docs(visual): the command-hierarchy edges (ask/notify/override) + fix §7 post-#168

[drewstone]

Folds the question/command-hierarchy design into architecture-visual.md as §8 (DESIGNED, not built), and fixes §7 (now stale after #168 deleted the program op-set).

§8 — the command hierarchy

The escalation model is a resumable effect with handlers (à la LangGraph interrupt() / algebraic effects / OTP supervisor-escalation) — not agent-to-agent messaging, so no A2A/bus. Three edges complete the atom, two of which already exist:

• ask (up, blocking, terminates at the first handler who answers) — the one genuinely-new edge (or a 3rd Settled kind {question}).
• notify (up, non-blocking, salience-filtered) — already the lifecycle hook stream; lets a higher agent see and countermand a decision.
• override (down) — already scope.send; the same edge carries the answer and the override.

Locks the principles: optimistic + async, never synchronous approval (else the tree serializes through the root); command one level deep (root→D1→W, composes, no skip-level); block-vs-settle-and-resume via the shipped sandbox-lineage; budget pauses while awaiting a human; a human answer is an oracle injection → off/held-constant in gated experiments. The PI/chat agent is the top handler.

Implementation stays gated on the verifier-grounded gate result + the PI repo defining the human-handler contract; this nails the seam so both repos build to it.

§7 — anti-staleness fix

The program op-set (runProgram/Program) was DELETED in #168 (the diverse@k gate runs on fanout in keystone-gate.ts, never runProgram). Updated the table: the kernel is now two layers — the Scope atom + the runLoop leaf backend — with no redundant third.

Docs-only; merges clean.

🤖 Generated with Claude Code

[drewstone]

Folds the question/command-hierarchy design into architecture-visual.md as §8 (DESIGNED, not built), and fixes §7 (now stale after #168 deleted the program op-set).

§8 — the command hierarchy

The escalation model is a resumable effect with handlers (à la LangGraph interrupt() / algebraic effects / OTP supervisor-escalation) — not agent-to-agent messaging, so no A2A/bus. Three edges complete the atom, two of which already exist:

• ask (up, blocking, terminates at the first handler who answers) — the one genuinely-new edge (or a 3rd Settled kind {question}).
• notify (up, non-blocking, salience-filtered) — already the lifecycle hook stream; lets a higher agent see and countermand a decision.
• override (down) — already scope.send; the same edge carries the answer and the override.

Locks the principles: optimistic + async, never synchronous approval (else the tree serializes through the root); command one level deep (root→D1→W, composes, no skip-level); block-vs-settle-and-resume via the shipped sandbox-lineage; budget pauses while awaiting a human; a human answer is an oracle injection → off/held-constant in gated experiments. The PI/chat agent is the top handler.

Implementation stays gated on the verifier-grounded gate result + the PI repo defining the human-handler contract; this nails the seam so both repos build to it.

§7 — anti-staleness fix

The program op-set (runProgram/Program) was DELETED in #168 (the diverse@k gate runs on fanout in keystone-gate.ts, never runProgram). Updated the table: the kernel is now two layers — the Scope atom + the runLoop leaf backend — with no redundant third.

Docs-only; merges clean.

🤖 Generated with Claude Code

[drewstone]

Folds the question/command-hierarchy design into architecture-visual.md as §8 (DESIGNED, not built), and fixes §7 (now stale after #168 deleted the program op-set).

§8 — the command hierarchy

The escalation model is a resumable effect with handlers (à la LangGraph interrupt() / algebraic effects / OTP supervisor-escalation) — not agent-to-agent messaging, so no A2A/bus. Three edges complete the atom, two of which already exist:

• ask (up, blocking, terminates at the first handler who answers) — the one genuinely-new edge (or a 3rd Settled kind {question}).
• notify (up, non-blocking, salience-filtered) — already the lifecycle hook stream; lets a higher agent see and countermand a decision.
• override (down) — already scope.send; the same edge carries the answer and the override.

Locks the principles: optimistic + async, never synchronous approval (else the tree serializes through the root); command one level deep (root→D1→W, composes, no skip-level); block-vs-settle-and-resume via the shipped sandbox-lineage; budget pauses while awaiting a human; a human answer is an oracle injection → off/held-constant in gated experiments. The PI/chat agent is the top handler.

