1.11.1 — live progress for backgrounded runs

Fixed

• A running Sol was no longer visible in Claude Code. 1.11.0 moved the
  launcher into the background and told Claude there was nothing to poll, so
  the Monitor it used to attach disappeared and a run became one silent shell
  command. SKILL.md now has Claude start a Monitor on the watcher for every
  launch, --resume and --research.

Added

• sol-watch.py --run <run-dir> follows a whole launcher run: every
  worker's event log, through a stall relaunch or a --resume replacing the
  log, until the launcher has written summary.json. It closes with one run:
  line per worker and exits 0 only if all ended ok or no-changes. A batch
  whose launcher died unclassified, or never started, is reported as such.

Changed

• The watcher is quieter. A search that finds nothing (rg/grep exit 1) is
  no longer a failed line, a heredoc body that mentions a test runner is no
  longer a verification run, and a JSON final report is not printed as the plan.

1.11.0 — supervised corrections and research, structured reports

Every run now goes through the supervised launcher, and the brief and the report
are contracts in both directions.

Added

• Structured worker report. The launcher passes references/report-schema.json
  as --output-schema, and summary.json embeds the parsed result under report:
  done / partial / blocked, one entry per numbered acceptance criterion with
  evidence, commands run with exit codes, deviations, blocked checks, open
  questions. SOL_REPORT_SCHEMA= (empty) returns to a free-text report.
• diff.patch per worker (diff_path in summary.json): everything the
  worker changed against the pre-run commit, including files it created. Review
  reads this instead of git diff, which omits untracked files.
• --research: a read-only run with live web search under the same watchdog
  and stall retry as an implementation run. Effort is SOL_RESEARCH_EFFORT
  (xhigh), independent of SOL_EFFORT. Needs no clean tree.
• Failure detail in summary.json: error (last error event), stderr_tail,
  token usage, plus top-level model and mode.
• Brief template blocks: <environment> (sandbox, network, known-blocked
  checks), <when_blocked> (stop and report rather than guess), numbered
  criteria (AC1…), and a slot in <action_safety> for the user's standing
  rules, which Codex cannot read from CLAUDE.md.

Changed

• --wait blocks and supervises. It used to poll once and return 75. It now
  takes over the watchdog and the stall retry when the launching call is gone,
  and returns 75 only after SOL_WAIT_TIMEOUT (540s). The retry budget is
  counted per worker, so a takeover neither skips nor repeats a retry.
• Model resolution. With SOL_MODEL unset the launcher follows the Codex
  config's default model when that is a Sol model, and falls back to
  gpt-6.1-sol otherwise. The model is recorded per run. check-codex.sh
  resolves it the same way.
• SKILL.md routes corrections and research through the launcher (--resume,
  --research) and tells Claude to launch in the background. The hand-rolled
  codex exec resume --last is gone, along with file paths that predated the
  launcher.

Fixed

• A stall relaunch dropped the brief's reference images. The relaunch was a
  second copy of the worker command without the -i arguments; both now share
  one command.
• An in-place worker that committed its own work was reported no-changes.
  A clean tree was the whole test. It is now ok with the commit recorded, and
  the default brief tells Sol not to commit.
• A stalled --resume orphaned its codex process. Job control was only
  enabled on the launch path, so the watchdog's group kill reached the wrapper
  alone.

1.10.0 — GPT-6.1 Sol, high by default

Changed

• The implementer is now GPT-6.1 Sol (gpt-6.1-sol). The default model in
  sol-parallel.sh (SOL_MODEL), check-codex.sh, and every command in
  SKILL.md and the README moved from gpt-6-sol. The README's worked example
  still shows its original gpt-5.6-sol transcript.
• SOL_EFFORT defaults to high again, reverting the xhigh default from
  1.9.0. A stalled worker now steps down to medium. Set SOL_EFFORT=xhigh
  for algorithmically hard briefs; research mode still runs at xhigh.

1.9.0 — GPT-6 Sol, xhigh by default

Changed

• The implementer is now GPT-6 Sol (gpt-6-sol). The default model in
  sol-parallel.sh (SOL_MODEL), check-codex.sh, and every command in
  SKILL.md and the README moved from gpt-5.6-sol. The README's worked
  example keeps its original gpt-5.6-sol transcript, since that is the run
  that was actually recorded.
• SOL_EFFORT defaults to xhigh again. The stall ladder and the
  300-second first-event budget from 1.8.1 stay, so a stalled xhigh launch
  costs at most that budget and then relaunches at high. Set
  SOL_EFFORT=high for mechanical briefs where speed matters more.

