugin reload integration; full connector dedup flows
Expand /status so it reports the current working directory, whether the CLI is operating on a live REPL or resumed session file, how many Claude config files were loaded, and how many instruction memory files were discovered. This makes status feel more like an operator dashboard instead of a bare token counter while still only surfacing metadata we can inspect locally.
Constraint: Status must only report context available from the current filesystem and session state
Rejected: Include guessed project metadata or upstream-only fields | would make the status output look richer than the implementation actually is
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep status additive and local-truthful; avoid inventing context that is not directly discoverable
Tested: cargo fmt --manifest-path ./rust/Cargo.toml --all; cargo clippy --manifest-path ./rust/Cargo.toml --workspace --all-targets -- -D warnings; cargo test --manifest-path ./rust/Cargo.toml --workspace
Not-tested: Manual interactive comparison of REPL /status versus resumed-session /status
Normalize Agent subagent aliases to Claude Code style built-in names, expose richer handoff metadata, teach ToolSearch to match canonical tool aliases, and polish NotebookEdit so delete does not require source and insert without a target appends cleanly. These are small parity-oriented behavior fixes confined to the tools crate.\n\nConstraint: Must not touch unrelated dirty api files in this worktree\nConstraint: Keep the change limited to rust/crates/tools\nRejected: Rework Agent into a real scheduler | outside this slice and not a small parity polish\nRejected: Add broad new tool surface area | request calls for small real parity improvements only\nConfidence: high\nScope-risk: narrow\nReversibility: clean\nDirective: Keep Agent built-in type normalization aligned with upstream naming aliases before expanding execution semantics\nTested: cargo test -p tools\nNot-tested: integration against a real upstream Claude Code runtime
Improve top-level help and shared slash-command help so the implemented surface is easier to discover, with explicit resume-safe markings and concrete examples for saved-session workflows. This keeps the command registry authoritative while making the CLI feel less skeletal and more like a real operator-facing tool.
Constraint: Help text must reflect the actual implemented surface without advertising unsupported offline/runtime behavior
Rejected: Separate bespoke help tables for REPL and --resume | would drift from the shared command registry
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Add new slash commands to the shared registry first so help and resume capability stay synchronized
Tested: cargo fmt --manifest-path ./rust/Cargo.toml --all; cargo clippy --manifest-path ./rust/Cargo.toml --workspace --all-targets -- -D warnings; cargo test --manifest-path ./rust/Cargo.toml --workspace
Not-tested: Manual UX comparison against upstream Claude Code help output
Extend --resume so operators can run multiple safe slash commands in sequence against a saved session file, including mutating maintenance actions like /compact and /clear plus useful local /init scaffolding. This brings resumed sessions closer to the live REPL command surface without pretending unsupported runtime-bound commands work offline.
Constraint: Resumed sessions only have serialized session state, not a live model client or interactive runtime
Rejected: Support every slash command under --resume | model and permission changes do not affect offline saved-session inspection meaningfully
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep --resume limited to commands that can operate purely from session files or local filesystem context
Tested: cargo fmt --manifest-path ./rust/Cargo.toml --all; cargo clippy --manifest-path ./rust/Cargo.toml --workspace --all-targets -- -D warnings; cargo test --manifest-path ./rust/Cargo.toml --workspace
Not-tested: Manual interactive smoke test of chained --resume commands in a shell session
Extend the Rust tools crate with NotebookEdit, Sleep, and PowerShell support. NotebookEdit now performs real ipynb cell replacement, insertion, and deletion; Sleep provides a non-shell wait primitive; and PowerShell executes commands with timeout/background support through a detected shell. Tests cover notebook mutation, sleep timing, and PowerShell execution via a stub shell while preserving the existing tool slices.\n\nConstraint: Keep the work confined to crates/tools/src/lib.rs and avoid staging unrelated workspace edits\nConstraint: Expose Claude Code-aligned names and close JSON-schema shapes for the new tools\nRejected: Stub-only notebook or sleep registrations | not materially useful beyond discovery\nRejected: PowerShell implemented as bash aliasing only | would not honor the distinct tool contract\nConfidence: medium\nScope-risk: moderate\nReversibility: clean\nDirective: Preserve the NotebookEdit field names and PowerShell output shape so later runtime extraction can move implementation without changing the contract\nTested: cargo fmt; cargo test -p tools\nNot-tested: cargo clippy; full workspace cargo test
Add a genuinely useful /init command that creates a starter CLAUDE.md from the current repository shape without inventing unsupported setup flows. The scaffold pulls in real verification commands and repo-structure notes for this workspace, and it refuses to overwrite an existing CLAUDE.md.
This keeps the command honest and low-risk while moving the CLI closer to Claude Code's practical bootstrap surface.
