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-039 (proposed): investigate/reproduce the orphaned-shell one-core spin
observed 2026-07-08 (single hard-looping thread after the hosting terminal
went away); detail file records the observation, repro/diagnosis approach
(reusing the G-036 dump tooling pipeline) and the G-038 coordination point
(dead console = failure branch of the console-EOF restart decision).
- G-036 acceptance reworked per user: standalone minimal repro waived as
moot (root cause proven and fixed upstream); remaining actionable is a
has_bug-style check reporting tcludp < 1.0.13 on Tcl 9 Windows through the
existing check/help machinery (simple version test, no behavioural probe).
Index and detail-file acceptance kept in sync.
- AGENTS.md User Preferences: cmd.exe truncates a heavily populated PATH -
tools resolvable in PowerShell may be not-found under cmd /c; use absolute
exe paths there and suspect truncation before absence.
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
Motivating transcript: piped `error xxxx` then interactive `set errorInfo`
shows tcllib virtchannel boot noise (caught `unset timer`) instead - the
restart boundary discards run-1's code interp, and bare ::errorInfo is
clobbered by internal caught errors anyway. Detail file records the two
mechanisms and design points (capture at the repl error-reporting site,
tsv/exitinfo handover, punk-owned record + one-line first-prompt notice,
quiet on clean pipes).
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
Live CloseHandle breakpoint (WinDbgX conditional bp on the sockListLock
handle value) caught the closer: Tcl thread-exit-handler walk in a dying
Thread-extension worker, tail-calling CloseHandle. The tcludp fossil diff
1.0.12 -> current names it: 1.0.12 Windows UDP_ExitProc ("called at thread
exit") closes the process-global waitForSock and sockListLock events, so the
first udp-loaded thread to die destroys the lock for every other udp-loaded
thread. Recycled handle values turn subsequent INFINITE waits into permanent
blocks (unrecycled values fail fast - why standalone probes stayed green
while punkshell run-2 startup churn wedged 4/4). Upstream already fixed in
1.0.13/trunk (UdpThreadExitProc per-thread + ExitSockets at process exit).
Remedy: upgrade bundled tcludp to 1.0.13+; loose ends recorded (punksys 8.6
udp version check, residual trunk weaknesses as optional upstream tickets).
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
Instrumented traffic-enabled specimen: wedged workers blocked (not spinning)
in WaitForSingleObject(sockListLock, INFINITE) inside tcludp event-source
procs (disassembly maps both wait sites to the sockListLock static; the
SocketThread released properly and parked on waitForSock). Handle-table data
from the full dump proves both tcludp global event objects were CloseHandle`d
mid-session: the statics still hold handle values now recycled to an
IoCompletion object and a manual-reset event (tcludp only creates auto-reset
events). Waiters starve forever on unsignalable objects. Only ExitSockets
closes these handles; the sole udp-loaded thread that died mid-session (run-1
shellfilter-run worker, terminated in the run transition) is the prime
trigger suspect. Console dependence dissolved - the CONIN$ restart was simply
the only mid-session udp-thread-death scenario. Minimal-repro die-variants
(release / terminate-mimic / full punk env) do not yet reproduce; next
candidates recorded (resultvar-terminate, thread::cancel, live CloseHandle
breakpoint).
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
Dump pipeline proven (procdump + WinDbgX scripted -z/-c/-logo). Uninstrumented
specimen stacks: workers indistinguishable wedged-vs-idle without heartbeat
timers (all in notifier MsgWait) - round 2 needs the instrumented build to map
wedged tids to dump threads. Structural finds: tcludp runs a global helper
thread (observed parked in WaitForSingleObjectEx inside tcl9udp1012.dll);
udp_tcl.c uses an auto-reset event as a pseudo-lock taken INFINITE in every
udp-owning thread's event-loop setup/check procs; checkProc recvfroms
non-blocking sockets per pass but does not decrement packetNum on the error
path (WSAECONNRESET from ICMP port-unreachable), while setupProc forces
zero block-time poll mode whenever packetNum>0 - a stuck packetNum locks the
owning thread's event loop into permanent poll mode. Blocking-recvfrom theory
ruled out (FIONBIO at open). Round-2 plan recorded (wedged-tid stack, MsgWait
timeout argument, per-thread CPU to discriminate spin vs block).
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
TEMP_REFERENCE checkouts confirmed current (Tcl core 2026-07-07, tcludp
2026-07-05). core-9-0-branch carries no tclWinConsole.c fix since 9.0.2, so
the wedge - if in the console driver - is likely still present upstream.
