0.6.0 — framework hardening (streaming, tools, parsing, packaging, facade)

[chiruu12]

Framework-hardening release for 0.6.0. Makes the agent runtime robust under streaming interruptions, hung tools, and corrupt state, and makes the package friendlier to import/embed. No breaking changes; DBs upgrade automatically. Poker/simulation is intentionally out of scope (separate repo).

Reviewable per-commit — each commit is one scope:

Commit	Scope	What
830554a	B6/B7/B8	Brace-matching structured parser + StructuredParseError (no more invalid model_construct()); structured-first provider error heuristics; warn on malformed tool-call args
b88dde5	C9	Optional [anthropic] [openai] [mcp] [cli] extras (core deps unchanged) + friendly MissingDependencyError
61f4709	A2/A3/A4	Per-tool tool_timeout; surfaced silent failures (memory recall, suffering/persona restore, plugin init) with explicit fallbacks; single active generated goal per agent
32d3e30	A1/A5	Streaming falls back to non-streaming on no-DONE / mid-stream error; stream closed on cancel; documented one-generation budget overshoot
2c9162f	C10	with / async with Hive() + lazy .hive init; version bump 0.6.0 + CHANGELOG + docs

Notable decisions

• A3 uses a re-check guard in the existence loop, not a unique index — subgoals, delegation, and schedules legitimately create multiple active goals per agent, so a hard constraint would break them.
• A5 keeps the post-generation budget check (the real gate) and documents the inherent one-generation overshoot. A projected pre-stop was implemented and rejected: estimating with the full output cap falsely refuses small-budget agents that emit short replies (a behavior regression).
• Three original audit findings were verified as non-issues and left untouched: the nudge SELECT-vs-UPDATE race (already fixed by-id), the daemon cycle-count "race" (atomic SQL UPDATE), and per-operation store connections (deliberate).

Verification

• uv run pytest — 1058 passed (added tests for streaming fallback, tool timeout, memory-recall warning, corrupt-checkpoint resume, existence guard, optional extras, context managers, robust parsing)
• uv run ruff check src/ tests/ + ruff format --check src/ — clean
• uv run mypy src/ — clean
• uv run mkdocs build --strict — clean
• Manual smoke: streaming fallback, per-tool timeout, missing-extra error, sync/async context managers + lazy init

Remaining audit findings filed as backlog issues (see linked issues).

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[chiruu12]

Backlog (remaining audit findings, deferred from this release):

• Plugin loader has no sandbox or validation #42 Plugin loader sandbox/validation
• Crisis detection lacks hysteresis (rapid oscillation resets suffering) #43 Crisis-detection hysteresis
• Agent profile cache is never invalidated on YAML mtime change #44 Profile cache mtime invalidation
• TF-IDF semantic-memory cache can stay corrupt after a torn write #45 TF-IDF cache torn-write robustness
• Verify daemon shutdown task-teardown completeness #46 Daemon shutdown task-teardown audit

[greptile-apps]

Greptile Summary

This PR hardens the Hive agent runtime across five well-scoped commits: streaming now falls back gracefully on interruption, tools have per-call timeouts, JSON parsing uses a proper brace-matching extractor with typed errors, optional provider extras are declared, and the Hive facade gains context-manager support with lazy init.

• Streaming (_generate_message): accumulated text is returned on mid-stream failure; an empty stream retries non-streaming; CancelledError propagates cleanly after aclose() in finally.
• Tool timeout (_run_tool in run()): asyncio.wait_for wraps each tool call when tool_timeout > 0, converting hangs into tool-error results fed back to the model.
• Structured parsing (parse_structured_response): _extract_json_object walks string literals correctly and raises StructuredParseError (with .raw) on ValidationError; StructuredTaskResult.parsed is now T | None instead of model_construct().

Confidence Score: 4/5

Safe to merge with one pre-existing unresolved defect: run_once() dispatches tools without the asyncio.wait_for guard that run() now has, so hung tools still block the coroutine indefinitely when callers use the direct API path.

The daemon path is fully protected by tool_timeout and the streaming fallback is well-guarded. However, run_once() — the main entry point for direct API callers — still calls tool.call() with no timeout wrapper, a gap noted in the previous review that this PR did not close.

src/hive/runtime/agent.py — specifically the inline tool-dispatch loop inside run_once() around line 626.

Important Files Changed

Filename	Overview
src/hive/runtime/agent.py	Core runtime changes: streaming fallback logic, per-tool asyncio.wait_for timeout in _run_tool, synthesized GenerateResult for interrupted streams, and budget guard added before run_once wrap-up. The tool_timeout guard is not applied inside run_once's inline dispatch loop (pre-existing issue noted in prior review).
src/hive/runtime/structured.py	New _extract_json_object correctly handles nested objects, quoted braces, and escape sequences. parse_structured_response raises StructuredParseError on ValidationError. Looks correct.
src/hive/api.py	Adds sync/async context managers and lazy _get_store init. HiveStore uses per-operation connections so no explicit close is needed. aexit correctly calls synchronous stop().
src/hive/errors.py	Adds MissingDependencyError, StructuredParseError, and require_dependency helper. Clean hierarchy: both new errors subclass HiveError and the relevant builtin for backward-compatible except clauses.
src/hive/agents/existence.py	Re-check guard added between goal validation and goal save to prevent duplicate generated goals from concurrent cycles. TOCTOU window remains but is intentional (unique index would break subgoals/delegation).
src/hive/daemon/loop.py	tool_timeout passed from config to both Agent branches (persona/no-persona). Suffering-state and persona-restore now have explicit exception handlers with clean fallbacks instead of silent failures.
src/hive/daemon/setup.py	ensure_hive_dirs() cleanly extracted from initialize_hive() to be loop-safe (no asyncio.run). initialize_hive now delegates to ensure_hive_dirs. Refactor is correct and idempotent.
src/hive/models/openai.py	_is_response_format_unsupported replaces fragile string matching with structured error field inspection (param, body.error.code) + string fallback. Malformed streaming tool-call JSON now logs a warning.
src/hive/config.py	tool_timeout: float = 60.0 added to DaemonConfig with a >= 0 validator. Clean addition consistent with existing validators.
src/hive/runtime/types.py	StructuredTaskResult.parsed changed from T to T

_{Reviews (2): Last reviewed commit: "fix(ci,review): format tests; estimate i..." | Re-trigger Greptile}

[chiruu12]

Addressed Greptile's two P2s in fa2e8c2:

• Interrupted-stream usage: _synthesize_stream_result now estimates output tokens/cost from the streamed text (~4 chars/token) so an interrupted generation stays visible to budget tracking instead of being zeroed.
• Extras pin duplication: provider/CLI extras are now unversioned aliases (["anthropic"] etc.); the version pins live only in [project.dependencies], which stays authoritative — no drift.

Also formatted tests/ (CI runs ruff format --check src/ tests/). Copilot review was skipped (quota).

[chiruu12]

Rebased onto main after #40 merged — PR is now a clean 6-commit hardening diff (the stacked #40 commits are gone). All CI green (build, lint & type-check, tests 3.11/3.12/3.13). Both Greptile P2s are addressed in code:

• interrupted-stream usage is now estimated from the streamed text (no longer zeroed for budgets)
• provider/CLI extras are unversioned aliases (no duplicated pins; dependencies stays authoritative)

Ready for your review.