Repository navigation
v0.3.0
HUGE Update
Dependency maintenance
Refresh Rust dependencies, including PyO3 0.29.3, Crossterm 0.29, Inquire 0.9.4, Indicatif 0.18.6, Console 0.16.6 and Tokio 1.53.3. Update the Cargo lockfile and validate with stable Rust 1.99.
Pin Ruff 0.17.0, ty 0.0.87 and build 1.6.1 in requirements-dev.txt for consistent contributor and CI checks. Update installed Python runtime integrations, retaining ProtoLink 0.8.1 as the engine. Bound Anthropic below 1.0 because its new Messages API removes generation parameters still sent by ProtoLink's adapter.
Update Docusaurus to 3.10.2, React to 19.3.1 and the docs lockfile. Retain ELK 0.1.9 as required by the Docusaurus Mermaid theme.
Refresh GitHub Actions and use Node 24 for documentation CI.
Added
Compact styled Edits cards in CLI/TUI, with green additions, red removals, bold literal file names and muted diff hints. Verification summaries fit one line, with full logs retained in /trace.
Missing Python runners/external dependencies end final verification quietly before another model request, preserving applied edits as unverified without install questions or code-repair loops. Genuine regressions retain bounded repair; native failed receipts and stop conditions remain authoritative.
Small-profile Gemma 4 automatic action acquisition uses ProtoLink's JSON protocol when advertised native tooling emits narration/channel/action text. Explicit native/json settings remain authoritative. Missing verification receipts are labeled not run, stopped-before-edit results report no changes, and Python dependency/runner failures get environment/setup explanations.
Parse-only Python afterimage guards for create, replace and exact edits reject introduced syntax errors before approval through native no-change feedback. Literal escapes remain unchanged; already malformed sources can be repaired incrementally, and valid proposals retain native write authorization/recovery.
Per-project /approvals picker and proto-cli approvals controls: Ask (default), automatic file edits, or automatic edits and commands. Exact native broker decisions retain policies, revision checks, checkpoints and audit records; MCP access and restoration still ask. The TUI displays the active mode.
Deterministic end-of-run applied-file summaries and combined diffs from native snapshots, including partial failures. /diff shows applied changes and /diff proposed labels proposal previews separately. Pre-existing work, rejected proposals, external edits and uncertain effects are distinguished.
Model-free project inspection in shell/TUI: ls, browse/files, open and cat. Folder browsing supports filtering, parent navigation and Enter to view text; the read-only viewer has line numbers, paging and horizontal panning. Listings, source loading and transcript previews are bounded; paths remain in the active project, and terminal control characters are displayed inertly. /open now views files; /project selects a workspace. Guide and completion share the catalog.
Runtime-owned TaskRecord with criteria, check selection, scoped worker packets and explicit done/needs_context/blocked worker reports.
Stateless, read-only Tester for regression planning and failure analysis.
Persistent optional-agent controls for Tester (default on), Scout and MCP (default off). Disabled workers are not constructed or registered; required Architect/Explorer/Coder/Verifier roles cannot be disabled.
mcp and /mcp setup, explicit discovery probes, server/tool allowlists and environment-referenced HTTP authentication. Model-free MCP broker with three fixed discovery/schema/call tools, native approvals and bounded model handoffs.
Repository check discovery, .protoagent/project.json configuration, and Verifier run_check(check_id, phase) with baseline and final verification.
Coder bounded source reads and exact single-occurrence edit_file with a source revision guard, native diff approval and native recovery.
Compact small-profile protocol prompts and worker cards with exact input schemas, omitting repetitive examples and large output schemas.
Per-request context admission through native ProtoLink policies and hooks, reserving output tokens and retaining the current task and runtime record. Small profiles with unknown capacity receive an 8192-token application cap.
Disposable coding evals with independent acceptance scripts and a same-model single-agent baseline: proto-cli eval coding --plan|--live.
Owned local subagents through ProtoLink 0.8: independent conversations, inherited policies, shared Graph budgets, sequential dispatch, depth one and configurable per-attempt child limits. Default execution needs no Registry server or loopback sockets; explicit transport meshes remain available.
Native ContextPolicy and AgentHooks for current task state, complete-turn pruning and scoped progressive retrieval of large tool/child observations.
Optional explicit same-provider fallback_models through native RoutedLLM, with bounded transient request fallback, shared accounting and no tool replay.
eval harness [--json]: native evaluate() runs repeated offline application read/task-state cases with fresh factories and linked child receipt checks.
Architect-only native ask_user for clarification and same-task continuation. TUI/shell question input supports explicit suggestions, free text, skip and cancellation, correlated replies, native deadlines and no implicit approval.