Current driver architecture recorded (process-global gConsoleLock, console
handle list shared across all Tcl threads, per-handle SRWLOCKs, reader/writer
threads with CONDITION_VARIABLE wakeups + Tcl_ThreadAlert fan-out) as the
credible wedge surface. Worker-side hypothesis sharpened: std-channel init on
a worker's first channel creation (udp_open / log-file open) attaches it to
the shared console machinery when the process stdin handle is the console -
fits which workers wedge vs survive, though probes have not yet reproduced.
Probe 'adopt' variant added (main thread closes stdin, CONIN$ adopted into
the stdin slot, SetStdHandle, active reader) - clean, ruling out the last
easy stdin-arrangement candidate. Approach reordered: native thread stacks
of a wedged in-context specimen promoted to step 1 (no debugger installed
yet - tool options noted).
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
Index entry recorded active per user. The detail file carries the full
investigation record: diagnosis chain of the piped-stdin exit/quit hang,
evidence and scoring caveats, ruled-out repro ingredients, remaining
candidates, and inline tooling listings (standalone probe script and the
hidden-console batch scorer with its 2026-07-08 baseline: punk902z run-2
syslog workers wedged 4/4, punksys 8.6 alive 4/4). Upstream reporting is the
user's manual step, gated on re-verifying against a Tcl 9 built from current
core sources.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
New suite src/tests/modules/punk/libunknown/testsuites/shadowing/
(7 tests, all pass) pins the .tm same-version shadowing rules verified
experimentally 2026-07-06/07: tcl:™️:add PREPENDS each argument (last
arg ends at the head of tcl:™️:list); the head of tcl:™️:list wins
exact-version ties (single and separate add calls); a higher version
wins from any position (order only arbitrates exact ties); and
punk::libunknown preserves all of it (scanner parity). Shipped behaviour
depends on these rules - the runtests testinterp tm ordering (bootsupport
over vendormodules) and punk_main''s package-mode block precedence that
G-033''s proj:/kit-first design builds on - so a Tcl or libunknown change
that shifts the tie-break now goes red here instead of silently
mis-resolving.
Goal G-035 (proposed): characterise mixed .tm / pkgIndex.tcl provision of
the same package - same or differing versions, under standard package
unknown, punk::libunknown and punk::packagepreference, across
scan-trigger orderings, forget/re-require crossing forms, and 8.6/9.
Converts the standing informal rule ("avoid mixing provision forms for
one package - unexpected behaviour even with libunknown''s improvements")
from folklore into either substantiated AGENTS.md guidance with named
failure modes, or retirement if mixing proves predictable. Discovery
work, so goal-specified rather than written now; extends the shadowing
suite when implemented.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
Design review outcomes (2026-07-07):
- Syntax: proj is not a fifth ordered token (category error - the dash
list is an ordered sequence of path blocks whose order IS the
same-version tie-break dial, per the verified tm rules and punk_main''s
own comment). Decided form: [proj:]<block>[-<block>...] - the prefix
scopes WHICH project root dev/src resolve against (cwd walk-up to the
nearest VCS repo root vs exe-relative), staying outside the ordered
list whose semantics are untouched. Explicit prefix = no silent rebind
on cd. Rejected: auto-detection, an explicit flag (less convenient),
token-attached qualifiers.
- Shadowing: visitor mode makes cross-lineage same-version-different-code
ties structurally expected (binary and visited project snapshots are
different vintages by definition). Order only arbitrates exact ties -
dev-versioned, higher-versioned and kit-absent modules come from the
project in either order. Canonical documented invocation is
proj:internal-src (kit-first: the visiting shell''s infrastructure
survives ties; for a standalone downloaded binary the kit is its only
trustworthy lineage and internal is location-independent); proj:src
documented as the faithful-vintage variant. Trimmed kits (G-019)
compose gracefully - kit-first only arbitrates ties.
- Required boot notice: detected root + effective precedence.
- packagemode help text drafted and persisted in the detail file as the
designated G-032 source (live only when the prefix ships); G-032 detail
cross-references it.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
G-033: an explicit `proj` package-mode token lets an installed (visitor)
punkshell binary detect the punk project containing the cwd - walking up
to the nearest VCS repo root, the marker punk::repo::is_project_root
already uses - and set up source/dev library resolution against THAT
project, so `punkshell proj shell` explores a project interactively even
when the project builds no shell-capable binary (or no binary at all).