Constraint: /init must be non-destructive and must not overwrite an existing CLAUDE.md
Constraint: Generated guidance must come from observable repo structure rather than placeholder text
Rejected: Interactive multi-step init workflow | too much unsupported UI/state machinery for this Rust CLI slice
Confidence: high
Scope-risk: moderate
Reversibility: clean
Directive: Keep generated CLAUDE.md templates concise and repo-derived; do not let /init drift into fake setup promises
Tested: cargo fmt; cargo clippy --workspace --all-targets -- -D warnings; cargo test --workspace
Not-tested: manual /init invocation in a separate temporary repository without a preexisting CLAUDE.md
Add minimal runtime-side remote session and upstream proxy primitives that model enablement, session identity, token loading, websocket endpoint derivation, and subprocess proxy environment shaping.
This intentionally stops short of implementing the relay or CA download path. The goal is to land real request/env foundations that future remote integration work can build on while preserving the fail-open behavior of the upstream implementation.
Constraint: Must keep the slice minimal and real without pulling in relay networking yet
Constraint: Verification must pass with runtime fmt, clippy, and tests
Rejected: Implement full upstream CONNECT relay now | too large for the current bounded slice
Rejected: Hide proxy state behind untyped env maps only | would make later integration and testing brittle
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep remote bootstrap logic fail-open; do not make proxy setup a hard dependency for normal runtime execution
Tested: cargo fmt --all; cargo clippy -p runtime --all-targets -- -D warnings; cargo test -p runtime
Not-tested: live CCR session behavior; relay startup; CA bundle download and trust installation
Improve resumed-session parity by letting top-level --resume execute shared read-only commands such as /help, /status, /cost, /config, and /memory in addition to /compact. This makes saved sessions meaningfully inspectable without reopening the interactive REPL.
Also add a genuinely useful /memory command that reports the Claude instruction memory already discovered by the runtime from CLAUDE.md-style files in the current directory ancestry. The command stays honest by surfacing file paths, line counts, and a short preview instead of inventing unsupported persistent memory behavior.
Constraint: Resume-path improvements must operate safely on saved sessions without requiring a live model runtime
Constraint: /memory must expose real repository instruction context rather than placeholder state
Rejected: Invent editable or persistent chat memory storage | no such durable feature exists in this repo yet
Confidence: high
Scope-risk: moderate
Reversibility: clean
Directive: Reuse shared slash parsing for resume-path features so saved-session commands and REPL commands stay aligned
Tested: cargo fmt; cargo clippy --workspace --all-targets -- -D warnings; cargo test --workspace
Not-tested: manual resume against a diverse set of historical session files from real user workflows
Add runtime OAuth primitives for PKCE generation, authorization URL building, token exchange request shaping, and refresh request shaping. Wire the API client to a real auth-source abstraction so future OAuth tokens can flow into Anthropic requests without bespoke header code.
This keeps the slice bounded to foundations: no browser flow, callback listener, or token persistence. The API client still behaves compatibly for current API-key users while gaining explicit bearer-token and combined auth modeling.
Constraint: Must keep the slice minimal and real while preserving current API client behavior
Constraint: Repo verification requires fmt, tests, and clippy to pass cleanly
Rejected: Implement full OAuth browser/listener flow now | too broad for the current parity-unblocking slice
Rejected: Keep auth handling as ad hoc env reads only | blocks reuse by future OAuth integration paths
Confidence: high
Scope-risk: moderate
Reversibility: clean
Directive: Extend OAuth behavior by composing these request/auth primitives before adding session or storage orchestration
Tested: cargo fmt --all; cargo clippy -p runtime -p api --all-targets -- -D warnings; cargo test -p runtime; cargo test -p api --tests
Not-tested: live OAuth token exchange; callback listener flow; workspace-wide tests outside runtime/api
Add in-REPL session restoration and read-only config inspection so the CLI can recover saved conversations and expose Claude settings without leaving interactive mode. /resume now reloads a session file into the live runtime, and /config shows discovered settings files plus the merged effective JSON.
The new commands stay on the shared slash-command surface and rebuild runtime state using the current model, system prompt, and permission mode so existing REPL behavior remains stable.
