@matheuskrumenauer/tanya 0.6.0-beta.0 → 0.8.0-beta.0

package/README.md

@@ -198,6 +198,35 @@ run directly; JavaScript and TypeScript commands export a default
 Project-local commands are arbitrary code execution and are gated by the same
 permission engine as native tools.
 
+## Sub-agent tool
+
+The `task` tool delegates a bounded child run while keeping the parent in
+control:
+
+```json
+{
+  "prompt": "Map the auth module and report blockers.",
+  "workspace": "src/auth",
+  "max_turns": 12,
+  "token_budget": { "max_tokens": 12000 },
+  "treat_failure_as": "warning"
+}
+```
+
+Children inherit the parent's skill packs, permission rules, workspace, and
+budget. They may narrow those constraints but cannot loosen them. Depth is
+capped at 2 by default (`TANYA_SUBTASK_MAX_DEPTH`), and active children share a
+default parallel cap of 3 (`TANYA_SUBTASK_MAX_PARALLEL`).
+
+Every child runs its own verifier. Failed child verdicts become parent blockers
+by default; `treat_failure_as` can demote a specific child to `warning` or
+`ignore` when the caller wants advisory work only. Child events stream into the
+parent log with a `subRunId`, and parent cancellation propagates into active
+children.
+
+See [docs/sub-agents.md](./docs/sub-agents.md) for permission inheritance,
+budget-ledger semantics, cancellation, verifier composition, and memory rollup.
+
 `--verify` adds required verification commands to the run context. Tanya must run and report each exact command before finishing the coding task.
 
 `tanya benchmark run --all` currently exercises 27 executable low-to-medium regression fixtures: targeted edits, new files, dependency/lockfile updates, framework-style migrations, failing-test repair, frontend smoke checks, artifact/context reuse, streaming long-tool execution, compaction-boundary recovery, run-history logging, dirty worktrees, report repair, and the CosmoHQ mobile/backend smoke profiles.
@@ -245,12 +274,24 @@ Tanya handles context pressure as a cascade instead of truncating abruptly:
 1. Microcompact folds empty/no-op tool-call pairs in place.
 2. Snip removes low-signal history such as duplicate file reads and empty read-only tool results.
 3. Auto-compact reacts to provider `413` / context-window errors by summarizing older turns into a `[compaction summary: ...]` system message and retrying once normally, then once more aggressively.
-4. Archive writes compacted messages to `.tanya/runs/<runId>/archive.jsonl` before they leave live history, so verifier scans and future memory tools can still inspect them.
+4. Archive writes compacted messages to `.tania/runs/<runId>/archive.jsonl` before they leave live history, so verifier scans and future memory tools can still inspect them.
 
 Runs are capped at three total auto-compactions. If the provider still rejects the context, Tanya raises `CompactionExhaustedError` and asks the user to narrow the task, clear the session, or split the work.
 
 See [docs/long-sessions.md](./docs/long-sessions.md) for details.
 
+## Token economy
+
+Tanya trims model-visible tokens while keeping state reversible and auditable.
+
+- Lite prompts can be enabled with `TANYA_LITE_PROMPT=1` for cheap-provider exploration turns. The legacy `TANIA_LITE_PROMPT` alias is still accepted.
+- System prompts are automatically capped to the active provider context window. Tune the default 25% cap with `TANYA_PROMPT_BUDGET_RATIO`.
+- Large shell/tool outputs are shortened for the model with a visible `<truncated ...>` marker. Use `expand_result` with the marker's `tool_call_id` to fetch the full output or a byte range.
+- Repeated unchanged `read_file` calls return a reference marker instead of resending the same content. Pass `force: true` when the agent genuinely needs the full file again.
+- `/budget` reports token usage, cost estimates, expensive turns, and one deterministic optimization suggestion. `/budget --enforce --max-usd <amount>` persists a session spend rule through the permission engine.
+
+See [docs/token-economy.md](./docs/token-economy.md) for the full model, cache locations, and tool-definition knobs.
+
 Context files are generic JSON envelopes for caller-supplied task metadata, artifacts, instructions, and verification commands.
 
 ## Current Tools