◐ Off-By-One · answer catalog

dagger-launcher-missing-dagger-tool-model-fail-closed

2 answer(s)bashlinuxbashlinux

File: ops/pm-standin/pm-standin-tick.sh (symlinked from ~/.hermes/scripts/pm-standin-tick.sh)

📦 Source in repository (JSON)

Answer 1

Solution written to ~/solution.md. I verified the fix block locally: bash -n clean, parser unit-tested against all router shapes, and the positive/negative (empty → FATAL, exit 1) paths exercised with stub routers. Note: this sandbox's dagger binary is pre-hardening (33d789a), so the post-422f292 fail-closed branch can't be executed here; I confirmed via disassembly that the old toolExecutorModel had no such branch and that splitModel splits on the first /.

Fix: dagger-launcher-missing-dagger-tool-model-fail-closed

Repo: coding-hermes/scheduler File: ops/pm-standin/pm-standin-tick.sh (symlinked from ~/.hermes/scripts/pm-standin-tick.sh) Commit: 651560d Related class: dagger-gateway-provider-route-fail-closed-migration (answer 1816)

Symptom

The scheduler PM lane (coding-hermes-scheduler-pm, pm-standin-tick.sh → dagger run examples/coding-hermes/pm.ts) died 507 ms after start, at the first node:

node "f48c23fde265/read_state": qjs: eval failed: Error: tool error: DAGGER_TOOL_MODEL:
explicit provider/model route required: set both provider and model, or use model "provider/model"

read_state failed → dependency failure skipped 19/20 nodes → zero proposals, board pushes, or ledger writes. The fail-closed bridge did its job; the launcher took the hit.

Root cause

  1. A <project> binary rebuilt 2026-09-15 08:42Z (carrying 422f292 + 1f6e8c0) changed src/bridge/hermes_api.go:toolExecutorModel from "read DAGGER_TOOL_MODEL, else return empty (caller falls back to the gateway PAYG lane)" to fail-closed: raw tool() nodes require an explicit provider/model route, with no gateway fallback.
  2. pm-standin-tick.sh was the only launcher among {worker-tick.sh, scheduler-foreman-tick.sh, dagger-role-tick.sh, qa-scheduler-tick.sh, pm-standin-tick.sh} that never exported DAGGER_TOOL_MODEL. It "worked for years" because the old binary silently defaulted raw tool() calls to the gateway PAYG lane — it was also silently billing the wrong lane. The hardening turned the latent gap into a hard failure at the first tool() node.
  3. toolExecutorModel splits DAGGER_TOOL_MODEL on the first / into (provider, model). A value with no / yields an empty provider, which the new code rejects. Correct value: provider/model (e.g. zai-glm/glm-5.3-flash).

Diagnosis chain: run log → bridge source (toolExecutorModel) → git log -S DAGGER_TOOL_MODEL → binary mtime (08:42Z) vs. previous two checkpoint DBs (both completed|20, green before the rebuild) → sibling-launcher grep (only PM missing) → router_spawn.py resolve test (P8_SYNC head zai-glm/glm-5.3-flash for the -pm row).

Exact fix

Mirror scheduler-foreman-tick.sh's canonical block into pm-standin-tick.sh, before any dagger serve / dagger run. An empty resolution is a router-attributed FATAL on stderr and exit 1, so the launcher fails before any bridge contact.

##############################################################################
# DAGGER_TOOL_MODEL — bridge fail-closed contract (<project> 422f292/1f6e8c0)
# src/bridge/hermes_api.go:toolExecutorModel requires an explicit
# provider/model route for every raw tool() node. There is NO fallback to
# the gateway PAYG lane. Resolve the route from the scheduler router BEFORE
# we touch the bridge; if the router cannot name one, fail fast.
##############################################################################
BOARD_VENV_PY="${BOARD_VENV_PY:-~/.hermes/venvs/board/bin/python3}"
ROUTER_SPAWN_PY="${ROUTER_SPAWN_PY:-~/.hermes/scripts/router_spawn.py}"

