The architect now prioritizes the roadmap + an internal scan into a single
ranked, issue-linked queue in internal/docs/PRIORITIES.md (re-ranked each run),
and the hourly increment loop works the top open item from that queue instead
of independently guessing each hour — falling back to its own judgment only
when the queue is empty. So work is roadmap-driven by default and the human can
redirect by reordering the file or its issues.
- New internal/docs/PRIORITIES.md (seeded from the roadmap + open issues).
- architecture-review.yml: architect owns/re-ranks PRIORITIES.md and keeps each
top item backed by a scoped issue.
- continuous-improvement.yml: pick the top open queue item; close both the
priority issue and the run tracker.
- CONTINUOUS_IMPROVEMENT.md: document the architect → queue → increment pipeline.
Co-authored-by: Claude <noreply@anthropic.com>
With branch protection + "Allow auto-merge" now enabled on master, Codex
enables GitHub auto-merge on its own PR (gh pr merge --squash --auto) right
after opening it, so the PR lands the moment the required CI checks pass —
no polling sweep, and the green-CI gate is enforced by GitHub instead of by
gh pr checks in a cron. Removes auto-merge-codex.yml and updates the dispatch
and AGENTS.md accordingly.
Claude-Session: https://claude.ai/code/session_01CmdEY7pYmV5zzwCjNJ4ykL
Co-authored-by: Claude <noreply@anthropic.com>
Codex was pushing to a generic branch (e.g. "work"), which the auto-merge
sweep ignores (it only matches codex/* branches with the codex label) and
which collides across runs. The dispatch and AGENTS.md now instruct Codex to
create a unique codex/increment-<issue> branch and pass --label codex to
gh pr create, so every increment is isolated and actually swept.
Claude-Session: https://claude.ai/code/session_01CmdEY7pYmV5zzwCjNJ4ykL
Co-authored-by: Claude <noreply@anthropic.com>
Codex's make_pr tool in the Cloud sandbox is a no-op stub — it records the
PR title/body and returns them "for downstream consumption" (the manual
"Create PR" click), but never pushes a branch or calls the GitHub API. That
is why "make_pr called, nothing happened": the tool literally cannot open a
PR. With the gh CLI now installed in the Codex setup and origin pointed at
the repo, the dispatch tells Codex to push and open the PR itself
(git push + gh pr create) instead of relying on make_pr.
Claude-Session: https://claude.ai/code/session_01CmdEY7pYmV5zzwCjNJ4ykL
Co-authored-by: Claude <noreply@anthropic.com>
Codex derives its PR branch name from the triggering issue's context, so
re-commenting on a single tracker issue (#3024) every hour collapsed every
increment onto one branch name. The first increment opened a PR; the rest
collided on the occupied branch and silently failed to open one — which is
why repeated runs produced "make_pr called, nothing happened."
Open a unique issue per run and dispatch Codex there, so each increment gets
its own branch and a clean PR. The dispatch asks Codex to "Closes #<issue>"
so each tracking issue auto-closes when its PR merges. The explicit
branch-name request (which Codex ignored in favor of the issue-derived name)
is dropped.
Claude-Session: https://claude.ai/code/session_01CmdEY7pYmV5zzwCjNJ4ykL
Co-authored-by: Claude <noreply@anthropic.com>
* fix: harden flow step execution against nil Run and cancellation
A step with no Run function panicked the run; it now returns a clear
configuration error. The retry loop also kept retrying after the run's
context was canceled or its deadline passed — it now stops immediately
and surfaces the context error, preserving cancellation/deadline
semantics for durable workflow runs. Adds regression coverage for both.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CmdEY7pYmV5zzwCjNJ4ykL
* ci: give each Codex dispatch a unique branch to avoid collisions
Codex derives its PR branch name from the dispatch text. The hourly loop
posted an identical generic prompt every run, so every increment resolved
to the same branch name — the previous increment's branch (until auto-
merged and deleted) blocked the next PR from opening. Include the run
number in the prompt and explicitly request a fresh
codex/improvement-<run_number> branch, so each increment is isolated.
auto-merge-codex.yml still matches (codex/* prefix) and deletes on merge.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CmdEY7pYmV5zzwCjNJ4ykL
---------
Co-authored-by: Claude <noreply@anthropic.com>
* thesis: position Go Micro as complementary to LangChain, not competing
Add a 'Where we fit' section: 'harness' has two layers — the intra-agent
harness (single-model runtime: prompts, tools, context, sandbox, the Ralph
loop) that LangChain/LangGraph/deepagents/Claude Code own and we do NOT
compete with, and the operational harness (the distributed substrate agents
operate inside: services-as-tools, discovery, durable runs, observability,
interop, the services->agents->workflows lifecycle) which is our focus. They
stack and interoperate via MCP/A2A; we make those agents better neighbours,
not obsolete.
* ci: gate Codex dispatch on CODEX_TRIGGER_TOKEN secret
Codex ignores @codex comments authored by the github-actions bot, so the
hourly dispatch was producing no PRs while still posting to the tracker
issue. Gate the comment step on the presence of CODEX_TRIGGER_TOKEN: the
workflow now no-ops until a PAT for a Codex-followed account is set, then
activates automatically with no further change.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CmdEY7pYmV5zzwCjNJ4ykL
---------
Co-authored-by: Claude <noreply@anthropic.com>
Add internal/docs/THESIS.md — the vision the autonomous loop steers by: a
holistic agent harness AND service framework encapsulating the lifecycle of
services -> agents -> workflows (workloads come after agents; the value is in
composing it into systems that do real work, on schedules and in loops).
Wire it in as the alignment criterion: the continuous-improvement charter and
the Codex dispatch prompt now require every increment to advance the North Star,
so improvements compound toward the thesis instead of drifting locally.
Co-authored-by: Claude <noreply@anthropic.com>
- continuous-improvement: cadence 12h -> hourly.
- add auto-merge-codex workflow: every 15 min, merge open PRs that are
codex-labelled AND from a codex/* branch once all checks are green
(gh pr checks). CI (build/test/lint/harnesses) is the only gate — no
human involvement. Scoped tightly so nothing else auto-merges.
Co-authored-by: Claude <noreply@anthropic.com>
A Claude Max subscription exposes no API key for CI, and Atlas Cloud models
can't run a coding agent — so the durable scheduler triggers Codex instead:
on a cadence it posts an @codex instruction on the tracker issue (#3024) to
run one improvement increment. Prefers a CODEX_TRIGGER_TOKEN PAT if set
(in case Codex ignores the Actions bot), else the default token.
Co-authored-by: Claude <noreply@anthropic.com>
Define the autonomous improvement loop (internal/docs/CONTINUOUS_IMPROVEMENT.md):
full autonomy with correctness (build/test/lint) as the only gate, work sourced
from roadmap + issues + an improvement radar + dogfooding, Claude Code driving
and Codex executing scoped tasks, with brand/positioning and breaking API kept
with the human.
Add a durable scheduled GitHub Action (.github/workflows/continuous-improvement.yml)
as the session-independent backbone — a safe no-op until an ANTHROPIC_API_KEY
secret is added.
Co-authored-by: Claude <noreply@anthropic.com>
Asim Aslam