Policy (GOALS.md maintenance rules, mirrored in root AGENTS.md and goals/AGENTS.md):
- agents apply goal-entry adds/edits only on user request, proposal-first: show the
proposed wording and get explicit approval before writing (replaces "only the user
adds or edits goals" - the user was applying agent-drafted edits by hand anyway)
- suggesting candidate goals is welcome and encouraged when grounded in the work at
hand; a suggestion is a flagged proposal, never a file edit; bulk goal surveys stay
out of bounds
- detail-file updates from user-directed work on a goal proceed without pre-approval,
reported in the completion summary; index stays canonical
- unchanged: achieved-flip on met acceptance (reported), no proposed->active flips
G-038 (user-approved):
- Acceptance gains a post-restart terminal-query clause: a code-interp query (e.g.
'help console') must succeed instead of being refused by the stale settled
can_respond=0 anchor
- detail file records the proven mechanism for the post-restart console failure
(piped phase settles can_respond=0; punkshell.tcl opens CONIN$ as a new channel
leaving the repl thread's stdin as the exhausted pipe; the anchored default console
keeps the dead pair; forced resettle honestly re-settles 0 via layer-1 eof), the
rebind requirements (re-point {in out} AND unsettle), and the hidden-console
repro/regression harness
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
G-036: the CloseHandle trap script is embedded in the detail file (was a
dangling reference to session-ephemeral scratchpad). G-038: the
repl::reopen_stdin race analysis from the G-036 investigation is written in
(the prior cross-reference pointed at content that existed only in
conversation), and the goal now explicitly owns that path's fate - it runs
for real on console-EOF at an interactive prompt, and the caller-driven
restart should cover that scenario so reopen_stdin can be retired. G-037:
twapi vendorlib(5.0b1)/kit-vfs(5.2.0) drift recorded - divergence exists in
both directions, so the sync step needs explicit authority semantics.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
User note recorded: with per-subshell configs (G-008/G-009), piping
`subshell <name>` then going interactive should reasonably land IN that
subshell (quit unwinding to the initial shell). Does not fall out of the
continuity changes automatically - the subshell stack is repl control flow
and unwinds at pipe-EOF under the current nested model. Options recorded:
re-enter captured subshell stack on restart (fights nesting) vs the natural
fit under G-002 non-nested subshells with console targeting (reattach
reopened console to innermost active subshell). Also noted: a -subshell
flag on the shell subcommand as the primary entry (G-032 launcher-args
candidate).
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
User identified the general least-surprise violation beyond the original
error-retention draft: `'set ::jjj blah' | punkexe` reaches an interactive
prompt where ::jjj does not exist - the piped-EOF restart silently swaps in
a fresh codethread/code interp, discarding variables, procs, packages, cwd
and errorInfo. Goal reworked to session continuity via the caller-driven
restart identified during G-036 (repl::start eof-restart done-mode without
codethread teardown; app-punkshell restarts on the reopened console without
repl::init). Error inspectability becomes an acceptance criterion that
likely falls out free (no re-init means no boot-noise re-stamping of
errorInfo); the standalone error-record mechanism is demoted to a recorded
fallback. Detail file renamed to match (slug change, ID preserved).
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
Julian Noble