History check recorded: cwd project-structure detection was never active
in the boot modes (src/dev are exe-relative; os is shallow cwd-only), but
the runtime detector find_project already exists and dev commands use it -
so the marker question (VCS repo root, not punkproject.toml which tomlish
lacks) is already answered; this aligns boot modes with it. Explicit token
gates it per user decision (no silent rebind on cd); boot chicken-and-egg
(find_project is in punk::repo, not loadable at the package-mode phase)
handled by a lean inline VCS-root walk-up or post-boot deferral. Reuses
the existing src-mode path machinery, parameterised on the detected root.
Not gated on the module.templates/punk::cap plugin machinery.
G-034: zip-based #modpod modules (e.g. punk::mix::templates) fail to mount
in the shell subcommand's code interp on Tcl 8.6 - `dev module.templates`
errors with `invalid command name vfs::RegisterMount` / no template
providers - while the same call works via the script subcommand /
main interp (reproduced 2026-07-07). The code interp lacks the vfs::zip
library that supplies zip mounting for pre-zipfs Tcl. Tcl 9 (zipfs)
unaffected.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
Root cause: Tk registers a main loop (Tcl_SetMainLoop); tclsh''s Tcl_Main
services it after the script, so `tclsh gui.tcl` behaves like wish. The
script subcommand sourced-then-exited, so a Tk script''s after-callbacks
never ran (scriptlib/tktimer.tcl flashed and died; shell was the
workaround). Confirmed: native tclsh tktimer 1 blocks ~1.35s; script did
0.38s.
Fix: after the script body (success paths only), if a Tk main loop is
registered (info exists ::tk_version && winfo exists .), tkwait window .
services the event loop until the main window closes, then exit - the
script-level equivalent of Tcl_Main. A script that exits from a callback
(tktimer countdown) terminates directly. Console scripts have no main
loop and exit at once; a script error exits at once (no hanging window).
Verified both generations: tktimer via script now blocks the countdown
(punk902z 1.49s, punksys 1.57s - Tk present in both) exit 0; tkhello_exit
immediate; tk-script-that-errors exit 1 in 0.37s (no hang); tkhello (no
self-close) blocks like wish; console script unaffected. Enables G-020
GUI automation via `script` instead of `shell` (noted in G-020 detail).
Project version 0.4.1.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
New lean app package src/lib/app-punkscript behind the (previously stub)
script subcommand in punk_main.tcl:
- `<commands> | <punkexe> script` runs piped stdin to EOF - no trailing
exit needed; `<punkexe> script <file> ?args?` runs a file with
conventional ::argv0/::argv (bare non-subcommand first arg still
reclassifies to the file form). Script may itself read stdin in the
file form.
- Default punk shell environment via shared definitions (package punk
registers deck aliases e.g dev; punk::aliascore::init adds the utility
aliases) - no local alias lists, no boilerplate in piped one-liners.
- Honest exit codes: 0 success, 1 on error with errorInfo on stderr,
script''s own exit honoured. No shellfilter stacks/transforms, no
logging side effects, no interactive fallback (terminal stdin with no
scriptname is a usage error, never a blocking read or shell).
- Stdin form echoes the script''s final result when non-empty (one-shot
eval ergonomics: return-value commands like `dev projects.work <glob>`
emit their table without a puts wrapper); file form keeps pure script
semantics.
Verified on both generations (punk902z/tcl9 from _build - bin deploy was
file-locked; punksys/tcl8.6 from bin): piped success/error/exit-code,
file form with args and stdin passthrough, missing-file error, motivating
one-liner `dev projects.work *tomlish*` emitting the project table with
exit 0, `src` mode combination (dev modules load), and `shell` subcommand
regression smoke. G-015 flipped to achieved 2026-07-07 per GOALS.md
rules; goal detail records the verification and the manual console
spot-check item. Project version 0.3.0 (new user-visible shell
capability - minor bump).
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
Not a mode of app-punkshell. The acceptance is defined by absence (no
shellfilter stacks, no logging side effects, no interactive fallback),
which a package that never loads that machinery guarantees by
construction; app-punkshell''s load-time behaviour (shellfilter stacks
with syslog/file targets writing punkshell_*.log into the caller''s cwd,
logging threads) is wrong for one-shot calls; isolation caps the blast
radius of shell-path machinery. Duplication mitigation: the default punk
shell module/alias environment definition is factored into a shared
module-level definition (punk::aliascore as seed) consumed by both app
packages - share definitions via modules, not launcher control flow.