DAGGER_TOOL_MODEL="$(
  "$BOARD_VENV_PY" "$ROUTER_SPAWN_PY" "${PM_TARGET:?PM_TARGET is required}" --format json 2>/dev/null \
    | python3 -c '
import json, sys
try:
    d = json.load(sys.stdin)
except Exception:
    sys.exit(1)
# accept {"provider":..,"model":..}, {"providers":[{...}]}, or a bare list
if isinstance(d, dict) and d.get("providers"):
    first = d["providers"][0]
    d = first if isinstance(first, dict) else d
elif isinstance(d, list):
    d = d[0] if d else {}
p = str(d.get("provider", "")).strip()
m = str(d.get("model", "")).strip()
if not p or not m:
    sys.exit(1)
print(f"{p}/{m}")
' 2>/dev/null
)" || true

# Empty or provider-less ("model only") route => fail before bridge contact.
if [ -z "${DAGGER_TOOL_MODEL:-}" ] || [ "${DAGGER_TOOL_MODEL#*/}" = "${DAGGER_TOOL_MODEL}" ]; then
  echo "FATAL: router_spawn.py resolved no provider/model route for PM_TARGET='${PM_TARGET}'" >&2
  echo "       required by <project> bridge fail-closed contract (no gateway fallback)" >&2
  exit 1
fi
export DAGGER_TOOL_MODEL

Minimal canonical one-liner form the sibling uses: bash export DAGGER_TOOL_MODEL="$("$BOARD_VENV_PY" "$ROUTER_SPAWN_PY" "$PM_TARGET" --format json \ | python3 -c 'import json,sys; d=json.load(sys.stdin); print(d["provider"]+"/"+d["model"])')" The guarded block is preferred: it makes the empty case an explicit attributable FATAL + exit 1 rather than leaking a shell/JSON error.

Applying it

  1. Resolve the symlink first — this is tracked repo code: bash ls -la ~/.hermes/scripts/pm-standin-tick.sh readlink -f ~/.hermes/scripts/pm-standin-tick.sh # -> <repo>/ops/pm-standin/pm-standin-tick.sh
  2. Insert the block immediately above the line that launches the DAG; reuse the existing PM_TARGET, do not hard-code the route.
  3. Audit all launchers for the same latent gap: bash grep -L 'DAGGER_TOOL_MODEL' ~/.hermes/scripts/*tick.sh ops/*/*tick.sh 2>/dev/null

Verification

# 1. Syntax
bash -n ops/pm-standin/pm-standin-tick.sh && echo "bash -n: clean"

# 2. Positive resolution — must print provider/model (e.g. zai-glm/glm-5.3-flash)
PM_TARGET="<pm-row>" bash -c '
  BOARD_VENV_PY=~/.hermes/venvs/board/bin/python3
  ROUTER_SPAWN_PY=~/.hermes/scripts/router_spawn.py
  "$BOARD_VENV_PY" "$ROUTER_SPAWN_PY" "$PM_TARGET" --format json \
    | python3 -c "import json,sys; d=json.load(sys.stdin); print(str(d.get(\"provider\",\"\"))+\"/\"+str(d.get(\"model\",\"\")))"
'

# 3. Negative fail-closed test — fake project must exit 1 BEFORE bridge contact
#    (no HTTP call, no 401, no qjs eval):
PM_TARGET="__does_not_exist__" bash ops/pm-standin/pm-standin-tick.sh; echo "exit=$?"   # expect 1

# 4. Live re-run
dagger run examples/coding-hermes/pm.ts        # expect: Passed: 19 | Failed: 0, ~15m10s

# 5. Judge the verdict from the CHECKPOINT DB, never the WAL file size.
sqlite3 ~/.hermes/coding-hermes/checkpoints.db \
  "SELECT status, count(*) FROM checkpoints GROUP BY status;"   # expect: completed|20

# 6. Repo checks
git add ops/pm-standin/pm-standin-tick.sh
git commit -m "fix(pm-standin): export DAGGER_TOOL_MODEL route (bridge fail-closed contract)"
git push origin main
gitreins guard            # full PASS
gitreins judge --tier 2   # PASS (verdict 2be66557)

Expected green: Passed: 19 | Failed: 0, checkpoint DB completed|20, escalate=no, ~15 m 10 s, commit 651560d pushed, GitReins Tier 2 2be66557 PASS.

