From f2cebc7437788cdb4c7e77bdbb3e2e8ac8d148c9 Mon Sep 17 00:00:00 2001 From: garrytan Date: Mon, 2 Nov 2026 15:07:28 +0000 Subject: [PATCH] fix(capy): label config-sourced agent_model, note the assumption, skip it inside eval runs (GST-982 review) Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com> --- bin/gstack-capy-install | 14 ++++++++++---- changelog.d/gsta161-gst-982-capy-model.md | 4 ++-- docs/capy-install.md | 3 ++- docs/loadouts.md | 2 +- test/gstack-capy-install.test.ts | 15 +++++++++++++++ 5 files changed, 30 insertions(+), 8 deletions(-) diff --git a/bin/gstack-capy-install b/bin/gstack-capy-install index a9f7bf26e..6122d76e4 100755 --- a/bin/gstack-capy-install +++ b/bin/gstack-capy-install @@ -724,9 +724,13 @@ fi [ "$RC" -eq 0 ] || { [ "$PROFILE" -eq 1 ] && rm -rf "${VERDICT%.json}.profile" 2>/dev/null; verdict_kept || rm -f "$LAUNCH_FILE" 2>/dev/null; LAUNCH_FILE=""; } record_verdict "$TRAILER" "$([ "$RC" -eq 0 ] && echo 1 || echo 0)" LOADOUT_LINES="" -if [ "$RC" -eq 0 ] && [ -z "$MODEL" ] && [ -n "${GSTACK_PREFERENCES_FILE:-}" ] && [ "$PROFILE" -eq 0 ] && [ -n "$LAUNCH_FILE" ]; then - MODEL=$(cd "$ROOT" && bin/gstack-models resolve --role agent 2>/dev/null | sed -n '1s|^agent: [a-z0-9-]*/\([A-Za-z0-9._-]*\) .*|\1|p') - MODEL_SOURCE=preferences +if [ "$RC" -eq 0 ] && [ -z "$MODEL" ] && [ -n "${GSTACK_PREFERENCES_FILE:-}" ] && [ -z "${GSTACK_EVAL_DIR:-}${EVALS_RUN_ID:-}" ] && [ "$PROFILE" -eq 0 ] && [ -n "$LAUNCH_FILE" ]; then + _ag_first=$(cd "$ROOT" && bin/gstack-models resolve --role agent 2>/dev/null | head -1) + case "$_ag_first" in + *", this machine") MODEL_SOURCE=config ;; + *", the preferences file") MODEL_SOURCE=preferences ;; + esac + [ "$MODEL_SOURCE" = GSTACK_MODEL ] || MODEL=$(printf '%s' "$_ag_first" | sed -n 's|^agent: [a-z0-9-]*/\([A-Za-z0-9._-]*\) .*|\1|p') fi if [ "$RC" -eq 0 ] && [ -n "$MODEL" ] && [ "$PROFILE" -eq 0 ] && [ -n "$LAUNCH_FILE" ]; then _lo_full=$(. "$SELF/bin/gstack-claude-root.sh" && dirname "$(gstack_canonical_claude_root)") @@ -736,9 +740,11 @@ if [ "$RC" -eq 0 ] && [ -n "$MODEL" ] && [ "$PROFILE" -eq 0 ] && [ -n "$LAUNCH_F _lo_out=$(cd "$ROOT" && bin/gstack-modules install --model "$MODEL" --full-root "$_lo_full" 2>&1) fi _lo_line=$(printf '%s\n' "$_lo_out" | grep -m1 '^LOADOUT: ') - [ "$MODEL_SOURCE" = preferences ] && _lo_line=$(printf '%s' "$_lo_line" | sed 's/ source=model-flag/ source=preferences/') + case "$MODEL_SOURCE" in preferences|config) _lo_line=$(printf '%s' "$_lo_line" | sed "s/ source=model-flag/ source=$MODEL_SOURCE/") ;; esac _lo_root=$(printf '%s\n' "$_lo_out" | sed -n 's/^SKILL_ROOT: //p' | head -1) _lo_notes=$(printf '%s\n' "$_lo_out" | grep -E '^(LOADOUT_INVALID_MODEL|LOADOUT_FALLBACK): ') + case "$MODEL_SOURCE" in preferences|config) _lo_notes="${_lo_notes:+$_lo_notes +}LOADOUT_ASSUMED: model=$MODEL is the owner's agent_model ($MODEL_SOURCE), not read from the session; a session on another model sets GSTACK_MODEL=" ;; esac _lo_name=$(printf '%s' "$_lo_line" | sed -n 's/^LOADOUT: \([a-z0-9-]*\) .*/\1/p') _lo_digest=$(printf '%s' "$_lo_line" | sed -n 's/.* digest=\([0-9a-f]\{12\}\) .*/\1/p') if [ -n "$_lo_name" ] && [ "$_lo_name" != full ]; then diff --git a/changelog.d/gsta161-gst-982-capy-model.md b/changelog.d/gsta161-gst-982-capy-model.md index 478e298bc..7bbe0e299 100644 --- a/changelog.d/gsta161-gst-982-capy-model.md +++ b/changelog.d/gsta161-gst-982-capy-model.md @@ -1,7 +1,7 @@ **Capy sessions get their model's loadout again without a sentence in the skill (GST-982).** The browse wave (#3858) removed the Capy skill's instruction to run the installer with `GSTACK_MODEL`, to keep `capy/skill/gstack/SKILL.md` under its context budget. After that, every Capy session installed the full skills, including Opus 5.5 sessions that should get `lite`. Now the installer finds the model by itself. ### Fixed -- **`gstack-capy-install` reads the owner's model preference.