docs(examples): prune + clarify the example set for tip-top DX by drewstone · Pull Request #368 · tangle-network/agent-runtime

drewstone · 2026-06-23T22:26:45Z

What

Make examples/ tip-top for DX: land a developer here, let them read ONE short example, and feel the power. Three confirmed defects fixed, the missing centerpiece added, and the index rewritten as an ordered "use this when" path. All edits stay inside examples/.

DELETE

examples/coder-loop/ — its .ts was refactored off runLoop onto worktreeLoopRunner, and its README still taught the deletedcoderProfile export, so it no longer demonstrated "the same runLoop kernel" it claimed. researcher-loop/ is now the primary runLoop teacher.

ADD — the centerpiece

examples/driver-loop/ — the one concept the request asked to SEE: a multi-round refine driver whose plan()reads the last worker's output from history and composes the next prompt from it ("the fold"), made visible with heavy plain-language comments. A second labeled section contrasts it with runMultishot so round vs shot sit side by side. Runs fully offline and is proven e2e:
- ROUND 0: [reject] "Shipped one-click restore for failed deploys." (no "rollback")
- driver folds the rejected output → ROUND 1: [PASS] "...with an instant rollback path..." → decision: pick-winner.
- Caught + fixed + documented a real footgun while building it: a decide() that returns a terminal value (fail) before refining stops the loop early — the example now teaches the non-terminal refine decision.

IMPROVE — stale-API docs + teaching comments

mcp-delegation/README + fleet-delegation/README: replace the deleteddelegate_code/delegate_research tools with the generic delegate verb + the MCP_ENABLE_DELEGATE=1 gate + the always-on delegate_feedback/delegation_status/delegation_history trio. (The .ts was already correct — only the READMEs lied.)
product-eval/README: teach runPersonaConversation (the facade evalPersona was deleted in 0.76.0 by chore(cleanup): delete dead eval-persona facade + orphaned topology module #367; the .ts already uses the canonical surface).
researcher-loop / ui-audit / self-improving-loop / supervisor-loop: inline-define round and shot the first time each appears and point at driver-loop/ as the example that actually shows the fold.
delegate: import from the @tangle-network/agent-runtime/loops subpath instead of a relative dist path; strip emoji from console output.

INDEX

examples/README.md rewritten as an ordered "use this when" path over tiers (cores → driver/supervisor → runLoop kernel → production runtime → self-improvement), with a vocabulary block up top, the missing driver-loop/supervise/delegate/intelligence-drop-in rows added, and coder-loop dropped.

Verification

pnpm build, pnpm run typecheck, pnpm run typecheck:examples, pnpm run lint (Biome) — all green.
pnpm run docs:check — freshness OK, no drift.
driver-loop and researcher-loop run offline; merges cleanly into origin/main.

🤖 Generated with Claude Code

Land a developer in examples/, let them read ONE short example, and feel the power. Three confirmed defects fixed plus the centerpiece added. - DELETE examples/coder-loop/: its .ts was refactored off runLoop to worktreeLoopRunner, and its README still taught the deleted coderProfile export, so it no longer demonstrated "the same runLoop kernel" it claimed. researcher-loop is now the primary runLoop teacher. - ADD examples/driver-loop/: the centerpiece. A multi-round refine driver whose plan() READS the last worker's output from history and COMPOSES the next prompt FROM it — "the fold" every supervisor is built on, made visible with heavy plain-language comments. A second labeled section contrasts it with multishot so round vs shot sit side by side. Offline, e2e-proven (round 0 rejected -> driver folds -> round 1 passes via the corrected prompt). - IMPROVE stale-API docs and teaching comments: - mcp-delegation/README + fleet-delegation/README: replace the deleted delegate_code/delegate_research tools with the generic delegate verb + the MCP_ENABLE_DELEGATE=1 gate + the always-on feedback/status/history trio. - product-eval/README: teach runPersonaConversation (evalPersona was deleted in 0.76.0); the .ts was already correct. - researcher-loop / ui-audit / self-improving-loop / supervisor-loop: inline-define round and shot the first time each appears, and point at driver-loop/ as the example that actually shows the fold. - delegate: import from the @tangle-network/agent-runtime/loops subpath instead of a relative dist path; strip emoji from console output. - INDEX examples/README.md: rewrite as an ordered "use this when" path over tiers (cores -> driver/supervisor -> runLoop kernel -> production runtime -> self-improvement), add a vocabulary block, add the missing driver-loop / supervise / delegate / intelligence-drop-in rows, drop coder-loop. All edits stay inside examples/; build + typecheck + typecheck:examples + Biome lint are green, and driver-loop and researcher-loop run offline.

tangletools

✅ Auto-approved PR — `21608ef3`

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-23T22:27:04Z}

…orker exchange shot = round = turn = one (driver prompts worker -> worker output -> driver) exchange; 'many shots' is the sequence where each output folds into the next prompt. Remove the runMultishot 'parallel shots' section: runMultishot is a multi-turn conversation, not a fanout, so it mislabeled the breadth axis. Point to researcher-loop for fanout instead.

tangletools

✅ Auto-approved PR — `56b143c1`

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-23T22:41:08Z}

drewstone · 2026-06-23T22:26:45Z

What

Make examples/ tip-top for DX: land a developer here, let them read ONE short example, and feel the power. Three confirmed defects fixed, the missing centerpiece added, and the index rewritten as an ordered "use this when" path. All edits stay inside examples/.