Constraint: /resume must update the live REPL session rather than only supporting top-level --resume
Constraint: /config should inspect existing settings without mutating user files
Rejected: Add editable /config writes in this slice | read-only inspection is safer and sufficient for immediate parity work
Confidence: high
Scope-risk: moderate
Reversibility: clean
Directive: Keep resume/config behavior on the shared slash command surface so non-REPL entrypoints can reuse it later
Tested: cargo fmt; cargo clippy --workspace --all-targets -- -D warnings; cargo test --workspace
Not-tested: manual interactive restore against real saved session files outside automated fixtures
Extend the Rust tools crate with concrete Agent and ToolSearch implementations. Agent now persists agent-handoff metadata and prompt payloads to a local store with Claude Code-style fields, while ToolSearch supports exact selection and keyword search over the deferred tool surface. Tests cover agent persistence and tool lookup behavior alongside the existing web, todo, and skill coverage.\n\nConstraint: Keep the implementation tools-only without relying on full agent orchestration runtime\nConstraint: Preserve exposed tool names and close schema parity with Claude Code\nRejected: No-op Agent stubs | would not provide material handoff value\nRejected: ToolSearch limited to exact matches only | too weak for discovery workflows\nConfidence: medium\nScope-risk: narrow\nReversibility: clean\nDirective: Keep Agent output contract stable so later execution wiring can reuse persisted metadata without renaming fields\nTested: cargo fmt; cargo test -p tools\nNot-tested: cargo clippy; full workspace cargo test
This replaces the single default pricing assumption with a small model-aware pricing table for Sonnet, Opus, and Haiku so CLI usage output better matches the selected model. Unknown models still fall back cleanly with explicit labeling.
The change keeps pricing lightweight and local while improving the usefulness of usage/cost reporting for resumed sessions and live turns.
Constraint: Keep pricing local and dependency-free
Constraint: Preserve graceful fallback behavior for unknown model IDs
Rejected: Add a remote pricing source now | unnecessary coupling and risk for this slice
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: If pricing tables expand later, prefer explicit model-family matching and keep fallback labeling visible
Tested: cargo fmt; cargo clippy --all-targets --all-features -- -D warnings; cargo test -q
Not-tested: Validation against live provider billing exports
This adds a dedicated session inspect command to the Rust CLI so users can inspect a saved session's path, timestamps, size, token totals, preview text, and latest user/assistant context without opening the underlying file by hand.
It builds directly on the new session list/resume flows and keeps the UX lightweight and script-friendly.
Constraint: Keep session inspection CLI-native and read-only
Constraint: Reuse the existing saved-session format instead of introducing a secondary index format
Rejected: Add an interactive session browser now | more overhead than needed for this inspect slice
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep session inspection output stable and grep-friendly so it remains useful in scripts
Tested: cargo fmt; cargo clippy --all-targets --all-features -- -D warnings; cargo test -q
Not-tested: Manual inspection against a large corpus of real saved sessions
This upgrades Rust session compaction so summaries carry more than a flat timeline. The compacted state now calls out recent user requests, pending work signals, key files, and the current work focus so resumed sessions retain stronger execution continuity.
The change stays deterministic and local while moving the compact output closer to session-memory style handoff value.
Constraint: Keep compaction local and deterministic rather than introducing API-side summarization
Constraint: Preserve the existing resumable system-summary mechanism and compact command flow
Rejected: Add a full session-memory background extractor now | larger runtime change than needed for this incremental parity pass
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep future compaction enrichments biased toward actionable state transfer, not just verbose recap
Tested: cargo fmt; cargo clippy --all-targets --all-features -- -D warnings; cargo test -q
Not-tested: Long real-world sessions with deeply nested tool/result payloads
This improves Rust prompt-building by deduplicating repeated CLAUDE instruction content, surfacing clearer project-context metadata, and truncating oversized instruction payloads so local rules stay useful without overwhelming the runtime prompt.
The change preserves ancestor-chain discovery while making the rendered context more stable, compact, and readable for downstream compaction and CLI flows.
Constraint: Keep existing CLAUDE.md discovery semantics while reducing prompt bloat
Constraint: Avoid adding a new parser or changing user-authored instruction file formats
Rejected: Introduce a structured CLAUDE schema now | too large a shift for this parity slice
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: If richer instruction precedence is added later, keep duplicate suppression conservative so distinct local rules are not silently lost
Tested: cargo fmt; cargo clippy --all-targets --all-features -- -D warnings; cargo test -q
Not-tested: Live end-to-end behavior with very large real-world CLAUDE.md trees
This adds a prompt-mode permission flow for the Rust CLI, surfaces permission policy details in the REPL, and improves tool output rendering with concise human-readable summaries before the raw JSON payload.
The goal is to make tool execution feel safer and more legible without changing the underlying runtime loop or adding a heavyweight UI layer.
Constraint: Keep the permission UX terminal-native and incremental
Constraint: Preserve existing allow and read-only behavior while adding prompt mode
Rejected: Build a full-screen interactive approval UI now | unnecessary complexity for this parity slice
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Keep raw tool JSON available even when adding richer summaries so debugging fidelity remains intact
Tested: cargo fmt; cargo clippy --all-targets --all-features -- -D warnings; cargo test -q
Not-tested: Manual prompt-mode approvals against live API-driven tool calls
Expand the shared slash registry and REPL dispatcher with real session-management commands so the CLI feels closer to Claude Code during interactive use. /permissions now reports or switches the active permission mode, /clear rebuilds a fresh local session without restarting the process, and /cost reports cumulative token usage honestly from the runtime tracker.
