Skip to content

implemented-features.md

Latest commit

 

History

History
74 lines (62 loc) · 3.35 KB

implemented-features.md

File metadata and controls

74 lines (62 loc) · 3.35 KB
Raw

Implemented capabilities (v1.0.0)

This is the maintained, user-visible capability inventory for the current release. Claims are backed by the deterministic regression suite and the offline walkthrough in demo-script.md. Source and test locations are mapped in capability-trace.md.

Model protocol and loop

  • Provider-neutral messages and native tool calling.
  • OpenAI-compatible, Anthropic, Google, and Ollama adapters, plus an offline deterministic provider for demos and CI.
  • Streaming text and tool-call fragments with normalized finish reasons.
  • Bounded provider retries, deadlines, cancellation, cleanup grace periods, and categorized errors.
  • AgentLoop lifecycle states, step/tool budgets, turn correlation IDs, and verification-aware completion.

Workspace tools and execution

  • Workspace-aware file listing, UTF-8 reads, text/regex search, repository map, symbols, definitions, references, and metadata.
  • Atomic file creation/writes, unified multi-file patches, previews, and transaction records with before/after hashes.
  • Bounded shell commands and named test profiles with argv validation, timeout, output limits, exit-code checks, cancellation, and process-tree termination.
  • Hard blocks for traversal, forbidden paths, unsafe commands, and secret exposure; failures preserve stdout, stderr, exit codes, and structured error metadata.

Modes, approvals, and context

  • Plan (read-only), Act (approved side effects), and Bypass (explicit trusted workspace) modes.
  • Interactive, automatic, or deny approval policies, including per-scope changes, execution, and evidence decisions.
  • Bounded /steer instructions injected at the next model boundary; steering never interrupts synchronous tool side effects and is cleared on cancel.
  • Live /status diagnostics expose lifecycle, timing, event cursor, provider/tool counters, and configured step/tool budgets.
  • Explicitly managed, bounded workspace memory (memory add/show/remove/clear) is atomically persisted and injected as untrusted context; models cannot mutate it implicitly.
  • Scoped AGENTS.md rules, explicit references, incremental context indexing, context search/show/complete, compaction, and health diagnostics.
  • Validated skills and lifecycle hooks with bounded prompt text, quotas, and no ability to bypass the execution boundary.

Sessions, review, and interfaces

  • Durable sessions, checkpoints, pause/resume/cancel/Esc controls, session listing/tree/import, status, diff, and hash-checked transaction undo.
  • Evidence-driven review covering secrets, forbidden paths, suspicious commands, Python syntax, tests, plans, hooks, and transaction state.
  • Review artifact export and verification bound to the workspace and hashes.
  • Human REPL plus strict JSON/JSONL CLI, rpc.describe, Python embedding, and a small Node JSONL client with shared contracts.
  • Local telemetry policy and redacted audit records; credentials stay in environment variables or ignored local state.

Reproducible checks

uv run pytest -q
uv run forgecode doctor
uv run forgecode tools
uv run forgecode provider health
uv run forgecode --workspace $demo run --demo --auto-approve

Use the full commands and expected evidence in demo-script.md; run the focused test modules associated with a changed subsystem before the complete suite.