DELETE

examples/coder-loop/ — its .ts was refactored off runLoop onto worktreeLoopRunner, and its README still taught the deletedcoderProfile export, so it no longer demonstrated "the same runLoop kernel" it claimed. researcher-loop/ is now the primary runLoop teacher.

ADD — the centerpiece

examples/driver-loop/ — the one concept the request asked to SEE: a multi-round refine driver whose plan()reads the last worker's output from history and composes the next prompt from it ("the fold"), made visible with heavy plain-language comments. A second labeled section contrasts it with runMultishot so round vs shot sit side by side. Runs fully offline and is proven e2e:
- ROUND 0: [reject] "Shipped one-click restore for failed deploys." (no "rollback")
- driver folds the rejected output → ROUND 1: [PASS] "...with an instant rollback path..." → decision: pick-winner.
- Caught + fixed + documented a real footgun while building it: a decide() that returns a terminal value (fail) before refining stops the loop early — the example now teaches the non-terminal refine decision.

IMPROVE — stale-API docs + teaching comments

mcp-delegation/README + fleet-delegation/README: replace the deleteddelegate_code/delegate_research tools with the generic delegate verb + the MCP_ENABLE_DELEGATE=1 gate + the always-on delegate_feedback/delegation_status/delegation_history trio. (The .ts was already correct — only the READMEs lied.)
product-eval/README: teach runPersonaConversation (the facade evalPersona was deleted in 0.76.0 by chore(cleanup): delete dead eval-persona facade + orphaned topology module #367; the .ts already uses the canonical surface).
researcher-loop / ui-audit / self-improving-loop / supervisor-loop: inline-define round and shot the first time each appears and point at driver-loop/ as the example that actually shows the fold.
delegate: import from the @tangle-network/agent-runtime/loops subpath instead of a relative dist path; strip emoji from console output.

INDEX

examples/README.md rewritten as an ordered "use this when" path over tiers (cores → driver/supervisor → runLoop kernel → production runtime → self-improvement), with a vocabulary block up top, the missing driver-loop/supervise/delegate/intelligence-drop-in rows added, and coder-loop dropped.

Verification

pnpm build, pnpm run typecheck, pnpm run typecheck:examples, pnpm run lint (Biome) — all green.
pnpm run docs:check — freshness OK, no drift.
driver-loop and researcher-loop run offline; merges cleanly into origin/main.

🤖 Generated with Claude Code

Land a developer in examples/, let them read ONE short example, and feel the power. Three confirmed defects fixed plus the centerpiece added. - DELETE examples/coder-loop/: its .ts was refactored off runLoop to worktreeLoopRunner, and its README still taught the deleted coderProfile export, so it no longer demonstrated "the same runLoop kernel" it claimed. researcher-loop is now the primary runLoop teacher. - ADD examples/driver-loop/: the centerpiece. A multi-round refine driver whose plan() READS the last worker's output from history and COMPOSES the next prompt FROM it — "the fold" every supervisor is built on, made visible with heavy plain-language comments. A second labeled section contrasts it with multishot so round vs shot sit side by side. Offline, e2e-proven (round 0 rejected -> driver folds -> round 1 passes via the corrected prompt). - IMPROVE stale-API docs and teaching comments: - mcp-delegation/README + fleet-delegation/README: replace the deleted delegate_code/delegate_research tools with the generic delegate verb + the MCP_ENABLE_DELEGATE=1 gate + the always-on feedback/status/history trio. - product-eval/README: teach runPersonaConversation (evalPersona was deleted in 0.76.0); the .ts was already correct. - researcher-loop / ui-audit / self-improving-loop / supervisor-loop: inline-define round and shot the first time each appears, and point at driver-loop/ as the example that actually shows the fold. - delegate: import from the @tangle-network/agent-runtime/loops subpath instead of a relative dist path; strip emoji from console output. - INDEX examples/README.md: rewrite as an ordered "use this when" path over tiers (cores -> driver/supervisor -> runLoop kernel -> production runtime -> self-improvement), add a vocabulary block, add the missing driver-loop / supervise / delegate / intelligence-drop-in rows, drop coder-loop. All edits stay inside examples/; build + typecheck + typecheck:examples + Biome lint are green, and driver-loop and researcher-loop run offline.

tangletools

✅ Auto-approved PR — `21608ef3`

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-23T22:27:04Z}

…orker exchange shot = round = turn = one (driver prompts worker -> worker output -> driver) exchange; 'many shots' is the sequence where each output folds into the next prompt. Remove the runMultishot 'parallel shots' section: runMultishot is a multi-turn conversation, not a fanout, so it mislabeled the breadth axis. Point to researcher-loop for fanout instead.

tangletools

✅ Auto-approved PR — `56b143c1`

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-23T22:41:08Z}

drewstone · 2026-06-23T22:26:45Z

What

Make examples/ tip-top for DX: land a developer here, let them read ONE short example, and feel the power. Three confirmed defects fixed, the missing centerpiece added, and the index rewritten as an ordered "use this when" path. All edits stay inside examples/.