Package shape previews the G-031 registration model, as app-tomlish
demonstrated.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
G-031: split the monolithic vfs main (punk_main.tcl: 1190-line apply
block - ~1060 lines boot machinery, ~100 lines dispatch tail) into a
layout-owned boot core (shipped via _vfscommon.vfs, sourced by
vfs-relative path since no package paths exist at boot; pull-updatable
per the G-027 ownership model) and a thin project-owned main declaring
app subcommands at commented customization points. Fork cost measured:
tomlish_main.tcl carries ~20 custom lines on a stale 500-line 2025
vintage copy; project_main.tcl (901 lines) same at template vintage -
boot improvements never reach derived projects. Acceptance: behaviour
parity across dispatch cases and vfs types, project subcommand added by
editing only the thin main (tomlish the proving ground), boot core
versioned/reportable (G-025).
G-032: the default dispatch defines subcommands via punk::args - tabled
-help enumerating built-in and project-registered subcommands, parsed
options with tabled usage errors (script/G-015 or version-report/G-025
as first candidates), G-030 degradation rules (guarded require; parsing
falls back to switch semantics, help to a plain list; rendering degrades
independently). Parked decision: the silent unknown-arg->script
reclassification should stop swallowing typos (candidate rule:
reclassify only when the argument names an existing file).
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
G-030 (proposed): make.tcl - the first surface a developer sees - parses
subcommands/options via punk::args and showcases the tabled usage for
help and argument errors; every interactive y/n prompt gains a declared
flag equivalent (interactive stdin prompts as today; flag proceeds;
non-interactive stdin without the flag fails fast with usage instead of
hanging or half-aborting) so agents drive make.tcl with arguments rather
than piped y/n; punk::args joins the bootstrap-tracked staleness set.
Degradation is a hard requirement: boot stays self-contained and the
repair commands (check/bootsupport/modules) keep working with plain
fallback help when the bootsupport punk::args or the ANSI rendering
stack is stale/unavailable - this week''s stale-fileutil and
stale-textblock incidents are the cautionary tales.
G-015 detail: recorded the runx -tcl result-shape observation (dict keyed
{result stdout stderr}, no exitcode key unlike exec mode).
G-020 detail: stamped the 2026-07-07 decision - acceptance gates on a
full testable Windows system without requiring other platforms first.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
Root AGENTS.md "Git Commit Conventions" renamed "Commit Conventions (any
VCS)": the Co-Authored-By/AI-attribution ban and the Assisted-by trailer
apply to git commits and fossil checkins alike (git parses trailers
structurally, fossil comments are free text - either way the trailer is
the last line of the message). Generated projects start fossil-only but
may adopt git; punkshell itself is git-primary today and could flip.
CLAUDE.md mirror section updated to match.
G-027 detail: documentation/goals ownership design input - layout-owned
child AGENTS.md docs ride the pull channel under the
unmodified-since-install classification; project-owned skeletons (thin
root AGENTS.md, GOALS.md, goals/AGENTS.md, empty GOALS-archive.md) are
seeded once and never pulled; commit conventions are part of the seed
(developer-changeable, VCS-neutral); layout-owned marker convention;
transitive chains inherit the parent''s modified docs as their baseline.
G-012 detail: AGENTS.md template payloads recorded as the same hazard
class as live .gitignore payloads - a template AGENTS.md under
src/project_layouts/ is binding DOX for agents walking the punkshell tree
though written for generated projects; needs the same inert storage +
materialize-at-generation treatment.
First field test: the tomlish project was seeded today with the thin root
AGENTS.md + CLAUDE.md + GOALS framework this design describes.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
src/tests becomes the single source of truth for module test suites: a
punkcheck-tracked make.tcl step generates the packaged test::<modulename>
#modpod modules from src/tests/modules/<namespacepath>/testsuites,
ending hand-maintained parallel copies under src/modules/test/.
The packaged form is a distributable in its own right: a user who
downloads a built module can optionally download the matching
test::<modulename> and verify the module''s behaviour on their own
system - package require + RUN interface, or a wrapped executable''s
-app test - with no source tree or harness. Acceptance proves the
standalone consumer scenario (runs without the project source tree),
suite-result parity with src/tests/runtests.tcl, and byte-for-byte
data fidelity (crlf roundtrip fixtures).
Motivated by the tomlish dual-copy interim (its modpod testsuites were
ported to src/tests while the modpod remains the live test::tomlish
build source - tomlish checkins 20bd00dc, feedb6f0).
