Skip to content

v0.337.0 - Twenty of the maintainer's rulings: solo releases, a default second view, and records that say what they are

Choose a tag to compare

@github-actions github-actions released this 12 Nov 20:48
· 1 commit to main since this release
b037cb0

2026-11-12. The maintainer ruled on the design questions the upstream-candidate triage could not settle on its
own, by interview, one option per question. This release builds the twenty rulings that call for a change.

Releases and gates

  • A team-only exposure is its own trial at L3. An exposure whose audience starts with team-only passes the
    release Evidence gate as n/a with that reason. An exposure to anyone else still needs outside evidence. This
    tightens one case: until now an L3 Evidence of n/a let a release through for any audience.
  • A service check can be scoped to one exposure. /mycelium:service-check run for one exposure writes the 15
    principles to that exposure's gates_basis.service_quality with a decision-log line, and the release gate reads
    that record first. services.yml stays the product-wide record and is not touched.
  • Viability asks "should we build it?", pointing at /mycelium:harms-check, in theory-gates.md, ice-score and
    ost-builder. Still four risks, as Cagan has them.
  • A blind second view runs by default after /mycelium:wardley-map emits a map and before a Four Risks gate is
    recorded as passed (devils-advocate Technique 7). It can only raise concerns, never pass anything; its result goes
    on the decision's note; the person may decline it, and the decline is recorded. devils-advocate now says once that
    a blind pass stands in for judgement, never for stake.
  • "Make compliance the cheapest option" is a design principle, and the pre-ship checklist gains a step for every
    new gate: compare the cost of complying with the cheapest bypass, exemption fields included.
  • diamond-progress asks the opposite question at an L1 or L2 target, as define-done does; a prompt, not a block.
  • A dismissal made from metadata, when the artifact is one call away, needs the artifact opened first (one new
    section in the operating contract).

What the agent is told

  • A "fix" ruling says "fix owed" instead of "muted until you rule", and the item stays in front of the person.
  • A shell command that writes a canvas file not read this session is reported, pointing to Edit or
    safe_replace.py. Nothing is undone.
  • Records past 256 KB, more than one Read call takes, are named with their 30-day growth, in canvas-health and
    at session start when one is (scripts/check_record_size.py). Nothing moves.
  • A retired hook label's cancellation no longer counts against the delivery SLO in dora-check.
  • Two warnings at write time (post-write-nudge.sh): an instrument saved while an unscored human task on the
    same diamond is in flight, and an outreach task started while go-to-market.yml is unread this session.
  • A diamond that is not complete and measures a delivery stream (dora-metrics) gets a WARN; diamond-rules says a
    stream lives in dora-metrics.yml, not on a diamond.
  • Confidence weights are guidance for a judgement; no code computes confidence from them, and the files now say
    so. theory_gates_status values are an enum in the diamonds schema.

What gets written down

  • Take-away files: /mycelium:handoff, /mycelium:user-interview and /mycelium:launch-tier write what the
    person will use later to .claude/handoffs/<date>-<slug>.md as a step, and say where.

  • acting_on on an opportunity records, per story, whether the person was acting on the need (under way, tried
    options) or living with it; validate_canvas counts them on a COVERAGE line, and ost-builder carries the counts into
    the L2 target comparison.

  • triggers on an L1 records trigger sets (when, then, rule out); wardley-map writes them, diamond-assess
    reports them. diamond-rules says why there is no separate container type.

  • jtbd-map and its schema say they carry part 1 of the job spec only, naming the four parts left out.

  • assumption-test: literature can answer claims about people in general, graded by what was measured on whom;
    claims about this product need a human.

  • Install: in a Claude Code Project, add the plugin under Project settings > Plugins; whether hooks then fire is
    not yet checked.

  • Alternatives considered. Each ruling was the maintainer's pick between options shown to him; the options he
    did not pick are recorded with the rulings in his project. Within them:

    • Team-only: a new boolean on the exposure (a token on the existing audience field says the same without a field).
    • Default blind pass: failing the gate when it does not run (the ruling made it a default, not a block).
    • The dismissal rule's place: inside Communication Rules (that contract section was within a few hundred
      characters of its part's cap, and past the cap no contract is delivered at all).
    • Record size: a threshold set by an environment variable (nothing asked for one; 256 KB is the size the Read tool refuses, which says why the line is there).
    • Retrodiction checks: gate-set entries (they are write-time nudges, so they run where the write happens).
    • acting_on: prose in the interview notes (no reader could count it).