Repository navigation
2026-11-13. Three things found wrong by a review of the plugin and of the maintainer's dogfood project.
-
Session start, the session-end summary and the reflexion count now count corrections the way the prompt
line does.preflight.shwas corrected on 2026-08-09 to count what_corrections_lib.pycalls an entry: a
dated heading of depth 2 to 4, or a dated bold bullet.session-start.sh,stop-check.shand
reconcile_reflexions.pystill counted###headings only. On the dogfood project one session read "244
corrections logged" at its start and "447 corrections in memory" at every prompt, from the same file. -
A correction written as
## <date>or as a dated bullet now answers a reflexion.
reconcile_reflexions.pycredits a correction against the command failures that prompted one. It did not see
those two shapes, so in a project that writes its entries that way the "unreconciled reflexions" line only
rose. The ledger (.claude/state/reflexion-ledger.json) now names the counter its baseline was taken with
(corrections_counter). A ledger written before this is carried over on the next run: entries dated after
the day the ledger was written are credited, older ones are not. On the dogfood project that moves the line
from 64 to 47. -
Pages that said what is no longer so.
AGENTS.mdand/mycelium:startsaid 45 skills (there are 63, and
neither now states a number). The hooks page was headed "5-Layer", listed six, and said hooks are registered
in.claude/settings.json; in plugin form they are registered in the plugin'shooks/hooks.json. The
glossary pointed atCLAUDE.mdfor the pre-task and pre-ship protocols, which live in
engine/agent-operating-contract.md. -
Tests.
test_correction_count_agreement.pynow runs each of the three hooks' patterns against the shared
fixture, and fails a hook that counts another way.test_reconcile_reflexions.pycovers the two entry shapes,
an undated heading, and the carry-over of an old ledger with and without a day on it. -
Alternatives considered.
- One shared shell function for the count: it needs a new sourced file in every hook, and a hook whose
source line fails would count nothing. Three copies held together by one test is whatpreflight.sh
already did. - Compare the old baseline with the new count: it would credit every entry the old count never saw (203 on
the dogfood project) in one step, and the line would read zero for work nobody did. - Carry over the old credit and nothing else: it leaves out the corrections written since the ledger, which
are the ones the old count wrongly ignored.
- One shared shell function for the count: it needs a new sourced file in every hook, and a hook whose
-
Not in this release.
stop-check.shcounts decision-log entries by###heading; nothing defines what
a decision entry is, so that count is left as it is.manifest.ymlstill describes a legacy copy that
Check 28 keeps in step; removing the copy is its own change.
A CI job that runs the bash and Python tests on macOS was added in this change and taken out again. On
its first run 47 of 86 bash test files failed under the bash 3.2 that macOS ships, 45 of them because
tests/validate-template.shdoes not parse there. Every hook and plugin shell script parses under 3.2.
Until the suite runs on that bash, no CI job runs the half of each hook branch that a Mac takes.