Skip to content

Releases

v0.2.0

Choose a tag to compare

@debedb debedb released this 19 Nov 00:43
5a6d63d

The first release since v0.1.1 on 2026-10-02, covering twenty-one commits.

Read this before upgrading

The license changed. This project is now AGPL-3.0-or-later; it was MIT
through v0.1.1 (#130). If you were depending on the MIT terms, v0.1.1 is the
last release under them.

Every line a run writes to its log is now timestamped (#140). A log parser
that matched from the start of a line needs the prefix added, or it will match
nothing. Two idioms that still work are in the README under "Reading a run log".

Nothing else in this range needs a config change or a hand migration.

Signal as a second interface

A signal-messenger destination delivers to Signal on its own account (#120),
and a transport: "gateway" switch routes that through a gateway service that
owns the account instead of a local subprocess (#128). A private group's id
stays in the gateway's config rather than the agent's: the destination names a
conversation alias. Signal's message ceiling is 2000 UTF-8 bytes, not
UTF-16 units, and the splitter and the limit check now share one counter, so
messages that fit are no longer refused (#127).

Approvals

A proposal can be resolved by a loopback link from the card, for deployments
where Slack buttons are not available (#124). One listener now serves several
workspaces, routing each click to the workspace that owns the proposal (#109).
A slack-draft action drafts on approval for pasting by hand (#108), and a
signal-mode destination raises one proposal per signal rather than per run
(#107). The button listener exits when its Socket Mode connection is gone, so
the supervisor restarts it instead of leaving every button dead on a process
that looks healthy (#143).

A log you can draw conclusions from

The run log now reports the score of every scored signal, which is what makes a
min_score gate measurable at all (#131), and names the waterfall rung that
produced it, so a drifting corpus can be told from a drifting scorer (#134).
The Reddit source says how many feeds answered and names the ones that did not,
rather than reporting an empty run as a quiet success (#135). The LLM scorer
got named bands instead of two anchors (#136), and a signal near the gate is
scored more than once where the noise can flip the outcome -- measured at plus
or minus 16 points on identical input at temperature zero (#137).

Sources and the model waterfall

The websearch source has a Firecrawl backend, with Brave as the fallback when
Firecrawl answers 402 (#141). The Reddit source fails closed when a
subreddit's rules cannot be read, rather than dropping the promotion guard
silently (#106). In the waterfall, a hung rung is cooled rather than retried
into the same hang, and a 402 parks a rung on its status alone (#91).

Also

Two public-preparation passes removed a leaking example, closed two unignored
directories, and settled the license file (#110, #129).

v0.1.1 — tweet length counts URLs as X does

Choose a tag to compare

@debedb debedb released this 03 Oct 17:07
2e621e3

One fix, found by a live failure rather than by review.

The September release-notes announcement was refused and never posted:

twitter-post: rendered text is 304 characters, over the 280 limit

It would have posted. The blurb was 208 characters and the permalink 95, but X wraps every link to t.co and counts it as 23 regardless of length, so X's own count was 232.

The check was a plain len(). The module docstring had anticipated the opposite error — something passing len() and still failing X's weighted count, rejected by the API with a 403, "which is just as loud" — but not this one. This direction is the worse of the two: the step runs inside an approved action, after a human has read the text and after earlier steps have already published, so a false refusal marks the whole proposal failed with the blog post already live.

Verified against the exact text refused: 305 raw, 232 weighted, accepted. A blurb with no URL is unchanged, two URLs count 23 each, and 300 plain characters is still refused.

X's other weighting rule — some scripts counting 2 characters each — is still left to the API to reject, and the docstring now says so rather than implying the check is complete.

v0.1.0 — first tagged release

Choose a tag to compare

@debedb debedb released this 01 Oct 04:50
e5f0eb0

First tagged release. The framework for small, single-purpose lead-finding and digest agents that run as short-lived, launchd-invoked processes — not daemons.

53 commits, and the first tag: the repo has been consumed by pinning a git SHA, so this is the point at which the declared version starts meaning anything.

Plugins

All resolved from config by type:.

kind
sources agentmail, github-org, github-releases, ics, multi, reddit, slack-workspace, websearch
scorers llm, regex
destinations approval, digest, fanout, slack
actions command, slack-post, twitter-post, wordpress-publish

What the framework owns

An entrypoint contract (agent run — one invocation, one pass, exits 0/nonzero), plugin construction from config, dedupe between runs, an LLM waterfall across several vendors with per-vendor budget parking, launchd scheduling from a workspace, and approval-gated actions: a destination records a proposal, and a separate pass executes it only after a trusted human says so — by reply, or by clicking a Block Kit button over Socket Mode.

The recurring theme

Several decisions look paranoid, and they are, for one reason: a broken source must never read as a quiet one.

  • reddit raises when every feed fails.
  • ics raises on a calendar containing no events at all.
  • websearch raises when every query fails.
  • multi raises when every child fails, however optional each one was.
  • The runner prints one summary line on every invocation, including a failing one, emitted from a finally.

The repo has twice been bitten by a source that reported success while extracting nothing. Most of these guards are that lesson written down rather than abstract defensiveness.

Known gaps, stated rather than discovered

  • An approval proposal resolved by a typed reply leaves no mark on its announcement, so it can look unanswered.
  • multi cannot honour mixed dedupe shapes among its children; it resolves toward re-delivery, on the grounds that a duplicate is visible while a silently emptied window is not.
  • A window digest's "nothing new" cannot distinguish an empty window from an exhausted dedupe store without the operator knowing which they configured.