hol-cigar

Local context graphs, exact token budgets, source citations and reviewed answers for Python applications. No HOL service, account, API key, daemon or database is required. CIGAR owns a Rust graph in a persistent local worker process.

Version 0.14.0 contains a broker for independent agent processes. See the packaged broker guide for host/client authority, provenance, reviews and failure semantics. Installed package tests have passed across seven native targets. The 24-hour installed soak and independent replay passed on macOS ARM64 with twelve Python/Node clients. Node client RSS kept growing within the cap. HUMIDOR adoption and broader Hiero/task evidence remain pending. Qualification applies to the exact artifacts named in the signed release receipts.

0.14.1 security candidate

The next patch release adds aggregate compilation limits and Python HTTP deadlines, and requires Node.js >=24.22.1 <25. See the 0.14.1 release notes. It is prepared for review and is not yet published. Historical qualification below applies to the named 0.14.0 artifacts. Upgrade and restart Node consumers now.

Install and check

Version 0.14.0 targets Python >=3.14 <3.15. The distribution is hol-cigar; the Python import is cigar_sdk. The corresponding npm package is @hol-org/cigar.

python3.14 -m pip install --upgrade 'hol-cigar==0.14.0'
python3.14 -m cigar_sdk.local_cli doctor
python3.14 -m cigar_sdk.local_cli demo
python3.14 -m cigar_sdk.examples.shared_views

Pip selects the matching wheel for your supported platform. Seven native targets are included. Each native wheel bundles its worker; no Rust compiler or separate worker installation is required. Exact archive hashes and qualification receipts accompany the GitHub release. See the bundled changelog for migration details. The supported protobuf requirement is >=6.33.5,<8, qualified at minimum/current versions. The environment also gets a cigar-context command. doctor verifies a real local compile; demo runs the complete workflow. Add --json for machine-readable results.

Five agents sharing one graph in one host

The trusted host creates one graph and gives each cooperating agent a scoped view:

from cigar_sdk import LocalContextGraph

with LocalContextGraph("run-unique-domain") as graph:
    view = graph.create_view({
        "id": "agent-1", "allowed_sources": ["shared-policy", "agent-1"],
        "writable_sources": ["agent-1"], "policy_revision": "host-policy-1",
    })
    view.replace_source("agent-1", [{"id": "task-1", "source": "agent-1", "text": "Retry once."}])
    result = view.compile({"required": ["task-1"], "max_tokens": 512})
    # Bind a draft to result["context"]["snapshot"]["id"], obtain independent reviews,
    # then call view.check_answer(result["context"], draft, reviews).

Import LocalContextGraph as shown below. Create five views on the same graph; each stores only its scope definition. The packaged shared_views example runs five concurrent scripted agents, checks citations/budgets and tests reviewed release and missing-review abstention without a model provider.

allowed in a view request can only narrow the host's source scope. Empty sources deny all; writable sources must also be readable. Source replacement rejects ID collisions with other sources. graph.revoke_view(id) or redefining the view invalidates old handles. Changes anywhere in its readable sources/edges require a new review; updates strictly outside the scope leave the review valid. The returned checked_graph_revision records when assessment occurred, not permission for a later external action. The host must coordinate revalidation with effect execution.

Keep the root graph, view definitions and reviewer verdicts outside agent control. Views are logical host scopes, not a sandbox or authenticated capability. Sharing one process is appropriate only inside a trusted application/privacy domain. Use a host-owned LocalContextBroker and one LocalContextClient per agent for separate processes; never share Python objects through fork(). The broker can persist admitted evidence in an explicitly configured local directory. Recovery creates a new authority epoch: reissue grants, recompile and obtain fresh reviews. Ordinary graph views remain in-memory handles. Native graph calls remain serialized. HUMIDOR remains responsible for scheduling and recovery; adopting the optional context-to-Honey bridge requires its separate integration qualification.

Existing root methods retain their exact 0.12 behavior, including conservative global-revision stale-review rejection. New view methods require a worker advertising context_views.v1; unsupported workers fail with IncompatibleWorker.

Previous improvements from 0.11.0 to 0.12.0

Version 0.12.0 strengthens integrity checks and worker lifecycle handling while reducing Python startup cost. Existing valid bundle IDs, context snapshots, cigar.context.v1, the worker protocol and all 69 Python public exports are preserved.