1.8.1 — high effort by default, in-place recovery fixed

[1.8.1] — 2026-09-02

Changed

• SOL_EFFORT now defaults to high, not xhigh. Stalls are
  effort-correlated (openai/codex#24260, #23807), and the cost is asymmetric: a
  stalled worker burns the entire first-event budget before it does anything at
  all. Observed on the same brief — an xhigh worker sat 900s without a single
  tool call, while high made its first call in 36s. With a compiler and a test
  suite in the loop, high verifies its own work, which is the property the
  effort was buying. xhigh stays available as an explicit opt-in for
  algorithmically hard briefs, and research mode still writes it out, because
  nothing there compiles. The stall ladder is unchanged, so a high launch now
  degrades to medium on a stall.
• SOL_FIRST_EVENT_TIMEOUT now defaults to 300 seconds, not 900. A real
  think before the first file read rarely runs past five minutes; 900 turned a
  single stall into fifteen lost minutes before the retry could even start.

Fixed

• An --in-place run could not be relaunched, resumed or cleaned up. Only
  the launch knew about --in-place; every path after it recomputed the
  workspace as $WORKTREE_ROOT/<slug>, a directory an in-place run never
  creates. The stall relaunch therefore ran codex exec -C <nonexistent>, died
  in about two seconds with a bare "No such file or directory", and recorded
  failed-launch, exit 1, elapsed 2 — a stall recovery that reads as a broken
  invocation, on precisely the runs the watchdog exists to save. Seen in
  production: an xhigh worker stalled at 900s and its high retry never ran.
  --resume failed the same way (the wrapper's cd fell straight through to
  its exit 2), and --cleanup pointed at a worktree that never existed.

  In-place mode is now remembered in the run directory rather than inferred from
  the command line. Each worker's workspace is written to
  <run-dir>/workers/<slug>/worktree at launch, next to brief, and read back
  instead of recomputed; <run-dir>/in-place marks the run itself, so --wait,
  --resume and --cleanup re-attach correctly without the caller retyping
  --in-place — which nothing in the output ever prompted for.

• The stall relaunch and --resume now check the workspace exists before
  launching and say workspace <path> missing with the slug, rather than
  leaving codex to fail with an opaque OS error two seconds later.

• --cleanup on an in-place run prints one line saying there are no worktrees
  or branches to remove and exits 0 without touching anything. The workspace
  there is the user's own checkout, which may itself be a linked worktree, so a
  path that reached git worktree remove would have deleted it.

Full changelog: https://github.com/ozankasikci/sol-skill/blob/main/CHANGELOG.md

1.8.0 — SOL_SANDBOX for toolchains the sandbox cannot reach

[1.8.0] — 2026-08-23

Added

• SOL_SANDBOX (default workspace-write) sets the -s policy for every codex
  invocation. It exists because SOL_CODEX_CONFIG structurally cannot reach the
  sandbox mode: an explicit -s flag beats -c sandbox_mode=, so setting it
  through the config channel is silently ignored — verified with both on one
  command, which still failed. Any non-default value warns on stderr, since
  danger-full-access removes all confinement rather than one restriction.
• SKILL.md now covers Docker specifically. Its daemon socket is a unix socket
  connect, which neither network_access nor writable_roots unblocks —
  verified, both leave docker ps at connect: operation not permitted.
  Verified end to end through the launcher: the same brief gets permission
  denied under the default sandbox and lists real containers under
  SOL_SANDBOX=danger-full-access.

Full changelog: https://github.com/ozankasikci/sol-skill/blob/main/CHANGELOG.md

1.7.0 — sandbox overrides for blocked toolchains

[1.7.0] — 2026-08-23

Added

• SOL_CODEX_CONFIG: extra -c key=value overrides, space separated, applied
  to all three codex invocations the launcher makes (launch, stall relaunch,
  and resume). Unset by default — relaxing the sandbox is the caller's call per
  repo, never the script's. Verified end to end against real codex: the same
  probe brief reports HTTP 000 without it and 200 with
  sandbox_workspace_write.network_access=true.
• SKILL.md Sandboxed toolchains: how to recognise a repo whose build cannot
  run under workspace-write at all, and why it matters more than it looks — a
  worker that cannot compile verifies by grep instead, reports green on grep
  evidence, and silently turns the role split into "Sol implements blind."
  Covers the worktree case specifically: a worktree's git dir lives outside the
  write root, so git commit fails deterministically, not intermittently.