Implementation stays gated on the verifier-grounded gate result + the PI repo defining the human-handler contract; this nails the seam so both repos build to it.

§7 — anti-staleness fix

The program op-set (runProgram/Program) was DELETED in #168 (the diverse@k gate runs on fanout in keystone-gate.ts, never runProgram). Updated the table: the kernel is now two layers — the Scope atom + the runLoop leaf backend — with no redundant third.

Docs-only; merges clean.

🤖 Generated with Claude Code

[drewstone]

Folds the question/command-hierarchy design into architecture-visual.md as §8 (DESIGNED, not built), and fixes §7 (now stale after #168 deleted the program op-set).

§8 — the command hierarchy

The escalation model is a resumable effect with handlers (à la LangGraph interrupt() / algebraic effects / OTP supervisor-escalation) — not agent-to-agent messaging, so no A2A/bus. Three edges complete the atom, two of which already exist:

• ask (up, blocking, terminates at the first handler who answers) — the one genuinely-new edge (or a 3rd Settled kind {question}).
• notify (up, non-blocking, salience-filtered) — already the lifecycle hook stream; lets a higher agent see and countermand a decision.
• override (down) — already scope.send; the same edge carries the answer and the override.

Locks the principles: optimistic + async, never synchronous approval (else the tree serializes through the root); command one level deep (root→D1→W, composes, no skip-level); block-vs-settle-and-resume via the shipped sandbox-lineage; budget pauses while awaiting a human; a human answer is an oracle injection → off/held-constant in gated experiments. The PI/chat agent is the top handler.

Implementation stays gated on the verifier-grounded gate result + the PI repo defining the human-handler contract; this nails the seam so both repos build to it.

§7 — anti-staleness fix

The program op-set (runProgram/Program) was DELETED in #168 (the diverse@k gate runs on fanout in keystone-gate.ts, never runProgram). Updated the table: the kernel is now two layers — the Scope atom + the runLoop leaf backend — with no redundant third.

Docs-only; merges clean.

🤖 Generated with Claude Code

[drewstone]

Folds the question/command-hierarchy design into architecture-visual.md as §8 (DESIGNED, not built), and fixes §7 (now stale after #168 deleted the program op-set).

§8 — the command hierarchy

The escalation model is a resumable effect with handlers (à la LangGraph interrupt() / algebraic effects / OTP supervisor-escalation) — not agent-to-agent messaging, so no A2A/bus. Three edges complete the atom, two of which already exist:

• ask (up, blocking, terminates at the first handler who answers) — the one genuinely-new edge (or a 3rd Settled kind {question}).
• notify (up, non-blocking, salience-filtered) — already the lifecycle hook stream; lets a higher agent see and countermand a decision.
• override (down) — already scope.send; the same edge carries the answer and the override.

Locks the principles: optimistic + async, never synchronous approval (else the tree serializes through the root); command one level deep (root→D1→W, composes, no skip-level); block-vs-settle-and-resume via the shipped sandbox-lineage; budget pauses while awaiting a human; a human answer is an oracle injection → off/held-constant in gated experiments. The PI/chat agent is the top handler.

Implementation stays gated on the verifier-grounded gate result + the PI repo defining the human-handler contract; this nails the seam so both repos build to it.

§7 — anti-staleness fix

The program op-set (runProgram/Program) was DELETED in #168 (the diverse@k gate runs on fanout in keystone-gate.ts, never runProgram). Updated the table: the kernel is now two layers — the Scope atom + the runLoop leaf backend — with no redundant third.

Docs-only; merges clean.

🤖 Generated with Claude Code

[drewstone]

Folds the question/command-hierarchy design into architecture-visual.md as §8 (DESIGNED, not built), and fixes §7 (now stale after #168 deleted the program op-set).

§8 — the command hierarchy

The escalation model is a resumable effect with handlers (à la LangGraph interrupt() / algebraic effects / OTP supervisor-escalation) — not agent-to-agent messaging, so no A2A/bus. Three edges complete the atom, two of which already exist:

• ask (up, blocking, terminates at the first handler who answers) — the one genuinely-new edge (or a 3rd Settled kind {question}).
• notify (up, non-blocking, salience-filtered) — already the lifecycle hook stream; lets a higher agent see and countermand a decision.
• override (down) — already scope.send; the same edge carries the answer and the override.