The performance comparison used installed packages on one macOS ARM64 host, Python 3.14.7, protobuf 6.33.5 in both environments, warm filesystem and bytecode caches, and 25 fresh-process samples per startup case. MB means 1,000,000 bytes; Python RSS excludes the child worker. Steady-state compile performance was essentially unchanged. These are measured results, not guarantees for every host.

Worker cleanup is now bounded and idempotent, preserves the original timeout or transport error, and exposes cleanup_complete. Inherited graphs fail promptly after fork(); create a new graph in the child. Worker verification still hashes the complete executable on every launch, using a bounded buffer.

The comparison also recorded tradeoffs: some native workloads had up to 7.4% higher median latency and one microsecond-scale p95 increased 26%. Native benchmark peak RSS increased up to 3.5%; wheel sizes increased 0.13–0.17%. These observations passed the defined median/RSS thresholds but do not establish zero degradation. Offline answer-review results matched 0.11.0; no reduction in real-model hallucinations is claimed. Workerless source builds now require explicit opt-in.

See the full comparison and qualification scope and the changelog for compatibility and migration details.

Local context graph (no server or model required)

from cigar_sdk import LocalContextGraph

with LocalContextGraph("my-project") as graph:
    graph.replace_source("src/auth.py", [
        {"id": "authorize", "source": "src/auth.py", "text": "def authorize(user):\n    return user.active\n"}
    ])
    result = graph.compile({"query": "authorize", "max_tokens": 1024, "reserve_tokens": 128})
    context_text = result["rendered"]  # place in your prompt's DATA/context role
    assert result["snapshot"]["stats"]["rendered_tokens"] <= 896

The persistent worker retains incremental indexes and its bounded exact o200k_base cache. upsert, remove, link/unlink, atomic replace_source, line-preserving chunks, compile, verify, delta, apply_delta, stats, and clear_cache expose the Rust core. Source replacement reuses unchanged indexed documents. The optional prompt_view(snapshot, max_tokens) returns a LocalContextPrompt with all selected text and short citation handles. Keep its citation map and full snapshot; use verify_prompt(prompt, snapshot) or resolve_citation("c1", prompt, snapshot) against the expected authorized snapshot. Its separate exact budget fails without truncation. Token savings depend on citation overhead. Input/request/snapshot types are exported as LocalDocument, LocalContextRequest, LocalContextSnapshot, etc. Local methods are synchronous and thread-serialized; the existing AsyncCigarClient remains the asynchronous remote client.

required IDs include their complete hard dependency/contradiction closure. Pass allowed on every call when source permissions differ ([] authorizes none; omission permits the whole caller-owned graph). excerpt_mode="query_windows" is opt-in; full text is the default. Source withdrawal preserves hard edges, so missing required evidence fails closed. Snapshot digests are integrity commitments, not signatures or authority. Deltas reduce transport/storage bytes, not stateless model prompt tokens. Token budgets include Rust-rendered citations but exclude the provider envelope; reserve that separately. Retrieval optimizations preserve the previous selected evidence. Answer review adds a host-enforced release contract; it does not establish real-model quality without a separately evaluated reviewer.

Use with or close() to release the worker. Each graph owns one process and privacy-local cache. Defaults: 30-second call timeout, 32 MiB request, 64 MiB response, 100k documents, 256 MiB indexed text, 2048 cache entries/8 MiB cached text. Customize limits and timeout. Cache clearing drops strings but does not guarantee memory zeroization. A LocalContextError has a content-free code; timeout/protocol/pipe failure closes the graph and never retries mutations. Invalid wire fields also close it. A lock-wait Busy error leaves the active call alone.

Create graphs in the process that uses them. After fork(), inherited graph operations and close() raise ForkedProcess before touching inherited locks or the parent's worker. Create a new graph in the child; use a spawn-based process pool where possible. close() is idempotent and preserves ordinary primary errors. If OS cleanup fails, cleanup_complete remains false and another close retries.

The native inventory covers macOS ARM64/x64, Linux glibc/musl ARM64/x64 and Windows x64. Installed tests and independent archive comparison pass for this inventory. The macOS ARM64 24-hour soak passed; downstream application qualification remains open. Use the receipt for your exact archive to identify completed checks.