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
Motivated by a make.tcl project deploy failure ("could not delete target
binary") where the locker turned out to be 7-Zip inspecting the exe -
identifying it took several dead ends because Windows has no cheap
file-to-process reverse lookup (name matching and Process.Modules scans
both miss plain file handles; openfiles needs a pre-enabled global flag;
handle.exe is external and elevation-bound).
G-028: a punkboot::utils helper returns the locking processes via the
Windows Restart Manager API (the installers-grade answer to "who is
blocking this file"), with twapi or cffi required lazily at call time -
punkboot::utils stays pure Tcl per the bootsupport no-compiled-extensions
contract, and unavailability degrades to the existing message. make.tcl
kit deploy failures append the locker report when determinable. Recurring
customer: G-023 versioned/dev binaries rewritten every build.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
Projects generated by dev project.new receive a layout snapshot that
drifts; the current remedy is a push from punkshell (project.new -force 1
-update 1). G-027: one command run inside the derived project pulls
make.tcl/build.tcl, bootsupport modules/libs and template payloads from
the originating punkshell project - .punkcheck records remain the
provenance basis, origin identity becomes durable to workdir moves (not
relative paths alone; projects.work discovery as re-locator), VCS
awareness on both ends (G-026 clean-checkout policy on the source,
uncommitted-local-modification detection on the target), and the
remote-pull question gets a recorded decision.
Detail file records the motivating incident (tomlish 2026-07-06: a stale
make.tcl whose bootsupport updater was disabled by the very drift it
needed to fix - broken bundled fileutil::traverse under Tcl 9) and the
constraint that the pull entrypoint must not depend on the freshness of
what it updates.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
make.tcl vendorupdate now walks up from each vendor source path to the
nearest fossil and/or git root and warns (non-fatal, once per VCS root)
when the checkout is dirty - vendored artifacts built from uncommitted
source have no committed provenance. Silent for unversioned or missing
locations; dual git+fossil roots report each VCS separately. First real
run flagged 8 dirty source projects (the tomlish 1.1.7/1.1.8 incident -
vendored content whose source history ended at 1.1.6 - motivated this,
remedied upstream by tomlish checkin 37b71e82).
Goal G-026 (proposed) carries the enforcement follow-on: abort with
explicit override, warn-only as configured mode, shared coverage of
bootsupport_localupdate, and a recorded decision on the residual
staleness gap. Abort-by-default now would block routine work given the
dirty-project count.
Project version 0.2.5; vendorupdate doc note in src/AGENTS.md.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
G-023: builds produce punk8|9-<major>-<minor>-<patch>.exe from the
punkproject.toml version plus punk8|9-dev.exe tracking the latest build;
plain punk8|9.exe created initially and replaced only by an explicit
release step. Disk growth tolerated for now; archival/deletion trigger
recorded as a later question.
G-024: src/runtime/mapvfs.config converted to toml parsed via tomlish -
explicit runtime/vfs/name/kittype mappings preserved, plus generative
scheme declarations so the G-023 versioned outputs are one succinct entry
rather than per-version lines. Entry-level validation errors; full
migration preferred over a deprecation window.
G-025: build stamps project version, runtime binary and vfs folder into
the kit; a version-report subcommand prints them machine-parseably, with
the same-named punk-module command as the single implementation (the
subcommand is a thin wrapper) so scripts get the report in-process. Report
keeps build stamp, live runtime facts and source-tree version distinct:
src-mode sessions report stamp and live versions separately when they
differ; unstamped contexts (tclsh src/make.tcl shell, plain tclsh with
punk modules) report stamp fields absent rather than fabricated.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
Goal G-022: scriptable safe fossil move/rename in dev repo (all checkouts
repointed, no phantom central config-db entries, no dangling old repo db,
project-name renamable with project-code preserved), then rename this
project fossil repo shellspy -> punkshell through that mechanism via a
G-015 piped script call.
New tests src/tests/modules/punk/mix/testsuites/repo/fossilmove.test pin
current behaviour in a FOSSIL_HOME-isolated sandbox with an isolation
guard: the file-copy + fossil test-move-repository sequence repoints the
checkout but leaves the old db file on disk, a phantom repo: entry in the
central config-db, and a stale ckout: back-reference; file moves never
touch project-name. GAP-marked tests flip when G-022 lands.
shellrun 0.1.4: add missing package require punk::args - the 0.1.3
migration left run/runout/runerr/runx calling punk::args::parse without
requiring it, which broke shellrun in bare interps and made every
src/tests/runtests.tcl run fail at test-file invocation. Full suite now
back to baseline (523 tests: 508 pass, 14 skip, exec-14.3 known failure).