The implementation keeps command parsing centralized in the commands crate and preserves the existing prompt-mode path while rebuilding runtime state safely when commands change session configuration.
Constraint: Commands must be genuinely useful local behavior rather than placeholders
Constraint: Preserve REPL continuity when changing permissions or clearing session state
Rejected: Store permission-mode changes only in environment variables | would not update the live runtime for the current session
Confidence: high
Scope-risk: moderate
Reversibility: clean
Directive: Keep future stateful slash commands rebuilding from current session + system prompt instead of mutating hidden runtime internals
Tested: cargo fmt; cargo clippy --workspace --all-targets -- -D warnings; cargo test --workspace
Not-tested: manual live API session exercising permission changes mid-conversation
Centralize slash command parsing in the commands crate so the REPL can share help metadata and grow toward Claude Code parity without duplicating handlers. This adds shared /help and /model parsing, routes REPL dispatch through the shared parser, and upgrades /status to report model and token totals.
To satisfy the required verification gate, this also fixes existing workspace clippy and test blockers in runtime, tools, api, and compat-harness that were unrelated to the new command behavior but prevented fmt/clippy/test from passing cleanly.
Constraint: Preserve existing prompt-mode and REPL behavior while adding real slash commands
Constraint: cargo fmt, clippy, and workspace tests must pass before shipping command-surface work
Rejected: Keep command handling only in main.rs | would deepen duplication with commands crate and resume path
Confidence: high
Scope-risk: moderate
Reversibility: clean
Directive: Extend new slash commands through the shared commands crate first so REPL and resume entrypoints stay consistent
Tested: cargo fmt; cargo clippy --workspace --all-targets -- -D warnings; cargo test --workspace
Not-tested: live Anthropic network execution beyond existing mocked/integration coverage
This adds a lightweight session home for the Rust CLI, auto-persists REPL state, and exposes list, search, show, and named resume flows so users no longer need to remember raw JSON paths.
The change keeps the old --resume SESSION.json path working while adding friendlier session discovery. It also makes API env-based tests hermetic so workspace verification remains stable regardless of shell environment.
Constraint: Keep session UX incremental and CLI-native without introducing a new database or TUI layer
Constraint: Preserve backward compatibility for the existing --resume SESSION.json workflow
Rejected: Build a richer interactive picker now | higher implementation cost than needed for this parity slice
Confidence: high
Scope-risk: moderate
Reversibility: clean
Directive: Keep human-friendly session lookup additive; do not remove explicit path-based resume support
Tested: cargo fmt; cargo clippy --all-targets --all-features -- -D warnings; cargo test -q
Not-tested: Manual multi-session interactive REPL behavior across multiple terminals
This adds token and estimated cost reporting to runtime usage tracking and surfaces it in the CLI status and turn output. It also upgrades compaction summaries so users see a clearer resumable summary and token savings after /compact.
The verification path required cleaning existing workspace clippy and test friction in adjacent crates so cargo fmt, cargo clippy -D warnings, and cargo test succeed from the Rust workspace root in this repo state.
Constraint: Keep the change incremental and user-visible without a large CLI rewrite
Constraint: Verification must pass with cargo fmt, cargo clippy --all-targets --all-features -- -D warnings, and cargo test
Rejected: Implement a full model-pricing table now | would add more surface area than needed for this first UX slice
Confidence: high
Scope-risk: moderate
Reversibility: clean
Directive: If pricing becomes model-specific later, keep the current estimate labeling explicit rather than implying exact billing
Tested: cargo fmt; cargo clippy --all-targets --all-features -- -D warnings; cargo test -q
Not-tested: Live Anthropic API interaction and real streaming terminal sessions
Extend the Rust tools crate with concrete TodoWrite and Skill implementations. TodoWrite now validates and persists structured session todos with Claude Code-aligned item shapes, while Skill resolves local skill definitions and returns their prompt payload for execution handoff. Tests cover persistence and local skill loading without disturbing the previously added web tools.\n\nConstraint: Stay within tools-only scope and avoid depending on broader agent/runtime rewrites\nConstraint: Keep exposed tool names and schemas close to Claude Code contracts\nRejected: In-memory-only TodoWrite state | would not survive across tool calls\nRejected: Stub Skill metadata without loading prompt content | not materially useful to callers\nConfidence: medium\nScope-risk: narrow\nReversibility: clean\nDirective: Preserve TodoWrite item-field parity and keep Skill focused on local skill discovery until agent execution wiring lands\nTested: cargo fmt; cargo test -p tools\nNot-tested: cargo clippy; full workspace cargo test