DELETE

examples/coder-loop/ — its .ts was refactored off runLoop onto worktreeLoopRunner, and its README still taught the deletedcoderProfile export, so it no longer demonstrated "the same runLoop kernel" it claimed. researcher-loop/ is now the primary runLoop teacher.

ADD — the centerpiece

examples/driver-loop/ — the one concept the request asked to SEE: a multi-round refine driver whose plan()reads the last worker's output from history and composes the next prompt from it ("the fold"), made visible with heavy plain-language comments. A second labeled section contrasts it with runMultishot so round vs shot sit side by side. Runs fully offline and is proven e2e:
- ROUND 0: [reject] "Shipped one-click restore for failed deploys." (no "rollback")
- driver folds the rejected output → ROUND 1: [PASS] "...with an instant rollback path..." → decision: pick-winner.
- Caught + fixed + documented a real footgun while building it: a decide() that returns a terminal value (fail) before refining stops the loop early — the example now teaches the non-terminal refine decision.

IMPROVE — stale-API docs + teaching comments

mcp-delegation/README + fleet-delegation/README: replace the deleteddelegate_code/delegate_research tools with the generic delegate verb + the MCP_ENABLE_DELEGATE=1 gate + the always-on delegate_feedback/delegation_status/delegation_history trio. (The .ts was already correct — only the READMEs lied.)
product-eval/README: teach runPersonaConversation (the facade evalPersona was deleted in 0.76.0 by chore(cleanup): delete dead eval-persona facade + orphaned topology module #367; the .ts already uses the canonical surface).
researcher-loop / ui-audit / self-improving-loop / supervisor-loop: inline-define round and shot the first time each appears and point at driver-loop/ as the example that actually shows the fold.
delegate: import from the @tangle-network/agent-runtime/loops subpath instead of a relative dist path; strip emoji from console output.

INDEX

examples/README.md rewritten as an ordered "use this when" path over tiers (cores → driver/supervisor → runLoop kernel → production runtime → self-improvement), with a vocabulary block up top, the missing driver-loop/supervise/delegate/intelligence-drop-in rows added, and coder-loop dropped.

Verification

pnpm build, pnpm run typecheck, pnpm run typecheck:examples, pnpm run lint (Biome) — all green.
pnpm run docs:check — freshness OK, no drift.
driver-loop and researcher-loop run offline; merges cleanly into origin/main.

🤖 Generated with Claude Code

Land a developer in examples/, let them read ONE short example, and feel the power. Three confirmed defects fixed plus the centerpiece added. - DELETE examples/coder-loop/: its .ts was refactored off runLoop to worktreeLoopRunner, and its README still taught the deleted coderProfile export, so it no longer demonstrated "the same runLoop kernel" it claimed. researcher-loop is now the primary runLoop teacher. - ADD examples/driver-loop/: the centerpiece. A multi-round refine driver whose plan() READS the last worker's output from history and COMPOSES the next prompt FROM it — "the fold" every supervisor is built on, made visible with heavy plain-language comments. A second labeled section contrasts it with multishot so round vs shot sit side by side. Offline, e2e-proven (round 0 rejected -> driver folds -> round 1 passes via the corrected prompt). - IMPROVE stale-API docs and teaching comments: - mcp-delegation/README + fleet-delegation/README: replace the deleted delegate_code/delegate_research tools with the generic delegate verb + the MCP_ENABLE_DELEGATE=1 gate + the always-on feedback/status/history trio. - product-eval/README: teach runPersonaConversation (evalPersona was deleted in 0.76.0); the .ts was already correct. - researcher-loop / ui-audit / self-improving-loop / supervisor-loop: inline-define round and shot the first time each appears, and point at driver-loop/ as the example that actually shows the fold. - delegate: import from the @tangle-network/agent-runtime/loops subpath instead of a relative dist path; strip emoji from console output. - INDEX examples/README.md: rewrite as an ordered "use this when" path over tiers (cores -> driver/supervisor -> runLoop kernel -> production runtime -> self-improvement), add a vocabulary block, add the missing driver-loop / supervise / delegate / intelligence-drop-in rows, drop coder-loop. All edits stay inside examples/; build + typecheck + typecheck:examples + Biome lint are green, and driver-loop and researcher-loop run offline.

tangletools

✅ Auto-approved PR — `21608ef3`

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-23T22:27:04Z}

…orker exchange shot = round = turn = one (driver prompts worker -> worker output -> driver) exchange; 'many shots' is the sequence where each output folds into the next prompt. Remove the runMultishot 'parallel shots' section: runMultishot is a multi-turn conversation, not a fanout, so it mislabeled the breadth axis. Point to researcher-loop for fanout instead.

tangletools

✅ Auto-approved PR — `56b143c1`

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-23T22:41:08Z}

drewstone · 2026-06-23T22:26:45Z

What

Make examples/ tip-top for DX: land a developer here, let them read ONE short example, and feel the power. Three confirmed defects fixed, the missing centerpiece added, and the index rewritten as an ordered "use this when" path. All edits stay inside examples/.