Project version 0.2.4 (changelog entry for the shellrun fix).
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
G-018: zig-built plain zip-based tclsh kits with no punk infrastructure
(Tcl 9 kit satisfies acceptance; 8.6 interim = plain non-kit builds, kits
when practical, non-gating). Notes record that wish executables are never
a build target - Tk ships as a loadable package.
G-019: dependency-scan-driven module trimming for punk executables via the
tclparser parse API (c-only library today; punk::lib pure-Tcl fallback is a
stub), with a declared-extras manifest for scan-invisible dynamic requires;
batteries-included builds preserved.
G-020: screen capture + input injection module with per-platform backends -
Windows first (treectrl loupe + twapi, all already shipped in punk9win kit);
X11 (Linux/FreeBSD) and macOS follow behind the same contract; Wayland-native
out of scope. Platform show-stopper assessment recorded in the detail file.
G-021: agent-drivable visual/UI verification - window match, snapshot to
PNG/base64, scripted interactions - through G-015 piped script calls, with
guidance to land alongside G-017. aloupe.tcl stays a standalone app.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
G-015: complete the punk executable `script` subcommand - run a script file
or piped stdin (scriptname optional) in an interp preloaded with the default
punk shell modules/aliases, honest exit codes, no shellfilter transforms, no
interactive-shell fallthrough. Replaces the fragile `...;exit | punkexe`
repl-pipe pattern (currently a stub in punk_main.tcl that only emits stderr
diagnostics). Lean bare-script execution deferred to a later `tclsh`
subcommand expansion.
G-016: extend `dev projects.work` to discover git-based projects alongside
fossil (git has no central registry - enumeration-source candidates recorded
in the detail file).
G-017 (index-only): once G-015 lands, direct agents via repo guidance to
locate sibling projects with a piped projects.work call instead of scanning
the filesystem.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
Scope gains the vendored tomlish module (canonical source in the external
tomlish project space). Goal line rules out ad-hoc toml parsing; acceptance
now requires the tomlish API procs punk::config consumes to carry punk::args
(PUNKARGS) documentation, added upstream and re-vendored before punk::config
implementation proceeds. Detail file records the 1.1.8 coverage gap: argdocs
exist only on the cmdline app layer, not the library API (from_toml, to_dict,
from_dict, update_tomlish_from_dict).
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
G-013: raw input mode as the repl default - per-keystroke debug
emissions (add_chunk frame, rhs editbuf view) individually gated and
repl-toggleable, raw editor parity for history/cursor movement,
debugrepl marked-line form retained for simple terminals with its
first-word activation mechanism under review; retires the tcl 8.6
idle-prompt query residue guarded in punk::console 0.7.1.
G-014: complete ::punk::config - toml files in the XDG config dir with
per-named-subshell override/inherit resolution; first consumer is the
G-013 editbuf-view startup default. Write-back persistence deferred.
Both proposed; detail files under goals/.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
Project layouts carry .gitignore files as payload for generated projects,
but git treats any nested .gitignore as live configuration scoped to its
directory - the templates self-censor, silently untracking their own
content in the outer repo (discovered via the git/fossil tracked-set
verification; stopgap force-adds in fc1c474c). G-012 [proposed] requires
payloads stored inert (neutral rename or fauxlink-encoded - mechanism an
implementation decision recorded in the detail file) with materialization
at project generation, plus a punkcheck-tracked make.tcl template-refresh
step generalising the existing sync_layouts (which covers only
build.tcl/make.tcl into custom/_project) to the vendor/punk layout base
and derived payloads. Acceptance covers a provably-inert nested-rule scan,
generated-project content equivalence, a working refresh after a
canonical-source edit, and retiring the per-file force-add exceptions.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
Tcl 8.6 windows console channels expose no -inputmode configure key, so
punk::console guards testing -inputmode/-mode classified a real 8.6 console
as a pipe. A twapi console handle for stdin is now treated as definitive
console detection at all three affected sites (stdin only - pipe probes
still never touch the process console; mintty-as-pipes has no console
handle and falls through to the env heuristics as before):
1. get_ansi_response_payload's raw-cycling gate (introduced 0.1.5) skipped
the raw cycle for 8.6 line-mode queries.
2. is_input_console_or_tty returned a false negative, making
settle_can_respond's layer-2 heuristic settle a real 8.6 console as
unable to respond.