Locks the principles: optimistic + async, never synchronous approval (else the tree serializes through the root); command one level deep (root→D1→W, composes, no skip-level); block-vs-settle-and-resume via the shipped sandbox-lineage; budget pauses while awaiting a human; a human answer is an oracle injection → off/held-constant in gated experiments. The PI/chat agent is the top handler.

Implementation stays gated on the verifier-grounded gate result + the PI repo defining the human-handler contract; this nails the seam so both repos build to it.

§7 — anti-staleness fix

The program op-set (runProgram/Program) was DELETED in #168 (the diverse@k gate runs on fanout in keystone-gate.ts, never runProgram). Updated the table: the kernel is now two layers — the Scope atom + the runLoop leaf backend — with no redundant third.

Docs-only; merges clean.

🤖 Generated with Claude Code

[drewstone]

Folds the question/command-hierarchy design into architecture-visual.md as §8 (DESIGNED, not built), and fixes §7 (now stale after #168 deleted the program op-set).

§8 — the command hierarchy

The escalation model is a resumable effect with handlers (à la LangGraph interrupt() / algebraic effects / OTP supervisor-escalation) — not agent-to-agent messaging, so no A2A/bus. Three edges complete the atom, two of which already exist:

• ask (up, blocking, terminates at the first handler who answers) — the one genuinely-new edge (or a 3rd Settled kind {question}).
• notify (up, non-blocking, salience-filtered) — already the lifecycle hook stream; lets a higher agent see and countermand a decision.
• override (down) — already scope.send; the same edge carries the answer and the override.

Locks the principles: optimistic + async, never synchronous approval (else the tree serializes through the root); command one level deep (root→D1→W, composes, no skip-level); block-vs-settle-and-resume via the shipped sandbox-lineage; budget pauses while awaiting a human; a human answer is an oracle injection → off/held-constant in gated experiments. The PI/chat agent is the top handler.

Implementation stays gated on the verifier-grounded gate result + the PI repo defining the human-handler contract; this nails the seam so both repos build to it.

§7 — anti-staleness fix

The program op-set (runProgram/Program) was DELETED in #168 (the diverse@k gate runs on fanout in keystone-gate.ts, never runProgram). Updated the table: the kernel is now two layers — the Scope atom + the runLoop leaf backend — with no redundant third.

Docs-only; merges clean.

🤖 Generated with Claude Code

[drewstone]

Folds the question/command-hierarchy design into architecture-visual.md as §8 (DESIGNED, not built), and fixes §7 (now stale after #168 deleted the program op-set).

§8 — the command hierarchy

The escalation model is a resumable effect with handlers (à la LangGraph interrupt() / algebraic effects / OTP supervisor-escalation) — not agent-to-agent messaging, so no A2A/bus. Three edges complete the atom, two of which already exist:

• ask (up, blocking, terminates at the first handler who answers) — the one genuinely-new edge (or a 3rd Settled kind {question}).
• notify (up, non-blocking, salience-filtered) — already the lifecycle hook stream; lets a higher agent see and countermand a decision.
• override (down) — already scope.send; the same edge carries the answer and the override.

Locks the principles: optimistic + async, never synchronous approval (else the tree serializes through the root); command one level deep (root→D1→W, composes, no skip-level); block-vs-settle-and-resume via the shipped sandbox-lineage; budget pauses while awaiting a human; a human answer is an oracle injection → off/held-constant in gated experiments. The PI/chat agent is the top handler.

Implementation stays gated on the verifier-grounded gate result + the PI repo defining the human-handler contract; this nails the seam so both repos build to it.

§7 — anti-staleness fix

The program op-set (runProgram/Program) was DELETED in #168 (the diverse@k gate runs on fanout in keystone-gate.ts, never runProgram). Updated the table: the kernel is now two layers — the Scope atom + the runLoop leaf backend — with no redundant third.

Docs-only; merges clean.

🤖 Generated with Claude Code