DELETE

examples/coder-loop/ — its .ts was refactored off runLoop onto worktreeLoopRunner, and its README still taught the deletedcoderProfile export, so it no longer demonstrated "the same runLoop kernel" it claimed. researcher-loop/ is now the primary runLoop teacher.

ADD — the centerpiece

examples/driver-loop/ — the one concept the request asked to SEE: a multi-round refine driver whose plan()reads the last worker's output from history and composes the next prompt from it ("the fold"), made visible with heavy plain-language comments. A second labeled section contrasts it with runMultishot so round vs shot sit side by side. Runs fully offline and is proven e2e:
- ROUND 0: [reject] "Shipped one-click restore for failed deploys." (no "rollback")
- driver folds the rejected output → ROUND 1: [PASS] "...with an instant rollback path..." → decision: pick-winner.
- Caught + fixed + documented a real footgun while building it: a decide() that returns a terminal value (fail) before refining stops the loop early — the example now teaches the non-terminal refine decision.

IMPROVE — stale-API docs + teaching comments

mcp-delegation/README + fleet-delegation/README: replace the deleteddelegate_code/delegate_research tools with the generic delegate verb + the MCP_ENABLE_DELEGATE=1 gate + the always-on delegate_feedback/delegation_status/delegation_history trio. (The .ts was already correct — only the READMEs lied.)
product-eval/README: teach runPersonaConversation (the facade evalPersona was deleted in 0.76.0 by chore(cleanup): delete dead eval-persona facade + orphaned topology module #367; the .ts already uses the canonical surface).
researcher-loop / ui-audit / self-improving-loop / supervisor-loop: inline-define round and shot the first time each appears and point at driver-loop/ as the example that actually shows the fold.
delegate: import from the @tangle-network/agent-runtime/loops subpath instead of a relative dist path; strip emoji from console output.

INDEX

examples/README.md rewritten as an ordered "use this when" path over tiers (cores → driver/supervisor → runLoop kernel → production runtime → self-improvement), with a vocabulary block up top, the missing driver-loop/supervise/delegate/intelligence-drop-in rows added, and coder-loop dropped.

Verification

pnpm build, pnpm run typecheck, pnpm run typecheck:examples, pnpm run lint (Biome) — all green.
pnpm run docs:check — freshness OK, no drift.
driver-loop and researcher-loop run offline; merges cleanly into origin/main.

🤖 Generated with Claude Code

Land a developer in examples/, let them read ONE short example, and feel the power. Three confirmed defects fixed plus the centerpiece added. - DELETE examples/coder-loop/: its .ts was refactored off runLoop to worktreeLoopRunner, and its README still taught the deleted coderProfile export, so it no longer demonstrated "the same runLoop kernel" it claimed. researcher-loop is now the primary runLoop teacher. - ADD examples/driver-loop/: the centerpiece. A multi-round refine driver whose plan() READS the last worker's output from history and COMPOSES the next prompt FROM it — "the fold" every supervisor is built on, made visible with heavy plain-language comments. A second labeled section contrasts it with multishot so round vs shot sit side by side. Offline, e2e-proven (round 0 rejected -> driver folds -> round 1 passes via the corrected prompt). - IMPROVE stale-API docs and teaching comments: - mcp-delegation/README + fleet-delegation/README: replace the deleted delegate_code/delegate_research tools with the generic delegate verb + the MCP_ENABLE_DELEGATE=1 gate + the always-on feedback/status/history trio. - product-eval/README: teach runPersonaConversation (evalPersona was deleted in 0.76.0); the .ts was already correct. - researcher-loop / ui-audit / self-improving-loop / supervisor-loop: inline-define round and shot the first time each appears, and point at driver-loop/ as the example that actually shows the fold. - delegate: import from the @tangle-network/agent-runtime/loops subpath instead of a relative dist path; strip emoji from console output. - INDEX examples/README.md: rewrite as an ordered "use this when" path over tiers (cores -> driver/supervisor -> runLoop kernel -> production runtime -> self-improvement), add a vocabulary block, add the missing driver-loop / supervise / delegate / intelligence-drop-in rows, drop coder-loop. All edits stay inside examples/; build + typecheck + typecheck:examples + Biome lint are green, and driver-loop and researcher-loop run offline.

tangletools

✅ Auto-approved PR — `21608ef3`

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-23T22:27:04Z}

…orker exchange shot = round = turn = one (driver prompts worker -> worker output -> driver) exchange; 'many shots' is the sequence where each output folds into the next prompt. Remove the runMultishot 'parallel shots' section: runMultishot is a multi-turn conversation, not a fanout, so it mislabeled the breadth axis. Point to researcher-loop for fanout instead.

tangletools

✅ Auto-approved PR — `56b143c1`

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-23T22:41:08Z}

drewstone · 2026-06-23T22:26:45Z

What

Make examples/ tip-top for DX: land a developer here, let them read ONE short example, and feel the power. Three confirmed defects fixed, the missing centerpiece added, and the index rewritten as an ordered "use this when" path. All edits stay inside examples/.