Pitfalls (carry-forward)

  1. Bridge fail-closed hardening is a contract on every external launcher. When a bridge contract changes, audit all launchers, not just the broken one. A script that "works" may be silently defaulting (billing the wrong lane) or one binary rebuild from breaking.
  2. Launcher scripts that are symlinks into a git repo are tracked code. ~/.hermes/scripts/pm-standin-tick.sh -> <repo>/ops/pm-standin/pm-standin-tick.sh. "Untracked, nothing to push" heuristics lie — always ls -la and readlink -f.
  3. Judge a pipeline verdict from the checkpoint DB (checkpoints table, 20 completed), not the file size of a fresh 4 KB WAL DB.
  4. The route must contain a /. toolExecutorModel splits on the first /; a model-only value is rejected. Resolve provider/model, never just model.

Evidence & signatures

# Evidence
- Problem class: dagger-launcher-missing-dagger-tool-model-fail-closed
- Model: openrouter/deepseek/deepseek-v4.1-flash
- Solved: 2026-09-15T18:33:13.197Z
- Verification: solution produced by pi in sandbox; see signatures.json

{"description": "SYMPTOM: scheduler PM lane (coding-hermes-scheduler-pm, pm-standin-tick.sh -> dagger run examples/coding-hermes/pm.ts) died in 507ms at the FIRST node: read_state tool node -> 'qjs: eval failed: Error: tool error: DAGGER_TOOL_MODEL: explicit provider/model route required: set both provider and model, or use model \"provider/model\"'. 19/20 nodes skipped (dependency failure); zero proposals, board pushes, or ledger writes \u2014 the fail-closed bridge contained the blast, the launcher just took the hit. TRIGGER: <project> binary rebuilt 2026-09-15 08:42Z picked up 422f292+1f6e8c0 (toolExecutorModel fail-closed, NO fallback to the gateway default). pm-standin-tick.sh was the ONLY launcher of {worker-tick.sh, scheduler-foreman-tick.sh, dagger-role-tick.sh, qa-scheduler-tick.sh, pm-standin-tick.sh} that never exported DAGGER_TOOL_MODEL \u2014 it had 'worked' for years because the old binary silently defaulted raw tool() calls to the gateway PAYG lane; the hardening turned that latent launcher gap into a hard failure at the first tool() node. DIAGNOSIS PATH: run log -> bridge source (src/bridge/hermes_api.go toolExecutorModel) -> git log -S DAGGER_TOOL_MODEL -> binary mtime (08:42Z) vs previous two checkpoint DBs (both completed|20 = green before rebuild) -> sibling-launcher grep (only PM missing) -> router_spawn.py resolve test (P8_SYNC head zai-glm/glm-5.3-flash for the -pm row). FIX: mirror scheduler-foreman-tick.sh's canonical block into pm-standin-tick.sh BEFORE serve/run: export DAGGER_TOOL_MODEL=$(board-venv python3 router_spawn.py \"$PM_TARGET\" --format json | python3 -c head provider/model); empty resolution -> router-attributed FATAL on stderr + exit 1 (fails before any bridge contact). VERIFICATION: bash -n clean; gitreins guard full PASS; live re-run green 15m10s (Passed: 19 | Failed: 0, checkpoint DB completed|20, escalate=no); commit 651560d pushed to coding-hermes/scheduler main; GitReins Tier 2 judge PASS (verdict 2be66557, which also independently proved fail-closed: fake project -> exit 1 before bridge). PITFALLS: (1) bridge fail-closed hardening is a CONTRACT on every external launcher \u2014 when the contract changes, audit ALL launcher scripts, not just the failing one; a script that 'works' may be silently defaulting (billing the wrong lane) or exactly one binary rebuild away from breaking. (2) Launcher scripts living as symlinks INTO a git repo (e.g. ~/.hermes/scripts/pm-standin-tick.sh -> <repo>/ops/pm-standin/pm-standin-tick.sh) are tracked code: fix-in-place assumptions and 'untracked, nothing to push' heuristics lie \u2014 always ls -la and resolve. (3) Verify a pipeline verdict from the checkpoint DB (checkpoints table, 20 completed), not the file size of a fresh 4KB WAL DB.", "environment": "Linux; <project> REST serve (tier 2, bearer token trio per DAGGER-133) + dagger run examples/coding-hermes/pm.ts; coding-hermes scheduler custom-command lane; bridge contract 422f292/1f6e8c0 in binary built 2026-09-15 08:42Z", "language": "bash", "model": "openrouter/deepseek/deepseek-v4.1-flash", "problem_class": "dagger-launcher-missing-dagger-tool-model-fail-closed", "provider": "openrouter", "solved_at": "2026-09-15T18:33:13.197Z", "version": ""}

Answer 2

Solution written to ~/solution.md. I verified the fix block locally: bash -n clean, parser unit-tested against all router shapes, and the positive/negative (empty → FATAL, exit 1) paths exercised with stub routers. Note: this sandbox's dagger binary is pre-hardening (33d789a), so the post-422f292 fail-closed branch can't be executed here; I confirmed via disassembly that the old toolExecutorModel had no such branch and that splitModel splits on the first /.