3. input_at_eof (introduced 0.1.2) took the pipe branch and performed its
probe read on the drained console channel. This was the line-mode query
killer: verified empirically against clean tclkits 8.6.13 and 8.6.17,
a read on a drained 8.6 console channel makes the channel driver park a
blocking cooked-mode ReadConsole; driver reads sample the console mode
when issued, not when data arrives, so the query's subsequent raw flip
cannot rescue it and the terminal response is swallowed until Enter.
Since get_ansi_response_payload calls input_at_eof immediately before
its raw cycle, every 8.6 line-mode query timed out (~500ms) with the
response later leaking to the line reader as phantom input - also the
source of increased ANSI artifacts (e.g. 'help env') once 0.7.0
brokering made code-interp queries actually reach the terminal.
Interactively verified on punksys (tcl 8.6.13, src launch): line-mode
get_cursor_pos and dec_get_mode succeed mid-command, brokered query loops
run clean, and typed-ahead lines are recovered without phantom input.
Known residue (documented in the G-007 detail notes): queries fired while
the repl reader is armed and idle at a line-mode prompt (including some
repl-init detection queries) still hit the parked-read limitation on 8.6 -
repl init ordering is a follow-up; raw mode is unaffected.
Tests: probes.test gains constraint-gated wiring tests for the twapi
console branch of is_input_console_or_tty and input_at_eof (skip in piped
runs, engage on a real console). Full suite passes at baseline (exec-14.3
only) under Tcl 9.0.3; probes suite also verified under Tcl 8.6.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
Flip G-007 (location-transparent punk::console) to achieved 2026-07-05 in the
GOALS.md index and detail file. Interactive verification on built punksys
(tcl 8.6) and punk902z (tcl 9) kits (punk::console 0.7.0, opunk::console
0.4.0; punk902z also rerun via src launch against dev modules): fact parity
and immediate cross-context visibility in both directions; brokered terminal
queries from the code interp succeed and cooperate with the repl reader in
raw mode with typed-ahead preserved; a chan-pipe console anchored from the
code interp registers its codethread as owner, operates locally, and clears
its registry entry on forget.
Detail-file Notes record the verification and characterize line-mode reader
cooperation (out of scope, no routing regression - brokered and owner-side
queries behave identically): on tcl 9, complete typed-ahead lines survive
with cosmetic garbling while incomplete typed-ahead starves the query raw
window and leaks responses as phantom input; on tcl 8.6 every line-mode
query fails that way, owner-side included (repro:
repl eval {punk::console::get_cursor_pos}) - candidate owner-side regression
to investigate separately. Also noted: the leaked stderr diagnostics are the
traffic G-011's per-console err channel would house.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
Acceptance now names the discovery fact (console_fact_get fact key err,
returning the effective err channel name) so the clause is testable like
G-007's named facts. Detail file additions: "every -console site" is
satisfied at the module's documented resolution choke points
(console_spec_resolve, hybrid_console_spec, opt_console_* helpers,
spec_to_channelpair, shared PUNKARGS fragments) rather than by per-site
audit; the two consumer-shaped acceptance clauses' division of labour is
stated (internal diagnostics migration = first consumer, external consumer
= public emit form); G-001 backends wanting a distinct diagnostic stream
must set err explicitly (default merges into out); and
previous_stty_state_stderr stays as-is (stty-state restoration, orthogonal
to diagnostic routing).
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
punk::console models two channels per console ({in out}); its own
diagnostics fall back to raw puts stderr and subshell/backend/theme goals
(G-001/G-008/G-009) inherit the dual-channel assumption. G-011 defines err
as an attribute of the canonical {in out} console identity: {in out err}
specs accepted everywhere -console is, a nullable err field on
opunk::Console, err defaulting to process stderr for the default console
and to the console's out channel elsewhere, console-directed diagnostics
and an emit-to-err path, and cross-thread discovery via console_fact_get.
Ownership/fact/mode-cache keys stay canonical {in out}, so the G-007 work
is untouched.
Detail file records the gap analysis, the rejected shapes (err-in-identity
triple key, scope-boundary-only, uniform err==out, unset-means-none) and
the G-001 sequencing interaction.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
Per-console facts are now visible to every thread of a punk session: the
default console {stdin stdout} keeps its legacy namespace variables as
authoritative local storage, with write traces mirroring every write
(including direct variable writes) into tsv punk_console_facts; non-default
consoles store facts only in tsv with owner-qualified keys (non-std channel
names are thread-local and would otherwise alias). console_fact_set from a
non-owner thread forwards to the owning thread (vt52-alias transport) and
always writes the tsv mirror itself (covers unreachable or older-version
owners). New console_fact_clear for tests/maintenance.
