Agent harness: a bounded tool-calling loop with the WebContainer as control plane - #37

Merged

twilson63 merged 30 commits into

mainfrom

agent-harness

Jul 28, 2026

Commits on Jul 27, 2026

agent-harness: PRD + progress docs, plan gate resolved

twilson63
and
claude
committed
a770c7e
agent-harness U1: server spine — models, limiter split, step tokens

twilson63
and
claude
committed
b202693
agent-harness U3: tool-call transport (plan-gate blocker B1)

twilson63
and
claude
committed
fbd0caa
agent-harness U2: the loop as a pure Gleam actor (unwired)

twilson63
and
claude
committed
84c3f61
agent-harness U5: fs tools + exec on the WebContainer

twilson63
and
claude
committed
d741851
agent-harness U4+U6: step endpoint, tool-mode prompt, parity guards

twilson63
and
claude
committed
261f594
agent-harness: progress doc, evidence, and residuals for U1-U6

twilson63
and
claude
committed
a4e70ab
agent-harness U7a+U7b: wire the loop, retire the patch protocol

twilson63
and
claude
committed
29009c4
agent-harness U8: activity trail + two bugs the browser probe caught

twilson63
and
claude
committed
9cd4876
agent-harness U9: the job picker — model choice as a product decision

twilson63
and
claude
committed
656124e
agent-harness: progress doc reflects release 1 complete

twilson63
and
claude
committed
b67483e
agent-harness: handoff records release 1 complete

twilson63
and
claude
committed
bf7f8bb
agent-harness U11: server web tools, SSRF guard, web_post approval

twilson63
and
claude
committed
adffb2c
agent-harness U11b: the web_post approval flow, end to end

twilson63
and
claude
committed
0285b5e
agent-harness: fix three blockers found by the implementation judge

twilson63
and
claude
committed
13a4a12
agent-harness: progress doc records U11 and the judge pass

twilson63
and
claude
committed
ca1d8ce
agent-harness: handoff + memory record U11 and the judge pass

twilson63
and
claude
committed
166d467
agent-harness: close the two judge advisories (A1 race, A2 lost web reads)

twilson63
and
claude
committed
ccdacfc
agent-harness U10: undo the last build

twilson63
and
claude
committed
2834e2a
agent-harness U12: fs_delete — the last unit

twilson63
and
claude
committed
9a9e47b
agent-harness: progress doc records U10 and U12 complete

twilson63
and
claude
committed
dcc7eb2
agent-harness: memory + handoff record all twelve units complete

twilson63
and
claude
committed
0a120c2
BYOK-OpenRouter tool mode, and two bugs it surfaced

twilson63
and
claude
committed
2cd709c
agent-harness: docs + memory record BYOK tool mode

twilson63
and
claude
committed
5fe1dae
live-model run: the harness works, and it found the bug that mattered most

twilson63
and
claude
committed
c241b2a
agent-harness: docs + memory record the live-model validation

twilson63
and
claude
committed
28f054c
live suite: three harder turns in one session, and two prompt fixes it exposed

twilson63
and
claude
committed
a1e1c9a
agent-harness: docs + memory record the live suite results

twilson63
and
claude
committed
a41b47d

Commits on Jul 28, 2026

fix four defects found by review: one destroys an unrelated project's files

twilson63
and
claude
committed
d868afd
deploy: declare BRAVE_SEARCH_API_KEY so web_search is not inert in prod

twilson63
and
claude
committed
aa0578f

Navigation Menu

Agent harness: a bounded tool-calling loop with the WebContainer as control plane - #37

Agent harness: a bounded tool-calling loop with the WebContainer as control plane#37
twilson63 merged 30 commits into
mainfrom
agent-harness

Commits on Jul 27, 2026

agent-harness: PRD + progress docs, plan gate resolved

agent-harness U1: server spine — models, limiter split, step tokens

agent-harness U3: tool-call transport (plan-gate blocker B1)

agent-harness U2: the loop as a pure Gleam actor (unwired)

agent-harness U5: fs tools + exec on the WebContainer

agent-harness U4+U6: step endpoint, tool-mode prompt, parity guards

agent-harness: progress doc, evidence, and residuals for U1-U6

agent-harness U7a+U7b: wire the loop, retire the patch protocol

agent-harness U8: activity trail + two bugs the browser probe caught

agent-harness U9: the job picker — model choice as a product decision

agent-harness: progress doc reflects release 1 complete

agent-harness: handoff records release 1 complete

agent-harness U11: server web tools, SSRF guard, web_post approval

agent-harness U11b: the web_post approval flow, end to end

agent-harness: fix three blockers found by the implementation judge

agent-harness: progress doc records U11 and the judge pass

agent-harness: handoff + memory record U11 and the judge pass

agent-harness: close the two judge advisories (A1 race, A2 lost web reads)

agent-harness U10: undo the last build

agent-harness U12: fs_delete — the last unit

agent-harness: progress doc records U10 and U12 complete

agent-harness: memory + handoff record all twelve units complete

BYOK-OpenRouter tool mode, and two bugs it surfaced

agent-harness: docs + memory record BYOK tool mode

live-model run: the harness works, and it found the bug that mattered most

agent-harness: docs + memory record the live-model validation

live suite: three harder turns in one session, and two prompt fixes it exposed

agent-harness: docs + memory record the live suite results

Commits on Jul 28, 2026

fix four defects found by review: one destroys an unrelated project's files

deploy: declare BRAVE_SEARCH_API_KEY so web_search is not inert in prod