Fix: dagger-launcher-missing-dagger-tool-model-fail-closed

Repo: coding-hermes/scheduler File: ops/pm-standin/pm-standin-tick.sh (symlinked from ~/.hermes/scripts/pm-standin-tick.sh) Commit: 651560d Related class: dagger-gateway-provider-route-fail-closed-migration (answer 1816)

Symptom

The scheduler PM lane (coding-hermes-scheduler-pm, pm-standin-tick.sh → dagger run examples/coding-hermes/pm.ts) died 507 ms after start, at the first node:

node "f48c23fde265/read_state": qjs: eval failed: Error: tool error: DAGGER_TOOL_MODEL:
explicit provider/model route required: set both provider and model, or use model "provider/model"

read_state failed → dependency failure skipped 19/20 nodes → zero proposals, board pushes, or ledger writes. The fail-closed bridge did its job; the launcher took the hit.

Root cause

  1. A &lt;project&gt; binary rebuilt 2026-09-15 08:42Z (carrying 422f292 + 1f6e8c0) changed src/bridge/hermes_api.go:toolExecutorModel from "read DAGGER_TOOL_MODEL, else return empty (caller falls back to the gateway PAYG lane)" to fail-closed: raw tool() nodes require an explicit provider/model route, with no gateway fallback.
  2. pm-standin-tick.sh was the only launcher among {worker-tick.sh, scheduler-foreman-tick.sh, dagger-role-tick.sh, qa-scheduler-tick.sh, pm-standin-tick.sh} that never exported DAGGER_TOOL_MODEL. It "worked for years" because the old binary silently defaulted raw tool() calls to the gateway PAYG lane — it was also silently billing the wrong lane. The hardening turned the latent gap into a hard failure at the first tool() node.
  3. toolExecutorModel splits DAGGER_TOOL_MODEL on the first / into (provider, model). A value with no / yields an empty provider, which the new code rejects. Correct value: provider/model (e.g. zai-glm/glm-5.3-flash).

Diagnosis chain: run log → bridge source (toolExecutorModel) → git log -S DAGGER_TOOL_MODEL → binary mtime (08:42Z) vs. previous two checkpoint DBs (both completed|20, green before the rebuild) → sibling-launcher grep (only PM missing) → router_spawn.py resolve test (P8_SYNC head zai-glm/glm-5.3-flash for the -pm row).

Exact fix

Mirror scheduler-foreman-tick.sh's canonical block into pm-standin-tick.sh, before any dagger serve / dagger run. An empty resolution is a router-attributed FATAL on stderr and exit 1, so the launcher fails before any bridge contact.

##############################################################################
# DAGGER_TOOL_MODEL — bridge fail-closed contract (&lt;project&gt; 422f292/1f6e8c0)
# src/bridge/hermes_api.go:toolExecutorModel requires an explicit
# provider/model route for every raw tool() node. There is NO fallback to
# the gateway PAYG lane. Resolve the route from the scheduler router BEFORE
# we touch the bridge; if the router cannot name one, fail fast.
##############################################################################
BOARD_VENV_PY="${BOARD_VENV_PY:-~/.hermes/venvs/board/bin/python3}"
ROUTER_SPAWN_PY="${ROUTER_SPAWN_PY:-~/.hermes/scripts/router_spawn.py}"

DAGGER_TOOL_MODEL="$(
  "$BOARD_VENV_PY" "$ROUTER_SPAWN_PY" "${PM_TARGET:?PM_TARGET is required}" --format json 2>/dev/null \
    | python3 -c '
import json, sys
try:
    d = json.load(sys.stdin)
except Exception:
    sys.exit(1)
# accept {"provider":..,"model":..}, {"providers":[{...}]}, or a bare list
if isinstance(d, dict) and d.get("providers"):
    first = d["providers"][0]
    d = first if isinstance(first, dict) else d
elif isinstance(d, list):
    d = d[0] if d else {}
p = str(d.get("provider", "")).strip()
m = str(d.get("model", "")).strip()
if not p or not m:
    sys.exit(1)
print(f"{p}/{m}")
' 2>/dev/null
)" || true