Full changelog: https://github.com/ozankasikci/sol-skill/blob/main/CHANGELOG.md

1.6.0 — every run under the watchdog

[1.6.0] — 2026-08-19

Added

• Reference images for workers. A brief may carry a sidecar,
  <run-dir>/tasks/NN-<slug>.images, one path per line with # comments; each
  entry resolves against the repo root and is passed to that worker as
  codex exec -i. A path that does not resolve fails the run at preflight,
  before any worktree exists. Verified end to end against real codex.
• --in-place: run a single brief in the repo itself — no worktree, no branch,
  no commit, changes left in the working tree exactly as a bare codex exec
  leaves them — but supervised by the inactivity watchdog, with stall retries,
  a summary.json status and a real exit code.

Changed

• Phase 2 now launches through the script in single-worker mode too. The
  skill used to end its stall guidance with "parallel mode does all of this
  automatically; single-worker mode is your job", which is the one thing this
  project has repeatedly established does not work: a flow depending on the
  model reliably doing something degrades to silence when it doesn't, and that
  is what the 1.2.0 revert was about.

  It cost hours in practice. A visual brief could not use the launcher, because
  workers could not take reference images, so the run went out on the
  unsupervised path and hung at xhigh with thread.started, turn.started
  and 101 bytes — silent for nearly three hours. SOL_FIRST_EVENT_TIMEOUT would
  have killed and retried it in fifteen minutes; it was simply unreachable from
  where the work had to run. The images sidecar removed that reason, and
  --in-place removes the remaining one, so the unsupervised path is now an
  explicitly-documented fallback rather than the default.

Full changelog: https://github.com/ozankasikci/sol-skill/blob/main/CHANGELOG.md

1.5.0 — image input and asset output

[1.5.0] — 2026-08-19

Added

• Image input. The brief command can now take -i <file> (repeatable) to attach
  screenshots, mockups, or a failing chart, and the brief template shows how to
  reference the attachment from <task>. Verified against a real run: Sol reads
  the image and describes its actual content.
• Guidance for asset-producing briefs. Codex ships a built-in image_gen tool —
  a model-side tool, invisible in codex exec --help, which is why this went
  unnoticed — so a brief may ask for a raster asset and get AI-generated pixels.
  Sol routes to code on its own when the visual is code-native: asked for a solid
  red circle it rendered with ImageMagick and said so; asked for a painterly
  title screen it used image_gen. Acceptance criteria for an asset cannot be a
  test command, so the template shows the checkable form instead.

Changed

• Phase 3 now covers binary artifacts. git diff reports only Binary files differ, so an image is reviewed by opening it and judging its content against
  the brief — the review phase exists precisely so the model that produced the
  artifact does not certify it, and that argument does not stop applying at the
  file-format boundary.

Fixed

• Phase 2 said a silent in-flight command was backstopped by "the 30-minute
  absolute cap", which 1.4.1 turned off by default. SOL_COMMAND_TIMEOUT is the
  backstop now, and the text says so.

Full changelog: https://github.com/ozankasikci/sol-skill/blob/main/CHANGELOG.md

1.4.2 — fix the watchdog being dead on Linux

[1.4.2] — 2026-08-17

Fixed

• The inactivity watchdog never worked on Linux. mtime_of tried BSD
  stat -f %m first and fell back to GNU stat -c %Y, but GNU's -f means
  --file-system: it prints a filesystem dump and exits 0, so the fallback
  never fired. The dump then reached [ "$last" -gt 0 ] as a syntax error, and
  every stall check — first-event, idle, and the new command budget — was dead
  on every Linux host from 1.4.0 onward. macOS was unaffected, which is why the
  local suite passed while CI failed. mtime_of now tries GNU first and
  validates the result is numeric before returning it.

  This mattered most on 1.4.1, which turned the absolute cap off and left Linux
  workers with no watchdog at all. Linux users on 1.4.0 or 1.4.1 should upgrade.

• A unit test now asserts mtime_of returns an epoch. The behavioural stall
  tests only exercise the platform they run on, so a BSD-first ordering passed
  every macOS run while being broken everywhere else; this asserts the contract
  itself, so either platform catches a regression.

Includes 1.4.1: the absolute SOL_WORKER_TIMEOUT is off by default and silence is bounded instead.
Full changelog: https://github.com/ozankasikci/sol-skill/blob/main/CHANGELOG.md