DELETE

examples/coder-loop/ — its .ts was refactored off runLoop onto worktreeLoopRunner, and its README still taught the deletedcoderProfile export, so it no longer demonstrated "the same runLoop kernel" it claimed. researcher-loop/ is now the primary runLoop teacher.

ADD — the centerpiece

examples/driver-loop/ — the one concept the request asked to SEE: a multi-round refine driver whose plan()reads the last worker's output from history and composes the next prompt from it ("the fold"), made visible with heavy plain-language comments. A second labeled section contrasts it with runMultishot so round vs shot sit side by side. Runs fully offline and is proven e2e:
- ROUND 0: [reject] "Shipped one-click restore for failed deploys." (no "rollback")
- driver folds the rejected output → ROUND 1: [PASS] "...with an instant rollback path..." → decision: pick-winner.
- Caught + fixed + documented a real footgun while building it: a decide() that returns a terminal value (fail) before refining stops the loop early — the example now teaches the non-terminal refine decision.

IMPROVE — stale-API docs + teaching comments

mcp-delegation/README + fleet-delegation/README: replace the deleteddelegate_code/delegate_research tools with the generic delegate verb + the MCP_ENABLE_DELEGATE=1 gate + the always-on delegate_feedback/delegation_status/delegation_history trio. (The .ts was already correct — only the READMEs lied.)
product-eval/README: teach runPersonaConversation (the facade evalPersona was deleted in 0.76.0 by chore(cleanup): delete dead eval-persona facade + orphaned topology module #367; the .ts already uses the canonical surface).
researcher-loop / ui-audit / self-improving-loop / supervisor-loop: inline-define round and shot the first time each appears and point at driver-loop/ as the example that actually shows the fold.
delegate: import from the @tangle-network/agent-runtime/loops subpath instead of a relative dist path; strip emoji from console output.

INDEX

examples/README.md rewritten as an ordered "use this when" path over tiers (cores → driver/supervisor → runLoop kernel → production runtime → self-improvement), with a vocabulary block up top, the missing driver-loop/supervise/delegate/intelligence-drop-in rows added, and coder-loop dropped.

Verification

pnpm build, pnpm run typecheck, pnpm run typecheck:examples, pnpm run lint (Biome) — all green.
pnpm run docs:check — freshness OK, no drift.
driver-loop and researcher-loop run offline; merges cleanly into origin/main.

🤖 Generated with Claude Code

Land a developer in examples/, let them read ONE short example, and feel the power. Three confirmed defects fixed plus the centerpiece added. - DELETE examples/coder-loop/: its .ts was refactored off runLoop to worktreeLoopRunner, and its README still taught the deleted coderProfile export, so it no longer demonstrated "the same runLoop kernel" it claimed. researcher-loop is now the primary runLoop teacher. - ADD examples/driver-loop/: the centerpiece. A multi-round refine driver whose plan() READS the last worker's output from history and COMPOSES the next prompt FROM it — "the fold" every supervisor is built on, made visible with heavy plain-language comments. A second labeled section contrasts it with multishot so round vs shot sit side by side. Offline, e2e-proven (round 0 rejected -> driver folds -> round 1 passes via the corrected prompt). - IMPROVE stale-API docs and teaching comments: - mcp-delegation/README + fleet-delegation/README: replace the deleted delegate_code/delegate_research tools with the generic delegate verb + the MCP_ENABLE_DELEGATE=1 gate + the always-on feedback/status/history trio. - product-eval/README: teach runPersonaConversation (evalPersona was deleted in 0.76.0); the .ts was already correct. - researcher-loop / ui-audit / self-improving-loop / supervisor-loop: inline-define round and shot the first time each appears, and point at driver-loop/ as the example that actually shows the fold. - delegate: import from the @tangle-network/agent-runtime/loops subpath instead of a relative dist path; strip emoji from console output. - INDEX examples/README.md: rewrite as an ordered "use this when" path over tiers (cores -> driver/supervisor -> runLoop kernel -> production runtime -> self-improvement), add a vocabulary block, add the missing driver-loop / supervise / delegate / intelligence-drop-in rows, drop coder-loop. All edits stay inside examples/; build + typecheck + typecheck:examples + Biome lint are green, and driver-loop and researcher-loop run offline.

tangletools

✅ Auto-approved PR — `21608ef3`

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-23T22:27:04Z}

…orker exchange shot = round = turn = one (driver prompts worker -> worker output -> driver) exchange; 'many shots' is the sequence where each output folds into the next prompt. Remove the runMultishot 'parallel shots' section: runMultishot is a multi-turn conversation, not a fanout, so it mislabeled the breadth axis. Point to researcher-loop for fanout instead.

tangletools

✅ Auto-approved PR — `56b143c1`

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-23T22:41:08Z}

drewstone · 2026-06-23T22:26:45Z

What

Make examples/ tip-top for DX: land a developer here, let them read ONE short example, and feel the power. Three confirmed defects fixed, the missing centerpiece added, and the index rewritten as an ordered "use this when" path. All edits stay inside examples/.