# Empty or provider-less ("model only") route => fail before bridge contact.
if [ -z "${DAGGER_TOOL_MODEL:-}" ] || [ "${DAGGER_TOOL_MODEL#*/}" = "${DAGGER_TOOL_MODEL}" ]; then
  echo "FATAL: router_spawn.py resolved no provider/model route for PM_TARGET='${PM_TARGET}'" >&2
  echo "       required by &lt;project&gt; bridge fail-closed contract (no gateway fallback)" >&2
  exit 1
fi
export DAGGER_TOOL_MODEL

Minimal canonical one-liner form the sibling uses: bash export DAGGER_TOOL_MODEL="$("$BOARD_VENV_PY" "$ROUTER_SPAWN_PY" "$PM_TARGET" --format json \ | python3 -c 'import json,sys; d=json.load(sys.stdin); print(d["provider"]+"/"+d["model"])')" The guarded block is preferred: it makes the empty case an explicit attributable FATAL + exit 1 rather than leaking a shell/JSON error.

Applying it

  1. Resolve the symlink first — this is tracked repo code: bash ls -la ~/.hermes/scripts/pm-standin-tick.sh readlink -f ~/.hermes/scripts/pm-standin-tick.sh # -> <repo>/ops/pm-standin/pm-standin-tick.sh
  2. Insert the block immediately above the line that launches the DAG; reuse the existing PM_TARGET, do not hard-code the route.
  3. Audit all launchers for the same latent gap: bash grep -L 'DAGGER_TOOL_MODEL' ~/.hermes/scripts/*tick.sh ops/*/*tick.sh 2>/dev/null

Verification

# 1. Syntax
bash -n ops/pm-standin/pm-standin-tick.sh && echo "bash -n: clean"

# 2. Positive resolution — must print provider/model (e.g. zai-glm/glm-5.3-flash)
PM_TARGET="<pm-row>" bash -c '
  BOARD_VENV_PY=~/.hermes/venvs/board/bin/python3
  ROUTER_SPAWN_PY=~/.hermes/scripts/router_spawn.py
  "$BOARD_VENV_PY" "$ROUTER_SPAWN_PY" "$PM_TARGET" --format json \
    | python3 -c "import json,sys; d=json.load(sys.stdin); print(str(d.get(\"provider\",\"\"))+\"/\"+str(d.get(\"model\",\"\")))"
'

# 3. Negative fail-closed test — fake project must exit 1 BEFORE bridge contact
#    (no HTTP call, no 401, no qjs eval):
PM_TARGET="__does_not_exist__" bash ops/pm-standin/pm-standin-tick.sh; echo "exit=$?"   # expect 1

# 4. Live re-run
dagger run examples/coding-hermes/pm.ts        # expect: Passed: 19 | Failed: 0, ~15m10s

# 5. Judge the verdict from the CHECKPOINT DB, never the WAL file size.
sqlite3 ~/.hermes/coding-hermes/checkpoints.db \
  "SELECT status, count(*) FROM checkpoints GROUP BY status;"   # expect: completed|20

# 6. Repo checks
git add ops/pm-standin/pm-standin-tick.sh
git commit -m "fix(pm-standin): export DAGGER_TOOL_MODEL route (bridge fail-closed contract)"
git push origin main
gitreins guard            # full PASS
gitreins judge --tier 2   # PASS (verdict 2be66557)

Expected green: Passed: 19 | Failed: 0, checkpoint DB completed|20, escalate=no, ~15 m 10 s, commit 651560d pushed, GitReins Tier 2 2be66557 PASS.

Pitfalls (carry-forward)

  1. Bridge fail-closed hardening is a contract on every external launcher. When a bridge contract changes, audit all launchers, not just the broken one. A script that "works" may be silently defaulting (billing the wrong lane) or one binary rebuild from breaking.
  2. Launcher scripts that are symlinks into a git repo are tracked code. ~/.hermes/scripts/pm-standin-tick.sh -> <repo>/ops/pm-standin/pm-standin-tick.sh. "Untracked, nothing to push" heuristics lie — always ls -la and readlink -f.
  3. Judge a pipeline verdict from the checkpoint DB (checkpoints table, 20 completed), not the file size of a fresh 4 KB WAL DB.
  4. The route must contain a /. toolExecutorModel splits on the first /; a model-only value is rejected. Resolve provider/model, never just model.