New console ownership registry (console_owner_register/get/forget, tsv
punk_console_owners keyed by canonical {in out} pair): ownership is captured
when an opunk::console instance is anchored - via the new pluggable
::opunk::console::lifecycle_callback (opunk::console 0.4.0, base class
unchanged) wired by ensure_object_integration - and by default_console.
Consult-time liveness validation clears entries for exited threads. For
{stdin stdout} first registration wins and only the owner's forget releases
the entry, so a thread anchoring a local view cannot steal ownership.
dec_has_mode/ansi_has_mode caches moved to tsv punk_console_modecache
(single-key atomic entries shared process-wide).
Infrastructure tsv arrays are now punk_-prefixed to avoid collisions with
application tsv usage in subshells: console -> punk_console (is_raw), with
call-site patches in punk::repl 0.2.1, punk::lib 0.2.1, punk::basictelnet
0.1.1 (rule recorded in src/modules/AGENTS.md; legacy repl/codethread_*/
zzzload_pkg* arrays left for a coordinated follow-up).
punk::console is now loadable in secondary threads without ::argv0
(powershell consolemode fallback guarded).
Tests: consolefacts gains ownership-registry and cross-thread visibility
coverage (worker-thread punk::console load, trace-mirror, forwarded set with
deadline-polled sync - see comments re runtests child-interp topology);
cleanups use console_fact_clear / tsv instead of poking removed internal
stores. Full suite passes (baseline exec-14.3 only).
G-007 flipped proposed -> active (user-confirmed); remaining acceptance work
is the choke-point brokering slice.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
Index entries in GOALS.md plus detail files, all proposed:
- G-007 location-transparent punk::console across repl and code
interps: punk-side tsv-backed ownership registry, fact/cache
visibility, choke-point brokering generalising the vt52 alias;
detail records rejected alternatives (audience namespace split,
owner field on ::opunk::Console, opunk-package-namespace registry)
- G-008 scoped console state for same-console subshells: activatable
per-subshell state sets (journal-on-write, re-emit on activation)
rather than a LIFO undo stack, so later tree switching works;
restore of tabstops/modes/palette on quit; opt-in, shared behaviour
stays default
- G-009 themed subshell profiles: poshinfo schemes bound to
restriction aspects (G-003 limits/hidden commands, punk::island) so
restricted subshells are visually distinct; terminal effects ride
the G-008 state set
- G-010 subshell suspend/resume and tree navigation: switch to any
live subshell (grandparent, cross-branch) via attachment handoff +
G-008 state activation; navigation commands governed by G-003
hide/expose; depends on G-002's non-nested model
Sequencing: G-007 -> G-008 -> G-009/G-010, with G-002 prerequisite
for G-010.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
Index entries in GOALS.md plus detail files. G-004 targets a binary-free
committed tree, achievable once G-005 (zig-based build of binary deps from
source) or G-006 (consent-gated download of pre-built artifacts) provides
the replacement supply of the previously committed binaries.
Assisted-by: harness=claude; primary-model=claude-fable-5; api-location=anthropic.com
G-001 pluggable console backends for non-detectable terminals (ssh channel, tk widget) via ::opunk::Console subclasses. G-002 non-nested subshell with console targeting and inter-subshell comms, including thread::send -async routing to the code sub-interp and the canonical launch API design (subshell as sole entry point, interp alias as context signal). G-003 configurable resource limits (command-count, time via interp limit) and sandboxing spectrum (interp hide/expose/alias, safe::interpCreate) on subshell interps, downstream of G-002's asymmetry fix.
Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent)
Assisted-by: harness=opencode; primary-model=unknown; api-location=unknown
Harness-agnostic technical project goal system. GOALS.md holds the canonical one-line index with format spec, status taxonomy, maintenance rules, and an agent-authoring workflow so agents asked to 'help me write a goal for X' ask the right questions. GOALS-archive.md is the achieved-goals archive stub. goals/AGENTS.md is the DOX child doc covering naming (G-<id>-<slug>.md), when a detail file is warranted, and the no-orphans/archive rule. goals/archive/.gitkeep tracks the archive dir.
Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent)
Assisted-by: harness=opencode; primary-model=unknown; api-location=unknown