DELETE

examples/coder-loop/ — its .ts was refactored off runLoop onto worktreeLoopRunner, and its README still taught the deletedcoderProfile export, so it no longer demonstrated "the same runLoop kernel" it claimed. researcher-loop/ is now the primary runLoop teacher.

ADD — the centerpiece

examples/driver-loop/ — the one concept the request asked to SEE: a multi-round refine driver whose plan()reads the last worker's output from history and composes the next prompt from it ("the fold"), made visible with heavy plain-language comments. A second labeled section contrasts it with runMultishot so round vs shot sit side by side. Runs fully offline and is proven e2e:
- ROUND 0: [reject] "Shipped one-click restore for failed deploys." (no "rollback")
- driver folds the rejected output → ROUND 1: [PASS] "...with an instant rollback path..." → decision: pick-winner.
- Caught + fixed + documented a real footgun while building it: a decide() that returns a terminal value (fail) before refining stops the loop early — the example now teaches the non-terminal refine decision.

IMPROVE — stale-API docs + teaching comments

mcp-delegation/README + fleet-delegation/README: replace the deleteddelegate_code/delegate_research tools with the generic delegate verb + the MCP_ENABLE_DELEGATE=1 gate + the always-on delegate_feedback/delegation_status/delegation_history trio. (The .ts was already correct — only the READMEs lied.)
product-eval/README: teach runPersonaConversation (the facade evalPersona was deleted in 0.76.0 by chore(cleanup): delete dead eval-persona facade + orphaned topology module #367; the .ts already uses the canonical surface).
researcher-loop / ui-audit / self-improving-loop / supervisor-loop: inline-define round and shot the first time each appears and point at driver-loop/ as the example that actually shows the fold.
delegate: import from the @tangle-network/agent-runtime/loops subpath instead of a relative dist path; strip emoji from console output.

INDEX

examples/README.md rewritten as an ordered "use this when" path over tiers (cores → driver/supervisor → runLoop kernel → production runtime → self-improvement), with a vocabulary block up top, the missing driver-loop/supervise/delegate/intelligence-drop-in rows added, and coder-loop dropped.

Verification

pnpm build, pnpm run typecheck, pnpm run typecheck:examples, pnpm run lint (Biome) — all green.
pnpm run docs:check — freshness OK, no drift.
driver-loop and researcher-loop run offline; merges cleanly into origin/main.

🤖 Generated with Claude Code

Land a developer in examples/, let them read ONE short example, and feel the power. Three confirmed defects fixed plus the centerpiece added. - DELETE examples/coder-loop/: its .ts was refactored off runLoop to worktreeLoopRunner, and its README still taught the deleted coderProfile export, so it no longer demonstrated "the same runLoop kernel" it claimed. researcher-loop is now the primary runLoop teacher. - ADD examples/driver-loop/: the centerpiece. A multi-round refine driver whose plan() READS the last worker's output from history and COMPOSES the next prompt FROM it — "the fold" every supervisor is built on, made visible with heavy plain-language comments. A second labeled section contrasts it with multishot so round vs shot sit side by side. Offline, e2e-proven (round 0 rejected -> driver folds -> round 1 passes via the corrected prompt). - IMPROVE stale-API docs and teaching comments: - mcp-delegation/README + fleet-delegation/README: replace the deleted delegate_code/delegate_research tools with the generic delegate verb + the MCP_ENABLE_DELEGATE=1 gate + the always-on feedback/status/history trio. - product-eval/README: teach runPersonaConversation (evalPersona was deleted in 0.76.0); the .ts was already correct. - researcher-loop / ui-audit / self-improving-loop / supervisor-loop: inline-define round and shot the first time each appears, and point at driver-loop/ as the example that actually shows the fold. - delegate: import from the @tangle-network/agent-runtime/loops subpath instead of a relative dist path; strip emoji from console output. - INDEX examples/README.md: rewrite as an ordered "use this when" path over tiers (cores -> driver/supervisor -> runLoop kernel -> production runtime -> self-improvement), add a vocabulary block, add the missing driver-loop / supervise / delegate / intelligence-drop-in rows, drop coder-loop. All edits stay inside examples/; build + typecheck + typecheck:examples + Biome lint are green, and driver-loop and researcher-loop run offline.

tangletools

✅ Auto-approved PR — `21608ef3`

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-23T22:27:04Z}

…orker exchange shot = round = turn = one (driver prompts worker -> worker output -> driver) exchange; 'many shots' is the sequence where each output folds into the next prompt. Remove the runMultishot 'parallel shots' section: runMultishot is a multi-turn conversation, not a fanout, so it mislabeled the breadth axis. Point to researcher-loop for fanout instead.

tangletools

✅ Auto-approved PR — `56b143c1`

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-23T22:41:08Z}

drewstone · 2026-06-23T22:26:45Z

What

Make examples/ tip-top for DX: land a developer here, let them read ONE short example, and feel the power. Three confirmed defects fixed, the missing centerpiece added, and the index rewritten as an ordered "use this when" path. All edits stay inside examples/.