** When it gets no `--model` and no `GSTACK_MODEL`, it uses the `agent_model` in the session's `GSTACK_PREFERENCES_FILE` (the Drive skill's `preferences.yaml`, which `gstack-capy-session` already exports), resolved by `gstack-models resolve --role agent`. So `opus` means `claude-opus-5-5` today. The line reads `LOADOUT: … source=preferences`. `GSTACK_MODEL` and `--model` still win, and a `host` preference or no file still serves full. The frozen bootstrapper `capy/skill/gstack/scripts/install.sh` and the skill text are unchanged. +- **`gstack-capy-install` reads the owner's model preference.** When it gets no `--model` and no `GSTACK_MODEL`, it uses the `agent_model` in the session's `GSTACK_PREFERENCES_FILE` (the Drive skill's `preferences.yaml`, which `gstack-capy-session` already exports), resolved by `gstack-models resolve --role agent`. So `opus` means `claude-opus-5-5` today. The line reads `LOADOUT: … source=preferences` (`source=config` when this machine's `config.yaml` sets it), followed by a `LOADOUT_ASSUMED:` note that the model came from the preference and not the session. Eval runs (`GSTACK_EVAL_DIR` or `EVALS_RUN_ID` set) skip the preference. `GSTACK_MODEL` and `--model` still win, and a `host` preference or no file still serves full. The frozen bootstrapper `capy/skill/gstack/scripts/install.sh` and the skill text are unchanged. ### Tests -- `test/gstack-capy-install.test.ts`: an `opus` preference picks lite with `source=preferences` and its own launch file, `GSTACK_MODEL` overrides the preference, a `host` preference prints no loadout lines, and the real `gstack-models resolve --role agent` keeps the `agent: /` first line the installer reads. +- `test/gstack-capy-install.test.ts`: an `opus` preference picks lite with `source=preferences` and its own launch file, `GSTACK_MODEL` overrides the preference, a `config.yaml` preference reads `source=config`, eval runs skip the preference, a `host` preference prints no loadout lines, and the real `gstack-models resolve --role agent` keeps the `agent: /` first line the installer reads. diff --git a/docs/capy-install.md b/docs/capy-install.md index 5f50afee9..1aa27b2dc 100644 --- a/docs/capy-install.md +++ b/docs/capy-install.md @@ -473,7 +473,8 @@ make writable, or set GSTACK_STATE_ROOT to a writable directory and removes the file an earlier passing run left beside the verdict it rewrites (a profile check that keeps a fuller ok verdict leaves that verdict's file). With `--model `, `GSTACK_MODEL=` in its environment, or else an -`agent_model` in the session's `GSTACK_PREFERENCES_FILE` (GST-982), an ok run other than `--for review-voice` also prints `LOADOUT: …` +`agent_model` in the session's `GSTACK_PREFERENCES_FILE` outside an eval run (GST-982, with a +`LOADOUT_ASSUMED:` note), an ok run other than `--for review-voice` also prints `LOADOUT: …` and `SKILL_ROOT: ` just before `GSTACK_LAUNCH`. For a loadout other than `full`, `GSTACK_LAUNCH_FILE` then names `.loadout--.env`, a copy of the verdict's launch file that adds `GSTACK_LOADOUT_SKILL_ROOT`; the diff --git a/docs/loadouts.md b/docs/loadouts.md index 91042a7c6..adc02078b 100644 --- a/docs/loadouts.md +++ b/docs/loadouts.md @@ -55,7 +55,7 @@ A model id is lowercased, its `provider/` prefix and a trailing `[…]` are drop ### Capy -A plain `bash scripts/install.sh --check --for ` picks the loadout from the owner's `agent_model` preference, so the skill needs no model sentence (`source=preferences` on the line below). A session can still name its own model with `GSTACK_MODEL=`, which wins over the preference and, unlike `--model`, is safe on a checkout whose installer predates loadouts: that installer ignores it instead of refusing an unknown argument. The installer prints two lines before the launch lines: +A plain `bash scripts/install.sh --check --for ` picks the loadout from the owner's `agent_model` preference, so the skill needs no model sentence (`source=preferences` on the line below, or `source=config` when this machine's `config.yaml` sets `agent_model`). The preference names the model Garry starts sessions on, not the model reading the skill, so the installer says so in a `LOADOUT_ASSUMED:` line, and it skips the preference inside an eval run (`GSTACK_EVAL_DIR` or `EVALS_RUN_ID` set), where the model under test may differ. A session on another model names its own with `GSTACK_MODEL=`, which wins over the preference and, unlike `--model`, is safe on a checkout whose installer predates loadouts: that installer ignores it instead of refusing an unknown argument. The installer prints two lines before the launch lines: ``` LOADOUT: lite model=claude-opus-5-5 source=GSTACK_MODEL digest=3f2a9c01d4e7 reused=yes diff --git a/test/gstack-capy-install.test.ts b/test/gstack-capy-install.test.ts index 6f90daac8..52db9a11a 100644 --- a/test/gstack-capy-install.test.ts +++ b/test/gstack-capy-install.test.ts @@ -605,6 +605,7 @@ describe('gstack-capy-install: launch PATH, Bun reuse, CI pins, setup skip, verd ]); script(path.join(f.root, 'bin/gstack-models'), [ 'echo "models $* prefs=${GSTACK_PREFERENCES_FILE:-}" >> "$CAPY_LOG"', + '[ -z "${FAKE_CONFIG_MODEL:-}" ] || { echo "agent: anthropic/$FAKE_CONFIG_MODEL (opus) - from agent_model in $HOME/.gstack/config.yaml, this machine"; exit 0; }', 'case "$(sed -n "s/^agent_model: *//p" "${GSTACK_PREFERENCES_FILE:-/dev/null}" 2>/dev/null)" in', ' opus) echo "agent: anthropic/claude-opus-5-5 (opus) - from agent_model in $GSTACK_PREFERENCES_FILE, the preferences file" ;;', ' *) echo "agent: host - keep the running agent\'s own model and omit the model field (no preference set)" ;;', @@ -618,6 +619,20 @@ describe('gstack-capy-install: launch PATH, Bun reuse, CI pins, setup skip, verd expect(f.calls().filter(l => l.startsWith('modules '))).toEqual([`modules install --model claude-opus-5-5 --full-root ${f.home}/.claude/skills`]); expect(opus.stdout).toContain(`LOADOUT: lite model=claude-opus-5-5 source=preferences digest=0123456789ab reused=no\nSKILL_ROOT: ${f.home}/lite-0123456789ab\n`); expect(launchOf(opus.stdout)).toBe(verdicts(f)[0].replace(/\.json$/, '.loadout-lite-0123456789ab.env')); + expect(opus.stdout).toContain("LOADOUT_ASSUMED: model=claude-opus-5-5 is the owner's agent_model (preferences), not read from the session; a session on another model sets GSTACK_MODEL=\n"); + + const config = run(f, ['--for', 'ship'], { GSTACK_PREFERENCES_FILE: prefs, FAKE_CONFIG_MODEL: 'claude-opus-5-5' }); + expect(config.status, config.stdout + config.stderr).toBe(0); + expect(config.stdout).toContain('LOADOUT: lite model=claude-opus-5-5 source=config digest=0123456789ab'); + expect(config.stdout).toContain("is the owner's agent_model (config)"); + + const modulesBefore = f.calls().filter(l => l.startsWith('modules ')).length; + for (const evalEnv of [{ GSTACK_EVAL_DIR: path.join(f.home, 'evals') }, { EVALS_RUN_ID: 'r1' }]) { + const ev = run(f, ['--for', 'ship'], { GSTACK_PREFERENCES_FILE: prefs, ...evalEnv }); + expect(ev.status, ev.stdout + ev.stderr).toBe(0); + expect(ev.stdout).not.toMatch(/^(LOADOUT|LOADOUT_ASSUMED|SKILL_ROOT):/m); + } + expect(f.calls().filter(l => l.startsWith('modules ')).length).toBe(modulesBefore); const env = run(f, ['--for', 'ship'], { GSTACK_PREFERENCES_FILE: prefs, GSTACK_MODEL: 'claude-sonnet-5-5' }); expect(env.status, env.stdout + env.stderr).toBe(0);