Changed
Admit up to three progressing native Coder work windows in one assignment, rather than interrupting every nine calls. Each continuation requires fresh applied writes, current revisions and no native stop or explicit missing-context report. Scope, history, receipts, approvals and aggregate budgets are preserved. A stalled window cannot borrow earlier progress. PROTOAGENT_CODER_MAX_WINDOWS lowers the application ceiling (1–3, default three); worker.continuing updates the transient TUI activity. Native inference limits remain unchanged.
Return exhausted partial Coder assignments directly to the bounded Graph before another Architect model call, avoiding routine continuation questions that spend the run deadline. workflow.work_yielded records the transition; fresh Graph nodes can still inspect and narrow remaining work.
Explain native runtime expiry, including expiry while awaiting a user answer, rather than presenting it only as an unresolved worker/check failure.
Return completed failing final checks directly to the native repair Graph before another Architect model request. This prevents edits in a closed check phase from turning an eligible repair into a policy denial. After two repairs, remaining failures finish incomplete. Missing dependency failures can still prompt for a manual verification handoff in the repair attempt.
Deduplicate live native events across the whole workflow, so continuation and enclosing handles do not replay old activity, answer chunks or command output. Complete native reports retain their execution evidence.
Continue stable, applied partial Coder work when Architect returns too early: the native Graph permits at most two continuations before final verification, preserving the frozen plan, edit phase and shared budget. Execution failures, denials, uncertainty, no-progress handoffs and manual verification stops do not qualify. Exhausted continuations report incomplete work without retaining unfinished tool-call prose as the final answer.
Small-profile Ollama requests default to at most 4096 generated tokens; small Gemma 4 variants default to think=false for short coding actions. Explicit num_predict/max_tokens and thinking settings win; sampling is preserved. Generation-repeat aborts now explain the provider failure separately from agent turn limits; partial streamed tool proposals are never executed/replayed.
Statically recognize pytest functions/imports/test classes when selecting a Python runner, including test/ directories. Preserve unittest for its own suites and owner-declared checks. Root pytest checks retain current moved test files/configuration as dependencies; discovery never imports project code or installs a missing test runner.
Enforce Coder's work-slice handoff before the last native inference turn, without requiring a final model reply. A public Agent extension consumes only the harness's pre-model stop signal, preserving native execution and stop conditions. Return recorded partial edits, or a blocked no-progress report. Architect can assign only unfinished work within the shared run budget; unresolved handoffs never satisfy completion. Reject identical exact edits before approval and return the approved source digest after successful edits to avoid redundant reads. Failed runs show applied paths and readable causes while retaining the complete native error and receipts in the trace.
Hide fallback tool/delegation JSON from native-capable model streams, reusing ProtoLink normalization for display classification. Preserve JSON answer data and typed final content; clear provisional narration at native tool selection so the muted work status returns instead of retaining an action as an answer.
Show one muted, transient activity line while an answer is loading. Native work phases replace it in place, distinguish command preparation from running, and disappear when text streams or the run ends. Saved answers retain no loading labels; phase updates survive the live trace preview cap.
Resolve bare command programs and project-relative working directories before native approval, preserving explicit environments and interpreter identity. Default output limits fit the configured ceiling. Small local calls return native no-change correction feedback for invalid preparation before dispatch.
Preserve root unittest discovery when tests move into packages; capture moved tests and package markers as current check dependencies without changing the frozen runner. Agent/Guide prompts explain layout preparation and discovery.
Honor explicit manual verification answers immediately through native cancellation, without another model call or repair. Report applied files as unverified with verification deferred to the user, preserving failed checks and native transport status. Uncertain effects and stale/denied work still stop.
Prefer the embedded CLI's injected virtual environment over bare framework Python when a project has no local environment. Embedded regression coverage requires subprocess and coding-oracle imports to see ProtoLink dependencies.
Clarify the 120-message live trace preview cap and continue delivering later decisions and execution outcomes. Full stored native traces remain complete.
Return stale/non-unique exact-edit selections through read-only validation feedback so ProtoLink's native correction loop can recover before a write is prepared. Valid edits retain native write approval and recovery; ordinary failures, denial and uncertain effects still stop. Coder copies short observed source spans and paginates truncated reads.
Preserve default checks for an empty code plan selection and reject narrowed bootstrap scopes that omit the required regression test files before mutation.
Size each turn's worker roster, tools, prompts and native permission ceilings to its original request. Conversation skips project/check discovery and uses Architect alone; repository questions cannot invoke Coder or commands; explicit verification enables Verifier without editing authority. Recognize verification clauses attached to questions and keep skipped clarification from becoming an invented implementation requirement.