DELETE

examples/coder-loop/ — its .ts was refactored off runLoop onto worktreeLoopRunner, and its README still taught the deletedcoderProfile export, so it no longer demonstrated "the same runLoop kernel" it claimed. researcher-loop/ is now the primary runLoop teacher.

ADD — the centerpiece

examples/driver-loop/ — the one concept the request asked to SEE: a multi-round refine driver whose plan()reads the last worker's output from history and composes the next prompt from it ("the fold"), made visible with heavy plain-language comments. A second labeled section contrasts it with runMultishot so round vs shot sit side by side. Runs fully offline and is proven e2e:
- ROUND 0: [reject] "Shipped one-click restore for failed deploys." (no "rollback")
- driver folds the rejected output → ROUND 1: [PASS] "...with an instant rollback path..." → decision: pick-winner.
- Caught + fixed + documented a real footgun while building it: a decide() that returns a terminal value (fail) before refining stops the loop early — the example now teaches the non-terminal refine decision.

IMPROVE — stale-API docs + teaching comments

mcp-delegation/README + fleet-delegation/README: replace the deleteddelegate_code/delegate_research tools with the generic delegate verb + the MCP_ENABLE_DELEGATE=1 gate + the always-on delegate_feedback/delegation_status/delegation_history trio. (The .ts was already correct — only the READMEs lied.)
product-eval/README: teach runPersonaConversation (the facade evalPersona was deleted in 0.76.0 by chore(cleanup): delete dead eval-persona facade + orphaned topology module #367; the .ts already uses the canonical surface).
researcher-loop / ui-audit / self-improving-loop / supervisor-loop: inline-define round and shot the first time each appears and point at driver-loop/ as the example that actually shows the fold.
delegate: import from the @tangle-network/agent-runtime/loops subpath instead of a relative dist path; strip emoji from console output.

INDEX

examples/README.md rewritten as an ordered "use this when" path over tiers (cores → driver/supervisor → runLoop kernel → production runtime → self-improvement), with a vocabulary block up top, the missing driver-loop/supervise/delegate/intelligence-drop-in rows added, and coder-loop dropped.

Verification

pnpm build, pnpm run typecheck, pnpm run typecheck:examples, pnpm run lint (Biome) — all green.
pnpm run docs:check — freshness OK, no drift.
driver-loop and researcher-loop run offline; merges cleanly into origin/main.

🤖 Generated with Claude Code

Land a developer in examples/, let them read ONE short example, and feel the power. Three confirmed defects fixed plus the centerpiece added. - DELETE examples/coder-loop/: its .ts was refactored off runLoop to worktreeLoopRunner, and its README still taught the deleted coderProfile export, so it no longer demonstrated "the same runLoop kernel" it claimed. researcher-loop is now the primary runLoop teacher. - ADD examples/driver-loop/: the centerpiece. A multi-round refine driver whose plan() READS the last worker's output from history and COMPOSES the next prompt FROM it — "the fold" every supervisor is built on, made visible with heavy plain-language comments. A second labeled section contrasts it with multishot so round vs shot sit side by side. Offline, e2e-proven (round 0 rejected -> driver folds -> round 1 passes via the corrected prompt). - IMPROVE stale-API docs and teaching comments: - mcp-delegation/README + fleet-delegation/README: replace the deleted delegate_code/delegate_research tools with the generic delegate verb + the MCP_ENABLE_DELEGATE=1 gate + the always-on feedback/status/history trio. - product-eval/README: teach runPersonaConversation (evalPersona was deleted in 0.76.0); the .ts was already correct. - researcher-loop / ui-audit / self-improving-loop / supervisor-loop: inline-define round and shot the first time each appears, and point at driver-loop/ as the example that actually shows the fold. - delegate: import from the @tangle-network/agent-runtime/loops subpath instead of a relative dist path; strip emoji from console output. - INDEX examples/README.md: rewrite as an ordered "use this when" path over tiers (cores -> driver/supervisor -> runLoop kernel -> production runtime -> self-improvement), add a vocabulary block, add the missing driver-loop / supervise / delegate / intelligence-drop-in rows, drop coder-loop. All edits stay inside examples/; build + typecheck + typecheck:examples + Biome lint are green, and driver-loop and researcher-loop run offline.

tangletools

✅ Auto-approved PR — `21608ef3`

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-23T22:27:04Z}

…orker exchange shot = round = turn = one (driver prompts worker -> worker output -> driver) exchange; 'many shots' is the sequence where each output folds into the next prompt. Remove the runMultishot 'parallel shots' section: runMultishot is a multi-turn conversation, not a fanout, so it mislabeled the breadth axis. Point to researcher-loop for fanout instead.

tangletools

✅ Auto-approved PR — `56b143c1`

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-23T22:41:08Z}

drewstone · 2026-06-23T22:26:45Z

What

Make examples/ tip-top for DX: land a developer here, let them read ONE short example, and feel the power. Three confirmed defects fixed, the missing centerpiece added, and the index rewritten as an ordered "use this when" path. All edits stay inside examples/.