Building a wheel from the portable source distribution without native staging requires explicit CIGAR_ALLOW_PORTABLE_WHEEL=1. This also applies to intentional pip install --no-binary hol-cigar source installs. These builds retain all SDK APIs, but local graphs require an explicit trusted absolute worker_path. Build it from the matching 0.14.1 Rust source using Rust 1.92+:

cargo build --locked --release -p cigar-context --features bpe,broker-persistence --bin cigar-context-worker

Pass the resulting executable to LocalContextGraph("my-project", worker_path="/absolute/path/to/cigar-context-worker"). No PATH discovery, install-time downloads, shell invocation, or automatic filesystem ingestion occurs. Bundled worker bytes are checked against their package manifest before launch; this is not a substitute for trusting the package publisher. The subprocess runs with your OS privileges and inherits its environment; it is not a security sandbox. It adds startup/IPC/memory overhead compared with embedding Rust directly. Reuse graphs to amortize initialization.

Complete workflow and diagnostics

The same complete example run by cigar-context demo is importable:

from cigar_sdk.examples.local_workflow import run_local_workflow

result = run_local_workflow()
print(result["status"])  # passed

It covers ingestion, dependencies, authorized compilation, compact citations, cache reuse, trusted fixture reviews, source updates, deltas and stale-review rejection. Its reviewer uses separately authored fixture facts. Replace it with an authenticated semantic reviewer for real answers. Reuse your application's graph across requests.

from cigar_sdk import get_local_context_capabilities

capability = get_local_context_capabilities()
print(capability["platform"], capability["worker_available"], capability["guidance"])

The capability API inspects platform and bytes without starting a worker. doctor also compiles synthetic context and verifies it. Missing/unexecutable workers report WorkerUnavailable, absent platform support reports UnsupportedPlatform, and altered worker bytes report WorkerIntegrity. These errors do not mean HOL services are missing. An explicitly supplied matching worker can be checked with python -m cigar_sdk.local_cli doctor --worker /absolute/path.

The installed cigar_sdk package includes AGENT_GUIDE.md, BROKER_GUIDE.md, SELECTION_EXPLANATIONS.md, SYNTAX_INGESTION.md and llms.txt. The agent integration guide explains explicit file ingestion, graph relationships, authorization and reviewed answers.

Compatible remote client

Choose AsyncCigarClient or CigarClient when connecting to an existing CIGAR server. Both expose all 45 frozen v1 operations, bounded deadlines, typed problems, resumable streams, pagination, fixed idempotency keys, safe retry and local bundle/delta verification. They accept your server's URL; HOL hosting is optional. CONTEXT_ABI remains cigar.context.v1.

from cigar_sdk import AsyncCigarClient, TypedOperationRequest, create_idempotency_key, models

async with AsyncCigarClient("https://cigar.example", bearer_token=token_provider) as client:
    result = await client.compile_context_bundle(
        TypedOperationRequest(
            models.CompileContextBundleRequest(plan_id=plan_id),
            idempotency_key=create_idempotency_key("compile"),
        )
    )

Every nominal request and response is validated against the frozen payload schema. Mutating retries reuse the exact caller-provided key and bytes. Effect dispatch is always one attempt. Synchronous and asynchronous streams are explicit context managers so callers can close the underlying response deterministically.

Token providers accept the remaining call timeout in seconds. Injecting a custom HttpTransport requires trust_custom_transport=True; the default transport ignores ambient proxies and refuses redirects. The wheel and source distribution both include the shared fixture, so cigar-qualify-bundle works from a clean installation.

Honey distributions also install cigar-agent-b-handoff. It accepts an existing recipient-bound handoff, then records one typed, evidence-backed result with independent idempotency keys. The Agent B bearer token is read only from CIGAR_AGENT_B_TOKEN, never a command-line argument. Run cigar-agent-b-handoff --help for the required handoff, plan, base-commit, revision, claim, evidence, and caller-generated acceptance/result idempotency keys. The example requests no follow-up capability and never dispatches an effect.

Remote HTTPS construction requires an explicit bearer_token value or provider. The SDK never discovers credentials from the URL, environment, project configuration, proxy settings, or a redirect target. Explicit cleartext loopback mode remains available only for local development.

Answer review

Call graph.review_keys(draft) to bind an AnswerDraft's claims to the snapshot provided to the generator. Obtain LocalClaimReview verdicts through a separate, trusted host review path, then call graph.check_answer(request, draft, reviews). Only display the assessed claims when decision == "release". The check recompiles current authorized context and rejects stale snapshots/reviews. Missing reviews, unresolved support, invalid citations or unreviewed explicit conflicts block release. confidence_bps is optional telemetry (0–10000), never permission. Keep reviews and policy outside model control; CIGAR does not run a semantic judge. See the core contract.