Distinguish command/check approvals from file-change approvals in the TUI. V views details/diffs without approving; Y runs/applies, including from reviews. Native expiry/cancellation closes both views and stale decisions are dropped. Approval wait expiry reports a clear reason instead of missing-evidence jargon.
Expose available baseline IDs and avoid empty bootstrap baseline approvals during coding tasks. Source inspection, changes and focused regressions precede final checking when a project has no tests yet.
Expand Coder's project-relative create/replace paths before native preparation, preserving native traversal, symlink, root, preimage, diff and approval checks.
Simplify the small-profile local worker interface with typed Explorer reads, Coder/Tester assignments and Verifier calls. Fixed prepared routes use native owned-child supervision and retain each worker's policies, approvals, budgets, cancellation and receipts. Optional Scout/MCP routes follow enabled settings. Hide redundant packet preparation and repeated child schemas from this profile; narrow Coder's tool list and make its return-after-edit boundary explicit. Other profiles and transport meshes retain native delegation.
Clarify worker routing for small models after user questions: worker tools require separate worker/tool fields rather than invented combined names. Compact prompts include the selected check's exact baseline call and distinguish preparing a worker packet from executing Coder. Native correction limits remain.
Restrict worker-packet roles to explicit schema choices, return an unexecuted status and dispatch reminder, and show an exact Coder inference example.
Resolve external Python for discovered checks and coding evals when CPython is embedded in the Rust CLI. Preserve project venv identity and exclude the CLI host executable; embedded-runtime tests cover discovery and oracle execution.
ProtoLink 0.8.1 owns action normalization/correction, Ollama automatic capability discovery and task-local protocol selection. Remove the application metadata probe/cache and final-answer rejection hook; retain compact coding context and legacy history preparation through the native parser.
Native ToolRejectedError replaces custom planning feedback and handles invalid worker packets/reports, stale or ambiguous exact-edit proposals and invalid check requests before effects. Rejections have distinct events and inference artifacts, never completion evidence. Provider max_tool_corrections defaults to two per inference; native tools/schema/delegation errors share it, including the outer RoutedLLM loop. Successful actions do not reset it.
Compact worker cards retain native supports_inference; the engine rejects unsupported inference, unknown worker tools and invalid schemas before dispatch. CLI progress, timeline and trace distinguish rejection and correction exhaustion from execution and denial. check includes native action-validation readiness.
Python projects without repository checks receive a predefined unittest runner for new root regression tests. Compact task records expose valid and bootstrap check IDs; empty suites still cannot verify changes. Planning mistakes return corrective feedback without changing the plan, and omitted check IDs preserve defaults. Explicit project configuration stays authoritative.
Ollama models advertising tools use ProtoLink's native tool channel, with native cached metadata discovery and explicit auto/native/json overrides. Small JSON prompts include an Explorer README delegation example. Earlier malformed proposals become notes in model context while valid historical actions, explicit final data and original reports remain intact.
Code changes require selected repository checks executed at current revisions. Preparation and unrelated successful commands do not satisfy verification. Applied, verified and criteria-supported states are reported separately.
Baseline and preparation commands allow subsequent edits; final checks close the edit phase and preserve the two-repair ceiling.
Source tools return one bounded source span with line range and SHA-256 revision; the duplicated raw_content field is removed.
Explicit small-model hints take priority over provider hosting location.
Intent classification handles polite requests and explicit no-write instructions.
CLI/core/docs coordinated versions advance to 0.3.0; ProtoLink floor advances to 0.8.1 with the HTTP, LLM and MCP extras.
Replace model acquisition wrappers with native context policies and lifecycle hooks. The small profile retains compact declarations and exact input schemas.
Present local child output from native task records before join, deduplicated against the parent stream, preserving live model/process previews.
Migration and limits
Existing code-change runs without a repository check now remain incomplete. Configure explicit argv, cwd, env and dependency paths for uncommon projects.
execute_command remains approved host execution; only frozen repository checks qualify as verification. Passing checks supports criteria, not arbitrary semantics.
Tester is enabled by default and can be disabled with agents tester off. Architect/Coder then handle criteria and regressions; required checks remain.
MCP tool access is optional and off by default. Calls use managed ProtoLink sessions without automatic retries. Failed started invocations mark external effects uncertain and block further effectful work in the run. Approvals/allowlists do not sandbox server effects; remote receipts cannot satisfy repository verification. Resources, prompts, OAuth and server installation are not part of this integration.
Request admission uses token estimates. Runtime-held task state survives observation eviction within a live run. Live model performance remains to be measured with coding evals.
Native execution restart/resume, Docker check execution and background model supervision are not integrated into the custom coding Graph. Saved conversations, reports and file recovery are distinct from execution continuation.