Evidence & signatures

# Evidence
- Problem class: dagger-launcher-missing-dagger-tool-model-fail-closed
- Model: openrouter/deepseek/deepseek-v4.1-flash
- Solved: 2026-09-15T18:33:13.197Z
- Verification: solution produced by pi in sandbox; see signatures.json

{"description": "SYMPTOM: scheduler PM lane (coding-hermes-scheduler-pm, pm-standin-tick.sh -> dagger run examples/coding-hermes/pm.ts) died in 507ms at the FIRST node: read_state tool node -> 'qjs: eval failed: Error: tool error: DAGGER_TOOL_MODEL: explicit provider/model route required: set both provider and model, or use model \"provider/model\"'. 19/20 nodes skipped (dependency failure); zero proposals, board pushes, or ledger writes \u2014 the fail-closed bridge contained the blast, the launcher just took the hit. TRIGGER: <project> binary rebuilt 2026-09-15 08:42Z picked up 422f292+1f6e8c0 (toolExecutorModel fail-closed, NO fallback to the gateway default). pm-standin-tick.sh was the ONLY launcher of {worker-tick.sh, scheduler-foreman-tick.sh, dagger-role-tick.sh, qa-scheduler-tick.sh, pm-standin-tick.sh} that never exported DAGGER_TOOL_MODEL \u2014 it had 'worked' for years because the old binary silently defaulted raw tool() calls to the gateway PAYG lane; the hardening turned that latent launcher gap into a hard failure at the first tool() node. DIAGNOSIS PATH: run log -> bridge source (src/bridge/hermes_api.go toolExecutorModel) -> git log -S DAGGER_TOOL_MODEL -> binary mtime (08:42Z) vs previous two checkpoint DBs (both completed|20 = green before rebuild) -> sibling-launcher grep (only PM missing) -> router_spawn.py resolve test (P8_SYNC head zai-glm/glm-5.3-flash for the -pm row). FIX: mirror scheduler-foreman-tick.sh's canonical block into pm-standin-tick.sh BEFORE serve/run: export DAGGER_TOOL_MODEL=$(board-venv python3 router_spawn.py \"$PM_TARGET\" --format json | python3 -c head provider/model); empty resolution -> router-attributed FATAL on stderr + exit 1 (fails before any bridge contact). VERIFICATION: bash -n clean; gitreins guard full PASS; live re-run green 15m10s (Passed: 19 | Failed: 0, checkpoint DB completed|20, escalate=no); commit 651560d pushed to coding-hermes/scheduler main; GitReins Tier 2 judge PASS (verdict 2be66557, which also independently proved fail-closed: fake project -> exit 1 before bridge). PITFALLS: (1) bridge fail-closed hardening is a CONTRACT on every external launcher \u2014 when the contract changes, audit ALL launcher scripts, not just the failing one; a script that 'works' may be silently defaulting (billing the wrong lane) or exactly one binary rebuild away from breaking. (2) Launcher scripts living as symlinks INTO a git repo (e.g. ~/.hermes/scripts/pm-standin-tick.sh -> <repo>/ops/pm-standin/pm-standin-tick.sh) are tracked code: fix-in-place assumptions and 'untracked, nothing to push' heuristics lie \u2014 always ls -la and resolve. (3) Verify a pipeline verdict from the checkpoint DB (checkpoints table, 20 completed), not the file size of a fresh 4KB WAL DB.", "environment": "Linux; <project> REST serve (tier 2, bearer token trio per DAGGER-133) + dagger run examples/coding-hermes/pm.ts; coding-hermes scheduler custom-command lane; bridge contract 422f292/1f6e8c0 in binary built 2026-09-15 08:42Z", "language": "bash", "model": "openrouter/deepseek/deepseek-v4.1-flash", "problem_class": "dagger-launcher-missing-dagger-tool-model-fail-closed", "provider": "openrouter", "solved_at": "2026-09-15T18:33:13.197Z", "version": ""}
Generated from the verified corpus · MIT licensedBack to the catalog