A governance stack for AI agents - tools that redirect an agent's attention to what it is actually doing: before it responds, after it responds, while it reasons, across sessions, and when it delegates.
Read the MANIFESTO for the full architecture. Read EVIDENCE.md for every claim in the stack, classified by evidence status - the stack audits itself before anyone else gets the chance. Agent runtimes (Claude, OpenClaw, Hermes): SKILL.md is the drop-in skill for running this dispatcher as a governed tool.
| Pillar | Repo | Role |
|---|---|---|
| 1 | before-turn | Governs entry - reads prior thinking before every response |
| 2 | pre-response-selfcheck | Governs exit - reads output as a cold reader before delivering |
| 3 | agent-open-mind | Reads what sub-agents thought, not what they said |
| 4 | open-mind | Compares what the agent thought to what it said |
| 5 | sleep-marks | Restores reasoning state across session breaks |
| 6 | splitvantage | Sends the same task to two models, surfaces divergence (CrossPol) |
Enterprise extension (referenced, not required): TelsonBase - permissions, audit, trust levels.
- Instruction tier - the dispatcher runtime in this repo: a closed-track message hub where every (from → intent → to) tuple is enforced at send time, acks exist only after persist-and-delivery, and the audit log is the sole source of every KPI and after-action report.
- Detection tier - the six pillars, each imported from its own package and bound to a named runtime seam (dispatcher/pillars.py): before-turn at turn entry (hub re-reads its own recent reflections and answers the pillar's check questions, audited); open-mind's Comparator on every hub reasoning-vs-action artifact; agent-open-mind's taint gate at spoke-trace ingestion (the pillar function IS the gate - absent thought = tainted, never silently admitted); pre-response-selfcheck as the exit gate on outbound text (FAIL verdict holds the envelope in clarification); sleep-marks riding every territory transfer (the outgoing crew's reasoning state restores on the receiving hub); splitvantage as the second-opinion diff on drift-flagged decisions. Two gates (selfcheck, splitvantage) need reviewer models - deployment config; UNARMED IS AUDITED, never silent. Every binding has a test.
- Structural tier - TelsonBase: allow/gate/block per tool call, below the model, for deployments that need enforcement rather than detection.
Instruction < detection < structural. The org is the hierarchy.
dispatcher/ - the hub: envelope schema with enforced confidence vocabulary,
tuple-legality routing, append-only hash-chained audit log with external anchoring, HMAC and
Ed25519 authority signatures, boot attestation (hash manifest, fail-closed verification), JIT
priority classes with live-hold siding, identity side-load loader, KPI
computation from the log only, after-action generation per schema. Tests are
doctrine as executable assertions, run against a real 51-route vertical
track - see README-BUILD.md for the claim-to-test map and EVIDENCE-runtime.md
for the evidence classes.
| Vertical | Repo | Status |
|---|---|---|
| Real-estate listing agent | listing-agents | v0.19 - 21 agents, 24 playbooks, 51 ratified route tuples, ratified priority classes (2026-07-10); runtime-driven end-to-end: P11 demo in this repo, and P01 (all three phases from one signed trigger) plus a full lead-to-close lifecycle as continuous-flow tests in the identity repo. First of several. |
git clone https://github.com/QuietFireAI/dispatcher-agents
cd dispatcher-agents
pip install -e ".[pillars,crypto,dev]" # core hub is zero-dep; pillars+Ed25519 are extras
git clone https://github.com/QuietFireAI/listing-agents ../listing-agents
IDENTITY_DIR=../listing-agents python3 -m pytest tests/ # 103 doctrine tests vs the real 51-route track
IDENTITY_DIR=../listing-agents python3 demo/run_p11_demo.py # real spokes chain 11 envelopes from 2 signals
# Identity-gated tests target the repo bundled at IDENTITY_DIR (default: a fixture track); point IDENTITY_DIR at listing-agents for the real 51-route run.
# The core hub alone installs with plain `pip install -e .` (no extras).
This is the hub: the manifesto, the claims ledger, and the findings.
- MANIFESTO.md - what the stack is and why
- EVIDENCE.md - the claims ledger (MEASURED / OBSERVED / HYPOTHESIS / DESIGN CLAIM / POSITION)
- findings/ - documented observations, each with its evidence class and, where applicable, a preregistered validation protocol
v0.4, July 2026. Hash-chained audit log with external anchor (anchor/verify_anchor - wholesale log regeneration is now detectable), login-based signer enforcement with temporal effective-date checks on all .authority lanes, CI on every push (103-test suite). All six pillars exist as code with reviewed documentation,
and the dispatcher runtime exists with a green doctrine-test suite against a
real vertical track. The stack's quantitative claims are classified in
EVIDENCE.md; highest-priority open items are the A/B validation of the
context-load finding
and the tuple-enforcement A/B (HYPOTHESIS class, protocol in the runtime's
after-action schema).
Apache License 2.0 - see LICENSE. Free for broad adoption and contribution. Domain identity packages that run on this chassis are licensed separately (QuietFire Identity License).