DELETE

examples/coder-loop/ — its .ts was refactored off runLoop onto worktreeLoopRunner, and its README still taught the deletedcoderProfile export, so it no longer demonstrated "the same runLoop kernel" it claimed. researcher-loop/ is now the primary runLoop teacher.

ADD — the centerpiece

examples/driver-loop/ — the one concept the request asked to SEE: a multi-round refine driver whose plan()reads the last worker's output from history and composes the next prompt from it ("the fold"), made visible with heavy plain-language comments. A second labeled section contrasts it with runMultishot so round vs shot sit side by side. Runs fully offline and is proven e2e:
- ROUND 0: [reject] "Shipped one-click restore for failed deploys." (no "rollback")
- driver folds the rejected output → ROUND 1: [PASS] "...with an instant rollback path..." → decision: pick-winner.
- Caught + fixed + documented a real footgun while building it: a decide() that returns a terminal value (fail) before refining stops the loop early — the example now teaches the non-terminal refine decision.

IMPROVE — stale-API docs + teaching comments

mcp-delegation/README + fleet-delegation/README: replace the deleteddelegate_code/delegate_research tools with the generic delegate verb + the MCP_ENABLE_DELEGATE=1 gate + the always-on delegate_feedback/delegation_status/delegation_history trio. (The .ts was already correct — only the READMEs lied.)
product-eval/README: teach runPersonaConversation (the facade evalPersona was deleted in 0.76.0 by chore(cleanup): delete dead eval-persona facade + orphaned topology module #367; the .ts already uses the canonical surface).
researcher-loop / ui-audit / self-improving-loop / supervisor-loop: inline-define round and shot the first time each appears and point at driver-loop/ as the example that actually shows the fold.
delegate: import from the @tangle-network/agent-runtime/loops subpath instead of a relative dist path; strip emoji from console output.

INDEX

examples/README.md rewritten as an ordered "use this when" path over tiers (cores → driver/supervisor → runLoop kernel → production runtime → self-improvement), with a vocabulary block up top, the missing driver-loop/supervise/delegate/intelligence-drop-in rows added, and coder-loop dropped.

Verification

pnpm build, pnpm run typecheck, pnpm run typecheck:examples, pnpm run lint (Biome) — all green.
pnpm run docs:check — freshness OK, no drift.
driver-loop and researcher-loop run offline; merges cleanly into origin/main.

🤖 Generated with Claude Code

Land a developer in examples/, let them read ONE short example, and feel the power. Three confirmed defects fixed plus the centerpiece added. - DELETE examples/coder-loop/: its .ts was refactored off runLoop to worktreeLoopRunner, and its README still taught the deleted coderProfile export, so it no longer demonstrated "the same runLoop kernel" it claimed. researcher-loop is now the primary runLoop teacher. - ADD examples/driver-loop/: the centerpiece. A multi-round refine driver whose plan() READS the last worker's output from history and COMPOSES the next prompt FROM it — "the fold" every supervisor is built on, made visible with heavy plain-language comments. A second labeled section contrasts it with multishot so round vs shot sit side by side. Offline, e2e-proven (round 0 rejected -> driver folds -> round 1 passes via the corrected prompt). - IMPROVE stale-API docs and teaching comments: - mcp-delegation/README + fleet-delegation/README: replace the deleted delegate_code/delegate_research tools with the generic delegate verb + the MCP_ENABLE_DELEGATE=1 gate + the always-on feedback/status/history trio. - product-eval/README: teach runPersonaConversation (evalPersona was deleted in 0.76.0); the .ts was already correct. - researcher-loop / ui-audit / self-improving-loop / supervisor-loop: inline-define round and shot the first time each appears, and point at driver-loop/ as the example that actually shows the fold. - delegate: import from the @tangle-network/agent-runtime/loops subpath instead of a relative dist path; strip emoji from console output. - INDEX examples/README.md: rewrite as an ordered "use this when" path over tiers (cores -> driver/supervisor -> runLoop kernel -> production runtime -> self-improvement), add a vocabulary block, add the missing driver-loop / supervise / delegate / intelligence-drop-in rows, drop coder-loop. All edits stay inside examples/; build + typecheck + typecheck:examples + Biome lint are green, and driver-loop and researcher-loop run offline.

tangletools

✅ Auto-approved PR — `21608ef3`

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-23T22:27:04Z}

…orker exchange shot = round = turn = one (driver prompts worker -> worker output -> driver) exchange; 'many shots' is the sequence where each output folds into the next prompt. Remove the runMultishot 'parallel shots' section: runMultishot is a multi-turn conversation, not a fanout, so it mislabeled the breadth axis. Point to researcher-loop for fanout instead.

tangletools

✅ Auto-approved PR — `56b143c1`

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-23T22:41:08Z}

tangletools previously approved these changes Jun 23, 2026
View reviewed changes

drewstone dismissed tangletools’s stale review via 56b143cJune 23, 2026 22:41

tangletools approved these changes Jun 23, 2026
View reviewed changes

drewstone merged commit cfacf1c into mainJun 23, 2026
1 check passed